Datasets:

Fischerboot

/

yannik-data

like 0

write a youtube script

I like planes I like boats I like planes I like boats. but what if I want that boaty planess in one?! oh no! I like planes and I like boats these are the words that Xyla Foxlin will say every morning! so anyways Its YouTube makers secret santa time yes! Loads of makers all group together we all make stuff for each other weve been doing it for like three years in a row now. This year I have Xyla Foxlin now she lives in LA thats gonna cause me problems later and basically you know from her videos shes made a boat before and I know shes got a pilots license from her social media stuff So I have a plan okay unfortunately were not making a flying boat because it is the th of December these videos are supposed to go on the th of December I have to post whatever I make to Los Angeles. I havent made it yet oh yes Im behind. Right so what Im thinking boats have oars we know boats have oars planes have propellers but boats can sort of have propellers as well in the term of props you know outboard motors. So so Im thinking well put propellers or outboard motor things in the oars. Genius Furze! Now Im not sure how its going to work because obviously when youre in a boat you push the Oar into the water and then you pull against the water and push the boat forward but if the oar is trying to push itself through the water does the oar just try and get out of your hands and disappear? I dont know but I cant think of anything better so lets make it! Ok Ive hit simply bearings hard its where I get all my bearings chains and gears from. This is not sponsored theyre just a nice company. So anyway I think Im going to go with this chain which is like mm chain so what Im thinking Colin what have you bought chain? Well basically oar propellers set in there chain around the shaft with a cog on it down through the um the pole the oar thing and then to some sort of central Hub where there will be a motor some some gears and stuff that will spin them both so they will both go around off the one motor. Right so Im going to get all this set out make myself some little templates do a little design just get on with it Colin! Its half ten now you got to get on with it! what is this.... Xylas arm slashing hair tangling super paddle... oh middle bit! Right thats the center bit now this middle Cog which is just over a bolt at the moment thatll be on a shaft which will be attached to whatever power source we choose. Now this chain goes up and over it back and forth and then you can see if I pull it look if the chains nice and tight it should go across the top of it and then pull back underneath it in equal amounts and therefore spinning the propellers at each end. Now the chain itself will slide down tube this is going to be noisy Colin yes it is! and then theyll go into these slots like that. both ends thats that right theres another plate going to go on top of that and then Ive got to weld them onto the things and hope that I can get to everything. So many complications. Right thats welded on. Cant get to them cogs very easily now. check. right the ends propellers now Ive started us off look Ive put some little thicker plates on the end. Ive welded them on and then whatever we make we can then bolt onto that and if its wrong or breaks we can then unbolt it. So we dont have to grind it off or cut it off or anything like that right so the ends about that! This is my concept people... yes we still need something that resembles a paddle even though weve got our propeller so Ive made this obviously its got a lot of holes in it or else youre just propellering against your paddle. Bolt this on here chain comes out round axle axillo skinny. but because its a little bit floppy Im gonna put like a secondary one on which will slide on bolt through it stick it off make it double skinned like the middle bit itll be a bit stiffer... yes! OK thats got it all together that was a nightmare. Now then its all connected up it spins around both go around in the same direction of course which is good now in the middle I have to shave a bit off the gears which is a complete pain cuz I had to try and get them out and end up in the end like plasma cutting some of my own spanners out of some . mil Steels we had to get down and slotting in but anyway its all good. Now the chain is tensioned via these little slots here so basically you put it all together and you just pull the whole thing back and tighten it up. That is good. Now the next thing we need to consider Colin is the power source so anyway I thought what did I think? Right and so I was thinking we need a nice motor thats got power and got speed I always found its quite a tricky thing to get hold of and I thought well wait a minute why dont we do something that she might already have and I was thinking... drill shell have a drill wont she! I dont think a drills fast enough but I know what Ive got thats a bit quicker grinder! oh!! haha! Oh God! its a bit quick! I think well use a grinder! Weve got no choice! Right lets somehow rig up a grinder on this thing and then see how much of a laceration machine I have produced! I cant help noticing this is all starting to look a bit big and a bit heavy! I dont think shes that big you know? No I think itll be all right shes not that small Ive seen her stood next to a rocket shes pretty tall actually! whatever whatever! Right... see what happens here! oh hahaha! If that sound does not translate onto video that is like belt of knives X I dont even think thats full belt! I think I can manage this. Right this sounds absolutely horrific and everything but does it actually work? Ok here we are in December down the river its freezing! Right Ive rigged up like this really crude trigger system put a drill on it just for the time being .Im gonna give that a test first. Right gonna put my coat back on well have a go. The only the only thing Ive got to attach the actual GoPro to is the oar. So the footage for this might be a little bit wobbly. Immediately Im returning to this one thats not a good! Ive dropped it! why did I build something that goes into blooming water in December! faster and faster! oh no! This is a design problem the chain is bringing water up and dribbling it in my oh no Ive got a wet... There is so much wrong with this but it sort of works in a weird way. I mean it spins around in circles Ive now got wet A! Just thinking about this Id be better with the thing the other way around. We need to do this and you have the triggers all wrong! oh God this is terrible! Im bloody soaked! right right! You know what I think this actually works! This is working! Right Im going to stick the angle grinder on it because this things A the batteries are dying and its not as fast as it could be. oh this is... I cant feel my hands! Just get back in the old boat! Right people oh my God check this out! ah! oh my God Im so bloody wet! The grinder is not quite powerful enough in terms of it spins too fast and hasnt got the torque but the paddle does work. okay we do have some success here. Thats it for today its cold Ive had enough! Now despite getting freezing cold and soaking wet there is some potential in this! Basically the grinder spins very fast its not quite got the torque it needs to as soon as you put it in the water it basically stops and the drill wasnt actually as bad as I thought it was going to. So Im thinking Im gonna go back for another attempt with some proper clothing on and a better drill with a bigger battery on ! Right back again look like Im searching for dead bodies but at least Ill be dry. Come on this is gonna work! Here we go people! this is working! If you put it in look it will work like an outboard motor and spin me around! oh my God its actually icy! t more speed for more effort! back to the workshop! We may not have invented a new extreme sport but there is potential in this. If you got the right motor built into the center fixed it in properly so it dont spin around and smack in the head. Put the batteries either side of it so you can hold the thing properly and it wasnt quite so heavy I think thered be a bit of potential here Now we aint got time to do that because this videos got to go on the internet tomorrow and obviously clearly Im not going to get all that done. Now Colin if its going on the internet tomorrow arent you supposed to be sending this to Zyla? yes I am but I did decide well a few days back now that this is probably not the best thing to just hand over to somebody for them to try with all the spinning blades of dome potential lacerations. Ive made her something else and sent that in tribute to this this is a dedication to her and her love of planes and boats youll have to go and see what the other gift is I thought it was a lovely and wonderful gift but now Ive thought about it and Ive looked at it since... it does look a bit kinky... so youll have to go and make your own minds up! hmm! But this is YouTube makers secret santa its not just about what Im making for other people people have made for me as well and I have mine here! lets undo it. OK its from Emily the Engineer now I know this because it says here on the box who its from. Now this will be interesting. She is like a Marvel addict she does a lot of stuff with Iron Man suits shes made her own Iron Man suits and stuff so Ive been quite interested to see what this is. Rightyho! Got notes Colin Merry Christmas I know you often enjoy wearing neck ties so I thought Id make you a device to make putting them on a little more interesting! So it is a tie launcher for launching ties onto me! this is what you gotta like about YouTube makers Secret Santa you just never know what youre gonna get but its very well wrapped. oh you star Emily look shes even got the old Safety Tie warning triangle on there. Right shes advised I try it with a mannequin first so Ive got Mrs yellow in its got a little laser pointer on the top so its put like a little dot where you want to fire the tie I like it! Right lets put a little COT cartridge in the back which is what its powered by and give it a test. Here we go. oh Ill stand in front of that! Right no more Mrs thingy come on Furze tie to the face! By the way Emily this gets a thumbs up from me! Right so take my tie off collars up collars up! Ready to accept! Come on timing. I have no idea if that nearly got me at all oh thats so close Yeah! it works!!! Fantastic! ah! That is brilliant! Right now of course YouTube Makers Secret Santa the whole point the whole reason we do this is so you can go on your little course going from one video to another so you can even go backwards you can go and see how Emily made this or you can go forwards and go and see what Zyla has made and go and see her reaction to her nonpresent because Ive kind of Ive kind of failed a little bit this year but anyway we dont think it was it but anyway . so go on the trail I think theres of us in it this year you know the reason why I do this is to try and spread the love to all the other makers so just go and check them all out things like this absolutely fantastic thanks for watching hope youve had a good year see you next year where of course there is more tunneling to do weve got a car park were going to connect it to the bunker and theres all sorts of things to do at the farm as well. boom for you

write a youtube script

Im making a functional tabletop bowling out this is going to be fully functional with the working pin Setter to start off I started by printing some pins relative to the scale of my in marble bowling ball the first subsystem I want to tackle is the lane and bumpers I think itd be cool to have a Deployable bumper system for those who arent so skilled I knew I needed some sort of locking mechanism that could be pulled up and stopped at the top of its position this is a locking parallelogram mechanism as the mechanism is lifted a pin rides along a chamfered edge to lock it into its top position the mechanism cant be moved until someone pulls a springloaded latch to release the mechanism and fold it back down I recreated my own version of the mechanism in solid works to fit the scale of my bowling lane and printed out the parts for testing my first design didnt work well but it never does its just for learning purposes I kept reiterating and adding small improvements with each print for instance the springloaded latch was a bit too Wiggly and needed more walls to contrain its movement it took me about six tries to get this perfect where the mechanism can be pushed into place and now it has a perfect satisfying release to fold it back down follow to see the rest of my build

write a youtube script

this is jaws jaws is not a nutcracker hes a nut obliterator hes times more powerful than a normal nutcracker so load up and strap on your body armor because with jaws you just might need it Music jaws was made because i thought it would be really funny to crack this kind of nut so well be testing that and of course sticking everything else we can think of in there if youre wondering why something like this should exist i mean come on its gonna be awesome three two one wow this all started when i made an explosive powered bat which as best i can tell broke the mlb home run distance record ever since ive been wanting to make a nutcracker using this technology and tis the season the bat was powered by these little explosive blanks which are used to drive nails directly into concrete they should be really good at driving nutcracker jaws into everything the red tip on these blanks means that theyre maximum power which is really cool but you know whats even cooler four blanks at the same time getting this thing working was pretty hard it really wants to just violently tear itself apart that reminds me i need to tell you do not try this at home if you dont know exactly what youre doing this is incredibly dangerous and even deadly theres a lot of engineering testing and precautions that im not even showing in the video that im doing to make this safe so live vicariously through me and if the friend asks you if you want to make something powder actuated just say no all right lets go back to the beginning a traditional nutcracker works by amplifying the force from your hand with a lever to squeeze a nut really hard and crack it with powder actuation i should be able to make a nutcracker that generates up to pounds of force thats like times stronger than this nutcracker and thats all thanks to these wonderful little blinks youre supposed to put them in a special tool which tickles them in just the right way they shoot out super high pressure gas which drives a nail into concrete my goal is to somehow use this expanding gas to actuate the nutcracker although im not going to use a lever because they really like to snap and besides this is so powerful i dont need no stinking lever if i put a piston over these blanks the expanding gas will shoot it up and crush these nuts this is similar to how my bat worked but theres a big difference this is the pushy bit from the bat that launches the ball its cal because the pistons are half of an inch diameter and this is what im going to be making the piston for the nutcracker out of cal inches this is going to make the nutcracker generate about times the force of the bat to understand why we have to talk about pressure pressurized gas is basically gas that wants to take up a lot of space smooshed down into a tiny little space and like a compressed spring it wants to expand except the gas wants to expand in every direction at the same time pressure is measuring how hard the gas is pushing on a given area if i have a bigger area it will push harder the problem is that when you start making the piston really big it really wants to tear itself apart and thats because the stress in the cylinder grows like this as you make the piston bigger so to try to make it strong enough i could just make everything thicker than a bowl of oatmeal but that isnt a very good strategy for one i could be wrong and it could explode in my face but also metal is really expensive and if i make it thicker than it needs to be im wasting material which is a tragedy this is why engineering is a thing if i can calculate how strong the nutcracker needs to be and how strong a given design is i can design it to be just strong enough and its just like they say any old schmuck can build a bridge but only an engineer can make a bridge that barely stands im planning to bolt together a stack of thick steel plates to make the piston when the pressures pushing the piston up its also pushing downward the opposing forces will try to rip the nutcracker in half and the bolts are the only thing keeping it together so its critical that the bolts are strong enough so lets talk about steel if i tighten the nut on this big bolt its hard to see but the bolt is getting longer and if i loosen the nut it springs back to its original length usually at least if youre anything like me youve cranked the nut to the point where it suddenly gets easy to turn and whats happening is that the bolt stretches so much and it starts to permanently deform we need to use bolts that are strong enough but the question is what size different size bolts take different amounts of force to stretch but if i take the force it took to stretch any given bolt and divide it by the bolts cross sectional area i always get the same number which is interesting and this number is actually a really important property of metals its called yield strength its telling me how many pounds per square inch of crosssection steel can take before it stretches i can use this to directly calculate what size bolt i need and this is one of the ways that engineers predict how somethings going to behave rather than having to physically test it so at this point we have a piston that were going to shoot upwards utilizing the farts of these shells the plan is for the piston to be the nutcrackers torso so were going to move the shells over to the nutcrackers back where theres also room for the firing pins and a hammer and all that stuff when it goes bang all the gas shoots down the tube and launches the piston upwards but once its done i need to let the gas out i could have a little hole that opens up like this and lets the gas out but this is going to be incredibly loud and nutcrackers are supposed to be used in the family room with everyone gathered around them loud explosions would totally spoil the mood and besides everyone knows that the only loud explosions allowed in these situations is vicious political debate so im going to add a tube to the piston that gets pulled up as the nutcracker fires when the jaw reaches the closed position itll open up a little hole that lets the gas shoot down the tube which will have a muffler on the end of it which will make it quiet this tube lets me solve another big problem which is stopping the piston remember its going to be a chunk of inch steel and if i fire the nutcracker with nothing in it its going to be going up to miles per hour if it slams into a solid steel wall at that speed it will destroy everything so im going to put a spring right here which has to be compressed before the piston hits the wall this will bring it to a gentle stop and by gentle i mean gs which would liquify your brain but its much better than hitting the wall this is the spring were going to be using the technical spec for this is beef supreme remember that gas that tried to rip our bolts in half well it also pushes out sideways which is a big problem we need to make the nutcracker thick enough so that this doesnt happen but how thick is thick enough just like the bolts our goal is to make sure that this tube never stretches to the point that it permanently deforms but figuring this one out is tricky if i pressurize the pipe and look at how much its stretching it varies in a complicated way across the pipe this problem is a lot easier to think about if we zoom way in on the pipe and look at a little tiny piece thats so small the variation across it is negligible we can imagine what happens to this little piece the pressure in the tube is pushing it outwards but this piece has little pieces next to it that pull on it to keep it from moving away it also has a piece behind it thats going to keep it from moving which is also held in place by pieces next to it and so on everything is pushing and pulling and squeezing on everything if this seems like a huge complicated mess youre right ill spare you the specific details but math makes it possible to precisely specify all these different things that happen in a little piece which you can use to figure out the stretch or the stress for any part in the tube then all we have to do is run the numbers until we find a tube thats thick enough so that it doesnt explode if youve ever wondered why anyone cares about calculus or differential equations this is why they are incredibly powerful tools for a huge number of problems i just ran the numbers and theyre kind of crazy to keep the nutcracker from exploding the walls have to be an inch thick which is really thick hopefully this reinforces why you dont want to mess with these things i would have never thought that the walls need to be that thick all we got to do now is design it ive been working on this for a couple days now and this is what ive come up with basically a giant block of steel which is what it takes to hold this thing together steel is pretty slow to machine so this is going to take a while if youve ever wondered what its like to make something like this look no further Music so Music man that took forever assembly is usually a lot easier than the fabrication this is the triggering mechanism these are the pins that ignite the shells Music all of these big steel parts are the piston i tried to reduce the number of seals but theres still a lot of seals all the nonstrength critical parts are d printed on the fuse one this is starting to look like a nutcracker i think hes ready to make his big debut Music sporting a classic military uniform with an avantgarde flare he has a face that says see what happens will the influencers of the world adopt his style i think that would be great but only time will Music tell all right lets be honest this is amazing this is so cool ah its huge theres a normal nutcracker i love this this is so good this little backpack that hes wearing is what holds the shells and does the firing and just like a real nutcracker you push this lever down to fire it to load it you open the breech and put the shells in here then when youre done you close the breech and then it and its armed my goal was to make it impossible to set this thing off when you dont intend to and theres a series of safeties to ensure this the first thing we have is the safety stick this physically prevents the nutcracker lever from being depressed it also goes inside of the nutcracker and physically blocks the hammer from reaching the shells so even if you somehow release the hammer it just cant go off if the slide isnt fully shut it blocks the hammer as well and the slide cant be fully shut unless the breech is fully locked this prevents it from being triggered with the breech partially locked or open which could be really dangerous and then if we decapitate him these bolts that hold the body together are designed to fail before the body itself this means that if somehow something went pop its gonna shoot this plate straight up rather than pieces of the body outwards which is much more dangerous thats a lot of engineering and precautions but i still dont trust my engineering that much so weve got body armor hard hat eye protection im cowering behind the barricade and it has remote start so i dont have to be anywhere near it a while back i had to fast for a couple of days and this is a reenactment of what it looked like when i finally got some food three two one this isnt nearly as powderized as i would have expected lets do something a little more quantitative metal nut versus two shells three two one what what some really precise measurements i think i can barely detect some squish but i shouldnt need calipers to see this it should be smooshed a lot more i spent a lot of time wandering in the wilderness and were just going to skip over that part this has just been driving me crazy ive been messing with this for multiple days now ive fired it a million times and it kind of works but the power just is not matching what it should be i thought there was a bunch of different problems i remade the chamber that holds the shells three times and im pretty sure i figured out the actual problem the gas from the shells comes out of this hole and goes into this hole to drive the piston and it would really hurt the performance of the nutcracker if these holes didnt line up no one could have guessed that right i mean come on it only took three tries to figure it out there is at least a quick hack to fix this three two one Music its better but it still isnt as powerful as i calculated the nut with three shells should be flat im just going to go for four i want to know if it can survive it i want to know if im gonna have to remake a bunch of parts three two one Music that was so loud even with the hearing protection i felt that in my soul it didnt even do anything with a nut its worse than three shells like what doesnt make any sense i dont understand what is wrong with you oh i see it looks like the pressures so high that it blew the tops of the shells off and squeezed them out into this little tiny gap in the breech which is surprising because the fit is really good on this its built like a swiss watch maybe a swiss watch made in china its pretty good unfortunately what happened here is pretty fundamental to the design of these parts i dont see any way to modify these to make it work i guess im going to go redesign and remake all these parts great its not that bad i just remade these parts and youre not going to believe it the firing pins dont line up with the shells so i get to remake them again all right it is that bad this doesnt fit this one should fit finally the new design bolts everything together so theres no gap to squeeze the shells into i just realized a big problem when this thing fires all that gas goes down through the muffler into the legs which are totally sealed it has nowhere to go i gotta find a way to get the hot gas out before i break something you know this is the best spot dont judge me Music all right i think hes finally ready for prime time and in this corner we have jaws weighing in at pounds he is a chonker and in the other corner is lil joe weighing less than two mice i have to say is this fair wait hes already devoured him is that even legal all right i think its working we get to do the fun part now we get to see what this little man is made of ive brought the wife in so she can render her opinions what do you think of jaws manicured mustache and eyebrows clearly had some dental work done theyre also metal hes got a grill and i really like his hair were going to start with what this thingy is designed for this is a rock hard macadamia nut what do you think is gonna happen to this nut its gonna explode and then the nut is gonna be smushed you heard it here first three two one wow Music it seems to have ejected all of the nut nut dust called it nice all right what do we got here some silly putty any predictions satisfying smoosh its not nearly as exciting it seems like it would be really fun boring snore i got jawbreakers but theyre too big red hot jawbreakers shatter it thats pretty much what i was expecting i think what about five jawbreakers same thing i ate too many jawbreakers oh i think im going to be sick oh disgusting its just all stuck to the roof of his mouth like peanut butter a glass marble this doesnt sound dangerous i just feel like everythings gonna shatter come on give him something hard a hardened steel ball bearing this is the hardest jawbreaker like split in half or something i dont know these bearings are really hard so were going with four shells im a little concerned about four if it does crack it could really come flying out fast so were gonna evacuate the wife get out of here doing four doing live three two one wow it just pressed the ball bearing into the jaws and the bottom the bearing is so hard its forming the steel jaws around it a literal jawbreaker it should be able to shatter this i think i just need harder jaws well save that one for later because right now were gonna do what weve all been waiting for metal nut versus four shells i think thats gonna get smooshed you think its gonna smush this three two one Music all right this looks great it looks like that cool s estee drawn middle school anyone whos cool in any grade it was one piece but with a little touch it snapped i cant believe it actually cracked the nut success my lifes work crack can i fire it uh yeah what do you want to smash what a battery i thought the same thing but thats a really bad idea do you have an extra lego men that i have i forgot that i had festivities very festive okay now we can lego man this little man went through santas court of law and has been deemed naughty oh yeah he goes into the nutcracker who is santas enforcement who is currently loaded with four shells two one Music Applause Music swap both arms all wow santa doesnt mess around dont get on the naughty list kids absolute carnage Music wow oh my gosh his poor head its like it split open in the back i think hes dead pretty crazy the big question i already married is do you think that this is an important invention a little i cant say no with him looking right at me all right hes not looking no i mean whats it gonna change the world id like to see it try see what she says behind his back its ridiculous i think hes very twofaced and very festive and very handsome do you like it yes you gotta ask the right questions this is my favorite kind of project its awesome arguably useless and definitely commercially not viable and that means that after all the engineering there isnt really anything for me to sell which makes projects like this really dependent on your support so im not trying to make you feel guilty if you enjoyed the video im happy but if you do want to help support projects like this there are a couple things that you can do so you might have noticed this awesome shirt that im wearing this is a blueprint of my unpickable lock and if you like it you can get one at stuffmatehere.shop another thing you can do is support these projects directly on patreon and the last thing you can do is take a minute to check out this video sponsor kiwico what kiwico does is every month they send you a crate that has everything that you need to do some kind of project and this one is an automaton it simulates santa coming to deliver all of your presents the reason i love these crates so much is they go beyond the project to teach you general concepts and in the case of this project santa is driven by a set of differently timed cams my unpickable lock actually had a bunch of cams people who see my projects like the unpickable lock ask me where did i learn to design stuff like that and it all started with a steady diet of kits and projects just like these theyre just a great investment in a kids future at this point pretty much all the kids in my extended family are getting kiwico crates if youre thinking about useful gifts that you could give a loved one you should check out kiwico and if you go to kiwico.com stuff made here theyll give you off your first month and they have a ton of different crates for different age groups and interests and thats it thank you kiwiko for sponsoring this video and thank you for taking the time to listen Applause Music Music you

write a youtube script

Im making a D printed tabletop bowling alley and these are the lock in place bumpers the bumpers deploy and lock into its top position to release the lock you press a latch and it folds down I was pretty happy with this and I think the mechanism works well but at the end of the day when its locked its pretty Wiggly the mechanism hinges around these M bolts one way to reduce how Wiggly the mechanism is to reduce the size of the hole around the bolts look how many holes I have across this model and theyre all across different sketches so I made a variable called m three hole and that is linked to all the different holes so now I only have to change the diameter here and itll update all the holes across the entire model I reprinted the mechanism and now its way less Wiggly look at this I know that may seem simple or obvious but Im always looking for ways to make my modeling more efficient now you can lie and say it took three times longer than it actually did

write a youtube script

The Gradient of Generative AI Release Methods and Considerations Irene Solaiman Hugging Face irenehuggingface.co Abstract As increasingly powerful generative AI systems are developed the release method greatly varies. We propose a framework to assess six levels of access to generative AI systems fully closed gradual or staged access hosted access cloudbased or API access downloadable access and fully open. Each level from fully closed to fully open can be viewed as an option along a gradient. We outline key considerations across this gradient release methods come with tradeoffs especially around the tension between concentrating power and mitigating risks. Diverse and multidisciplinary perspectives are needed to examine and mitigate risk in generative AI systems from conception to deployment. We show trends in generative system release over time noting closedness among large companies for powerful systems and openness among organizations founded on principles of openness. We also enumerate safety controls and guardrails for generative systems and necessary investments to improve future releases. Introduction What constitutes a robustly safe and responsible release of new AI systems from components such as training datasets to model access itself urgently requires multidisciplinary guidance. There is no overarching standard or standardsmaking body to form consensus for what constitutes responsible release. This is particularly true for generative AI systems which can be leveraged for a broad range of tasks and are sometimes referred to as generalpurpose . A subtype of foundation models these systems generate outputs across modalities such as text and image. They can be applied to both beneficial and harmful tasks. For example language models can be adapted to tasks such as grammar correction and translation or be used for phishing and spam. The relative novelty means new uses and misuses of these systems are still being discovered. Carefully considering release strategies for present and less powerful systems better prepares and sets precedent for the AI community and the broader affected community as increasingly powerful systems are deployed. Due to the fast pace of AI progress and release developers deployers researchers and policymakers must take action via community discussions guardrails and investments. The current state of generative AI system release is largely determined by the organizations developing systems. Developers are most likely to best understand the system but understanding impact and ripple effects requires multidisciplinary expertise that is rarely housed in one organization. Waiting for longterm evidence of consequences is infeasible for highrisk and powerful systems. A strictly closed and vertical process to commercialization can lead to concentrated power among highresourced organizations. An open process without ethical considerations can inflict and exacerbate risk and harm from misuse to bias. Many components make up a system throughout its lifecycle from training data to computing power. This paper will primarily focus on the cumulative release of a model and its components by outlining Preprint. Under review. arXiv.v cs.CY Feb key considerations in release release options along the gradient the timeline of released systems and necessary investments to improve safe releases. Previous Work Discussion about safe release has been ongoing in the AI research community but there is no standards body or default convener for these discussions. Initiatives such as the Partnership on AIs Publication Norms for Responsible AI and Stanford Universitys Call for Community Norms for the Release of Foundation Models have made space for discussion around the many options for system releases and the many components involved and available platforms that have made options complex. At the core of release considerations is the tension between openness that shares rather than concentrates power and closedness that minimizes potential harm and risk. Broadly safety and risk control development lags behind system development for example tools for detecting generated outputs underperform as systems become increasingly powerful . For researchers in natural language processing whether developers should be ethically responsible for downstream misuse of a publicly released system is contentious with about half of researchers believing professionals should be responsible. Parallel fields such as opensource software deployment can share informative lessons such as the ability for open source software communities to enable community research and crowdsourcing work such as discovering vulnerabilities . Examining the specific use case of opensource Deepfakes highlights the difficulty of managing downstream harms in realtime and the risk of safety controls being seen as futile . While generative AI research would also greatly benefit from this community insight lessons from software are not often directly applied due to substantial difference in functionality . Ultimately personal values around openness are a large factor in decision making and tensions can be further examined by output modality . What is Being Released? The parts of an AI system considered in a release can be broken into three broad and overlapping categories access to the model itself components that enable further risk analysis and components that enable model replication. Components are organized based on their most straightforward use. There is overlap among these components the same model cannot be replicated without its component for risk analysis such as its entire original training data even if all replication components are available. Conversely components for replication can also be analyzed for social impacts such as biases. . The Model Itself Access to the model itself includes the model weights and the ability to query adapt or otherwise examine and conduct further research into a model. The range of access is expanded according to the gradient in Figure . . Components for Risk Analysis These components are the parts of system development that could provide further insight into the model the models capabilities the decision making process on what data was collected and how and documentation of the process. Additionally this details system risks training data finetuning data and information on people and human crowdworkers involved in adapting the model through methods such as reinforcement learning with human feedback. This also includes evaluation results published results from any evaluations that researcher and developers may have run on the base model. These components may be withheld due to intellectual property IP rights consent or privacy concerns. . Components for Replication These components include a technical paper detailing the model training process and code used to train the model as these can ease replication efforts. This also includes training information such as configuration settings e.g. batch size and telemetry collected during training e.g. training loss. These components may be withheld for competition IP and misuse reasons. They are also high risk for misuse concerns as they can be repurposed or adapted to malicious or otherwise harmful use cases . Key Considerations in Release Deployers should weigh the following considerations when making release decisions. Risks and threats from increasingly powerful systems are difficult to enumerate and assess especially since malicious actors and their incentives are constantly evolving . Taxonomies of ethics and risks of specific systems can serve as a framework for potential harms. Specific considerations across all generative systems are listed below. . Concentration of Power One of the most prominent arguments for providing access to systems is to avoid concentrating the level of power that highresource organizations are collecting as one of the few groups capable of developing and deploying these systems. Large technology companies are able to create powerful AI systems because of their access to training data computing infrastructure and commercial capabilities for deploying that system. This monopolization also gives these highresource institutions more influence in AI development the behavior of these systems and the narrative and direction of the field . Although these companies may provide access or even opensource their systems contributions to system development are limited to people and resources working towards that companys interests . Large companies are often geographically concentrated in Western countries whereas systems are deployed globally which can asymmetrically impose cultural values . These companies can also punish pushback or dissent . The people most affected and exploited by AI systems are rarely found in large technology companies. They must be empowered to shape systems that also benefit them or to opt out of interaction with AI entirely . . Exacerbating Disparate Performance and Harmful Social Impacts The fewer perspectives that are incorporated into the system development process leads to higher likelihood the system performs disparately for different groups. AI systems can propagate harms such as exacerbating social inequity and harmful biases which can be further amplified in larger systems as scale increases . Means of measuring and mitigating risk in these systems are largely cultural and contextdependent . The many technical and social aspects of AI systems require robust research conducted with communities affected to ensure these systems benefit and do not exploit marginalized groups if the systems are to be deployed among these groups. . Malicious Use and Unintentional Misuse With more modalities of AI generation improving in output quality from high quality text to high quality images the potential for harmful use cases also increases. Malicious uses such as the creation of deepfake imagery AIgenerated disinformation and illegal and disturbing material can cause severe emotional harm at the individual level and destructive institutional harm at the societal level. Furthermore malicious actors have historically worked to circumvent safety controls. Threat modeling will necessarily differ by modality but as systems improve in types of outputs such as code generation potential harms can also broaden . While limiting access can prevent some malicious uses and is often a suggested action to minimize misuse systems can still be vulnerable to attacks with only querying functionality available . . Auditability The question of auditability addresses who is conducting audits and the level of access required to effectively examine an AI system. Auditing must be considered both pre and postdeployment as impacts from a system may not be detectable predeployment and when deployed impacts may be difficult to trace back to a specific system . The actors conducting and capable of conducting audits will likely require some level of technical skill even when numerous nocode tools are built. The size of the system and its components also determine auditability the datasets that large generative AI systems are trained on are not only difficult to analyze at scale but few tools exist to analyze large static datasets . Formal audits alone cannot be the only insight or governance of a system . . Accountability in Case of Harm In the case that an AI system harms or is connected to harming people who or what is to be held accountable is unclear. More open and deployed systems have a higher likelihood of a broader reach and therefore a higher chance of harm. Since harm is not explicitly defined and not always physical what constitutes harm can have a large range. The range may include encouraging physical harm propagating social harms such as identity stereotypes and more abstract harms such as lack of access to a system lowering opportunities for a specific group. Work to characterize sociotechnical harms can narrow the scope . . Value judgments for gating and limiting access A base generative AI system is capable of many types of content making content moderation complex . What constitutes appropriate outputs is influenced by religion cultural and personal beliefs. What content can and should be limited filtered and gated is also vague. For example sexual content may not be inherently unsafe to generate in some cultures but may be subject to local laws. Technical filters may not be able to distinguish between sexual content and nonsexual nudity and may not be able to distinguish between consensual and nonconsensual content. While most difficult without specific use cases or context specific applications face the same challenges. The Gradient of System Access Once considerations are taken into account the group determining release method must choose if the system and its components are publicly acknowledged and released. The below gradient of release options are based on five years of publicized generative AI systems. This gradient of options serves as a framework and does not fully capture the nuance of the many components and details in a system release. Figure shows the tradeoffs in considerations along the gradient as systems become more open they better enable audits and community research but are more difficult to control for risks. internal research only high risk control low auditability limited perspectives community research low risk control high auditability broader perspectives fully closed gradualstaged release hosted access cloudbasedAPI access downloadable fully open PaLM Google Gopher DeepMind Imagen Google MakeAVideo Meta GPT OpenAI Stable Diffusion Stability AI DALLE OpenAI Midjourney Midjourney GPT OpenAI OPT Meta Craiyon craiyon BLOOM BigScience GPTJ EleutherAI gated to public System Developer L e v el o f A c c e s s Considerations Figure Considerations and Systems Along the Gradient of System Access Below the gradient are examples of generative systems placed according to their original release method upon announcement for example GPT may fall under Downloadable today but was originally released as GradualStaged. . Fully Closed When all aspects and components of a system are inaccessible outside the developer organization or even closed outside a specific subsection of an organization the system is fully closed. At the furthest end of the spectrum the systems existence is unknown outside a select group within the developer organization even after full training. A fully closed system may or may not include some form of public announcement that the system exists. These systems can only be researched by the developer organization which is often a highresource organization such as the Alphabet companies Google and DeepMind. Some publiclyknown systems such as Googles Imagen and DeepMinds Gopher are examples. Public engagement may come from the system being deployed in a commercial application or the public calling out biases and notable social aspects of a system from public releases. These releases can be cherrypicked for example showing only nonhuman animals or human silhouettes which does not give robust insight to broad capabilities or social impacts such as biases. . GradualStaged Release This method refers to releasing a system in stages or gradually over a predetermined amount of time. The time between stages is intended for investments that minimize risk such as monitoring for malicious actor activity and conducting research on potential harms. In OpenAI stagereleased language model GPT in four sizes by increasing parameter count over nine months while conducting research internally and with external partners . This sparked debate among some but still is a recommended tactic among others . In Stability AIs Stable Diffusion initially approached a stage release by providing access to a hosted model before releasing the model weights. However model weights were leaked days after their initial hosted release. This exemplifies the need to inject safety protocols and prevent leakage during this approach. While there is no standardized time frame for staged releases generally substantial sociotechnical research requires multiple weeks months and sometimes years. . Gated to Public Access Including Paid and Free When providing access to a system without fully opening all components actors deciding release method may choose to place access limitations. Above the infrastructural limitation options namely hosting cloudbased access or fully downloadable access is the choice to make the release gated or public. Gating system access is a selective process by group of people used to block high risk or outofscope use cases. Limited access can make enforcing controls easier for example system deployers withhold the right to revoke access in a gated and hosted access setting. However gating downloadable systems is unreliable as a technical mechanism the network effect of researchers sharing within the same circles can provide a loophole to gating. The releasing organization cannot fully monitor whether users are sharing access through screensharing credentialsharing or simply sending components such as model weights to unauthorized users. This does not mean this is an ineffective guardrail as it still creates barriers to sharing the model. The deploying organization will still be making critical decisions that contribute to concentration of power. .. Hosted Access System deployers may provide access to the model itself by hosting the model on their own servers and allowing surfacelevel interfacing. Access can differ depending on the interfaces usability especially for users with minimal or no experience with these systems. Generally users are unable to perform tasks outside what is prescribed usually simple inputoutput probing. This method is specific only to model access not access to other system components. Examples include Midjourney which allows users to interact with its image generation model via Discord bot or web interface . When optimized for usability and dialogue as seen with OpenAIs ChatGPT broader perspectives can interact with and the model but raises misuse concern and ethical challenges . This method can also transition to API or downloadable as seen with OpenAIs DALLE was switching to APIbased access seven months later . While this method provides some model access it limits external research ability. .. CloudbasedAPI Access Cloudbased access or access provided via application programming interface API provides more insight and researchability into a model than Hosting but still allows for restrictive functionality. Some APIs only allow for querying such as OpenAIs original GPT release via API . Additional functionality can be added such as finetuning via API. Similar to Hosting this method is specific only to model access. Nonreleased components and system information can be determined via tools such as EleutherAIs evaluation harness used to determine GPT parameter sizes via OpenAIs API . This method is favorable for structured access where research is possible but can still be tracked and is unlikely to create a modified version . Cloudbased access can track users and their activity to monitor for risky behavior. This can also better enforce safety controls such as rate limiting. . .. Downloadable The main distinction between downloadable and fully open systems is the withholding of system components such as training dataset availability. Downloadable systems can also be gated. Downloadability does not inherently imply full access to any user granted access as the size of a model can limit who is capable of running a modal locally. Personal and standard consumer hardware is unlikely to support large and powerful models. The infrastructure needed to run large models creates an access barrier. In response industry and public initiatives are creating accessible infrastructure for researchers. Downloadable models better enable robust research but are difficult to track for potential misuse or harm. This method also eases user ability to erode or disable safety controls such as content filters. . Fully Open When all aspects of the system are accessible and downloadable including all components the system is fully open. These systems cannot be gated and by definition are fully public. For the purposes of this framework a basic level of accessibility and documentation across components qualifies a system release as fully open but releases may differ in documentation detail and levels of granularity. The most prominent fully open systems were developed by organizations founded on the principle of openness. EleutherAI is a decentralized collective also prioritizes transparency and has released all system components as seen in their GPTJ and GPTNeo language models and the Pile dataset . The BigScience global research community of over researchers developed the BLOOM language model in the open . Over working groups covered aspects from dataset creation to carbon footprint to modeling approach to optimize for a multilingual system created transparently . While openness does enable broader research that can engage many peoples it can also enable dangerous uses and model creation and controls can be difficult to enforce. Trends in System Releases We analyze release trends across prominent base generative AI systems this does not include finetuned or updated systems such as models that undergo reinforcement learning with human feedback. These figures are based on tracking and evaluation initiatives are not exhaustive and intend to show release trends over time. . Timelines for Large Language Models When examining systems by the original method of release over time trends seen in Figure show closing and limiting language model access as more common since GPTs staged release. Language models with fewer than six billion parameters have generally been towards the open end of the gradient but more powerful models especially from large companies tend to be closed. This can be due to their requiring deeper consideration and safeguards due to risk potential but Figure also illustrates the high number of large companies able to develop and close language models. Fully Closed GradualStaged Hosted CloudbasedAPI Downloadable Fully Open Figure Language Model Release Method By Parameter Count Over Time . Timelines for All Modalities As more generative modalities are developed from image to audio to video they face similar release decision challenges. Figure shows system release over this same time period. As there is no standard means to compare capabilities across modalities all levels of system capability are placed equally. Again trends show openness until GPTs staged released. This timeline also shows a sharp increase in the amount of systems developed and closed after . The systems most commonly toward the open end of the gradient are developed by smaller organizations founded with the intent to be open. Conversely many systems from large companies are becoming closed or have closed components. OpenAI is the most common company to restrict but not fully close or open access. Alphabet companies Google and DeepMind are most common among closed systems. Across modalities large companies have steered toward closedness. Open initiatives from large companies are shown to release a downloadable model trained on public datasets crafted by other organizations as seen with Metas OPTB . It is unclear at this time whether movements towards openness will pressure historically closed organizations to adjust their release strategies. Safety Controls and Guardrails A combination of controls and guardrails largely from the developer and deploying organizations but also from external researchers can complement each other in order to address the above considerations and risks. Many of these methods are pioneered and honed in research environments and outside developer organizations. Individually no one control can serve as a panacea. While it is possible to add controls and guardrails long after deployment these options are most effective when deployed simultaneously with system release. . Documentation and Transparency Structured documentation that clearly communicates critical information about each component of the system gives further insight to the system and can take many forms. Proposed approaches to documentation at dataset and model levels have proven successful without any enforcement mechanisms many releases across AI companies include some form of this documentation. Datasheets for datasets See Appendix A for logo and developer key GPT GPT Turing NLG Jukebox GPT DALLE BERT CTRL Gopher GPTNeo CPM M PanGu ERNIE . Wu Dao Jurassic Hyper CLOVA Yuan ERNIE Titan Cohere LM PAG nol XGLM LaMDA Megatron Turing GPT NLG GPT NeoX SW Chinchilla PaLM DALL E Lyrafr Imagen OPT Craiyon YaLM Minerva Midjourney BLOOM Stable Diffusion MakeAVideo Phenaki Audio LM ELMo GPTJ Galactica Chat GPT Anthropic LM Fully Closed GradualStaged Hosted CloudbasedAPI Downloadable Fully Open Figure Release Methods Over Time All Modalities communicates aspects of datasets such as creators motivations collection process and overall composition. Metas OPTB release included a datasheet in its appendix . Data statements for natural language processing are another popular tool more tailored to languagebased systems seen used by bias measurement dataset CrowSPairs . Model cards have been popular as seen in Googles PaLM OpenAIs GPT and GPT and Runway Research and Stability AIs Stable Diffusion . Model cards are deployed across Hugging Faces platform and have evolved to be interactive . System cards blend datasheets and model cards and have been used for DALLE . . Technical Tools Technical tools can address specific technical safety concerns but cannot be a substitute for addressing complex societal problems. In some cases technical tools can create new social harms and should therefore be vetted and combined with other guardrails. .. Rate Limiting Constricting the amount of outputs a user can generate via cloudbased access is a popular means of preventing attacks and harmful generations. Rate limiting also helps a system perform well and protect underlying infrastructure from being overloaded. This defensive measure can be enforced with common strategies such as a token bucket which tracks and limits usage according to a set number of tokens that can refresh or accumulate on a predetermined time frame. As an example OpenAIs DALLE s public API rate limits external users . This can be adapted for users whose applications have been cleared as safe. .. Safety and Content Filters Filters developed to trigger blank responses when given an unsafe input are popularly deployed across varying levels of access. This can help block illegal and egregious content. Developers selecting these trigger categories must make normative judgements about what input content blocks generation. Stable Diffusions safety filter was found to primarily prevent generations with sexually explicit content but not violence and gore which is a normative judgment about the safety of both categories in generated images. Blocking generations for socially sensitive topics can result in entire identity groups being blocked. Lessons from social media platform content moderation highlight harms such as community erasure especially among marginalized groups . .. Detection Models While methods to detect AI generated outputs can vary and include human detection detection models can be a helpful tool especially for less powerful generative systems. While the human eye alone can detect outputs from less powerful systems such as Craiyon for AIgenerated images detection models can have higher accuracy for more powerful systems. This is particularly important when models are deployed in highstakes settings . As system output quality improves that distinction becomes more difficult for both humans and AI detection models. Approaches to detection can be tailored to modality such as text and include human annotation . Detection models can also differ based on type of generation within a modality such as facial generations . .. Hardcoding Responses Predetermined safe outputs triggered for a given input can be hardcoded into a model interface. This can aid legal compliance or provide standardized responses for highrisk inputs. Similar to filtering determining trigger inputs or trigger categories requires normative judgements about what constitutes unsafe inputs and what constitutes appropriate outputs. This can not only lead to community erasure but also impose these normative beliefs onto users. .. Watermarking The concept of digital watermarking media can be transferred to AI systems to protect against model theft protect IP and more easily identify AIgenerated outputs. Encoding a unique identifier in generated outputs can aid in detecting media as AIgenerated and synthetic and trace the output to a specific model. Research strives ensure these watermarks are invisible to the humaneye do not affect output quality and tamperproof from model attacks and alterations like finetuning via methods such as embedding noise as watermarks . Different approaches to watermarking can be deployed for different needs from the embedding method to easily determine whether an output is synthetic to linking watermarks to a model owners identity for authentication purposes . There are no current prominent successful case studies as watermarking has not yet been publicly deployed at scale for large generative systems. .. Model Weight Encryption Encryption can be used in order to protect model weights often to protect from model stealing and to protect IP. This allows only an authorized user with the key to use the model. s proposed NNLock does not change model structure so as not to adversely affect model performance. proposed an obfuscation framework that only authorizes users with a trustworthy hardware device. notes many existing IP protection methods are not robust to model attacks and are not suitable for commercial purposes as they verify model ownership but not user identities. .. Updating Adapting or Retraining models Models can be adapted in a way that mitigates risk. Popular methods include finetuning for example finetuning GPT on valuestargeted datasets or finetuning LaMDA on annotated data to improve factual grounding . Another method is reinforcement learning with human feedback as seen with InstructGPT and its opensource replication effort at CarperAI . These methods result in new models different from their base models but often improved along a safety parameter. . Community and Platform Efforts Communitydriven approaches to risk mitigation leverage new and varied viewpoints. Bounty programs from bug bounties to bias bounties can raise unforeseen safety issues and strengthen trust in a system . Bias bounties by nature benefit from diverse perspectives. Community moderation communitybased content flagging and naming and shaming techniques on a platform enables users to determine and stop harmful content before it escalates. Monitoring and logging inputs by a user on the backend helps track trends in harmful or extremist behavior. . Organizational and Platform Policies Organizational and platform policies can guide and enforce safe human interaction with generative AI systems. These policies can have drawbacks they may protect from harm but also limit beneficial uses. For example limiting access in a region under active war can prevent disinformation generation but also general access. Internal risk policies should provide a process for what considerations must be weighed and how to evaluate each prior to determining release options. If the system is deployed on a platform or on a given interface a code of conduct for engaging with the platform and other users prevents direct harm on the platform at risk of losing platform access. Mandating user accounts on a platform helps track specific users and their activity which supports community and platform efforts. Sharing policies that outline what can and cannot be posted on other platforms or for uses outside of personal use prevents harmful content from spreading and inciting further harmful content. . Legal Recourse Legal measures such as licenses are an enforceable control when a user uses a system in a way the deployer prohibited. The Responsible AI License RAIL places behavioral use conditions on a model with the model owner owning the license and responsibility for pursuing enforcement if need be . Both BigSciences BLOOM model and Runway Research and Stability AIs Stable Diffusion use RAILs. Licenses are difficult to enforce for downloadable or fully open systems as model behavior and uses cannot be fully monitored. Legal enforcement can also be costly in terms of both time and financial resources. Example cases studies are examined in . Necessary Investments for Responsible Release Developers and researchers must listen and leverage multidisciplinary and often external expertise especially for guardrails. Policymakers must mandate safety where possible and technically feasible and provide resources for the underresourced. Regardless of level of access generative AI systems are capturing reflecting and amplifying aspects of society that require multiple perspectives in exploratory and risk control research. Since a system cannot be fully safe or unbiased for all groups of peoples and there is no clear standard for when a system is safe for broad public release further discourse across all affected parties is needed. Research and decisions made now will inform considerations for increasingly powerful systems across modalities in the future making early investment crucial. . Accessible Interfaces and Low and Nocode Tools In order to make generative systems accessible to the many peoples they affect means of interacting with a system such as a model demo are needed. A clean easily usable interface that accommodates disabilities and all levels of technical comfort significantly improves accessibility. This step towards further openness can push a system toward the far end of the gradient with less risk control and increased redteaming. Accessible interfaces with lowbarrier sharing can also better enable crossfield collaborations . Largescale probing can reveal flaws as seen with Metas Galactica language model which was released with a demo. The demo was retracted within three days due to the public naming risks such as disinformation generation . Both computer science training and low to nocode interfaces are necessary to streamline sociotechnical research. Moral experiments show varying approaches to ethical problems by background and culture which are urgently needed perspectives in building evaluating and deploying new AI systems. Effective design and user interface must be optimized for experts outside of computer science . . Closing Resource Gaps Resource gaps mainly between among major labs research groups and academia have widened . In addition to the gap hindering groups from developing systems at the same level of performance it also hinders the ability to build and run exploratory research projects. The monetary infrastructure and sometimes skills limitations can bar especially underrepresented groups from contributing to understanding and mitigating risks. Public sector investment at national and global levels can start to bridge this gap. Grants from developer labs can also sponsor thirdparty research but should have builtin mechanisms for also allowing critical research. Infrastructure grants for computer clusters can enable smaller research groups to engage with powerful systems. Skillbuilding requires longerterm investment. . Technical and Practical Ethics Training Increasing access to social scientists and the many multidisciplinary experts underrepresented in the AI research community is insufficient. The technical barriers to evaluating and improving or mitigating harms of AI systems can slow or hinder critical research. Conversely the lack of practical ethics and science and technology studies STS training among technical professionals prevents thoughtfully integrating societal guardrails from project conception throughout the development process. Training must be implemented at early stage education academic courses and curricula in computer science must integrate social and ethical considerations. Social sciences geared toward examining AI systems must foster technical understanding. . Expert Foresight Experts in relevant disciplines should be included while relative risk is low. As generative AI systems become a higher risk for specific applications and fields such as disinformation and medical advice correlating experts should be tapped in and begin foundational work in mitigating that risk. The rapid rate of AI development means substantial research should be planned in anticipation to prevent the trend of detection and mitigation trailing capabilities advancement. For example research conducted on radicalization risks of language models using GPT and GPT show that GPT has significantly higher risk potential in generating extremist text . Starting this research with less powerful systems can better inform future mitigation efforts. . Multidisciplinary Discourse Increasing access to social scientists and the many multidisciplinary experts underrepresented in this discussion is insufficient. Critically actors in this space must have some incentive to engage in frequent community discussion and be held accountable to commitments for safe releases. Googles public position on responsible AI practices encourages inhouse risk evaluation and mitigation but conflicts of interest can result in internal critics being unable to share or publish findings and dismissal . These initiatives can also be formed as an industry argument for selfregulation but ultimately lack external accountability. A new thirdparty convening body can help facilitate this discourse. Instead of relying on existing fora conversations can take lessons from social and abolitionist movements in how to include underrepresented and affected communities . . Enforcement Mechanisms for Unsafe Release Ultimately all actors involved in the release of a powerful AI system must have some incentive to conduct releases safely. But enforcing responsible release requires a definition for what constitutes responsible release. Responsible is distinct from safe and can emphasize meeting all possible enforceable safety guardrails pre and postrelease. Regulation can mandate that releases include system documentation and auditing for highrisk or highimpact releases. Updatable policies can recommend certain risk controls and guardrails and policy bodies can better fund risk research and development of further evaluations and controls. Conclusion The gradient of generative AI system release shows the complexity and tradeoffs of any one option. Releases must balance concentration of power and AI risks in addition to considering precedent for future releases as system capabilities increase. Developers and deployers regardless of release method preference must engage multidisciplinary experts and the AI community to better form norms for safe release. Existing and evolving risk controls and guardrails require developer deployer researcher and policymaker action and can mitigate some foreseeable harms but longterm investments in disciplines and discourse across the AI community and among affected peoples are necessary. Acknowledgments and Disclosure of Funding Thank you to Hugging Face for funding this research. Thank you to Joshua Achiam Stella Biderman Miles Brundage Clmentine Fourrier Yacine Jernite Margaret Mitchell Percy Liang and Sonja SchmerGalunder for their thoughtful feedback on earlier versions of this paper. Thank you to Johann Christensen for testing figures. References A. Abid A. Abdalla A. Abid D. Khan A. Alfozan and J. Y. Zou. Gradio Hasslefree sharing and testing of ML models in the wild. CoRR abs. . URL http arxiv.orgabs.. M. Alam S. Saha D. Mukhopadhyay and S. Kundu. Nnlock A lightweight authorization to prevent ip threats of deep learning models. J. Emerg. Technol. Comput. Syst. apr . ISSN . doi .. URL httpsdoi.org.. E. Awad S. Dsouza R. Kim J. Schulz J. Henrich A. Shariff J. F. Bonnefon and I. Rahwan. The moral machine experiment. Nature . URL httpswww.nature.comarticless. N. Benaich and I. Hogarth . URL httpswww.stateof.aireportlaunch. html. E. M. Bender and B. Friedman. Data statements for natural language processing Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics . doi .tacla. URL https aclanthology.orgQ. E. M. Bender T. Gebru A. McMillanMajor and S. Shmitchell. On the dangers of stochastic parrots Can language models be too big? . In Proceedings of the ACM Conference on Fairness Accountability and Transparency FAccT page New York NY USA . Association for Computing Machinery. ISBN . doi .. . URL httpsdoi.org... R. Benjamin. Race after technology Abolitionist Tools for the new jim code. Polity . Y. Benkler. Dont let industry write the rules for ai. Nature News May . URL https www.nature.comarticlesd. A. S. Bergman G. Abercrombie S. Spruit D. Hovy E. Dinan Y.L. Boureau and V. Rieser. Guiding the release of safer EE conversational AI through value sensitive design. In Proceedings of the rd Annual Meeting of the Special Interest Group on Discourse and Dialogue pages Edinburgh UK Sept. . Association for Computational Linguistics. URL httpsaclanthology.org.sigdial.. BigScience. Bigscience rail license v. . URL httpshuggingface.cospaces bigsciencelicense. A. Birhane P. Kalluri D. Card W. Agnew R. Dotan and M. Bao. The values encoded in machine learning research. In ACM Conference on Fairness Accountability and Transparency FAccT page New York NY USA . Association for Computing Machinery. ISBN . doi ... URL httpsdoi.org... S. Black S. Biderman E. Hallahan Q. Anthony L. Gao L. Golding H. He C. Leahy K. McDonell J. Phang M. Pieler U. S. Prashanth S. Purohit L. Reynolds J. Tow B. Wang and S. Weinbach. GPTNeoXB An opensource autoregressive language model. In Proceedings of the ACL Workshop on Challenges Perspectives in Creating Large Language Models . URL httpsarxiv.orgabs.. S. L. Blodgett S. Barocas H. Daum III and H. Wallach. Language technology is power A critical survey of bias in NLP. In Proceedings of the th Annual Meeting of the Association for Computational Linguistics pages Online July . Association for Computational Linguistics. doi .v.aclmain.. URL httpsaclanthology.org.aclmain.. F. Boenisch. A systematic review on model watermarking for neural networks. Frontiers in Big Data . ISSN X. doi .fdata... URL https www.frontiersin.orgarticles.fdata... R. Bommasani D. A. Hudson E. Adeli R. Altman S. Arora S. von Arx M. S. Bernstein J. Bohg A. Bosselut E. Brunskill E. Brynjolfsson S. Buch D. Card R. Castellon N. S. Chatterji A. S. Chen K. Creel J. Q. Davis D. Demszky C. Donahue M. Doumbouya E. Durmus S. Ermon J. Etchemendy K. Ethayarajh L. FeiFei C. Finn T. Gale L. Gillespie K. Goel N. D. Goodman S. Grossman N. Guha T. Hashimoto P. Henderson J. Hewitt D. E. Ho J. Hong K. Hsu J. Huang T. Icard S. Jain D. Jurafsky P. Kalluri S. Karamcheti G. Keeling F. Khani O. Khattab P. W. Koh M. S. Krass R. Krishna R. Kuditipudi and et al. On the opportunities and risks of foundation models. CoRR abs. . URL httpsarxiv.orgabs.. G. Brockman M. Murati P. Welinder and OpenAI. Openai api Jun . URL https openai.comblogopenaiapi. M. Brundage S. Avin J. Clark H. Toner P. Eckersley B. Garfinkel A. Dafoe P. Scharre T. Zeitzoff B. Filar H. S. Anderson H. Roff G. C. Allen J. Steinhardt C. Flynn S. . higeartaigh S. Beard H. Belfield S. Farquhar C. Lyle R. Crootof O. Evans M. Page J. Bryson R. Yampolskiy and D. Amodei. The malicious use of artificial intelligence Forecasting prevention and mitigation. CoRR abs. . URL httparxiv. orgabs.. M. Brundage S. Avin J. Wang H. Belfield G. Krueger G. Hadfield H. Khlaaf J. Yang H. Toner R. Fong T. Maharaj P. W. Koh S. Hooker J. Leung A. Trask E. Bluemke J. Lebensold C. OKeefe M. Koren T. Ryffel J. Rubinovitz T. Besiroglu F. Carugati J. Clark P. Eckersley S. de Haas M. Johnson B. Laurie A. Ingerman I. Krawczuk A. Askell R. Cammarota A. Lohn D. Krueger C. Stix P. Henderson L. Graham C. Prunkl B. Martin E. Seger N. Zilberman S. . higeartaigh F. Kroeger G. Sastry R. Kagan A. Weller B. Tse E. Barnes A. Dafoe P. Scharre A. HerbertVoss M. Rasser S. Sodhani C. Flynn T. K. Gilbert L. Dyer S. Khan Y. Bengio and M. Anderljung. Toward trustworthy ai development Mechanisms for supporting verifiable claims . URL httpsarxiv.orgabs. . L. Cai and Y. Zhu. The challenges of data quality and data quality assessment in the big data era. Data Sci. J. . A. Caliskan J. J. Bryson and A. Narayanan. Semantics derived automatically from language corpora contain humanlike biases. Science . doi .science. aal. URL httpswww.science.orgdoiabs.science.aal. N. Carlini F. Tramr E. Wallace M. Jagielski A. HerbertVoss K. Lee A. Roberts T. B. Brown D. Song . Erlingsson A. Oprea and C. Raffel. Extracting training data from large language models. CoRR abs. . URL httpsarxiv.orgabs. . CarperAI. Carperai an eleutherai lab announces plans for the first opensource instructiontuned language model. Oct . URL httpscarper.ai instructgptannouncement. A. Chakraborty A. Mondal and A. Srivastava. Hardwareassisted intellectual property protection of deep learning models. In Proceedings of the th ACMEDACIEEE Design Automation Conference DAC . IEEE Press . ISBN . A. Chowdhery S. Narang J. Devlin M. Bosma G. Mishra A. Roberts P. Barham H. W. Chung C. Sutton S. Gehrmann P. Schuh K. Shi S. Tsvyashchenko J. Maynez A. Rao P. Barnes Y. Tay N. Shazeer V. Prabhakaran E. Reif N. Du B. Hutchinson R. Pope J. Bradbury J. Austin M. Isard G. GurAri P. Yin T. Duke A. Levskaya S. Ghemawat S. Dev H. Michalewski X. Garcia V. Misra K. Robinson L. Fedus D. Zhou D. Ippolito D. Luan H. Lim B. Zoph A. Spiridonov R. Sepassi D. Dohan S. Agrawal M. Omernick A. M. Dai T. S. Pillai M. Pellat A. Lewkowycz E. Moreira R. Child O. Polozov K. Lee Z. Zhou X. Wang B. Saeta M. Diaz O. Firat M. Catasta J. Wei K. MeierHellstern D. Eck J. Dean S. Petrov and N. Fiedel. Palm Scaling language modeling with pathways . URL httpsarxiv.orgabs.. D. Contractor D. McDuff J. K. Haines J. Lee C. Hines and B. J. Hecht. Behavioral use licensing for responsible AI. CoRR abs. . URL httpsarxiv.orgabs .. A. Crisan M. Drouhard J. Vig and N. Rajani. Interactive model cards A humancentered approach to model documentation. In ACM Conference on Fairness Accountability and Transparency. ACM jun . doi ... URL httpsdoi.org .F.. L. Dabbish C. Stuart J. Tsay and J. Herbsleb. Social coding in github Transparency and collaboration in an open software repository. In Proceedings of the ACM Conference on Computer Supported Cooperative Work CSCW page New York NY USA . Association for Computing Machinery. ISBN . doi .. . URL httpsdoi.org... P. Dave and J. Dastin. Google told its scientists to strike a positive tone in ai research documents. Reuters . URL https www.reuters.comarticleusalphabetgoogleresearchfocus googletolditsscientiststostrikeapositivetoneinairesearchdocumentsidUSKBNXCB Y. Dou M. Forbes R. KoncelKedziorski N. A. Smith and Y. Choi. Scarecrow A framework for scrutinizing machine text. CoRR abs. . URL httpsarxiv.org abs.. C. Ebell R. BaezaYates R. Benjamins H. Cai M. Coeckelbergh T. Duarte M. Hickok A. Jacquet A. Kim J. Krijger J. Macintyre P. Madhamshettiwar L. Maffeo J. Matthews L. Medsker P. Smith and S. Thais. Towards intellectual freedom in an ai ethics global community. AI and Ethics . doi .s. A. Fan S. Ilic T. Wolf and M. Gall editors. Proceedings of BigScience Episode Workshop on Challenges Perspectives in Creating Large Language Models virtualDublin May . Association for Computational Linguistics. URL httpsaclanthology.org .bigscience.. L. Gao. On the sizes of openai api models May . URL httpsblog.eleuther.ai gptmodelsizes. L. Gao S. Biderman S. Black L. Golding T. Hoppe C. Foster J. Phang H. He A. Thite N. Nabeshima S. Presser and C. Leahy. The pile An gb dataset of diverse text for language modeling. CoRR abs. . URL httpsarxiv.orgabs.. T. Gebru J. Morgenstern B. Vecchione J. W. Vaughan H. M. Wallach H. D. III and K. Crawford. Datasheets for datasets. CoRR abs. . URL httparxiv. orgabs.. S. Gehrmann H. Strobelt and A. Rush. GLTR Statistical detection and visualization of generated text. In Proceedings of the th Annual Meeting of the Association for Computational Linguistics System Demonstrations pages Florence Italy July . Association for Computational Linguistics. doi .vP. T. Gillespie. Custodians of the internet Platforms content moderation and the hidden decisions that shape social media. Yale University Press . J. A. Goldstein G. Sastry M. Musser R. DiResta M. Gentzel and K. Sedova. Generative language models and automated influence operations Emerging threats and potential mitigations . URL httpsarxiv.orgabs.. D. Gragnaniello F. Marra and L. Verdoliva. Detection of AIGenerated Synthetic Faces pages . Springer International Publishing Cham . ISBN . doi . . URL httpsdoi.org.. O. L. Haimson D. Delmonaco P. Nie and A. Wegner. Disproportionate removals and differing content moderation experiences for conservative transgender and black social media users Marginalization and moderation gray areas. Proc. ACM Hum.Comput. Interact. CSCW oct . doi .. URL httpsdoi.org.. W. D. Heaven. Why metas latest large language model survived only three days online Nov . URL httpswww.technologyreview.com metalargelanguagemodelaionlysurvivedthreedaysgptscience. D. Hovy and S. L. Spruit. The social impact of natural language processing. In Proceedings of the th Annual Meeting of the Association for Computational Linguistics Volume Short Papers pages Berlin Germany Aug. . Association for Computational Linguistics. doi .vP. URL httpsaclanthology.orgP. G. Jawahar M. AbdulMageed and L. V. S. Lakshmanan. Automatic detection of machine generated text A critical survey. CoRR abs. . URL httpsarxiv.org abs.. P. Kalluri. Dont ask if artificial intelligence is good or fair ask how it shifts power. Nature News July . URL httpswww.nature.comarticlesd. N. Kaloudi and J. Li. The aibased cyber threat landscape A survey. ACM Comput. Surv. feb . ISSN . doi .. URL httpsdoi.org. . H. Khlaaf P. Mishkin J. Achiam G. Krueger and M. Brundage. A hazard analysis framework for code synthesis large language models . URL httpsarxiv.orgabs. . J. Kirchenbauer J. Geiping Y. Wen J. Katz I. Miers and T. Goldstein. A watermark for large language models . URL httpsarxiv.orgabs.. S. Kreps R. M. McCain and M. Brundage. All the news thats fit to fabricate Aigenerated text as a tool of media misinformation. Journal of Experimental Political Science . doi .XPS... P. Liang R. Bommasani K. Creel and R. Reich. The time is now to develop community norms for the release of foundation models . URL httpshai.stanford.edu newstimenowdevelopcommunitynormsreleasefoundationmodels. P. Liang R. Bommasani T. Lee D. Tsipras D. Soylu M. Yasunaga Y. Zhang D. Narayanan Y. Wu A. Kumar B. Newman B. Yuan B. Yan C. Zhang C. Cosgrove C. D. Manning C. R D. AcostaNavas D. A. Hudson E. Zelikman E. Durmus F. Ladhak F. Rong H. Ren H. Yao J. Wang K. Santhanam L. Orr L. Zheng M. Yuksekgonul M. Suzgun N. Kim N. Guha N. Chatterji O. Khattab P. Henderson Q. Huang R. Chi S. M. Xie S. Santurkar S. Ganguli T. Hashimoto T. Icard T. Zhang V. Chaudhary W. Wang X. Li Y. Mai Y. Zhang and Y. Koreeda. Holistic evaluation of language models . URL httpsarxiv.orgabs.. T. Liao. Foundation model tracker. httpsfoundationmodeltracker.com . Z. C. Lipton. Openai trains language model mass hysteria ensues Feb . URL httpswww.approximatelycorrect.com openaitrainslanguagemodelmasshysteriaensues. T. Mandel J. Best R. H. Tanaka H. Temple C. Haili K. Schlectinger and R. Szeto. Lets keep it safe Designing user interfaces that allow everyone to contribute to ai safety . URL httpsarxiv.orgabs.. K. McGuffie and A. Newhouse. The radicalization risks of GPT and advanced neural language models. CoRR abs. . URL httpsarxiv.orgabs. . J. Michael A. Holtzman A. Parrish A. Mueller A. Wang A. Chen D. Madaan N. Nangia R. Y. Pang J. Phang and S. R. Bowman. What do nlp researchers believe? results of the nlp community metasurvey . URL httpsarxiv.orgabs.. Microsoft. Microsoft turing academic program mstap Oct . URL httpswww.microsoft.comenusresearchcollaboration microsoftturingacademicprogram. Midjourney. Quick start guide . URL httpsmidjourney.gitbook.iodocs. P. Mishkin L. Ahmad M. Brundage G. Krueger and G. Sastry. Dalle preview risks and limitations. . URL httpsgithub.comopenaidallepreviewblob mainsystemcard.mdhttpsgithub.comopenaidallepreviewblob mainsystemcard.md. M. Mitchell S. Wu A. Zaldivar P. Barnes L. Vasserman B. Hutchinson E. Spitzer I. D. Raji and T. Gebru. Model cards for model reporting. CoRR abs. . URL httparxiv.orgabs.. S. Mohamed M. Png and W. Isaac. Decolonial AI decolonial theory as sociotechnical foresight in artificial intelligence. CoRR abs. . URL httpsarxiv.org abs.. E. Mostaque. Stable diffusion public release Sep . URL httpsstability.ai blogstablediffusionpublicrelease. NAIRRTF. The national artificial intelligence research resource task force nairrtf . URL httpswww.ai.govnairrtf. N. Nangia C. Vania R. Bhalerao and S. R. Bowman. Crowspairs A challenge dataset for measuring social biases in masked language models. CoRR abs. . URL httpsarxiv.orgabs.. T. T. Nguyen C. M. Nguyen D. T. Nguyen D. T. Nguyen and S. Nahavandi. Deep learning for deepfakes creation and detection. CoRR abs. . URL httparxiv. orgabs.. S. U. Noble. Algorithms of oppression how search engines reinforce racism. New York University Press . OpenAI. Gpt model card Nov . URL httpsgithub.comopenaigptblob mastermodelcard.md. OpenAI. Gpt model card Sep . URL httpsgithub.comopenaigptblob mastermodelcard.md. OpenAI. Dalle api now available in public beta Nov . URL httpsopenai.com blogdalleapinowavailableinpublicbeta. OpenAI. Chatgpt Optimizing language models for dialogue Nov . URL https openai.comblogchatgpt. OpenAI. Whats the rate limit for the dalle api? how can i request an increase? . URL httpshelp.openai.comenarticles whatstheratelimitforthedalleapihowcanirequestanincrease. L. Ouyang J. Wu X. Jiang D. Almeida C. L. Wainwright P. Mishkin C. Zhang S. Agarwal K. Slama A. Ray J. Schulman J. Hilton F. Kelton L. Miller M. Simens A. Askell P. Welinder P. Christiano J. Leike and R. Lowe. Training language models to follow instructions with human feedback . URL httpsarxiv.orgabs.. A. Ovadya and J. Whittlestone. Reducing malicious use of synthetic media research Considerations and potential release practices for machine learning. CoRR abs. . URL httparxiv.orgabs.. C. Procope A. Cheema D. Adkins B. Alsallakh N. Green E. McReynolds G. Pehl E. Wang and P. Zvyagina. Systemlevel transparency of machine learning. . URL httpsai.facebook.comresearchpublications systemleveltransparencyofmachinelearning. J. W. Rae S. Borgeaud T. Cai K. Millican J. Hoffmann H. F. Song J. Aslanides S. Henderson R. Ring S. Young E. Rutherford T. Hennigan J. Menick A. Cassirer R. Powell G. van den Driessche L. A. Hendricks M. Rauh P. Huang A. Glaese J. Welbl S. Dathathri S. Huang J. Uesato J. Mellor I. Higgins A. Creswell N. McAleese A. Wu E. Elsen S. M. Jayakumar E. Buchatskaya D. Budden E. Sutherland K. Simonyan M. Paganini L. Sifre L. Martens X. L. Li A. Kuncoro A. Nematzadeh E. Gribovskaya D. Donato A. Lazaridou A. Mensch J. Lespiau M. Tsimpoukelli N. Grigorev D. Fritz T. Sottiaux M. Pajarskas T. Pohlen Z. Gong D. Toyama C. de Masson dAutume Y. Li T. Terzi V. Mikulik I. Babuschkin A. Clark D. de Las Casas A. Guy C. Jones J. Bradbury M. Johnson B. A. Hechtman L. Weidinger I. Gabriel W. S. Isaac E. Lockhart S. Osindero L. Rimell C. Dyer O. Vinyals K. Ayoub J. Stanway L. Bennett D. Hassabis K. Kavukcuoglu and G. Irving. Scaling language models Methods analysis insights from training gopher. CoRR abs. . URL httpsarxiv.orgabs.. I. D. Raji A. Smart R. N. White M. Mitchell T. Gebru B. Hutchinson J. SmithLoud D. Theron and P. Barnes. Closing the ai accountability gap Defining an endtoend framework for internal algorithmic auditing. In Proceedings of the Conference on Fairness Accountability and Transparency FAT page New York NY USA . Association for Computing Machinery. ISBN . doi ... URL httpsdoi.org... I. D. Raji E. M. Bender A. Paullada E. Denton and A. Hanna. AI and the everything in the whole wide world benchmark. CoRR abs. . URL httpsarxiv.org abs.. I. D. Raji P. Xu C. Honigsberg and D. Ho. Outsider oversight Designing a third party audit ecosystem for ai governance. In Proceedings of the AAAIACM Conference on AI Ethics and Society AIES page New York NY USA . Association for Computing Machinery. ISBN . doi ... URL httpsdoi.org... A. Ramesh P. Dhariwal A. Nichol C. Chu and M. Chen. Hierarchical textconditional image generation with clip latents . URL httpsarxiv.orgabs.. J. Rando D. Paleka D. Lindner L. Heim and F. Tramr. Redteaming the stable diffusion safety filter . URL httpsarxiv.orgabs.. R. Reich. Now ai can write students essays for them will everyone become a cheat? Nov . URL httpswww.theguardian.comcommentisfreenov aistudentsessayscheatteachersplagiarismtech. R. Rombach and P. Esser. Stablediffusion license Aug . URL httpsgithub.com CompVisstablediffusionblobmainLICENSE. R. Rombach A. Blattmann D. Lorenz P. Esser and B. Ommer. Highresolution image synthesis with latent diffusion models. In Proceedings of the IEEECVF Conference on Computer Vision and Pattern Recognition CVPR pages June . J. Rubinovitz. Bias bounty programs as a method of combatting bias in ai . URL httpsrubinovitz.com biasbountyprogramsasamethodofcombatting. C. Saharia W. Chan S. Saxena L. Li J. Whang E. Denton S. K. S. Ghasemipour B. K. Ayan S. S. Mahdavi R. G. Lopes T. Salimans J. Ho D. J. Fleet and M. Norouzi. Photorealistic texttoimage diffusion models with deep language understanding . URL https arxiv.orgabs.. G. Sastry. Beyond release vs. not release . URL httpscrfm.stanford.edu commentarysastry.html. R. Shelby S. Rismani K. Henne A. Moon N. Rostamzadeh P. Nicholas N. Yilla J. Gallegos A. Smart E. Garcia and G. Virk. Sociotechnical harms Scoping a taxonomy for harm reduction . URL httpsarxiv.orgabs.. T. Shevlane. Structured access an emerging paradigm for safe ai deployment . URL httpsarxiv.orgabs.. T. Shevlane and A. Dafoe. The offensedefense balance of scientific knowledge Does publishing AI research reduce misuse? CoRR abs. . URL httparxiv. orgabs.. T. Simonite. It began as an aifueled dungeon game. it got much darker May . URL httpswww.wired.comstoryaifueleddungeongamegotmuchdarker. U. Singer A. Polyak T. Hayes X. Yin J. An S. Zhang Q. Hu H. Yang O. Ashual O. Gafni D. Parikh S. Gupta and Y. Taigman. Makeavideo Texttovideo generation without textvideo data . URL httpsarxiv.orgabs.. I. Solaiman and C. Dennison. Process for adapting language models to society palms with valuestargeted datasets . I. Solaiman M. Brundage J. Clark A. Askell A. HerbertVoss J. Wu A. Radford and J. Wang. Release strategies and the social impacts of language models. CoRR abs. . URL httparxiv.orgabs.. P. Staff. Publication norms for responsible ai Dec . URL httpspartnershiponai. orgworkstreampublicationnormsforresponsibleai. L. Stark D. Greene and A. Hoffmann. Critical Perspectives on Governance Mechanisms for AIML Systems pages . . ISBN . doi . . Z. Talat A. Nvol S. Biderman M. Clinciu M. Dey S. Longpre S. Luccioni M. Masoud M. Mitchell D. Radev S. Sharma A. Subramonian J. Tae S. Tan D. Tunuguntla and O. Van Der Wal. You reap what you sow On the challenges of bias evaluation under multilingual settings. In Proceedings of BigScience Episode Workshop on Challenges Perspectives in Creating Large Language Models pages virtualDublin May . Association for Computational Linguistics. doi .v.bigscience.. URL httpsaclanthology.org.bigscience.. R. Thoppilan D. De Freitas J. Hall N. Shazeer A. Kulshreshtha H.T. Cheng A. Jin T. Bos L. Baker Y. Du Y. Li H. Lee H. S. Zheng A. Ghafouri M. Menegali Y. Huang M. Krikun D. Lepikhin J. Qin D. Chen Y. Xu Z. Chen A. Roberts M. Bosma V. Zhao Y. Zhou C.C. Chang I. Krivokon W. Rusch M. Pickett P. Srinivasan L. Man K. MeierHellstern M. R. Morris T. Doshi R. D. Santos T. Duke J. Soraker B. Zevenbergen V. Prabhakaran M. Diaz B. Hutchinson K. Olson A. Molina E. HoffmanJohn J. Lee L. Aroyo R. Rajakumar A. Butryna M. Lamm V. Kuzmina J. Fenton A. Cohen R. Bernstein R. Kurzweil B. AgueraArcas C. Cui M. Croak E. Chi and Q. Le. Lamda Language models for dialog applications . URL httpsarxiv.orgabs.. M. Trajtenberg. Ai as the next gpt a politicaleconomy perspective. Working Paper National Bureau of Economic Research January . URL httpwww.nber.orgpapers w. L. R. Varshney N. S. Keskar and R. Socher. Limits of detecting text generated by largescale language models. CoRR abs. . URL httpsarxiv.orgabs. . A. Vee. Automated trolling The case of gptchan when artificial intelligence is as easy as writing. Interfaces Essays and Reviews in Computing and Culture Vol. Charles Babbage Institute University of Minnesota pages . B. Wang. MeshTransformerJAX ModelParallel Implementation of Transformer Language Model with JAX. httpsgithub.comkingoflolzmeshtransformerjax May . L. Weidinger J. Mellor M. Rauh C. Griffin J. Uesato P.S. Huang M. Cheng M. Glaese B. Balle A. Kasirzadeh Z. Kenton S. Brown W. Hawkins T. Stepleton C. Biles A. Birhane J. Haas L. Rimell L. A. Hendricks W. Isaac S. Legassick G. Irving and I. Gabriel. Ethical and social risks of harm from language models . URL httpsarxiv.orgabs .. M. Whittaker. The steep cost of capture. Interactions nov . ISSN . doi .. URL httpsdoi.org.. J. Whittlestone and A. Ovadya. The tension between openness and prudence in AI research. CoRR abs. . URL httparxiv.orgabs.. D. G. Widder D. Nafus L. Dabbish and J. Herbsleb. Limits and possibilities for ethical ai in open source A study of deepfakes. In ACM Conference on Fairness Accountability and Transparency FAccT page New York NY USA . Association for Computing Machinery. ISBN . doi ... URL httpsdoi.org... B. Workshop T. L. Scao A. Fan C. Akiki E. Pavlick S. Ilic D. Hesslow R. Castagn A. S. Luccioni F. Yvon M. Gall J. Tow A. M. Rush S. Biderman A. Webson P. S. Ammanamanchi T. Wang B. Sagot N. Muennighoff A. V. del Moral O. Ruwase R. Bawden S. Bekman A. McMillanMajor I. Beltagy H. Nguyen L. Saulnier S. Tan P. O. Suarez V. Sanh H. Laurenon Y. Jernite J. Launay M. Mitchell C. Raffel A. Gokaslan A. Simhi A. Soroa A. F. Aji A. Alfassy A. Rogers A. K. Nitzav C. Xu C. Mou C. Emezue C. Klamm C. Leong D. van Strien D. I. Adelani D. Radev E. G. Ponferrada E. Levkovizh E. Kim E. B. Natan F. De Toni G. Dupont G. Kruszewski G. Pistilli H. Elsahar H. Benyamina H. Tran I. Yu I. Abdulmumin I. Johnson I. GonzalezDios J. de la Rosa J. Chim J. Dodge J. Zhu J. Chang J. Frohberg J. Tobing J. Bhattacharjee K. Almubarak K. Chen K. Lo L. Von Werra L. Weber L. Phan L. B. allal L. Tanguy M. Dey M. R. Muoz M. Masoud M. Grandury M. ako M. Huang M. Coavoux M. Singh M. T.J. Jiang M. C. Vu M. A. Jauhar M. Ghaleb N. Subramani N. Kassner N. Khamis O. Nguyen O. Espejel O. de Gibert P. Villegas P. Henderson P. Colombo P. Amuok Q. Lhoest R. Harliman R. Bommasani R. L. Lpez R. Ribeiro S. Osei S. Pyysalo S. Nagel S. Bose S. H. Muhammad S. Sharma S. Longpre S. Nikpoor S. Silberberg S. Pai S. Zink T. T. Torrent T. Schick T. Thrush V. Danchev V. Nikoulina V. Laippala V. Lepercq V. Prabhu Z. Alyafeai Z. Talat A. Raja B. Heinzerling C. Si D. E. Tasar E. Salesky S. J. Mielke W. Y. Lee A. Sharma A. Santilli A. Chaffin A. Stiegler D. Datta E. Szczechla G. Chhablani H. Wang H. Pandey H. Strobelt J. A. Fries J. Rozen L. Gao L. Sutawika M. S. Bari M. S. Alshaibani M. Manica N. Nayak R. Teehan S. Albanie S. Shen S. BenDavid S. H. Bach T. Kim T. Bers T. Fevry T. Neeraj U. Thakker V. Raunak X. Tang Z.X. Yong Z. Sun S. Brody Y. Uri H. Tojarieh A. Roberts H. W. Chung J. Tae J. Phang O. Press C. Li D. Narayanan H. Bourfoune J. Casper J. Rasley M. Ryabinin M. Mishra M. Zhang M. Shoeybi M. Peyrounette N. Patry N. Tazi O. Sanseviero P. von Platen P. Cornette P. F. Lavalle R. Lacroix S. Rajbhandari S. Gandhi S. Smith S. Requena S. Patil T. Dettmers A. Baruwa A. Singh A. Cheveleva A.L. Ligozat A. Subramonian A. Nvol C. Lovering D. Garrette D. Tunuguntla E. Reiter E. Taktasheva E. Voloshina E. Bogdanov G. I. Winata H. Schoelkopf J.C. Kalo J. Novikova J. Z. Forde J. Clive J. Kasai K. Kawamura L. Hazan M. Carpuat M. Clinciu N. Kim N. Cheng O. Serikov O. Antverg O. van der Wal R. Zhang R. Zhang S. Gehrmann S. Mirkin S. Pais T. Shavrina T. Scialom T. Yun T. Limisiewicz V. Rieser V. Protasov V. Mikhailov Y. Pruksachatkun Y. Belinkov Z. Bamberger Z. Kasner A. Rueda A. Pestana A. Feizpour A. Khan A. Faranak A. Santos A. Hevia A. Unldreaj A. Aghagol A. Abdollahi A. Tammour A. HajiHosseini B. Behroozi B. Ajibade B. Saxena C. M. Ferrandis D. Contractor D. Lansky D. David D. Kiela D. A. Nguyen E. Tan E. Baylor E. Ozoani F. Mirza F. Ononiwu H. Rezanejad H. Jones I. Bhattacharya I. Solaiman I. Sedenko I. Nejadgholi J. Passmore J. Seltzer J. B. Sanz L. Dutra M. Samagaio M. Elbadri M. Mieskes M. Gerchick M. Akinlolu M. McKenna M. Qiu M. Ghauri M. Burynok N. Abrar N. Rajani N. Elkott N. Fahmy O. Samuel R. An R. Kromann R. Hao S. Alizadeh S. Shubber S. Wang S. Roy S. Viguier T. Le T. Oyebade T. Le Y. Yang Z. Nguyen A. R. Kashyap A. Palasciano A. Callahan A. Shukla A. MirandaEscalada A. Singh B. Beilharz B. Wang C. Brito C. Zhou C. Jain C. Xu C. Fourrier D. L. Perin D. Molano D. Yu E. Manjavacas F. Barth F. Fuhrimann G. Altay G. Bayrak G. Burns H. U. Vrabec I. Bello I. Dash J. Kang J. Giorgi J. Golde J. D. Posada K. R. Sivaraman L. Bulchandani L. Liu L. Shinzato M. H. de Bykhovetz M. Takeuchi M. Pmies M. A. Castillo M. Nezhurina M. Snger M. Samwald M. Cullan M. Weinberg M. De Wolf M. Mihaljcic M. Liu M. Freidank M. Kang N. Seelam N. Dahlberg N. M. Broad N. Muellner P. Fung P. Haller R. Chandrasekhar R. Eisenberg R. Martin R. Canalli R. Su R. Su S. Cahyawijaya S. Garda S. S. Deshmukh S. Mishra S. Kiblawi S. Ott S. Sangaroonsiri S. Kumar S. Schweter S. Bharati T. Laud T. Gigant T. Kainuma W. Kusa Y. Labrak Y. S. Bajaj Y. Venkatraman Y. Xu Y. Xu Y. Xu Z. Tan Z. Xie Z. Ye M. Bras Y. Belkada and T. Wolf. Bloom A bparameter openaccess multilingual language model . URL httpsarxiv.org abs.. M. Xue C. He J. Wang and W. Liu. DNN intellectual property protection Taxonomy methods attack resistance and evaluations. CoRR abs. . URL https arxiv.orgabs.. J. Zhang Z. Gu J. Jang H. Wu M. P. Stoecklin H. Huang and I. Molloy. Protecting intellectual property of deep neural networks with watermarking. In Proceedings of the on Asia Conference on Computer and Communications Security ASIACCS page New York NY USA . Association for Computing Machinery. ISBN . doi ... URL httpsdoi.org... K. A. Zhang L. Xu A. CuestaInfante and K. Veeramachaneni. Robust invisible video watermarking with attention. CoRR abs. . URL httparxiv.orgabs .. S. Zhang S. Roller N. Goyal M. Artetxe M. Chen S. Chen C. Dewan M. Diab X. Li X. V. Lin T. Mihaylov M. Ott S. Shleifer K. Shuster D. Simig P. S. Koura A. Sridhar T. Wang and L. Zettlemoyer. Opt Open pretrained transformer language models . URL httpsarxiv.orgabs.. A Developer and Logo Key AI Labs AI Sweden Alibaba Allen Institute for AI Anthropic Baidu Beijing Academy of AI BigScience Cohere Craiyon DeepMind EleutherAI Google Huawei Inspur LightOn Meta Microsoft Midjourney Naver NVIDIA OpenAI Salesforce Stability AI Tsinghua University Yandex Developer Logo Developer Logo Developer Logo

write a youtube script

arXiv.v eess.SY Jul Distribution System Reconfiguration to Mitigate Load Altering Attacks via Stackelberg Games Sajjad Maleki Student Member IEEE Subhash Lakshminarayana Senior Member IEEE Charalambos Konstantinou Senior Member IEEE and E. Veronica Belmega Senior Member IEEE AbstractThe integration of IoTcontrollable devices in power systems such as smart electric vehicle charging stations heat pumps etc. despite their apparent benefits raises novel cybersecurity concerns. These vulnerabilities in these devices can be leveraged to launch loadaltering attacks LAAs that ca n potentially compromise power system safety. In this paper we analyze the impact of LAAs on the voltage profile of distribution systems. We derive closedform expressions to quantify the attack impact. Using the insights derived from this analysis we propose a method to mitigate LAAs based on reconfiguring the distribution system as a reactive defense approach. We stud y optimal defense strategies using a noncooperative sequential game theory approach that is robust to LAAs. The proposed solution takes the potential errors in the attack localization into account. Our results show that attacks launched on the deepest nodes in the distribution network result in the highest detrimental impact on the grid voltage profile. Furthermore the propose d gametheoretic strategy successfully mitigates the effect of the attack while ensuring minimum system reconfiguration. Index TermsDistribution system Cybersecurity Loadaltering attack LAA Reconfiguration Stackelberg game. NOMENCLATURE Parameters p p Active ZP load coefficients q q Reactive ZP load coefficients p p p Active ZIP load coefficients q q q Reactive ZIP load coefficients I Identity matrix b pre ij Component of the adjacency matrix before reconfiguration in i th row and j th column M Disjunctive parameter N Number of buses of the distribution system pd qd Nominal active and reactive power demands by a single attacked device p li q li Active and reactive load demands in bus i rij xij Resistance and reactance of the line from bus i to j vnom Nominal voltage Sets S. Maleki is with the School of Engineering University of Warwick CVAL UK and ETIS UMR CY Cergy Paris Universite ENSEA CNRS F Cergy France. S. Lakshimnarayana Corresponding author is with the School of Engineering University of Warwick. C.Konstantinou is with the CEMSE Division King Abdullah University of Science and Technology KAUST. E. V. Belmega is with Univ. Gustave Eiffel CNRS LIGM F MarnelaVallee France and ETIS UMR CY Cergy Paris Universite ENSEA CNRS F Cergy France. Emails sajjad.malekiwarwick.ac.uk subhash.lakshminarayanawarwick.ac.uk charalambos.konstantinoukaust.edu.sa veronica.belmegaesiee.fr. This work has been supported in part by PhD Cofund WALLEE project between the University of Warwick UK and CY Cergy Paris University France and in part by King Abdullah University of Science and Technology KAUST under Award No. RFSOFP. The work is the extension of a prior work accepted at IEEE PES General Meeting . D i Set of buses engaged in the unique path connecting the bus i to the root bus L L s Set of lines without switches L s Set of lines with switches N N f Subset of buses which are not substations N a Subset of buses under attack N f Subset of buses which are substation Variables vji Auxiliary voltage variable for MILP Coefficients matrices for the proposed closedform expressions i Parent bus of the bus i Probability of the bus to be under attack U Vector of the square of voltages of the system while no LAA U A Vector of square of voltages of the system under LAA a r a Players actions in equilibrium B Adjacency matrix of the distribution network after reconfiguration bij Component of B in i th row and j th column c n Number of attacked devices in a critical attack in bus n I ii Current flowing through the branch i i L natt The bus under attack pa q a Active and reactive powers raised by LAA in bus i p i p i Active and reactive load demands at rated voltage in bus i p fi q fi Active and reactive powers flowing from substation bus i pij qij Active and reactive power flows from bus i to j r a Followers best response to attackers strategy s i Apparent power in bus i s zip i s zpi Apparent power in bus i with ZIP and ZP load models v i Voltage in bus i I. INTRODUCTION InternetofThings IoT enabled devices offer enhanced enduser convenience improved efficiency and flexibility t o power systems for load peak management which has driven a notable surge in their adoption. However beyond their evident benefits these devices also present potential vulnerabilities serving as entry points for cyber attackers to exploit and compromise the security of power systems. Specifically loadaltering attacks LAA in power networks with high IoTenabled device penetration pose a significant cybersecurity threat . The concept of LAAs was first introduced in in which adversaries turn a group of IoTcontrollable electrical loads into bots and turn them onoff simultaneously to harm the stability of the system. The manipulation of loads disrupts the balance between power generation and demand leading to frequency instabilities in transmission networks . In distribution networks LAA can result in elevated line flows causing higher voltage drops leading to voltage constraint violation . Furthermore the data required to launch successful LAAs can be obtained from publicly available information . A. Literature Survey LAAs have gained significant interest over the last few years. The majority of existing works focus on the transmission system and the associated frequency control loops. We divide these works into two groups attack impact analysis and viability and attack mitigation. Attack Impact Analysis and Viability Researchers in and investigated the impact of LAAs on transmission systems with varying levels of inertia. These studies identified several effects encompassing line failure frequency deviation disruption in grid restarting tieline failure and increased operational costs. These works did not consider the inherent protection features embedded in power grids such as N scheduling etc. The research in examined the effects of LAA under a more realistic setting consisting of protection and loadshedding schemes and showed that LAA can still cause outages and islands. The authors of and examine the feasibility and impact of LAAs under low demand and low inertia conditions experienced during the COVID pandemic respectively and show that under these conditions it can be easier to launch LAAs that lead to unsafe grid conditions. A study on impacts of attacks on IoTenabled devices on interconnected transmission and distribution TD systems was carried out in . Reference proposed the socalled dynamic LAA DLAA in which the adversary toggles the compromised load devices on and off continually guided by a feedback control loop in response to the systems frequency fluctuations. An analytical framework was introduced in to analyze the impact of LAA on transmission systems and identify the nodes from which an attacker can launch the most effective attacks using the theory of secondorder dynamical systems. In a rareevent sampling algorithm was proposed that uncovers the spatial and temporal distribution of impactful DLAAs while taking N security constraint into account. Attack Mitigation Another stream of research investigates the mitigation of LAAs. The existing mitigation methods can be categorized into i offline or ii online methods. Offline methods Offline defenses try to install preventive measures to stop the detrimental impact of LAAs. For instance reference proposes algorithms to determine the operating points for generators in a way to prevent line overloads caused by potential botnettype attacks against IoT load devices. In the authors proposed the optimal placement of energy storage units to mitigate the effect of LAAs. Reference presented a mitigation framework based on securing the most critical loads that can prevent the destabilizing effects of LAA. This method finds the minimum magnitude of loads needed to be protected in order to guarantee frequency stability in the event of DLAAs. While the works above focus on transmission systems proposed a mitigation approach tailored specifically for distribution networks. Their research focuses on identifying optimal locations for deploying soft open points SOPs and refining their operation to mitigate the effects of attacks on voltage deviations from nominal value. Online methods Despite the effectiveness of the offline methods these measures may be too costly as the preventive features must be enabled irrespective of whether an attack occurs or not e.g. uneconomic generator operating points to cover for LAAs. Online methods on the other hand involve determining defensive actions to counter the effects of LAAs once the attack is launched via reactive measures. In a cyberresilient economic dispatch method has been introduced to mitigate LAAs based on altering the frequency droop control parameters of inverterbased resources to counter the destabilizing effects of LAAs. While EVs can be a source of security threat to target the power grid the work in build a defensive algorithm against LAA using PEVs. In this framework PEVs are designed as feedback controllers that can mitigate the impacts of LAA based on H and H norms. To analyze the manoeuvres of a strategic attacker initiating DLAA introduces a multistage game approach. In this game the defensive actions involve load shedding and the ultimate objective is to achieve a strategic balance between DLAA and the necessary amount of load shedding reaching a Nash equilibrium NE. B. Contributions In this paper we start with a rigorous investigation of the impact of LAAs on distribution networks. Then we introduce a Stackelberg noncooperative game to analyze the interaction between the defender operator and the attacker and propose an LAA mitigation technique in distribution networks based on network reconfiguration via switching different lines onoff. Finally we show the effectiveness of our proposed mitigation method by examining it on two standard test cases the and bus grids. Despite the growing literature on LAAs and mitigation techniques a majority of the previous studies have focused solely on transmission systems and frequency stability except for which investigated LAAs in distribution systems. Compared to our work provides analytical insights into the impact of LAAs and identifies the most vulnerable nodes of the system which guides the design of the mitigation strategy. Furthermore our work considers defending against a strategic and knowledgeable attacker while accounting for errors in the attack localization and for attacker resource constraints. In summary there exists a lack of comprehensive investigations into the effects of LAAs on distribution networks and potential mitigation methods. To address this gap we first present closedform expressions to determine the voltage profile of distribution systems under LAA considering ZIP load models and calculate the minimum required devices to be manipulated for a successful LAA. This analysis provides insights into the effect of the attacks spatial location and the number of devices required to cause the system to violate voltage safety standards. Based on these insights we next develop a mitigation algorithm relying on network reconfiguration that dynamically alters the spatial location of the attack within the distribution network topology . A major advantage is that our proposed mitigation method does not require installing any additional devices in the system and leverages the existing flexibility of the network reconfiguration. Additionally to take into account a strategic attacker and to defend against the worstcase attacks we model the attackerdefender interaction using a Stackelberg game framework with the attacker as the leader and the defender as the follower. The sequential game is relevant for reactive defenses in which mitigation is triggered only when an attack is detected note that cyberattacks are rare events and developing a robust reactive defense strategy provides the operator significant cost savings. Moreover drawing from the insights presented in we account for noisy system measurements and uncertainties associated with attack detectionlocalization into the utility functions. Reconfiguration of distribution systems is a wellstudied research topic. In the reconfiguration is modeled as a mixed integer secondorder conic programming MISOCP optimization problem to minimize power losses and increase system reliability. In this paper reconfiguration is applied in the context of mitigating cyber attacks by combining it with a gametheoretic formulation. Since the payoff computation in the game requires solving the optimization problem repeatedly we reformulate the network reconfiguration problem as a mixed integer linear programming MILP optimization by using the linearized distribution flow LinDistFlow and ZP approximations. The resulting formulation provides significant computational speed up in calculating the games payoffs. Additionally we integrate the ZIP load model into the constraints to take the voltage dependency of load demand into account. The proposed mitigation also bears similarities with the movingtarget defense approach that has received considerable attention in the past few years . However the proposed strategy applies the reconfiguration reactively rather than the proactive approach used in these past works and is specifically tailored to mitigate the effects of LAAs. Fig. briefly illustrates the proposed gametheoretic interaction of attacker and defender in this paper. This paper significantly extends our preliminary work in terms of using the insights derived from attack impact analysis to design a mitigation strategy to counter LAAs. The key contributions are summarized as follows Deriving closedform expressions to obtain the bus voltages of the distribution system in the presence of voltagedependent loads. Based on these results analysing the effects of LAAs launched at different locations in the distribution network and obtaining the minimum number of IoTcontrollable devices required for a detrimental attack. Proposing a mitigation strategy to counter the effects of LAA based on the distribution network reconfiguration. Fig. Summary of the proposed attackerdefender interaction. Formulating a sequential game approach to the model strategic attackers while considering the potential uncertainties associated with the localization of LAAs. Validating the proposed framework by extensive simulations using the IEEE bus and bus distribution systems. II. PRELIMINARIES In this section we introduce the system load and power flow models for the distribution systems considered in this research. A. Distribution System Model The distribution system under study is represented by the connected directed graph G N L where N . . . N denotes the set of buses and L denotes the set of branches. This graph has a radial structure hence it is a tree. Except for bus which is the root each bus is referred to as the child of its parent bus which is the adjacent bus closer to bus by one branch. Thus the set of branches is defined as L i i i i N where i represents the parent of bus i. In this configuration bus represents the generator bus. We denote by Dk the set of buses which forms the unique path connecting bus to bus k excluding bus and including bus k. The depth of each bus represents the distance in terms of the number of branches between that bus and the root bus. B. Load Model This subsection introduces the load models which are implemented in the rest of the paper. ZIP Load Model The power demand under the ZIP load model is given in as follows s zip i vi p i ppvipv i jq i qqviqv i where k k k where k p q. The ZIP load model captures the voltage dependency of realworld loads. ZP Approximation Based on the ZIP model is a function of both vi and v i . This causes the optimization tasks involving the power flow in the presence of ZIP loads to become nonconvex and complex. To tackle this problem have provided an approximate model for ZIP loads given by s zp i vi p i p p v i jq i q q v i where p p p q q q p p p and q q q while s zp i vi p zp i vi jqzp i vi. The new coefficients in the ZP model are obtained by the binomial approximation method. The ZP approximation is valid as long as the voltage is close enough to the nominal value i.e. while vi vnom . the ZP approximation is valid . C. Power Flow Equations Branch Flow Model The branch flow model encapsulates the complete AC power flow with the equations describing the system state as follows X kik sik sii zii Iii si where vi vi zii Iii sii vi I ii. Note that superscript denotes the conjugate of a complex number. Linearized Distribution Flow Linearized distribution flow LinDistFlow simplifies the branch flow model described in by neglecting branch power losses and is widely adopted in several distribution grid studies. The power flow equations under this model are given by X kikL pik pji p L i X kikL qik qji q L i v i v j rijpji xij qji. In this formulation and are active and reactive power balances in each bus and is the equation for finding subsequent voltage profile. D. Reconfiguration of Distribution System Network reconfiguration involves modifying the distribution networks topology by adjusting the openclosed state of its switches. In this subsection a set of MILP optimization constraints is introduced primarily to determine the configuration that maintains the nodal voltages closest to their nominal values. We modify the formulation in to accommodate the ZP approximation of the loads to capture its voltage dependency. We also change the power flow model to the LinDistFlow. As we show in Section V these approximations result in a MILP the network reconfiguration problem and provide significant computational speedups compared to the MISOCP. Connectivity Constraints First we present the connectivity constraints that determine the connection between the nodes while keeping the overall graph radial bij bif f N f bij bji i j L Ls bij bji yij i j Ls yij X jijL bji . i N N f Equation implies that the substation buses can not have a parent bus forces the lines without a switch to be always connected while lets the lines with switches to be either open or close and forces nonsubstation buses to have exactly one parent bus. Power Flow Constraints Below we present the optimization problems power flow constraints which are taken from the DistFlow and ZP models pij M bij qij M bij X jijL pij p f i i N f X jijL qij q f i i N f X jN pji pij p ZP i i N N f X jN qji qij q ZP i i N N f p ZP j p l j v j a p ZP j q l j v j b vi v i vi v ij M bij v ij v i rijpij xij qij v i X jN v ji. i N N f Equations represent the power flow constraints is ZIP load constraint and are voltage constraints. The reason behind using the auxiliary variable of v is to make the optimization following disciplined convex programming DCP rules of Python. Additionally we consider all of the normally connected lines to have switches thus they could be switched off if necessary. E. LoadAltering Attack Model Some major manufacturers of highwattage IoTcontrollable devices have acknowledged the presence of security vulnerabilities in their products . The concept presented in the LAAs involves attackers leveraging these vulnerabilities to transform a group of such devices into bots and toggle them on and off. This coordinated action is intended to disrupt the stability of the system. Based on this to implement the LinDistFlow while there is such an attack in the system and change into P jN pji pij p L i i N N f i N a P jN pji pij p L i p a i N N f i N a and P jN qji qij q L j i N N f i N a P jN qji qij q L j q a i N N f i N a . In the following based on the provided models and context we first analyze the impact of LAA on distribution systems. Then we propose a gametheoretic mitigation scheme for these attacks. III. EFFECTS OF LAA ON DISTRIBUTION SYSTEMS In this section we analyze the impact of LAA on distribution networks. Our objective is to derive closedform expressions for the voltage profile of the network with voltagedependant loads under LAA and for the minimum number of compromised load devices required to cause nodal voltage safety violations. It is worth noting that system voltages under LAA can also be computed by solving the power flow equation through an iterative approach such as the backwardforward sweep BFS technique. However unlike closedform equations which we derive in this section the application of iterative techniques does not yield analytical insights into the impact of LAAs on the distribution network. Furthermore the closedform expressions obtained in this section are essential in the design of the defense strategies to mitigate LAAs. A. Closedform Approximation of Nodal Voltages To derive the closedform expressions for the system voltages under LAAs we make two approximations i employing LinDistFlow formulations and ii utilizing the ZP model. Without LAA First we model the distribution system without LAA and analyze the power flow equations in . Integrating into results in vk s v X iDk riip zp ii xiiq zp ii where p zp ii pii p pv i q zp ii qii q qv i . Next we perform a variable change uk v k which results in a set of linear equations which can be written in matrix form as follows UN NN U N where U is the vector of squares of voltages and nn is the matrix with entries i X mDi rmmp mm p xmmq mm q ik Pi c rccp k p xccq k q if i Dk ik otherwise where i k N. We rewrite the system of linear equations of as INN NN UN N in which N k NN ik for i ... N and k N . With LAA Here we analyze the voltage profile of the network under LAA. For this we integrate the introduced LAA in Section IIE into . v a k q v k pA a rka qA a xka in which k P iDk riip zp i i xiiq zp ii rka P iDaDk rii and xka P iDaDk xii. This change results in a new set of coefficient matrices. To calculate the attacked systems square of voltages vector U A we solve U A N A NN U A N . To obtain the coefficients matrices and are dragged into LinDistFlow as the ZP model is imposed on them. The final results are A ik ik a ik for i and k A NN A N A NN in which for i and k and a i X cDiDa p A a p rcc q A a qxcc a ik riip a p xiiq a q if i Da A ik otherwise. B. Analytical Insights into the Attack Impact Next we use the closedform expressions obtained above to derive analytical insights into the attack impact. First note that when there is an attack in a leaf bus the last bus of each branch rka and xka have the highest possible values. As a result the voltage drop resulted from p a rka q axka in is higher and obtained voltages shrink. In conclusion attacks on the leaf buses yield the most detrimental effects. While this result is somewhat expected for distribution systems we further obtain the minimum number of attacked devices which leads to voltage safety violations. We call such a threat the critical attack. For this we consider the voltage of the leaf bus as a known variable vth. The new unknown variable is p a and based on attacked device type we can find q a via q a qd pd p a . So the new set of coefficients for obtaining voltages of buses except for the leaf one and the active power of the critical attack is forming d d d d in which d i i v thia if i a i v thia if i a d ik P cDi rcc p qd pd xcc q if k a i Da d ik if k a i Da ik otherwise. Hence we can solve the linear system of equations given by I a d X d in which X is a vector with the same dimension as U where all elements of it are the same as U except for one in which X contains p a instead of ua since we already know ua v th. Additionally I a is the identity matrix except for the element a a which equals . This gives the bus voltages as well a the required p a . Then we can find the number of required devices in bus n using cn p a pd . The closedform expressions provided in this section are not only used for identifying the worst effect of the LAAs but also in the following sections they inspire our mitigation method to find the optimal defensive action. Additionally the introduced critical attack has been implemented in finding the optimal action of the strategic attacker later in this paper. IV. MITIGATING LAA VIA RECONFIGURATION In this section we introduce a novel technique to mitigate LAAs by reconfiguring the distribution system topology. We exploit a sequential gametheoretic interaction in which following the LAA launched by the attacker the defender reconfigures the network to react optimally to the threat. A. Mitigation Design and Intuition The intuition behind the proposed defense technique of reconfiguring the distribution network lies in the analytical insights derived in Section III. Based on our analysis recall that LAAs targeting the leaf buses of the distribution network lead to the greatest attack impact in terms of the deviation of the voltage from the nominal values. For instance as represented in Fig. in the base configuration of the IEEE bus system an LAA targeting node the leaf bus of the longest branch is the most impactful attack. Note that in Fig. dashed lines present the normally open lines disconnected lines which could be connected in case of reconfiguration. In this context reconfiguring the distribution networks topology e.g. closing the link between Bus and Bus in bus grid changes the position of the leaf buses thus alleviating the attack impact. It is worth noting that the proposed mitigation leverages the preexisting capabilities of the distribution network e.g. devices enabling network reconfiguration are primarily installed to reduce power losses andor voltage deviations and hence does not require new infrastructure. Furthermore in the proposed scheme the system will be reconfigured only when an attack is detected thus avoiding unnecessary reconfigurations to mitigate attacks note that cyberattacks are somewhat rare events. The attack detection module can be based on existing modeldriven or datadriven approaches for detecting LAAs. The reader can refer to past works including and in this area for more details. B. Stackelberg Game for Attack Mitigation We model the strategic interaction of the attacker and the defender via a noncooperative game. Under this formulation the attacker first chooses a bus to launch an LAA and subsequently the defender exploits the flexibility of the system to reconfigure it and mitigate the attack. This successive interaction is modeled by a Stackelberg game. The sequential game models this reactive mitigation framework. A Stackelberg game with two players consists of a leader and a follower. The leader always commits their action first to maximize their own objective function by anticipating the followers strategic reaction. Then given the action of the leader the follower picks their best i.e. optimal response to maximize their own objective function. Since LAAs are rare incidents in the network we propose a reactive mitigation method instead of a preventive one. In this approach the defender waits for the attacker and responds optimally. Our Stackelberg game can be formally defined as H A DSA SDFA FD in which A and D are the players attacker and defender SA SD denote the set of available actions of players and FA FD denote their reward or payoff functions. The attackers set of actions SA is launching an LAA in any of the buses one attack on one bus at a time by assumption that causes voltage constraint violation. Thus the set of actions of the attacker is SA N L in which N L N denotes the set of load buses. The set of defense actions is the set of all possible system reconfigurations discussed in Section IID such that SD . . . NB denotes the set of indices of all possible reconfiguration matrices B bij ijN NN whose entries meet the constraints in equations and NB represents the number of such matrices. Henceforth we denote by Bd bij dijN the adjacency matrix of the network as a result of a specific defense d SD. The attacker wants to maximize the voltage deviation as a result we define FAd a X nN  v nom v nd a   as the attackers objective function. The reason behind using the squares of the voltages in the above equation is that it greatly simplifies the computation of the Stackelberg equilibrium involving a MILP as will be discussed later. Moreover increasing the distance between the squares of voltages is equivalent to increasing the distance between the voltages given that vn vnom hence leading to a consistent reward for the attacker. The defenders objective is to reconfigure the system to minimize the square of voltage deviation above. Achieving this goal with minimal changes can decrease the maintenance requirements for switches and reduce the likelihood of switching failures. To take this factor into account we add a penalty term and the resulting reward function of the defender is F perf D X nN  v nom v nd a   pend where the first part of the F perf D represents the sum of the square of voltage deviations in all buses and pend X N i X N j  b pre ij bij d   is the penalty term for enforcing a system reconfiguration with the minimum switches possible. The above reward F perf D is relevant when the defender is capable of perfect attack localization. However due to noisy measurements the defender might not be able to do this. Instead we assume that the defender is only able to locate a neighbourhood of the attack i.e. a connected cluster of buses which contains the bus under attack. We further assume that the defender has a favourite candidate bus under attack denoted by natt but does not discard the attack possibilities of other buses in the located cluster. To model this we define a discrete probability vector . . . N with entries if natt Anatt if Anatt otherwise. Above represents the likelihood that the defender assigns to bus P being under attack such that and N . Additionally . denotes the likelihood of the defenders favourite candidate. The subset Anatt is the set of buses directly adjacent to natt along with the buses directly adjacent to those all these buses are considered as the other potential candidates by the defender. Their likelihood is the remaining probability split equally between the Anatt other candidate buses. For instance if in the base configuration of the system of Fig. the defender chooses as favourite candidate bus natt with a likelihood of then A and the probabilities are . . and N A. Taking this uncertainty of detecting precisely the attack location into account at the defenders end results in the following reward FDd a X N X nN  v nom v nd a  pend which represents the expected reward over this uncertainty. Obtaining the optimal attack and defense requires computing the Stackelberg equilibrium. As discussed earlier in this paper the attacker commits the attack first and then the defender reacts. Definition . The best response of the defender to an action a SA is defined as ra arg max dSD FDd a. Definition . A profile of actions a d SA SD is a Stackelberg equilibrium iff FAra a FAra a a SA d ra . Intuitively the attacking action at the Stackelberg equilibrium is the one which maximizes the attackers reward when the defender responds with the best reaction. Furthermore the defenders best reaction to a is its Stackelberg equilibrium action. A Stackelberg equilibrium is ensured to exist if the defenders optimal response exists for every action of the leader. Assuming that normally open points exist in the distribution system this ensures that at least one system reconfiguration is possible and that the discrete feasible set in is nonvoid leading to the existence of the solution. Algorithm describes the method of finding the players actions at the Stackelberg equilibrium. First we need to find the best reaction ra for each attack a SA which requires running a reconfiguration optimization Algorithm Computing the Stackelberg equilibrium. Data A DSA SDFA FD ... N Result ra a Compute ra a SA by solving ra arg maxdSD FDd a s.t. and as a MILP optimization problem via SCIPY and CVXPY Compute FAra a a SA Choose a arg maxaSA FAra a Compute ra arg maxdSD FDd a s.t. and by solving the defined MILP optimization problem for each of the attacks see equation . Note that the constraints of this optimization are linear in the square of the voltages hence justifying our choice of the distance between the squares of voltages in the reward function leading to a linear program instead of a quadratic one obtained by a simple variable change ui v i . The resulting MILP optimizations are carried out in Python and the optimization modeling language is CVXPY. The solver used for these optimizations is SCIPY. Then find the reward of the attacker for ra a a SA. Finally the action corresponding to the maximum of FAra a is selected as the attacking strategy in our Stackelberg formulation. Additionally ra corresponding to the attacking action at the Stackelberg equilibrium a is the optimum defensive strategy at the Stackelberg equilibrium d ra . C. ResourceConstrained Attacker Drawing from our discussions in Section IIIB it is evident that each bus has a distinct critical attack leading to voltage constraint violation. Given the attackers tendency for launching such a critical attack their potential action will not only occur across different buses but also vary in magnitude. To accommodate this feature we define a new game H A DSA SDF A FD in which the attacker launches the critical attack which we call them the resourceconstrained attacker. Note that the critical attack in each bus is the minimum number of required devices to be manipulated in that bus to cause voltage constraint violation. Indeed cn for each attack is unique and varies with the attack location. In this modified game F A comprises two components the total nodal voltage deviation and the attack magnitude. The attacker seeks to maximize the former while minimizing the latter. However these two terms cannot be simply summed due to their disparate physical characteristics. Therefore we propose the following reward F Ad a F norm A d a cnorma in which and F norm A d a FAd a P iSA FAd i c norma ca P iN L ci . TABLE I Number of compromised devices required to cause voltage safety violations in the IEEE bus network during different hours of the day. Time Att. location bus Least load Peak load Air Conditioner Resistive heater Note that the parameter trades off between the two components of the objective function. If the attacker only cares about maximizing the harm caused in the voltage profile similar to the case using and if the attacker only cares about minimizing the attacked devices. The rest of the components of H are the same as H. The process of computing the Stackelberg equilibrium is similar to H and we only need to plug in the attackers new objective function F d a. V. RESULTS AND DISCUSSIONS Our simulations are conducted using the IEEE bus and bus systems. The ZIP load coefficients of all the buses are set to the residential loadtype F introduced in . A. Critical Attack Here we conduct spatial analyses to determine the most effective location for launching LAAs. To quantify this we use the load profile obtained from for as the base load without LAAs which contains the hourly electric power demand in New York US. To mimic this load profile in the IEEE bus test system we project the ratio of load changes at different hours of the day onto the nominal load of the test network. Table I presents the number of devices required for the critical attack in three of the leaf buses of the bus test case during the different hours of the day. These numbers are computed via the equations in Section IIIB. We can see that the attack on the bus with the highest depth requires fewer devices to be manipulated. This conclusion confirms our insights from Section IIIA. Furthermore it also shows the dependency on the type of load air conditioner resistive load etc. and the associated ZIP load coefficients. Since the results in Table I are obtained by the approximation discussed in Section III we evaluate how effective these attacks are when considering the full AC power flow model. To evaluate the extent of the errors between the voltages computed using the analytical results and the full AC power flow model we compare the obtained voltage profile with the results of BFS. Fig. shows the voltage profile calculated by the two methods during the peak load demand and the corresponding critical attack. We remark that the actual errors of computing the nodal voltages with our approximations never exceed thus proving the validity of the analytical results. B. LAA Mitigation Next we examine the proposed LAA mitigation method. The base load profiles without LAAs of bus and bus test cases are and of their nominal load profiles Fig. Voltage profile of the attacked on Bus bus test case with the proposed closedform equations and the accurate model. TABLE II Normally open lines. bus bus obtained by MATPOWER. Except for the resourceconstrained attacker results we consider LAA attacks of magnitude p a i q a i . p.u. in which i is the index of the attacked bus. This magnitude of the attack is significant enough to cause voltage safety violations in all the load buses except those adjacent to the root and hence needs to be mitigated. Table II represents the candidate lines which can be openedclosed during the system reconfiguration in the two bus systems. Three scenarios are considered next i accurate attack localization ii errors in attack localization and iii errors in attack localization and resourceconstrained attacker. i Accurate Attack Localization Here we set natt and N L N a . The results for the bus grid are presented in Tables III and IV. In all figures red buses represent the attacked bus attackers action and green lines are the ones with altered state been opened or closed as the defenders reaction. Based on the results the attack is launched on Bus . Fig. shows the voltage profiles of the grid under attack before and after reconfiguration. The result shows that the proposed mitigation method is able to return the voltage profile within the constraints hence mitigate the effects of LAA. Although the LAA on Bus causes the greatest impact in terms of voltage deviations a strategic attacker that can anticipate the defenders action chooses to launch the attack on Bus instead to maximize their payoff. Fig. and Table III illustrate the players actions at the Stackelberg equilibrium for the bus grid in which attack is launched on Bus and lines and changed their state to reconfigure the network. ii Errors in Attack Localization In this scenario we consider . of certainty about the location of the attack then the remaining . probability of the attack location is split equally between the other buses in the neighbourhood see Section IV. The result for the bus grid is presented in Tables III and IV. In this scenario the defense action should keep the voltage within the desired constraints assuming an LAA in any of the candidate buses. This uncertainty causes a system reconfiguration that necessitating more switching. Fig. represents the voltage profile of the bus grid after the reconfiguration if any of the three suspect buses are attacked. We can notice that the attack will be successfully mitigated and the voltage constraint will not be violated in any case. Note that the players actions for the bus system are the same as in scenario i. Fig. Voltage profile of the attacked bus grid before and after mitigation reconfiguration. Fig. IEEE bus system configuration after the attack and reconfiguration. iii ResourceConstrained Attacker Finally we present the resourceconstrained attacker introduced in Section IVC with . in their objective function. The uncertainty level in the attack localization is the same as in the previous scenario. The result for the bus system is presented in Tables III and IV. We observe that in this scenario the attacker chooses to attack Bus as the impact of load alteration on this bus will be the greatest as the number of compromised devices is taken explicitly into account. The new configuration selected by the defender and illustrated in Table IV tackles the voltage constraint violation by this attacker. Again the bus case maintains the same Stackelberg equilibrium actions as the past two scenarios. Table V represents the total voltage deviations P iN vnom vi in the grid for different scenarios. Note that in scenario ii the voltage deviation of the bus grid drops but the defender needs to commit more switching which is not desirable. Furthermore a portion of the reduced voltage deviation in scenario iii should be attributed to the smaller attack launched by the resourceconstrained attacker. C. Significance of GameTheoretic Approach We also consider a nonstrategic attacker that does not anticipate the defenders actions. In this case first the attacker launches the attack which maximizes the total voltage deviation regardless of potential defensive reactions. Subsequently the defender solves the optimal reconfiguration problem to mitigate the attack. The results of this approach for the bus test case are presented in Table VI. We remark that both scenarios i and ii result in the same output. Compared to Table V for the strategic attacker the defender always benefits. Indeed in scenario i the total voltage deviation is dropped from . p.u. to . p.u. while in ii the number of switching is reduced from to . Note that mitigation with Fig. Voltage profiles of bus grid attacked at any of suspect buses after mitigation. TABLE III Attacked bus and closedopen lines at the Stackelberg equilibrium of each scenario. Scenario Attacked bus Closed lines Opened lines bus bus bus bus bus bus i ii iii less switching is preferred even with slightly higher voltage deviation. Note that the strategic attacker in the bus grid picks the deepest bus Bus as the victim which is similar to the nonstrategic attackers choice. D. Approximation Efficiency Finally to show the efficiency of the approximations introduced in Section IID we compare the computation time for the proposed MILP model proposed in IID with that of the similar MISOCP model. Exhibited in Table VII for the bus test case a single MISOCP optimization takes . s while one MILP optimization requires . s which means a reduction of over in calculation time while for the bigger system this reduction is even more significant . It is important to note that as we aim to find the Stackelberg equilibrium solving the optimization must be iterated for each potential attack to uncover all potential rewards for players. Thus the proposed approximations provide significant computation time reduction to compute the Stackelberg equilibrium. VI. CONCLUSIONS In this paper we propose a set of closedform expressions for the power flow of distribution systems to determine bus voltages in the presence of voltagedependent loads with or without LAA which we exploit to investigate the impact of LAAs. Then we introduce a sequential gametheoretic approach to mitigate LAAs in distribution systems in a reactive manner by network reconfiguration with minimum possible switching. Furthermore we take into account the uncertainties in the attack localization by introducing a probability distribution over the potentially attacked nodes. Finally by introducing a hybrid objective function we considered the resourceconstrained attacker. REFERENCES S. Maleki S. Pan E. V. Belmega C. Konstantinou and S. Lakshminarayana The impact of load altering attacks on distribution systems with ZIP loads arXiv preprint arXiv. . TABLE IV Stackelberg equilibrium configurations of the bus grid Red buses are the location of the attack. Scenario i Accurate attack localization Scenario ii Error in localization the attack Scenario iii Resourceconstrained attacker TABLE V Total voltage deviations and switching required under each scenario. Scenario Number of switching Total voltage deviation p.u. bus bus bus bus i . . ii . . iii . . TABLE VI Nonstrategic attackers preferred actions obligated switching numbers and total voltage deviations. Scenario Attacked bus Total voltage deviation p.u. Number of switching i . ii . S. Amini F. Pasqualetti and H. MohsenianRad Dynamic load altering attacks against power system stability Attack models and protection schemes IEEE Transactions on Smart Grid vol. no. pp. . A.H. MohsenianRad and A. LeonGarcia Distributed internetbased load altering attacks against smart power grids IEEE Transactions on Smart Grid vol. no. pp. . S. Soltan P. Mittal and H. V. Poor BlackIoT IoT botnet of high wattage devices can disrupt the power grid in th USENIX Security Symposium pp. . S. Lakshminarayana S. Adhikari and C. Maple Analysis of IoTbased load altering attacks against power grids using the theory of secondorder dynamical systems IEEE Transactions on Smart Grid vol. no. pp. . A. Dabrowski J. Ullrich and E. R. Weippl Grid shock Coordinated loadchanging attacks on power grids The nonsmart power grid is vulnerable to cyber attacks as well in Proceedings of the rd Annual Computer Security Applications Conference pp. . Z. Liu and L. Wang A robust strategy for leveraging soft open points to mitigate load altering attacks IEEE Transactions on Smart Grid vol. no. pp. . S. Acharya Y. Dvorkin and R. Karri Public plugin electric vehicles grid data Is a new cyberattack vector viable? IEEE Transactions on Smart Grid vol. no. pp. . B. Huang A. A. Cardenas and R. Baldick Not everything is dark and gloomy Power grid protections against IoT demand attacks in USENIX Security Symposium pp. . J. Ospina X. Liu C. Konstantinou and Y. Dvorkin On the feasibility of loadchanging attacks in power systems during the COVID pandemic IEEE Access vol. pp. . S. Lakshminarayana J. Ospina and C. Konstantinou Loadaltering attacks against power grids under COVID lowinertia conditions IEEE Open Access Journal of Power and Energy vol. pp. . Y. Dvorkin and S. Garg IoTenabled distributed cyberattacks on transmission and distribution grids in North American Power Symposium NAPS. IEEE pp. . M. P. Goodridge S. Lakshminarayana and A. Zocca Uncovering loadaltering attacks against N secure power grids A rareevent sampling approach IEEE Transactions on Power Systems arXiv. . S. Soltan P. Mittal and H. V. Poor Protecting the grid against MAD attacks IEEE Transactions on Network Science and Engineering vol. no. pp. . A. Di Giorgio A. Giuseppi F. Liberali A. Ornatelli A. Rabezzano and L. R. Celsi On the optimization of energy storage system placement for protecting power transmission grids against dynamic load altering attacks in Proc. Mediterranean Conference on Control and Automation. IEEE pp. . TABLE VII Optimization times in seconds. Op. type bus grid bus grid MISOCP . . MILP . . Z. Chu S. Lakshminarayana B. Chaudhuri and F. Teng Mitigating loadaltering attacks against power grids using cyberresilient economic dispatch IEEE Transactions on Smart Grid vol. no. pp. . M. A. Sayed R. Atallah C. Assi and M. Debbabi Electric vehicle attack impact on power grid operation International Journal of Electrical Power Energy Systems vol. p. . M. A. Sayed M. Ghafouri R. Atallah M. Debbabi and C. Assi Protecting the future grid An electric vehicle robust mitigation scheme against load altering attacks on power grids Applied Energy vol. p. . Y. Guo L. Wang Z. Liu and Y. Shen Reinforcementlearningbased dynamic defense strategy of multistage game against dynamic load altering attack International Journal of Electrical Power Energy Systems vol. p. . H. Haghighat and B. Zeng Distribution system reconfiguration under uncertain load and renewable generation IEEE Transactions on Power Systems vol. no. pp. . S. F. Santos M. Gough D. Z. Fitiwi J. Pogeira M. Shafiekhah and J. P. Catalao Dynamic distribution system reconfiguration considering distributed renewable energy sources and energy storage systems IEEE Systems Journal vol. no. pp. . H. Jahangir S. Lakshminarayana C. Maple and G. Epiphaniou A deeplearningbased solution for securing the power grid against load altering threats by iotenabled devices IEEE Internet of Things Journal vol. no. pp. . Q. Li J. Zhang J. Zhao J. Ye W. Song and F. Li Adaptive hierarchical cyber attack detection and localization in active distribution systems IEEE Transactions on Smart Grid vol. no. pp. . R. S. Rao S. V. L. Narasimham M. R. Raju and A. S. Rao Optimal network reconfiguration of largescale distribution system using harmony search algorithm IEEE Transactions on Power Systems vol. no. pp. . T. Van Cutsem and C. Vournas Voltage stability of electric power systems. Springer Science Business Media . S. Lakshminarayana E. V. Belmega and H. V. Poor Movingtarget defense against cyberphysical attacks in power grids via game theory IEEE Transactions on Smart Grid vol. no. pp. . C. Liu J. Wu C. Long and D. Kundur Reactance perturbation for detecting and identifying FDI attacks in power system state estimation IEEE Journal of Selected Topics in Signal Processing vol. no. pp. . F. U. Nazir B. C. Pal and R. A. Jabr Approximate load models for conic OPF solvers IEEE Transactions on Power Systems vol. no. pp. . M. Farivar and S. H. Low Branch flow model Relaxations and convexificationpart I IEEE Transactions on Power Systems vol. no. pp. . M. Baran and F. F. Wu Optimal sizing of capacitors placed on a radial distribution system IEEE Transactions on Power Delivery vol. no. pp. . J. A. Taylor and F. S. Hover Convex models of distribution system reconfiguration IEEE Transactions on Power Systems vol. no. pp. . A. Bokhari A. Alkan R. Dogan M. DiazAguil o F. De Leon D. Czarkowski Z. Zabar L. Birenbaum A. Noel and R. E. Uosef Experimental determination of the ZIP coefficients for modern residential commercial and industrial loads IEEE Transactions on Power Delivery vol. no. pp. . New York Independent System Operator NYISO load data httpswww.nyiso.comloaddata accessed .

write a youtube script

Im making a tabletop bowling lane with an automatic pin Setter last time we built a lock and place latch mechanism for the bumpers my actual Lane was just these big D printed blocks I installed alignment pins to minimize the lip between each block somehow I always start with the worst ideas first I dont know why a better idea would just be to use some wood so I went to Lowes and picked up a T long plank I then D printed these risers that attached to it I made these Corner blocks to align the wood to the plas I then flipped the whole thing over and drilled some pilot holes I then installed some threaded inserts into the wood Im not really a wood guy I dont know what the hell Im doing with those inserts installed I can now easily attach my plastic parts to the plank and it makes it easy to take apart frequently after the risers were attached it was time to build out two sets of the bumper mechanisms this part was freaking tedious but we got it done then I attached the bumpers to the Lane I also added some gutters I think Im going to add some mesh or something on the inside of the bumpers because the pins were getting stuck under them Im really excited to build out the automated pin Setter next especially because manually resetting the pins for each shot right now is not fun follow to see the rest of my build

write a youtube script

LaRa Efficient LargeBaseline Radiance Fields Anpei Chen Haofei Xu Stefano Esposito Siyu Tang Andreas Geiger University of Tbingen Tbingen AI Center ETH Zrich httpsapchenstu.github.ioLaRa Abstract. Radiance field methods have achieved photorealistic novel view synthesis and geometry reconstruction. But they are mostly applied in perscene optimization or smallbaseline settings. While several recent works investigate feedforward reconstruction with large baselines by utilizing transformers they all operate with a standard global attention mechanism and hence ignore the local nature of D reconstruction. We propose a method that unifies local and global reasoning in transformer layers resulting in improved quality and faster convergence. Our model represents scenes as Gaussian Volumes and combines this with an image encoder and Group Attention Layers for efficient feedforward reconstruction. Experimental results demonstrate that our model trained for two days on four GPUs demonstrates high fidelity in reconstructing radiance fields and robustness to zeroshot and outofdomain testing. Keywords D Reconstruction D Transformer Radiance Fields Introduction The ability to reconstruct the shape and appearance of objects from multiview images has long been one of the core challenges for computer vision and graphics. Modern D reconstruction techniques achieve impressive results with various applications in visual effects ecommerce virtual and augmented reality and robotics. However they are limited to small camera baselines or dense image captures . In recent years the computer vision community has made great strides towards highquality scene reconstruction. In particular StructurefromMotion and multiview stereo emerged as powerful D reconstruction methods. They identify surface points by aggregating similarities between point features queried from source images and are able to reconstruct highly accurate surface and texture maps. Despite these successes geometry with viewconsistent textures is not the only aspect required in applications of D reconstruction. The reconstruction process should also be able to recover viewdependent appearance. To this end neural radiance fields and neural implicit surfaces investigate volumetric representations that can be learned from multiview captures without explicit feature matching. Their followups improve efficiency arXiv.v cs.CV Jul Chen et al. and quality but mostly require perscene optimization and dense multiview supervision. Several recent works thus investigate feedforward models for radiance field reconstruction while relaxing the dense input view requirement. While feedforward designs vary they commonly utilize local feature matching which however limits them to smallbaseline reconstruction since feature matching generally relies on substantial image overlap and reasonably similar viewpoints. Geometryaware transformers have also been adapted to address largebaseline problems but they often suffer from blurry reconstructions due to the lack of D inductive biases. Recent large reconstruction models learn the internal perspective relationships through context attention enabling largebaseline reconstruction. However the transformers are unaware of epipolar constraints and instead are tasked to implicitly learn spatial relationships which requires substantial data and GPU resources. In this work we present LaRa a feedforward reconstruction model without the requirement of heavy training resources for the task of bounded radiance fields reconstruction from unstructured fewviews. The core idea of our work is to progressively and implicitly perform feature matching through a novel volume transformer. We propose a Gaussian volume as D representation in which each voxel comprises a set of learnable Gaussian primitives. To obtain the Gaussian volume from image conditions we progressively update a learnable embedding volume by querying features in D. Specifically we utilize a DINO image feature encoder to obtain image tokens and lift D tokens to D by unprojecting them to a shared canonical space. Next we propose a novel Group Attention Layer architecture to enable local and global feature aggregation. Specifically we divide dense volumes into local groups and only apply attention within each group inspired by standard feature point matching. The grouped features and embeddings are fed to a crossattention sublayer to implicitly match features between feature groups of the feature volume and embedding volume which is followed by a D CNN layer to efficiently share information across neighboring groups. After passing through all attention layers the volume transformer outputs a Gaussian volume and is then decoded as D Gaussian parameters using a coarsetofine decoding process. By incorporating efficient rasterization our method achieves highresolution renderings. We demonstrate our methods efficiency and robustness for providing photorealistic novel view synthesis results using only four input images. We find that our model achieves zeroshot generalization to significantly outofdistribution inputs. Moreover our reconstructed radiance fields allow highquality mesh reconstruction using offtheshelf depthmap fusion algorithms. Finally our model achieves highquality reconstruction results using only AG GPUs within a span of days. LaRa Efficient LargeBaseline Radiance Fields Related Work Multiview stereo. Multiview stereo reconstruction aims to generate detailed D models by reasoning from images captured from multiple viewpoints which has been studied for decades . In recent years multiview stereo networks have been proposed to address MVS problems. MVSNet utilizes a D Convolutional Neural Network for processing a cost volume. This cost volume is created by aggregating features from a set of adjacent views employing the planesweeping technique from a reference viewpoint facilitating depth estimation and enabling superior D reconstructions. Subsequent research has built on top of this foundation incorporating strategies such as iterative plane sweeping point cloud enhancement confidencedriven fusion and the usage of multiple cost volumes to further refine reconstruction accuracy. However all of these works require large image overlap for faithful feature matching. Fewshot Radiance fields. The Radiance field representation has revolutionized the reconstruction field emerging as a promising replacement for traditional reconstruction methods. Despite the promising achievement in perscene sparse view reconstruction training a feedforward radiance field predictor has gained popularity. MVSNeRF proposed to combine a cost volume with volume rendering allowing appearance and geometry reconstruction only using a photometric loss. The following works are proposed to advance reconstruction quality and efficiency. Similarly to standard MVS methods they are limited to small camera baselines. Recently several works have explored feedforward models for fewshot input by capitalizing on largescale training datasets and model sizes. They leverage crossview attention to globally reason about D scenes and output D representation e.g. triplane IBplanes for radiance field reconstruction. Concurrent work by LGM and GRM introduces fewshot D reconstruction models that produce highresolution D Gaussians using a transformer framework. While these methods achieve impressive visual results training becomes expensive and less practical for the academic community. Unlike some recent single view reconstruction methods our work focuses on fewshot reconstruction since singleview input can be efficiently lifted to multiview by multiview generative models . LaRa Largebaseline Radiance Fields Our goal is to reconstruct the geometry and viewdependent appearance of bounded scenes from sparse input views using limited training resources. Given M images I I . . . IM with camera parameters . . . M our method reconstructs radiance fields as a collection of D Gaussians which is used to synthesize novel views and extract meshes. Our model is a function f of a discrete radiance field of voxel positions v and outputs a Gaussian volume VG Chen et al. Fig. Pipeline. Our method represents objects as dense voxels filled with D Gaussian primitives. We first construct D feature volumes Vf by lifting D DINO features to a canonical volume modulated by Plcker rays Section .. We then apply a volume transformer to reconstruct a Gaussian volume VG from the feature and embedding volumes Section .. We use a coarsetofine decoding process to regress D Gaussian primitive parameters Section . followed by rasterization for efficient rendering. VG Gk i K k fv I where G k i represents the primitives within ith voxel and k is the index of K primitives. The output Gaussian volume VG can be utilized for decoding into radiance fields. Our work considers sparse input views in which the camera rotates around a bounded region within a hemisphere. Our approach is designed to handle unstructured views and is flexible to accommodate various numbers of views see supplementary material. Figure shows an overview of our method. In the following we first describe how we model objects using Gaussian Volumes in which each voxel stores multiple Gaussian primitives Section .. Next we introduce how to infer the primitive parameters from multiview inputs Section .. For rendering we explore a coarsefine decoding process to enable efficient rendering with rich texture details Section .. Finally we discuss how we train our model from largescale image collections Section .. . D Representation We utilize a D voxel grid as our D representation consisting of volumes an image feature volume Vf to model image conditions an embedding volume Ve describes D prior learned from data and a Gaussian volume VG represents the radiance field. Image feature volume. We construct a feature volume for each input view by lifting the D image features to a canonical volume defined in the center of the LaRa Efficient LargeBaseline Radiance Fields scene. We use the DINO image encoder to extract perview image features and inject Plcker ray directions into the features via adaptive layer norm . Unlike previous works that modulate camera poses to image features using extrinsic and intrinsic matrices Plcker rays are defined by the cross product between the camera location and ray direction offering a unique ray parameterization independent of object scale camera position and focal length. After modulation we obtain M perview image feature maps. We further lift the D maps to D by backprojecting the feature maps to a canonical volume therefore resulting in M feature volumes Vf RWWWO with O channels. Embedding volume. Inspired by prior works we construct a learnable embedding volume Ve RWWWC for modeling prior knowledge. D reconstruction is generally underconstrained in sparse view settings hence prior knowledge is critical for faithful reconstructions. We propose to leverage a D embedding volume to model and learn prior information across objects which acts as a D object template that greatly reduces the solution space. The embedding volume is aligned with the image feature volume allowing for efficient cross attention see Section .. Gaussian volume. To achieve efficient rendering we propose to use dense primitives as an object representation and output a set of D Gaussians from the image feature volume and embedding volume. However predicting a set of dense unordered point sets without D supervision is always a challenge for neural networks. To this end we introduce a dense Gaussian volume representation that can effectively model points densely near the objects surface while being suitable for modern network architectures by facilitating prediction and generation. Specifically our Gaussian volume comprises K learnable Gaussian primitives per voxel where each primitive can move freely within a constrained spherical region centered at the voxels center. For primitive modeling we borrow the shape and appearance parametrization from D Gaussian splatting for better surface modeling. Each Gaussian has an opacity tangent vectors ttu tv a scaling vector S su sv controlling the shape of the D Gaussian and spherical harmonics coefficients for viewdependent appearance. Furthermore we substitute the primitives position with an offset vector incorporating a scaled sigmoid activation function. Consequently the position of Gaussian primitive k in voxel vi is expressed as p k i vi r k i where r signifies the maximum displacement range of the primitive. In this way primitives are restricted to neighborhoods of uniformly distributed local centers. The inclusion of offset modeling allows each voxel to effectively represent adjacent regions that require it. This reduces unnecessary capacity in empty space and enhances the representational capacity compared to the standard dense volume. We refer the reader to the supplementary material for more details on D Gaussian splatting. . Volume Transformer To predict the Gaussian volume we propose a volume transformer architecture to perform attention between volumes. Selfattention and crossattention modules Chen et al. Fig. Volume Transformer. We aggregate the embedding volume Ve and feature volume Vf through a series of Group Attention Layers that progressively match features. In each layer the volumes are first unfolded into local groups. Subsequently a layer normalization is applied followed by a GroupCrossAttn sublayer. This is followed by another normalization and an MLP layer. The output is reshaped back to the original embedding volume shape processed by a D convolution layer and forwarded to the next layer. To connect the output of the sublayers we use residual connections. as commonly used in transformers are inefficient for volumes since the number of tokens grows cubically with the resolution of the D representation. Nave applications thus result in long training times and large GPU memory requirements. In addition geometry constraints and regional matching play crucial roles in the context of D reconstruction which should be considered in the attention design. We now present our novel volume transformer containing a set of group attention layers that progressively update the embedding volume. Our group attention layers contain three sublayers see Figure group crossattention a multilayer perceptron MLP and D convolution. Given the image feature volume and embedding volume we first unfold these D volumes i.e. Vf and Ve into G local token groups along each axis. We then apply a crossattention layer between the corresponding groups of embedding tokens Vgj e and image feature tokens V g f where j denotes the index of the layer starting from and Vg e g Ve. Figure illustrates the unfolding for G and highlights the corresponding groups. The next sublayer is an MLP similar to the original transformer . The updated embedding groups V gj e G g are reassembled into the original volume shape resulting in V j e which are subsequently processed by a D convolutional layer to share information between groups and enable the intramodel connections within the spatially organized voxels. In summary dot bV text egj text GroupCrossAttnleft text LNleft bV text egjright bV gtext fright bV text egj text ddot bV text egj text MLPleft text LNleft dot bV text egjright right dot bV text egj text bV text ej text DCNNleft text LNleft ddot bV text ejright right ddot bV text ej text . To incorporate information from multiple views we flatten and concatenate the image feature tokens from multiview feature volumes. It is important to note LaRa Efficient LargeBaseline Radiance Fields Fig. Coarsefine decoding. Top row A coarse decoding module transforms the voxel features Vi G into K D Gaussian parameters representing shape specifically t S and appearance denoted as SHcoarse. This step is followed by a splatting procedure. On the bottom a fine decoding module aggregates rendering buffers i.e. RGB depth and alpha maps from the coarse module volume feature and source images for appearance enhancement. It projects the centers of primitives onto these buffers applies crossattention with the voxel features Vi G and produces residual spherical harmonics SHresiduals. These residuals are added to the coarse spherical harmonics for a refined splatting process. that different groups are processed simultaneously by the group attention layer across the batch dimension. This parallel processing allows for a larger training batch size within the attention sublayer reducing the number of training steps required. In addition using a D convolution layer increases inference efficiency compared to the popular selfattention layer. Also we also apply layer norms LM between the sublayers. Finally the output embedding volume Vj e serves as input for the subsequent jth group attention layer. After passing through all in our experiments group attention layers we use a D transposed CNN to scale up the updated embedding volume V e VG TransposeDCNN V e . Now we have a Gaussian volume VG each Gaussian voxel is a D feature vector Vi G R B representing the primitives associated with the voxel. . CoarseFine Decoding We obtain D Gaussian primitive shape and appearance parameters from the Gaussian volume so we introduce a coarsefine decoding process to better recover texture details. Instead of using a single network and sampling scheme to reason about the scene we simultaneously optimize two decoding modules one coarse and one fine. For the coarse decoding module we feed Gaussian volume features to a lightweight MLP and output a set of K Gaussian parameters per voxel. We employ the efficient D splatting technique to form highresolution renderings including RGB depth opacity and normal maps. During training we render M input views and M novel views for supervision. Chen et al. Despite the fact that the coarse renderings can already provide faithful depthsgeometries the appearance tends to be blurred as shown in e of Figure . This is because the image texture can easily be lost after the DINO encoder and the Group Attention layers. To address this problem we propose a fine decoding module to guide fine texture prediction. Specifically we project the primitive centers p k i onto the coarse renderings i.e. RGB image I depth image D and accumulation alpha map A to contain the coarse renderings for each primitive using the camera poses cX bp ki left bI bp kihat bI bp ki hat bD bp ki hat mathrm bA bp kiright bPhi left cP left bp ki bpi right oplus left bI hat bI hat bD hat mathrm bA right right text where P denotes the point projection is a concatenation operation along the channel dimension and is a bilinear interpolation in D space. In practice the depth features can change significantly across different scenes. To mitigate scaling discrepancies we replace the rendering depth D pk i with a displacement feature    D pk i zpk i    that compares the rendered depth for input views and the depth zpk i of a primitive allowing for occlusionaware reasoning. We then apply a pointbased crossattention layer to establish relationships between the features of a point Xpk i and the primitive voxel. The results of this crossattention process are then fed into an MLP which is tasked with predicting the residual spherical harmonics text SHiktextit residuals text MLPleft text CrossAttnleft cX bp ki bV itext e right right text text SHiktextit fine text SHiktextit coarse text SHiktextit residuals text . Intuitively the fine decoding module attempts to learn a geometryaware texture blending process based on multiview images primitive features and rendering buffers from the coarse module. Furthermore both coarse and fine modules are differentiable and updated simultaneously. Thus the fine renderings can further regularize the coarse predictions. Splatting. Our work takes advantage of Gaussian splatting to facilitate efficient highresolution image rendering. We follow the original rasterization process and further output depth and normal maps by integrating the z value and the normal of the primitives. . Training Our LaRa is optimized across scenes via gradient descent minimizing simple image reconstruction objectives between the coarse and fine renderings i.e. I and the groundtruth images i.e. I cL cL text MSEcI hat cI cL text SSIMcI hat cI cL text Reg text label eqloss where LMSE is the pixelwise L loss LSSIM is the structural similarity loss which are applied on both coarse and fine RGB outputs. LaRa Efficient LargeBaseline Radiance Fields Regularization terms. We find that only applying the photometric reconstruction losses is adequate for rendering. However the consistency across views is low because of the strong flexibility of the discrete Gaussian primitives. To encourage the primitives to be constructed on the surface we follow D Gaussian splatting that utilize a selfsupervised distortion loss Ld and a normal consistency loss Ln to regularize the training. Specifically we concentrate the weight distribution along the rays by minimizing the distance between the rayprimitive intersections inspired by MipNeRF . Given a ray ux of pixel x we obtain its distortion loss by cL text d sum ijomega iomega jzizjtext label eqnlossdistortion where i i Giux Qi j j Gj ux is the blending weight of the ith intersection and zi the depth of the intersection point. As D Gaussians explicitly model the primitive normals we can align their normals ni with the normals N derived from the depth maps via the loss cL text n sum i omega i bn itop bN text . label eqnlossnormal Therefore our regularization term for the ray ux is given by LReg dLdnLn. We set d and n . in our experiments. Implementation Details We briefly discuss our implementation including the training and evaluation dataset network design optimizer and mesh extraction. Datasets. We train our model on multiview synthetic renderings of objects based on the Objaverse dataset which includes scenes with a traintest split of . Each scene contains circular views with an image resolution of . To ensure sufficient angular coverage of the input views we employ the classical Kmeans algorithm to cluster the cameras into clusters. During training we randomly choose two views from each cluster for every iteration in which the first images share the same camera poses as the input views while the remaining images are novel view outputs. We employ the eight output images for supervision and leverage the loss objectives outlined in Eq. to update the network. We present our indomain evaluation using the Objaverse datasets test set consisting of scenes. To assess our models crossdomain applicability we conducted tests on the Google Scanned Objects dataset which contains scans of real objects and on the hydrants and teddy bears from the CoD test set totaling objects. To examine our models performance on zeroshot reconstruction task we use the generative multiview dataset from InstantD which comprises scenes generated from text prompts. Chen et al. Network. We developed LaRa using PyTorch Lightning and conduct our training on NVIDIA AG GPUs over a period of days for the fast model and . days for the base model with a batch size of per GPU. We use DINObase for encoding M multiview images at a resolution of . We use a volume resolution of W with C channels for the image feature volume and a resolution of W with C channels for the embedding volume dividing both into G groups for the group attention layers. Our group attention network consists of layers producing a Gaussian volume of size . We choose K primitives for each voxel and constrain the offset radius to r in our experiments. The total number of trainable parameters is million. Training. The optimization is carried out using the AdamW optimizer starting with a learning rate of and following a cosine annealing schedule with a period of epochs. Our final model is trained for epochs comprising iterations for each epoch. We observe that applying the regularization loss from the start can slow down the convergence regarding the shape. This is because regularization objectives tend to encourage thinner surfaces which may result in premature convergence to local minima if the shapes are noisy. In our experiments we thus enable regularization after the first epochs. Mesh extraction. To obtain a mesh from reconstructed D primitives we generate RGBD maps by rendering along three circular video trajectories at elevations of and . Inside the scene bounding box we construct a signed distance function volume and apply truncated SDF TSDF fusion to integrate the reconstructed rgb and depth maps allowing for efficient textured mesh extraction. In our experiments we use a resolution of and set a truncation threshold of . for TSDF fusion. Experiments We now present an extensive evaluation of LaRa our largebaseline radiance field. We first compare with previous and concurrent works on indomain and zeroshot generalization settings. We then analyze the effect of local attention regularization term and renderer. . Comparison We compare our method against MVSNeRF MuRF and the concurrent work LGM . The first two methods are key representatives of feature matchingbased methods and the latter shares a conceptually similar approach of using Gaussian primitives for largebaseline settings. It is worth noting that while existing feedforward radiance field reconstruction methods are capable of being evaluated in largebaseline settings retraining these methods to establish a new largebaseline benchmark on the Objaverse dataset is both time and GPU LaRa Efficient LargeBaseline Radiance Fields Fig. Rendering results of unseen scenes. The top two rows compare our reconstructions with MVSNeRF LGM on CoD . We also show the view synthesis results for Gobjaverse GSO and generative multiview datasets arranged from top to bottom. Note that visual results from MuRF are not shown due to their lack of content appearing as white images. The above results are reconstructed using input views. Chen et al. Table Quantitative results of novel view synthesis. Our fast model is trained for epochs days on GPUs while ours is trained for epochs taking . days. Gobjaverse GSO CoD Method PSNR SSIM LPIPS PSNR SSIM LPIPS PSNR SSIM LPIPS MVSNeRF . . . . . . . . . MuRF . . . . . . . . . LGM . . . . . . . . . Oursfast . . . . . . . . . Ours . . . . . . . . . intensive. Here we retrain MVSNeRF and the current stateoftheart feedforward radiance field reconstruction method MuRF . Appearance. Table shows quantitative results PSNR SSIM and LPIPS comparisons. Our method achieves clearly improved rendering quality for both indomain generation Gobjaverse testing set and zeroshot generalization GSO and CoD datasets. As shown in Figure MVSNeRF fails to provide faithful reconstructions on the largebaseline setting and tends to produce floaters within the reconstruction regions since the cost volume is extremely noisy in the sparse view scenarios resulting in a challenge for its convolution matching network to distinguish the surface. MuRF quickly overfits the white background and produces empty predictions for all inputs. Instead of predefining and constructing the feature similarity as network input our method injects volume features to the intermiddle attention layer and implicitly and progressively matches them through the attention mechanism between the volume feature and updated embeddings achieving clearer and overall better reconstructions. Our approach is robust to scene scale and can generalize to real captured images such as those in the CoD dataset thanks to our canonical modeling and projectionbased feature lifting. In contrast LGM leverages a monocular prediction and fusion technique that requires a reference scene scale and a constant cameraobject distance to avoid focal length and distance ambiguity. This requirement significantly limits its generalizability to real data. As shown in Table and Figure LGM provides faithful reconstructions in datasets with a strict constant cameraobject distance such as GSO but fails to generalize to unconstraint multiview data such as in Objaverse and CoD datasets and exhibits serious distortions. Our model trained on AG GPUs for days demonstrates superior results compared to the LGM model trained on A G GPUs GPUs RAM GPU hours and on the same synthetic Objaverse dataset . Furthermore our approach also performs well for generative multiview images where textures are not consistent across views. In this comparison we only present a qualitative analysis due to the absence of ground truths as illustrated in the bottom rows of Figure . Our method offers detailed texture and smooth surface reconstruction. We invite the reader to our Appendix for more results. LaRa Efficient LargeBaseline Radiance Fields Fig. Zeroshot reconstruction results. Our approach achieves faithful surface reconstruction for images generated from text. These images are produced using a pretrained texttoimage model . Table Depth reconstruction. We evaluate geometry quality within the mask by measuring L error and the accuracy in terms of the percentage of pixels that fall below thresholds of . . .. Method Abs err Acc . Acc . Acc . MVSNeRF . . . . LGM . . . . Oursfast . . . . Ours . . . . Geometry. We evaluate the quality of our geometry reconstruction by comparing the depth reconstructions on novel views generated by a weighted sum of the z values of the primitives. As shown in Table our approach achieves significantly lower L errors and higher geometry accuracy other baselines. In Figure we also visualize geometry reconstruction by extracting meshes using TSDF. In addition our trajectory video rendering views at a resolution of together with mesh extraction is highly efficient as it does not require finetuning and can be performed in just seconds. . Ablation Study We now analyze the contributions of individual elements of our model design. To reduce the training cost we reduce the training from to epochs for ablations. Fig. Ablation study on a Shell scene. We report the PSNR for each example at the top. Here the fast model corresponds to the full model detailed in Table . Chen et al. Table Ablations. See Section . for descriptions. Gobjaverse GSO Design PSNR SSIM LPIPS PSNR SSIM LPIPS Geo a G . . . . . . ... b G . . . . . . ... c wo LReg . . . . . . ... d DGS . . . . . . ... e coarse . . . . . . ... f SH order . . . . . . ... g full model . . . . . . ... h views . . . . . . ... i views . . . . . . ... Effect of local attention. We first evaluate the contribution of our group partition using different group numbers. Here G is equivalent to the standard crossattention layer however using such group size can lead to much higher compute time for the same number of iterations i.e. days on As for epochs. Therefore our ablation starts with groups for acceptable training time. As shown in ablations a b and g in Table and Figure the image synthesis and geometry quality are consistently improved with a larger group number thanks to the local attention mechanism. Effect of regularization term. We further evaluate the regularization term introduced in Eq. and Eq. . We observe a marked improvement in the average rendering score when disabling the regularization. Although this provides a stronger model capability for modeling details this may cause floaters near the surfaces as shown in c and d of Figure which leads to inconsistent freeviewpoint video rendering see Appendix video. In contrast our approach is able to reconstruct hard surfaces. Effect of renderer. We also compare D Gaussian splatting with D Gaussian splatting in our framework as shown in c and d. They achieve similar rendering quality and we choose DGS to facilitate surface regularization and mesh extraction. Furthermore to evaluate the effectiveness of the coarsefine decoding we conduct an evaluation of the coarse outputs shown in row e. Our fine decoding is able to provide richer texture details. Effect of renderer. We also compare D Gaussian splatting with D Gaussian splatting in our framework as shown in c and d. They achieve similar rendering quality and we choose DGS to facilitate surface regularization and mesh extraction. Furthermore to evaluate the effectiveness of the coarsefine decoding we conduct an evaluation of the coarse outputs shown in row e. Our fine decoding is able to provide richer texture details. LaRa Efficient LargeBaseline Radiance Fields Effect of input views. Our approach is highly efficient and compatible with different numbers of input views. In prior experiments we utilize views for both training and inference as our standard configuration. We evaluate our method in testing views as shown in rows hj and g using the full model in row g. Conclusion We have presented LaRa a novel method for bounded radiance fields reconstruction from largebaseline inputs. Our central idea is to match image features and embedding volume through unified local and global attention layers. By integrating this with a coarsefine decoding and splatting process we achieve high efficiency for both training and inference. In future work we plan to explore how to enlarge the batch size perGPU and volume resolution without increasing GPU usage. In addition we hope to investigate how to extend it to handle unbounded scenes and decompose the radiance field into its constituent physical material and lighting components. Limitations and Discussions. Our LaRa demonstrates a remarkable efficiency feedforward model that achieved highfidelity allaround novelview synthesis and surface reconstruction from sparse largebaseline images. However our approach struggles to recover highfrequency geometry and texture details mainly due to the low volume resolution. Enhancing our approach with techniques such as gradient checkpointing or mixedprecision training can potentially increase training batch size as well as volume resolution. We have also noticed that our method can yield inconsistent rendering results when the geometry is incorrectly estimated or when reconstructing multiview inconsistent inputs as demonstrated in the comparison video. This occurs because our method utilizes secondorder Spherical Harmonic appearance modeling. While such modeling can capture viewdependent effects it also introduces a stronger ambiguity between geometry and appearance. We believe that incorporating our method with a physicallybased rendering process can potentially address this issue. In addition our work assumes posed inputs but estimating precise camera poses for sparse views is a challenge in practice. Incorporating a pose estimation module into the feedforward setting is an orthogonal direction to our work. Acknowledgements We thank Bozidar Antic for pointing out a bug that resulted in an improvement of about dB. Special thanks to BinBin Huang and Zehao Yu for their helpful discussion and suggestions. We would like to thank Bi Sai Jiahao Li Zexiang Xu for providing us with the testing examples of InstantD and Jiaxiang Tang for helping us to construct a comparison with LGM. Chen et al. References . Anciukeviius T. Manhardt F. Tombari F. Henderson P. Denoising diffusion via imagebased rendering. In ICLR . Barron J.T. Mildenhall B. Tancik M. Hedman P. MartinBrualla R. Srinivasan P.P. MipNeRF A multiscale representation for antialiasing neural radiance fields. In ICCV . Caron M. Touvron H. Misra I. Jgou H. Mairal J. Bojanowski P. Joulin A. Emerging properties in selfsupervised vision transformers. In ICCV . Charatan D. Li S. Tagliasacchi A. Sitzmann V. pixelsplat d gaussian splats from image pairs for scalable generalizable d reconstruction . Chen A. Xu Z. Geiger A. Yu J. Su H. TensoRF Tensorial radiance fields. In ECCV . Chen A. Xu Z. Wei X. Tang S. Su H. Geiger A. Dictionary fields Learning a neural basis decomposition. ACM Trans. Graph. . Chen A. Xu Z. Wei X. Tang S. Su H. Geiger A. Factor fields A unified framework for neural fields and beyond. arXiv.org . Chen A. Xu Z. Zhao F. Zhang X. Xiang F. Yu J. Su H. MVSNeRF Fast generalizable radiance field reconstruction from multiview stereo. In ICCV . Chen R. Han S. Xu J. Su H. Pointbased multiview stereo network. In ICCV . Chen Y. Xu H. Wu Q. Zheng C. Cham T.J. Cai J. Explicit correspondence matching for generalizable neural radiance fields. arXiv.org . Chen Y. Xu H. Zheng C. Zhuang B. Pollefeys M. Geiger A. Cham T.J. Cai J. Mvsplat Efficient d gaussian splatting from sparse multiview images. arXiv.org . Cheng S. Xu Z. Zhu S. Li Z. Li L.E. Ramamoorthi R. Su H. Deep stereo using adaptive thin volume representation with uncertainty awareness. In CVPR . Chibane J. Bansal A. Lazova V. PonsMoll G. Stereo radiance fields SRF Learning view synthesis for sparse views of novel scenes. In CVPR . De Bonet J.S. Viola P. Poxels Probabilistic voxelized volume reconstruction. In ICCV . Deitke M. Schwenk D. Salvador J. Weihs L. Michel O. VanderBilt E. Schmidt L. Ehsani K. Kembhavi A. Farhadi A. Objaverse A universe of annotated D objects. In CVPR . Deng K. Liu A. Zhu J.Y. Ramanan D. Depthsupervised NeRF Fewer views and faster training for free. In CVPR . Dosovitskiy A. Beyer L. Kolesnikov A. Weissenborn D. Zhai X. Unterthiner T. Dehghani M. Minderer M. Heigold G. Gelly S. Uszkoreit J. Houlsby N. An image is worth words Transformers for image recognition at scale. In ICLR . Downs L. Francis A. Koenig N. Kinman B. Hickman R. Reymann K. McHugh T.B. Vanhoucke V. Google scanned objects A highquality dataset of D scanned household items. In ICRA . Du Y. Smith C. Tewari A. Sitzmann V. Learning to render novel views from widebaseline stereo pairs. In CVPR . Falcon W. The PyTorch Lightning team PyTorch Lightning https github.comLightningAIlightning LaRa Efficient LargeBaseline Radiance Fields . FridovichKeil S. Yu A. Tancik M. Chen Q. Recht B. Kanazawa A. Plenoxels Radiance fields without neural networks. In CVPR . Furukawa Y. Ponce J. Accurate dense and robust multiview stereopsis. PAMI . Goesele M. Snavely N. Curless B. Hoppe H. Seitz S.M. Multiview stereo for community photo collections. In ICCV . Gu X. Fan Z. Zhu S. Dai Z. Tan F. Tan P. Cascade cost volume for highresolution multiview stereo and stereo matching. In CVPR . Hernndez Esteban C. Schmitt F. Silhouette and stereo fusion for D object modeling. Computer Vision and Image Understanding . Hong Y. Zhang K. Gu J. Bi S. Zhou Y. Liu D. Liu F. Sunkavalli K. Bui T. Tan H. LRM Large reconstruction model for single image to D. In ICLR . Huang B. Yu Z. Chen A. Geiger A. Gao S. d gaussian splatting for geometrically accurate radiance fields. ACM SIGGRAPH . Im S. Jeon H.G. Lin S. Kweon I.S. DPSNet Endtoend deep plane sweep stereo. In ICLR . Jaegle A. Gimeno F. Brock A. Vinyals O. Zisserman A. Carreira J. Perceiver General perception with iterative attention. In Meila M. Zhang T. eds. ICML . Johari M.M. Lepoittevin Y. Fleuret F. GeoNeRF Generalizing NeRF with geometry priors. In CVPR . Karras T. Laine S. Aila T. A stylebased generator architecture for generative adversarial networks. In CVPR . Kerbl B. Kopanas G. Leimkhler T. Drettakis G. D Gaussian splatting for realtime radiance field rendering. ACM Trans. on Graphics . Kolmogorov V. Zabih R. Multicamera scene reconstruction via graph cuts. In ECCV . Kulhnek J. Derner E. Sattler T. Babuka R. ViewFormer NeRFfree neural rendering from few images using transformers. In ECCV . Kutulakos K.N. Seitz S.M. A theory of shape by space carving. International journal of computer vision . Li J. Tan H. Zhang K. Xu Z. Luan F. Xu Y. Hong Y. Sunkavalli K. Shakhnarovich G. Bi S. InstantD Fast texttoD with sparseview generation and large reconstruction model. In ICLR . Lin H. Peng S. Xu Z. Yan Y. Shuai Q. Bao H. Zhou X. Efficient neural radiance fields for interactive freeviewpoint video. In SIGGRAPH Asia . Liu R. Wu R. Hoorick B.V. Tokmakov P. Zakharov S. Vondrick C. Zero to Zeroshot one image to D object. In ICCV . Long X. Guo Y.C. Lin C. Liu Y. Dou Z. Liu L. Ma Y. Zhang S.H. Habermann M. Theobalt C. et al. WonderD Single image to D using crossdomain diffusion. arXiv.org . Loshchilov I. Hutter F. Decoupled weight decay regularization. In ICLR . Luo K. Guan T. Ju L. Huang H. Luo Y. PMVSNet Learning patchwise matching confidence aggregation for multiview stereo. In ICCV . Mildenhall B. Srinivasan P.P. Tancik M. Barron J.T. Ramamoorthi R. Ng R. NeRF Representing scenes as neural radiance fields for view synthesis. In ECCV . Miyato T. Jaeger B. Welling M. Geiger A. GTA A geometryaware attention mechanism for multiview transformers. In ICLR Chen et al. . Mller T. Evans A. Schied C. Keller A. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. on Graphics . NguyenPhuoc T. Li C. Theis L. Richardt C. Yang Y. HoloGAN Unsupervised learning of D representations from natural images. In ICCV . Niemeyer M. Barron J. Mildenhall B. Sajjadi M.S.M. Geiger A. Radwan N. RegNeRF Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR . Niemeyer M. Mescheder L. Oechsle M. Geiger A. Differentiable volumetric rendering Learning implicit D representations without D supervision. In CVPR . Peebles W. Xie S. Scalable diffusion models with transformers. In ICCV . Reizenstein J. Shapovalov R. Henzler P. Sbordone L. Labatut P. Novotny D. Common objects in D Largescale learning and evaluation of reallife D category reconstruction. In ICCV . Schnberger J.L. Zheng E. Frahm J.M. Pollefeys M. Pixelwise view selection for unstructured multiview stereo. In ECCV . Schnberger J.L. Frahm J.M. Structurefrommotion revisited. In CVPR . Seitz S.M. Curless B. Diebel J. Scharstein D. Szeliski R. A comparison and evaluation of multiview stereo reconstruction algorithms. In CVPR . Shi R. Wei X. Wang C. Su H. ZeroRF Fast sparse view reconstruction with zero pretraining arXiv. . Shi Y. Wang P. Ye J. Mai L. Li K. Yang X. MVDream Multiview diffusion for D generation. In ICLR . Snavely N. Seitz S.M. Szeliski R. Photo tourism exploring photo collections in D. ACM Trans. on Graphics . Somraj N. Karanayil A. Soundararajan R. SimpleNeRF Regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia . Sun C. Sun M. Chen H. Direct voxel grid optimization Superfast convergence for radiance fields reconstruction. In CVPR . Szymanowicz S. Rupprecht C. Vedaldi A. Splatter image Ultrafast singleview d reconstruction. CVPR . Tang J. Chen Z. Chen X. Wang T. Zeng G. Liu Z. LGM Large multiview Gaussian model for highresolution D content creation. arXiv.org . Truong P. Rakotosaona M.J. Manhardt F. Tombari F. SPARF Neural radiance fields from sparse and noisy poses. In CVPR . Vaswani A. Shazeer N. Parmar N. Uszkoreit J. Jones L. Gomez A.N. Kaiser L. Polosukhin I. Attention is all you need. In NIPS. pp. . Venkat N. Agarwal M. Singh M. Tulsiani S. Geometrybiased transformers for novel view synthesis. arXiv.org . Wang G. Chen Z. Loy C.C. Liu Z. SparseNeRF Distilling depth ranking for fewshot novel view synthesis. In ICCV . Wang P. Liu L. Liu Y. Theobalt C. Komura T. Wang W. NeuS Learning neural implicit surfaces by volume rendering for multiview reconstruction. In NeurIPS . Wang P. Tan H. Bi S. Xu Y. Luan F. Sunkavalli K. Wang W. Xu Z. Zhang K. PFLRM Posefree large reconstruction model for joint pose and shape prediction. In ICLR LaRa Efficient LargeBaseline Radiance Fields . Wang Q. Wang Z. Genova K. Srinivasan P. Zhou H. Barron J.T. MartinBrualla R. Snavely N. Funkhouser T. IBRNet Learning multiview imagebased rendering. In CVPR . Wang S. Leroy V. Cabon Y. Chidlovskii B. Jerome R. Dustr Geometric d vision made easy. CVPR . Wu R. Mildenhall B. Henzler P. Park K. Gao R. Watson D. Srinivasan P.P. Verbin D. Barron J.T. Poole B. Holynski A. ReconFusion D reconstruction with diffusion priors arXiv. . Xu C. Dong Y. Zuo Q. Zhang J. Ye X. Geng W. Zhang Y. Gu X. Qiu L. Zhao Z. Qing R. Jiayi J. Dong Z. Bo L. Gbuffer Objaverse Highquality rendering dataset of Objaverse httpsaigcd.github.iogobjaverse . Xu H. Chen A. Chen Y. Sakaridis C. Zhang Y. Pollefeys M. Geiger A. Yu F. MuRF Multibaseline radiance fields. In CVPR . Xu Y. Tan H. Luan F. Bi S. Wang P. Li J. Shi Z. Sunkavalli K. Wetzstein G. Xu Z. Zhang K. DMVD Denoising multiview diffusion using D large reconstruction model. In ICLR . Yao Y. Luo Z. Li S. Fang T. Quan L. MVSNet Depth inference for unstructured multiview stereo. In ECCV . Yao Y. Luo Z. Li S. Shen T. Fang T. Quan L. Recurrent MVSNet for highresolution multiview stereo depth inference. In CVPR . Yariv L. Kasten Y. Moran D. Galun M. Atzmon M. Ronen B. Lipman Y. Multiview neural surface reconstruction by disentangling geometry and appearance. In NIPS . Yinghao X. Zifan S. Wang Y. Hansheng C. Ceyuan Y. Sida P. Yujun S. Gordon W. Grm Large gaussian reconstruction model for efficient d reconstruction and generation . Yu A. Ye V. Tancik M. Kanazawa A. pixelNeRF Neural radiance fields from one or few images. In CVPR . Yu Z. Chen A. Huang B. Sattler T. Geiger A. Mipsplatting Aliasfree d gaussian splatting. CVPR

write a youtube script

arXiv.v cs.CC Jul Determination Problems for Orbit Closures and Matrix Groups Rida Ait El Manssour George Kenison Mahsa Shirmohammadi and James Worrell IRIF CNRS Universit Paris Cit Paris France Liverpool John Moores University Liverpool UK University of Oxford Oxford UK Abstract Computational problems concerning the orbit of a point under the action of a matrix group occur in numerous subfields of computer science including complexity theory program analysis quantum computation and automata theory. In many cases the focus extends beyond orbits proper to orbit closures under a suitable topology. Typically one starts from a group and several points and asks questions about the orbit closure of the points under the action of the group e.g. whether two given orbit closures intersect. In this paper we consider a collection of what we call determination problems concerning groups and orbit closures. These problems begin with a given variety and seek to understand whether and how it arises either as an algebraic group or as an orbit closure. The how question asks whether the underlying group is sgenerated meaning it is topologically generated by s matrices for a given number s. Among other applications problems of this type have recently been studied in the context of synthesising loops subject to certain specified invariants on program variables. Our main result is a polynomialspace procedure that inputs a variety V and a number s and determines whether V arises as an orbit closure of a point under an sgenerated commutative matrix group. The main tools in our approach are rooted in structural properties of commutative algebraic matrix groups and lattice theory. We leave open the question of determining whether a variety is an orbit closure of a point under an algebraic matrix group without the requirement of commutativity. In this regard we note that a recent paper Nos gives an elementary procedure to compute the orbit closure of a point under finitely many matrices. manssouririf.fr g.j.kenisonljmu.ac.uk mahsairif.fr jbwcs.ox.ac.uk Introduction Orbit Problems. The computational study of orbits of matrix groups stretches back many decades. One of the most fundamental problems in this area is determining for a given field F whether a given pair of vectors u v F d lie in the same orbit under the action of a finitely generated subgroup G of the general linear group GLdF that is whether u G v. For cyclic matrix groups over Q this problem reduces to the KannanLipton orbit problem. For such groups there exists a polynomialtime algorithm for deciding whether a given pair of vectors lie in the same orbit KL KL. The extension of the KannanLipton orbit problem to a group with a finite generating set M . . . Ms over Q and number fields is known to be decidable when the generating set comprises commuting matrices. However this problem becomes undecidable in the general case Bab. The exploration of this problem over finite fields was instrumental in leading Babai to introduce the concept of interactive proofs Bab. Orbit Closures. For many applications including in program analysis and geometric complexity theory it makes sense to study orbit closures in lieu of orbits proper Br Br DJK Hru. Given a field F and a group G GLdF acting on F d the orbit closure of v F d denoted by G v is the closure of the orbit G v in the Zariski topology. We note that over the field C if the group in question is Zariski closed then the Zariski closure of an orbit coincides with its closure in the Euclidean topology. The orbitclosure containment problem asks given vectors u v F d and a group G GLdF to determine whether u lies in G v whereas the orbitclosure intersection problem asks whether G v G u . Orbit closures feature in complexity theory program analysis quantum computation and automata theory. A striking application is geometric complexity theory in which many recent studies centred around the formulation of the VP VNP problem in terms of orbitclosure containment with respect to the action of GLdQ on polynomial rings Br Br. In certain applications such as nonconvex optimisation problems noncommutative rational identity testing and graph isomorphism Bl Br DM FS one considers the orbit closure of a point v under a linear algebraic group G that is explicitly presented via a given set of equations that is G is the zero set of a given polynomial ideal hP . . . Pki. In other applications such as quantum computing and program analysis DJK Hru one wishes to compute the orbit closure of a group G that is implicitly presented via a finite set of topological generators that is G hM . . . Msi. We refer to these settings as explicit and implicit presentations of the orbitclosure problems respectively. Given a polynomial ideal hP . . . Psi Kx defining an algebraic matrix group G GLdF with K an effective subfield of F as highlighted in Bl orbitclosure containment and intersection can easily be checked in existential fragment of the first order theory of the field F. By Koirans seminal results on Hilberts Nullstellensatz over Q Koi the explicit orbitclosure problems over Q are both in AM assuming the generalized Riemann hypothesis GRH. Furthermore answering a question posed by Wigderson it is shown in Bl that the explicit orbitclosure containment problem is NPhard and shown in Bl Theorem .. that this problem over R is polynomialtime equivalent to the existential theory over the reals ETR . The primary application of the orbit problem in that work is the formulation of the slice rank of tensors in the union of orbitclosures. In Br several other applications of the orbitclosure problems are identified in combinatorial optimisation and dynamical systems specifically where the underlying group is assumed to be commutative. Notably the results in Br fully resolve the problem for a subclass of commutative In Bl the problem is simply called orbitclosure containment. groups namely tori. However the complexity of explicit orbitclosure problems for commutative groups remains open. In the implicit orbitclosure problems the main challenge in computing an orbit closure lies in computing the Zariski closure of the group in question. Given a set M . . . Ms of matrices over a number field K the first algorithm to compute the polynomial ideal defining hM . . . Msi was introduced in DJK. The complexity bound of the algorithm therein is not known to be elementary largely due to an iterative call to a subprocedure that takes the quotient of linear algebraic groups and the inherent difficulty of forming such quotients Nos Appendix C. An elementary algorithm for implicit orbitclosure can be derived from a linearisation technique of MSa together with a recent result in Nos. The key result in Nos is a quantitative structure lemma for algebraic matrix groups providing an upper bound on the degree of the polynomials defining hM . . . Msi. The abovementioned complexity bounds for the implicit orbitclosure problem are in the order of seven fold exponential time when the generating matrices are over Q. It remains a challenging open problem to close the complexity gap borrowing the lower bound from the explicit orbitclosure problems. Further algorithms to compute the Zariski closure of matrix groups and semigroups are presented in Gra Hru Hru. Implicit orbitclosure problem in quantum computation automata theory and program analysis have been the subject of extensive interest after the resolution of decade long open problems such as the equivalence problem for deterministic topdown treetostring transducers SMK and the threshold problem for quantum automata DJK a natural version of the languageemptiness problem. Another application of closure problem in quantum setting is to test whether a finite set of quantum gates is universal Nos. Orbit closures feature in program analysis when one wants to automatically compute polynomial invariants of certain classes of loop programs CK Hru Kin MSb SSM the task of automatic invariant generation has been considered as the most important task in program verification Bey. . Determination Problems In this paper we investigate a series of determination problems related to groups and their orbit closures. These problems start with a given variety and examine whether it can be realized as an algebraic group or as an orbit closure with the constraint that the underlying group is topologically sgenerated. We define an algebraic group G GLdQ to be topologically sgenerated if there is a set S GLdQ of matrices with cardinally s such that G hSi. As noted in Proposition algebraic groups are always topologically generated by a finite set. In this context determining whether a variety Z Qd arises as an orbit closure under the action of G is in principle straightforward. In such a case G is necessarily a subgroup of the group SymZ A GLdQ AZ Z of matrices that fix Z and so we may assume without loss of generality that G equals the latter group. But SymZ is definable in firstorder logic over Q and hence the question of whether Z arises as the orbit closure of a point under SymZ reduces to the decision problem for this theory. It further holds by Proposition that SymZ has a Zariski dense subgroup that is finitely generated as a group. Whence Z is the orbit closure of a point under some finitely generated matrix group if and only if it is the orbit closure of a point under SymZ. In comparison it is more challenging to determine whether a given variety is the orbit closure of a topologically sgenerated group than simply determining whether it is an orbit closure tout court. Our main determination problems are as follows. Let x xijijd be a tuple of variables. The Group Determination Problem asks given s N and a family of polynomials f . . . fm Qx of total degree at most b to determine whether their zero locus Z GLdQ is an sgenerated matrix group. The Orbitclosure Determination Problem asks given s N and a family of polynomials f . . . fm Qx of total degree at most b determine whether their zero locus Z GLdQ is the orbit closure of some point v Qd under the action of an sgenerated matrix group. In our complexity analysis we refer to the tuple d m b as the parameters of the problem instances omitting the parameter s. This is without loss of generality as by Proposition Proposition and Remark the minimum number s of topological generators for the groups we study commutative groups is upper bounded by d. This paper focuses on addressing the complexity of the determination problems for commutative groups. The extension of our results to the case of general matrix groups appears to be challenging. To approach the above version of the orbit determination problem we rely on the observation that with respect to a convenient basis an orbit closure itself carries the structure of a matrix group. We then use basic structural results about semisimple and unipotent linear algebraic groups to identify when the above group is the closure of a commutative group. OrbitClosure Determination Problem. We reduce our determination problems to satisfiability problems of a fragment of the firstorder theory of algebraically closed fields of characteristic zero. The following theorem gives a complexity bound on the decision problem for this theory. Theorem CG. Consider a firstorder sentence in the language of rings that mentions m polynomials in d variables with total degree at most b and with k quantifier alternations. The truth of such a sentence in Q can be decided in time mb d k . Following BPR Remark . the truth of firstorder sentences over Q with a fixed number of alternations can be decided in space d log b O by rewriting into the firstorder theory of realclosed fields. Recall that the existential fragment is NPhard and in AM assuming GRH Koi. The following theorem is our main contribution Theorem . The orbitclosure determination problem for commutative matrices with the parameters d m b can be decided in time mb polyd and in space bounded by d log b O . Below we give a brief and informal overview of the main ingredients of the proof of Theorem . Let G be a commutative algebraic group. It is known that the subset of semisimple matrices in G denoted by Gs forms an algebraic subgroup likewise the set of unipotent matrices in G denoted by Gu forms an algebraic subgroup. Let Z be the zero locus of the input polynomials. If Z is the orbitclosure of a point v under the action of G it can be written as Gu Gs v. Given that both Gs and Gu are sgenerated by Remark G will be sgenerated. By Proposition and Lemma it follows from the commutativity of G and the rational parameterisation of Gu that Z Gu Gs v. The commutativity of Gs entails the existence of a matrix P and diagonal invertible matrices Di i s such that P GsP hDi i si. By Lemma we show that P can be chosen such that Pv is a zeroone vector. The algorithm guesses the zeroone vector Pv. The semisimple group hDi i si Pv is a union of toric varieties. Denote by the associated lattice of the defining ideal I of these toric varieties. By Proposition since hDi i si Pv as a linear algebraic group has s topological generators the torsion subgroup of Z d is sgenerated. As a consequence of Proposition the upper bound b on the degree of the defining polynomials of Z carries over to a generating set of I. Considering this degree bound b the algorithm guesses a lattice Z k that is generated by vectors whose entries have absolute value at most b and has at most s elementary divisors not equal to one. By Remark there exist unipotent matrices Ui i s that topologically generate the unipotent subgroup Gu of G. Furthermore the equality Gu expPs i ti log Ui t . . . ts Q holds by Proposition . Now we are ready to write an sentence in the theory of algebraically closed fields. The existential quantifiers encode the possible choices of the matrices P and U . . . Us while the equality of Z and expPs iti log Uih t . . . ts Q h hDi i si Pv is encoded by a sentence with parameters P and U . . . Us. The algorithm returns yes meaning that Z is an orbit closure of a point v under the action of the group G if the above sentence is satisfiable. By Theorem the truth of such a sentence can be decided in time mb polyd . Then the overall complexity bound follows from the fact that the number of choices of the lattice and vector Pv is at most b d . So ends our informal overview of the proof of Theorem the detailed proof can be found in Section . Example . Let Z Q be the zero set of the input ideal I hF Fi where F x x x and F xx x xx. Our nondeterministic procedure in Theorem shows that Z hMi v where M and v . An account of the procedural steps taken to produce M and v is given in Example . OrbitClosure vs. Group Determination. En route to proving Theorem on orbitclosure determination we consider a simpler variantnamely group determination. The two problems bear many similarities especially in the case when the input polynomial ideal I Qx . . . xd defines a variety that is a union of toric varieties. Let us first consider the group determination setting. In this setting the sought group G can be topologically generated by diagonal matrices. By standard results we associated a lattice with the input ideal I. If the torsion subgroup of Z d is sgenerated then by Proposition the minimal number of topological generators of G is s. However in the setting of orbitclosure determination this lower bound on the number of generators may no longer hold. This phenomenon is witnessed by the following example. Example . Let be the lattice associated with the ideal I hx x x x i Qx x x. The torsion subgroup of Z is generated. By Proposition the variety defined by the ideal I is generated as an algebraic subgroup of diagonal matrices in GLQ and in fact is the minimal number of generators. An example of such generators are and . By comparison the variety defined by I arises as the orbit closure of v under the action of hMi where M and v . . Further Applications OrbitClosure Determination and Loop Synthesis. A related area of research to the orbitclosure determination problem is the synthesis of simple linear loops which are branchfree loops characterized by linear update assignments. As explained above while implicit orbitclosure problems are primarily used for invariant generation in program analysis loop synthesis focuses on designing programs that conform to a given polynomial invariant. Recent works HBK Hit Hum KKV have focused on synthesising deterministic linear loops to ensure that a specified set of polynomial equalities holds among program variables at each loop iteration. This guarantees that the synthesised loops adhere to certain polynomial invariants. Geometrically this amounts to finding an infinite orbit of a cyclic matrix group that lies inside a given variety . Humenberger et al. HBK Hum give a method based on constraint solving that synthesises a loop satisfying a given polynomial invariant based on a usersupplied template. Recent work of Hitarth et al. Hit solves a version of the loop synthesis problem in which the polynomial invariant is given by a single quadratic equation. The authors of KKV synthesise simple linear loops whose polynomial invariants are specified by ideals generated by pure difference binomials. The zero set of such an ideal is a union of toric varieties. The synthesis procedure in KKV relies on a construction by Galuppi and Stanojkovski GS Proposition which demonstrates that for every toric variety V one can construct a diagonal rational matrix M such that Mn n Z V . In terms of the results we present the special case of Proposition and Theorem with s i.e. when the sought for group is required to be cyclic takes a variety Z Qd and determines whether Z is equal to the orbitclosure of the initial program variables under the action of a simple linear loop whose update assignments satisfy certain conditions. This is strictly stronger than the synthesis task in above references which asks only that the infinite orbit closure of the synthesised loop be contained in Z. Furthermore by lifting the restriction that the underling group is cyclic we extend the scope from deterministic loops to nondeterministic loops. Group Determination and Matrix Completion. The group determination problems for cyclic groups can also be seen as a type of matrix completion problem. Recall that the latter asks to determine whether a partially given matrix can be completed subject to some polynomial constraints on the entries e.g. lower bounds on the rank IKS. In the determination setting we ask instead to complete the matrix subjected to polynomial constraints on all powers of the matrix. Matrix completion has applications in areas such as combinatorial structures and perfect matching algorithms BFS Edm IKS Lov. Arguably the most interesting application for matrix completion is that of polynomial identity testing IKS arising from the fact that every arithmetic formula can be written as the determinant of a matrix of the linear forms Val. Moreover an arithmetic formula is nonzero if and only if the corresponding matrix can attain full rank. Algebraic Background Let Q denote the field of algebraic numbers and write Qx . . . xd for the ring of polynomials with coefficients in Q over the variables x . . . xd. A polynomial ideal I is an additive subgroup of Qx . . . xd that is closed under multiplication by polynomials in Qx . . . xd. Given a finite collection of polynomials S Qx . . . xd we denote by hSi the ideal generated by S. Notably loop synthesis includes a nontriviality condition asking that the orbit of the synthesised loop be infinite. For otherwise the synthesis problem reduces to that of polynomial equation solving see Hit Remark .. Such an assumption naturally aligns with that of a wandering point of an arithmetic dynamical system Ben Sil. An algebraic set or variety is the set of common zeroes of a finite collection of polynomials. By Hilberts basis theorem every polynomial ideal I Qx . . . xd is finitely generated. Thus the set V I x Q d fx for all f I is a variety. The Zariski topology on Qd has as its closed sets the varieties in Qd . Given a set E Qd we denote by E the closure of E under the Zariski topology. Here E is given by the smallest algebraic set that contains E. Given an ideal I QX where X xijijd and matrix M Qdd we write M I for the ideal fMX QX f I. Clearly V M I A Qdd MA V I. . Linear Algebraic Groups Recall that a matrix M Qdd is called nilpotent if there exists n N such that Mn d M is unipotent if M Idd is nilpotent where Idd is the d d identity matrix and M semisimple if it is diagonalisable over Q. We further say that M is upper triangular if all entries below the main diagonal are zero. We use the term upper unitriangular to refer to an upper triangular matrix whose entries along the main diagonal are all ones. We write GLdQ for the group of d d invertible matrices with entries in Q. We identify GLdQ with the variety M y Qd Q detM y . Under this identification matrix multiplication is a polynomial map GLdQ GLdQ GLdQ and by Cramers rule matrix inversion is also a polynomial map GLdQ GLdQ. A linear algebraic group G is a Zariskiclosed subgroup of GLdQ. Recall that the identity component of G denoted by G is its irreducible component containing the identity. We say that G is topologically generated by S GLdQ if G is the smallest Zariski closed subgroup of GLdQ that contains S. In this case we write G hSi. In case S is finite with s elements we say that G is sgenerated. If G is a commutative algebraic group then the subset of semisimple matrices in G denoted by Gs forms an algebraic subgroup likewise the set of unipotent matrices in G denoted by Gu forms an algebraic subgroup. Proposition . Let G be an algebraic group then G is topologically generated by a finite set. Proof. Below we denote by dim E the dimension of the variety E that is the maximal length of distinct nonempty irreducible subvarieties of E. Let U be the linear algebraic group hGui. It was proven that U can topologically be generated by dim U element see Nos Proof of Lemma . Moreover GU is an algebraic group which consists only of semisimple elements. Recall that the quotient of a linear algebraic group by its normal subgroups are linear algebraic groups of higher dimensions Hum Section .. Therefore GU GU is a torus and by GS Proposition it is generated. Let h G such that GU hhi and let H hh Ui G . We have GU HU which implies dim H dim G . Since G is connected we have G H hence G is topologically generated by at most dim U elements. In order to topologically generate G it is sufficient to take the topological generators of G and one element from any other connected component of G. Hence G is topologically generated by dim U GG elements. The ddimensional multiplicative group over Q is defined as G d m G d mQ n x Q d x xd o . Evidently this is a commutative group with respect to the pointwise multiplication. We identify Gd m with the subgroup of diagonal matrices in GLdQ via the map that sends a . . . ad Qd to the diagonal matrix a . . . ad GLdQ. Given a subgroup Z d define H x G d m v x v x vd d . The map H yields an isomorphism between subgroups of Z d and algebraic subgroups of Gd m. This implies that Gd m is topologically generated by any dtuple g . . . gd of multiplicatively independent elements of Q. It also follows that the vanishing ideal I Qx . . . xd of an algebraic subgroup of Gk m is a socalled pure binomial ideal that is an ideal generated by polynomials of the form x x d d x x d d where . . . d and . . . d are nonnegative integers. For future reference we note a mere binomial ideal is one that is generated by polynomials of the form x x d d x x d d where Q. The following proposition BG Proposition .. shows how to recover the generators of a binomial ideal from the defining equations of an algebraic subgroup of Gd m. Proposition . Let G be a subgroup of Gd m defined by polynomial equations fix P aix for i . . . m. Let Li Z d be the set of exponents of the monomials appearing in fi . Then G H where Z d is generated by where Li for some i . . . m. The following is an amalgamation of standard results in Diophantine geometry concerning the number of generators of a subgroup of Gd m cf. BG Chapter we include a sketch proof for the readers convenience. Proposition . Let have rank r and elementary divisors d . . . dr where di di for i . . . r . Then the following are equivalent for all s . . . r . H is sgenerated . d drs . the torsion subgroup of Z d is sgenerated. Proof. There exists a basis u . . . ud of Z d such that is generated by the vectors du . . . drur. The map Gd m Gd m defined by x x u . . . x ud is a Zariskicontinuous group automorphism of Gd m that maps H to the group G d dr G dr m where k denotes the group of all kth roots of unity for k a positive integer. Clearly H is sgenerated if and only if G is sgenerated. Write F d dr . We note that G F Gdr m is sgenerated if and only if F is sgenerated. In particular if S F is a generator of F then S g is a topological generator of G for any topological generator g of Gdr m . But from the fact that di di for i . . . r we see that F d dr is sgenerated if and only if d drs showing the equivalence of Items and . For the equivalence of Items and we note that the torsion subgroup of Z d is ZdZ ZdrZ which is isomorphic to F. For unipotent A GLdQ and nilpotent B GLdQ define logA X d k k A I k k and expB X d k Bk k! . Let G GLdQ be a commutative subgroup of unipotent matrices. Recall that L logA A G is a linear subspace of Qd consisting of nilpotent matrices Bor Chapter II Section .. Moreover exp L G and log G L yield polynomial isomophisms between L and G as algebraic groups. Taken together these observations lead to the following proposition. Proposition . Let G be a commutative subgroup of unipotent matrices and L the associated linear subspace of nilpotent matrices as above. Then G has a topological generator of cardinality s if and only if L is spanned by a set of s matrices as a Qvector space. Proof. For all A . . . As G we have the following equivalences A . . . As is a topological generator of G A n A n A ns s n . . . ns Z is dense in G Xs i ni logAi n . . . ns Z is dense in L Xs i ti logAi t . . . ts Q L as desired. Commutative Group Determination Recall that the group determination problem asks given s N and a family of polynomials f . . . fm Qxijijd of total degree at most b to determine whether their zero locus Z GLdQ is an sgenerated matrix group. In this section we first demonstrate a procedure for this problem subject to the constraint that the underlying group is semisimple commutative Proposition . Next we generalise this result by lifting the requirement that the matrices are semisimple Proposition . Proposition . The group determination problem for commutative semisimple matrices with the parameters d m b can be decided in time mb polyd and in space bounded by d log b O . Proof. Given semisimple commutative matrices M . . . Ms GLdQ there exists P GLdQ such that Di P MiP are diagonal matrices. Let G be the subgroup of Gd m defined by G g G d m g hDi i si . Then Z hMi i si if and only if P ZP g g G. Now P ZP is the zero set of polynomials of degree at most b and hence by Proposition Z hMi i si if and only if the group G has the form H for some lattice Z d that is generated by vectors having supremum norm at most b. The decision procedure is thus as follows . Guess a lattice Z d whose generators have norm at most b and has at most s elementary divisors not equal to one . Determine whether there exists P GLdQ such that P ZP g g H. Step amounts to checking the truth in Q of the sentence P GLdQ A GLdQ P AP Z ij aij a . . . add H with respect to the theory of algebraically closed fields. By Theorem this can be done in time mb polyd . The claimed running time for the overall procedure follows from the fact that the number of possibilities for the lattice is at most b d . By the following remark a commutative algebraic group G is sgenerated if its unipotent subgroup Gu and semisimple subgroup Gs are both sgenerated. Remark . Let U hUi i si be a commutative unipotent sgenerated algebraic group and S hSi i si be a semisimple commutative sgenerated algebraic group. Suppose moreover that the matrices Ui Si i s are pairwise commutative. Then we have hUi Si i si hSiUi i si. This relies on the fact that if A GLdQ is semisimple and B GLdQ is unipotent then both A and B lie in the Zariski closure of the subgroup generated by their product AB see Hum Section .. The next proposition generalises the procedure witnessed in Proposition . In Proposition we consider the group determination problem for sgenerated commutative algebraic groups. Key to our generalisation is the determination of a matrix P GLdQ and properties associated with the semisimple and unipotent subgroups of the group P ZP. Proposition . The group determination problem for commutative matrices with the parameters d m b can be decided in time mb polyd and in space bounded by d log b O . Proof. Given commutative matrices M . . . Ms GLdQ let P GLdQ be such that there exist diagonal matrices Di and upper unitriangular matrices Ui i s where P MiP DiUi and moreover Di and Ui commute. Then we can recover hDi i si as the set of diagonal matrices in P hMi i si P. It follows that hDi i si is the zero locus of a system of polynomials of degree at most b. By Proposition hDi i si g g H for some lattice Z d that is generated by vectors of norm at most b. Note that we can also recover hUi i si as the set of upper unitriangular matrices in P hMi i si P. The decision procedure is as follows. Guess a lattice Z d whose generating vectors have entries of absolute value at most b and has at most s elementary divisors not equal to one. Next determine whether there exists P GLdQ such that . G P ZP is a commutative group of upper triangular matrices . A G A diagonal g g H . logA A G A unipotent is a linear variety of dimension at most s. Item checks that G is a commutative matrix group. In this case both the set Gs of semisimple matrices in G and the set Gu of unipotent matrices in G form subgroups of G. Next Items and respectively check that Gs and Gu are sgenerated relying on Propositions and . But this implies that G itself is sgenerated as noted in Remark . The existence of P satisfying Items to reduces to checking the truth in Q of an sentence in the theory of algebraically closed fields. The existential quantifiers correspond to the possible choices of P while the universal quantifiers range over entries of the group G defined in Item . For a fixed choice of the truth of such a formula can be decided in time mb polyd by Theorem . Given that the number of possible choices of the lattice is at most b d the claimed complexity bound immediately follows. OrbitClosure Determination Recall the aforementioned orbitclosure determination problem. The problem asks given s N and a family of polynomials f . . . fm Qx of total degree at most b to determine whether their zero locus Z GLdQ is the orbit closure of some point v Qd under the action of an sgenerated matrix group i.e. determine whether there exist matrices M . . . Ms GLdQ and vector v Qd such that Z is the Zariski closure of the orbit M Ms s v . . . s Z. The main contributions of this section are the procedures for certain cases of the orbitclosure determination problem Proposition and Theorem . The procedure in Proposition makes the additional assumptions that the generators of the matrix group are semisimple and pairwise commutative. The procedure in Theorem lifts the requirement that the generators are semisimple. We illustrate the procedures with a worked example Example at the close of this section. In the work that follows it is convenient to employ the next lemma which intuitively speaking describes that orbit closures under the action of semisimple commutative groups are isomorphic to the orbit closure of . Lemma . Let G hMi i si Q dd be a semisimple commutative algebraic group and v Q d . There exist P GLdQ diagonal matrices D . . . Ds GLdQ and T dk with the following properties . Mi P DiP . T TT Idk . Pv T. Proof. Since the matrices Mi are commutative there exist a matrix Q GLdQ and diagonal matrices D . . . Ds GLdQ such that Mi QDiQ. Let R GLdQ be a diagonal matrix such that RQv d . We define P RQ. Since R is diagonal it commutes with all Di matrices. Thus we have Mi QDiQ P DiP. It remains to determine the matrix T. We write Pv ei eik as a sum of standard unit vectors of Qd . We define T to be the d k matrix with columns ei . . . eik . It follows that T TT Idk and Pv T. We move onto the first of the two main orbitclosure results in this section. Proposition . The orbitclosure determination problem for commutative semisimple matrices with the parameters d m b can be decided in time mb polyd and in space bounded by d log b O . Proof. Suppose that Z hMi i si v for semisimple commutative matrices M . . . Ms GLdQ and a vector v Qd . By Lemma there exist matrices P GLdQ and T dk such that Di PMiP are diagonal and moreover T TT Idk and Pv T hold. For all i s denote by D i GLkQ the diagonal matrix uniquely defined by the requirement that DiT T D i . Write G g G k m g hD i i si . Then P Z hDi i si T ThD i i si T g g G where the penultimate equality relies on the fact that image of a Zariskiclosed set under an injective linear map is again Zariski closed. Note that G can alternatively be written as g Gk m Ug P Z and is thereby defined by polynomials of total degree at most b. It follows from Proposition that G H for some lattice Z k that is generated by vectors whose entries have absolute value at most b. Conversely suppose that P Z T g g H for some matrices P GLdQ and T dk such that U TU Idk and lattice as above. Then Z hMi i si v where v P T and Mi P DiP with Di GLdQ being any diagonal matrices such that DiT T gi for some topological generators g . . . gn of H. In summary the decision procedure is as follows . Guess k . . . d and T dk such that T TT Idk . Guess a lattice Z k that is generated by vectors whose entries have absolute value at most b and has at most s elementary divisor not equal to one . Determine whether there exists P GLdQ such that P Z T g g H. Step can be reduced in polynomial time to checking the truth of a sentence in the theory of algebraically closed fields the existential quantifiers correspond to the matrix P while the universal quantifiers are used to encode the equation P Z T g g H. By Theorem the truth of such a sentence can be decided in time mb polyd . The claimed overall complexity bound now follows from the fact that there are at most b d choices of the lattice and matrix U. We include a worked example that demonstrates the constructive subroutines in Proposition . Example . Let Z Q be the zero set of the ideal I hx y xyxyi. In this example we construct a matrix M GLQ and vector v Q such that Z is precisely the Zariski closure of the orbit Mnv n Z. Suppose that in Steps and of the procedure in Proposition we guess H x y G m x y and the matrix T Id. For Step we want to find all invertible matrices P a b c d GLQ such that V P I T g g H. This is equivalent to the requirement that the two equations ax by cx dy ax bycx dy ax by cx dy and x y are multiples of one another. Therefore the equations defining P comprise the following ideal JP ha c ac b d b d bd ab cd ad bc a ci ha c c b d bc cdb d i. One choice of P is . Thus we associate with Z the Zariski closure of the orbit Mnv n Z where M P P and v P . The following lemma is crucial in generalizing the procedure in Proposition for orbitclosure determination of semisimple commutative groups to the general setting. Lemma . Let G GLdQ be a commutative algebraic group and v Qd . Then G v Gu Gs v where Gu and Gs are the subgroups of unipotent and semisimple elements in G respectively. Proof. We assume that G is sgenerated. Write G hMi i si and Gu hUi i si. Define V Gs v. For every w V we define the polynomial map w Q n Gu V t . . . ts expXs i ti log Ui w . Since w is a continuous map w V is Zariski closed. Let H T wV w V . We note that H is Zariski closed and is equal to t . . . ts Q n expXs i ti log Ui V V . Observe that for every pair t . . . tst . . . t s H and a b N we have at . . . ts bt . . . t s H. But since H is Zariski closed this property implies that H is a linear vector space. The following claim is crucial for the continuation of the proof. Claim . For all t . . . ts Qd such that expPs i ti log Ui V V we have that expXs i ti log Ui V V. Proof of the claim. Suppose that expPs i ti log UiV V . Then there exists w Gs v such that expPs i ti log Ui w V . By commutativity of G we get Gs expXs i ti log Ui w expXs i ti log Ui Gs w Gs V V. Thus expPs i ti log Ui V V holds. It implies expPs i ti log Ui V V concluding the proof of the claim. Denote by W the orthogonal space of H meaning that Qd W H. Let us introduce the following map W V Gu V t . . . ts w expXs i ti log Ui w . The map is a polynomial map and onetoone correspondence. To prove this assume that expXs i ti log Uiw expXs i t i log Uiw holds for some t . . . tst . . . t s W and w w V . Then expXs i ti t i log Ui V V . By the above and Claim we have expPs itit i log UiV V . This implies that t . . . ts t . . . t s which in turns shows that w w . Since is a onetoone correspondence between W V and Gu V and since W V is Zariski closed we have that W V Gu V is Zariski closed. This completes the proof. The following theorem is our main contribution which provides a decision procedure for the orbitclosure determination problem for commutative groups. The generalization of this result to the case of general matrix groups appears to be challenging. Theorem . The orbitclosure determination problem for commutative matrices with the parameters d m b can be decided in time mb polyd and in space bounded by d log b O . Proof. Suppose that Z G v for some commutative algebraic group G hMi i si GLdQ and v Qd . Let Gu be the group of unipotent elements of G and Gs be the group of semisimple elements of G. Then by Lemma we have Z Gu Gs v. We apply Lemma to the semisimple group Gs. Let P GLdQ and T dk be matrices such that P GsP hDi i si where Di are diagonal invertible matrices and the conditions T TT Idk and Pv T are satisfied. For all i s denote by D i GLkQ be the diagonal matrix uniquely defined by the requirement that DiT T D i . Furthermore write G g G k m g hD i i si and Gu hUi i si. Then we have Z Gu Gs v hUi i si P hDi i si Pv hUi i si P T hD i i si expPs iti log Ui P T g g G t . . . ts Q where the first equality follows from Lemma and the last equality follows Proposition . Note that G g Gk m P T g Z. It follows that G is defined by polynomials of total degree at most b and hence has the form H for some lattice Z k that is generated by vectors whose entries have absolute value at most b and has at most s elementary divisors not equal to one. Conversely if Z expPs iti log UiP T g g H t . . . ts Q for some matrix P GLdQ unipotent matrices U . . . Us GLdQ that are commutative matrix T dk and lattice as above then Z hMi i si v where Mi P UiDiP and Di i s GLdQ are any diagonal matrices such that DiT T gi for some topological generators g . . . gs of H. In summary the decision procedure is as follows . Guess k . . . d and T dk such that T TT Idk . Guess a lattice Z k that is generated by vectors whose entries have absolute value at most b and has at most s elementary divisor not equal to one . Return yes if there exists P GLdQ and unipotent commutative matrices U . . . Us GLdQ such that Z expPs i ti log Ui P T g g H t . . . ts Q. Step can be reduced to checking the truth of an sentence that is with a quantifier prefix comprising a block of existential quantifiers a block of universal quantifiers followed by a single existential quantifier with respect to the theory of algebraically closed fields. The existential quantifiers encode the possible choices of the matrices P and U . . . Us while the equality of Z and expPs i ti log Ui P T g g H t . . . ts Q is encoded by a sentence with parameters P and U . . . Us namely z Q d z Z t . . . ts s.t. T TP exp Ps iti log Ui z H . By Theorem the truth of such a sentence can be decided in time mb polyd . Then the overall complexity bound follows from the fact that the number of choices of the lattice and matrix T is at most b d . Example below applies the procedure in Theorem to the variety we first saw in Example in the Introduction. The calculations involved in the preparation of Examples and were performed in Macaulay GS. Example . Let Z Q be the zero set of the ideal I hF Fi where F x x x and F xx x xx. Below we shall construct a matrix M GLdQ and vector v Qd such that Z is the Zariski closure of the orbit Mnv n Z. Suppose that in Steps and of the procedure in Theorem we guess that H x y G m x y T and U has the form . For Step we would like to find the set of invertible matrices P pij ij GLQ such that there exists for which V P I exptlog UT g g H t Q. Note that exptlog U t tt t thus exptlog UT x x tt x tx x x . The ideal defining exptlog UT g g H t Q is H hx x xx x xxi. Consider the ideal IP hFP X FP Xi QP pij yhdet Py ix x x x. By applying Algorithm ContainmentIso and eliminating we obtain the following ideal defining the set of admissible choices of P JP hp p p p p p p p p p p pp pp pp pp p pp p pp p pp p p pi. One may choose for example P and . Thus we associate with Z the Zariski closure of the orbit Mnv n Z where M P UDP and v P T . Algorithms to Compute Generators In Section we gave an algorithm to determine whether a given variety is the Zariski closure of a commutative matrix group. The method there can also be used to find a set of generators of such a group using the fact that the theory of algebraically closed fields admits quantifier elimination. In this section we provide two alternative algorithms for cyclic groups to compute a generator relying instead on Grbnerbasis techniques. The first algorithm finds a semisimple generator if one exists while the second algorithm finds a generator in the general case. Let I Qx . . . xd be an ideal and denote by I the radical of I defined as I f Qx . . . xd f n I for some n N. By Hilberts Nullstellensatz the ideal of all polynomials that vanish on V I Qd is I. The ideal I is primary if for all f g Qx . . . xd if fg I then f I or g n I for some n N. Recall that the radical of a primary ideal is necessarily prime. A polynomial ideal I can be written as the intersection of primary ideals giving the socalled primary decomposition of I. It is known that there exists a unique irredundant primary decomposition I T i Qi that is a finite set Q . . . Q of primary ideals such that the prime ideals Qi are all distinct and T ij Qi Qj holds for all j . The prime ideals the Qi s are called the associated primes of I. An associated prime Q of the ideal I is called minimal if it does not contain any other associated primes of I. Both algorithms take as input a variety Z given as the zero set of an ideal I QX X xijijd. We will assume that it has already been verified that Z is a commutative subgroup of GLdQ. This verification entails first checking that Z is closed under matrix multiplication which entails closure under matrix inverse which amounts to showing that FXY p IX IY The algorithm ContainmentIso inputs two ideals I and I and outputs the locus of points P for which P I P I. This algorithm thereby solves a generalisation of the ideal membership algorithm since it determines the containment of an ideal into another after a change of variables. Clearly ContainmentIso can also be applied to determine equality after a change of variables since P I P I if and only if P I P I and P I P I. See KMM Algorithm . for more details. Cyclic Groups Semisimple Generator Input An ideal I QX X xijijd with q minimal associated primes such that V I is a commutative linear algebraic group. Output Determine whether there exists a semisimple matrix M GLdQ such that V I hMi. If yes output such a matrix M. Line Define the ideal J QP X Q as follows J h FP XQ . . . FkP XQ P Q Idd xijij i where P pijijd X xijijd and Q qijijd. Line Write J p J QX. Compute the primary decomposition J T sS Ps. Line Check whether all primary components Ps of J are binomials using Grbner basis computation return no if this test fails. Line Let P be one of the primary component of J such that Idd V P. Line Following Proposition we can construct a rational diagonal matrix D for P such that all the entries of q D lie in Qq. Write Dq q D. Line Check whether for all i . . . q the ideal Di q P is a primary component of J return no if this test fails. Line Write Iq T iq Di q P. Line Check whether J T Sd MIqM where M is the permutation matrix corresponding to Sd return no if this test fails. Line Define the ideal J hFQDqP . . . FkQDqP P Q Ii QP. Pick P V J. Line Check whether I P I qP return no if this test fails. Return yes together with the matrix P D qP . Figure A procedure for the group determination problem of cyclc groups specific to semisimple generators. for all polynomials F in I where X xijijd Y yijijd. Commutativity is captured by showing that XY Y X p IX IY . . Semisimple Generator In the following we describe a procedure that given an ideal I QX X xijijd determines whether there exists a semisimple matrix M GLdQ such that I is the vanishing ideal of the group hMi and which moreover outputs such an M in case the answer is yes. We show that if such an M exists then it can be chosen such that its eigenvalues lie in the number field Qq where q is a primitive qth root of unity and q is the number of minimal associated primes of I. Let the input ideal I be generated by a finite collection of polynomials F . . . Fk QX with q minimal associated primes. Write Z V I for the zero locus of I assumed to be a commutative linear algebraic group. The general procedure of the algorithm is depicted in Figure . The ideal J defined in Line is an ideal of the ring QP X Q where the relations xijij and P Q Idd ensure that every point P X Q V J comprises a diagonal matrix X and an invertible matrix P with P Q satisfying PXP Z. The aim is to find a single such point P X Q V J satisfying Z hPXPi. Subsequently the radical ideal V J defined in Line contains all diagonal conjugates of each matrix in Z. In particular for each matrix M Z not only one single diagonal matrix D satisfying M P D P lies in V J but all diagonal matrices of the form MDM with the permutation Sd also lie in V J. Due to this fact we cannot simply employ Proposition to construct a generator for J. Instead in Line we isolate a primary component P of J containing Idd. In the following line we apply Proposition to the binomial ideal P and construct a diagonal matrix D such that V P hDi. Since V P is connected the matrix D can be chosen rational and such that the entries of q D lie in Qq. The assertion in Line verifies whether the orbit of Dq rotates between the primary components of J this ensures that V Iq hDqi is included in V J where Iq is defined in Line . Next our procedure checks whether J equals to the intersection of MIqM . The necessity of the latter test is due to the abovementioned fact that V J contains all diagonal conjugates of each matrix in Z see Example . The rest of the algorithm is straightforward. Example . Let F z w F xy w and F y yww y w. Consider the following ideal as an input to the procedure in Figure I hF F Fi Q x z w y . The ideal I is prime meaning that q and V I is a commutative linear algebraic group. The output of our procedure shows that there exists M such that V I hMi and such that the eigenvalues of M lie in Q. Following the algorithms the ideal J defined in Line has two primary components P hw z y xi and P hw z x yi. Since I V P P we can pick any of these ideals as P in Line . Following Proposition in Line we may construct diagonal matrices D and D such that V P h i and V P h i Clearly matrices D and D are conjugates under permutation of diagonals implying that the assertion in Line holds. The above is an indication that permutation of matrices arising from one choice of P under M are suitable for other possible choices of P and the necessity of the check in Line . Following Line for Dq defining the ideal J hFQDqP FQDqP FQDqPi QP we have that J hp p p pi Q p p p p . Subsequently one choice for a semisimple generator of V I is the following matrix M M . . General Generator We employ the algorithm from the previous subsection to provide a procedure that given an algebraic set Z GLdQ determines whether there exists a matrix M GLdQ such that Z hMi and which moreover outputs such an M in the affirmative case. Let the input ideal I be generated by a finite collection of polynomials F . . . Fk QX with q minimal associated primes. Write Z V I for the zero locus of I that is by assumption a commutative linear algebraic group. Our algorithm first calls a modified variant of the procedure in Figure with the input ideal I to check whether the subgroup Gs of all semisimple matrices in Z is onegenerated. The modification is as follows the assertion in Line is omitted as this assertion requires that Z is generated with a single semisimple matrix and the algorithm outputs Dq and the ideal J defining the locus point of suitable P. Analogous to the preceding settingwhere the subgroup of semisimple matrices was onegenerated our algorithm proceeds by verifying that the subgroup Gu of all unipotent matrices is onegenerated. For this purpose it checks whether V I hX Idd n i is a commutative linear algebraic group and whether V I hX Idd n i is one dimensional. The algorithm returns no if either of the subgroups Gs or Gu is not onegenerated. Otherwise the procedure defines the ideal H QP X Q by H h FP XQ . . . FkP XQ P Q Idd xijjii i where P pij ijd X xijijd and Q qij ijd. It returns yes together with the matrix P D qXP where P X J H. References Bab L. Babai R. Beals J.Y. Cai G. Ivanyos and E. M. Luks. Multiplicative Equations over Commuting Matrices. In Proceedings of the Seventh Annual ACMSIAM Symposium on Discrete Algorithms January Atlanta Georgia. pp. . Bab L. Babai. Trading Group Theory for Randomness. In Proceedings of the th Annual ACM Symposium on Theory of Computing. ACM pp. . doi ... Ben R. Benedetto P. Ingram R. Jones M. Manes J. H. Silverman and T. J. Tucker. Current trends and open problems in arithmetic dynamics. In Bull. Amer. Math. Soc. N.S. . pp. . doi .bull. Bey D. Beyer T. A. Henzinger R. Majumdar and A. Rybalchenko. Invariant synthesis for combined theories. In International Workshop on Verification Model Checking and Abstract Interpretation. Springer. pp. . BFS J. F. Buss G. S. Frandsen and J. O. Shallit. The Computational Complexity of Some Problems of Linear Algebra. In Journal of Computer and System Sciences . June pp. . doi .jcss... BG E. Bombieri and W. Gubler. Heights in Diophantine Geometry. New Mathematical Monographs. Cambridge University Press . Bl M. Blser C. Ikenmeyer V. Lysikov A. Pandey and F.O. Schreyer. On the orbit closure containment problem and slice rank of tensors. In Proceedings of the ACMSIAM Symposium on Discrete Algorithms SODA. SIAM. pp. . Bor A. Borel. Linear Algebraic Groups. Springer New York . doi .. BPR S. Basu R. Pollack and M.F. Roy. Algorithms in real algebraic geometry. Second. Vol. . Algorithms and Computation in Mathematics. SpringerVerlag Berlin pp. x. Br P. Brgisser J. M. Landsberg L. Manivel and J. Weyman. An Overview of Mathematical Issues Arising in the Geometric Complexity Theory Approach to VPVNP. In SIAM J. Comput. . pp. . doi .. Br P. Brgisser M. L. Doan V. Makam M. Walter and A. Wigderson. Polynomial Time Algorithms in Invariant Theory for Torus Actions. In th Computational Complexity Conference CCC . Vol. . Leibniz International Proceedings in Informatics LIPIcs. Dagstuhl Germany Schloss Dagstuhl LeibnizZentrum fr Informatik . doi .LIPIcs.CCC... Br P. Brgisser. Completeness classes in algebraic complexity theory. In arXiv preprint arXiv. . CG A. L. Chistov and D. Y. GrigorEv. Complexity of quantifier elimination in the theory of algebraically closed fields. In International Symposium on Mathematical Foundations of Computer Science. Springer. pp. . CK J. Cyphert and Z. Kincaid. Solvable Polynomial Ideals The Ideal Reflection for Program Analysis. In Proc. ACM Program. Lang. .POPL pp. . doi .. DJK H. Derksen E. Jeandel and P. Koiran. Quantum automata and algebraic groups. In J. Symb. Comput. . pp. . DM H. Derksen and V. Makam. Algorithms for orbit closure separation for invariants and semiinvariants of matrices. In Algebra Number Theory . Nov. pp. . doi .ant.... Edm J. Edmonds. Systems of distinct representatives and linear algebra. In Journal of Research of the National Bureau of Standards Section B Mathematics and Mathematical Physics B. Oct. p. . doi .jres.b.. FS M. A. Forbes and A. Shpilka. Explicit Noether Normalization for Simultaneous Conjugation via Polynomial Identity Testing. In Approximation Randomization and Combinatorial Optimization. Algorithms and Techniques. Berlin Heidelberg Springer Berlin Heidelberg pp. . Gra W. A. de Graaf. Computation with Linear Algebraic Groups. Chapman HallCRC Monographs. CRC Press . GS D. R. Grayson and M. E. Stillman. Macaulay a software system for research in algebraic geometry. Available at httpwww.macaulay.com. GS F. Galuppi and M. Stanojkovski. Toric varieties from cyclic matrix semigroups. In Rend. Istit. Mat. Univ. Trieste . doi .. HBK A. Humenberger N. S. Bjrner and L. Kovcs. AlgebraBased Loop Synthesis. In Integrated Formal Methods th International Conference IFM Lugano Switzerland November Proceedings. Vol. . Lecture Notes in Computer Science. Springer pp. . doi .. Hit S. Hitarth G. Kenison L. Kovcs and A. Varonka. Linear Loop Synthesis for Quadratic Invariants. In st International Symposium on Theoretical Aspects of Computer Science STACS . Vol. . Leibniz International Proceedings in Informatics LIPIcs. Dagstuhl Germany Schloss Dagstuhl LeibnizZentrum fr Informatik . doi .LIPIcs.STACS... Hru E. Hrushovski J. Ouaknine A. Pouly and J. Worrell. Polynomial Invariants for Affine Programs. In Proceedings of the rd Annual ACMIEEE Symposium on Logic in Computer Science. LICS . Oxford United Kingdom Association for Computing Machinery pp. . doi ... Hru E. Hrushovski J. Ouaknine A. Pouly and J. Worrell. On strongest algebraic program invariants. In Journal of the ACM . pp. . Hum A. Humenberger D. Amrollahi N. Bjrner and L. Kovcs. AlgebraBased Reasoning for Loop Synthesis. In Form. Asp. Comput. . July . doi .. Hum J. E. Humphreys. Linear Algebraic Groups. Vol. . Graduate Texts in Mathematics. Springer . doi .. IKS G. Ivanyos M. Karpinski and N. Saxena. Deterministic Polynomial Time Algorithms for Matrix Completion Problems. In SIAM Journal on Computing . Jan. pp. . doi .. Kin Z. Kincaid J. Cyphert J. Breck and T. W. Reps. Nonlinear reasoning for invariant synthesis. In Proc. ACM Program. Lang. .POPL . doi .. KKV G. Kenison L. Kovcs and A. Varonka. From Polynomial Invariants to Linear Loops. In Proceedings of the International Symposium on Symbolic and Algebraic Computation ISSAC Troms Norway July . ACM pp. . doi ... KL R. Kannan and R. J. Lipton. The Orbit Problem is Decidable. In Proceedings of the th Annual ACM Symposium on Theory of Computing April Los Angeles California USA. ACM pp. . doi ... KL R. Kannan and R. J. Lipton. Polynomialtime algorithm for the orbit problem. In J. ACM . pp. . doi ... KMM L. Katthn M. Michaek and E. Miller. When is a Polynomial Ideal Binomial After an Ambient Automorphism? In Foundations of Computational Mathematics . doi .s. Koi P. Koiran. Hilberts Nullstellensatz Is in the Polynomial Hierarchy. In J. Complex. . . long version DIMACS report pp. . doi .jcom... Lov L. Lovsz. On determinants matchings and random algorithms. In Fundamentals of Computation Theory FCT Proceedings of the Conference on Algebraic Arthmetic and Categorial Methods in Computation Theory BerlinWendischRietz Germany September . AkademieVerlag Berlin pp. . MSa M. MllerOlm and H. Seidl. A note on Karrs algorithm. In International Colloquium on Automata Languages and Programming. Springer. pp. . MSb M. MllerOlm and H. Seidl. Computing polynomial program invariants. In Inf. Process. Lett. . pp. . doi .J.IPL.... Nos K. Nosan A. Pouly S. Schmitz M. Shirmohammadi and J. Worrell. On the Computation of the Zariski Closure of Finitely Generated Groups of Matrices. In Proceedings of the International Symposium on Symbolic and Algebraic Computation. pp. . Sil J. H. Silverman. The arithmetic of dynamical systems. Vol. . Graduate Texts in Mathematics. Springer New York pp. x. doi .. SMK H. Seidl S. Maneth and G. Kemper. Equivalence of Deterministic TopDown TreetoString Transducers is Decidable. In IEEE th Annual Symposium on Foundations of Computer Science FOCS Berkeley CA USA October . IEEE Computer Society pp. . doi .FOCS... SSM S. Sankaranarayanan H. Sipma and Z. Manna. Nonlinear loop invariant generation using Grbner bases. In Proceedings of the st ACM SIGPLANSIGACT Symposium on Principles of Programming Languages POPL Venice Italy January . ACM pp. . doi ... Val L. G. Valiant. Completeness classes in algebra. In Proceedings of the Eleventh Annual ACM Symposium on Theory of Computing. STOC . Atlanta Georgia USA Association for Computing Machinery pp. . doi ...

write a youtube script

Topological Generalization Bounds for DiscreteTime Stochastic Optimization Algorithms Rayna Andreeva r.andreevasms.ed.ac.uk School of Informatics University of Edinburgh UK Benjamin Dupuis benjamin.dupuisinria.fr INRIA Departement dInformatique de lEcole Normale Superieure PSL Research Univ France Rik Sarkar rsarkarinf.ed.ac.uk School of Informatics University of Edinburgh UK Tolga Birdal tbirdalimperial.ac.uk Deparment of Computing Imperial College London UK Umut Simsekli umut.simsekliinria.fr INRIA Departement dInformatique de lEcole Normale Superieure PSL Research Univ France Abstract We present a novel set of rigorous and computationally efficient topologybased complexity notions that exhibit a strong correlation with the generalization gap in modern deep neural networks DNNs. DNNs show remarkable generalization properties yet the source of these capabilities remains elusive defying the established statistical learning theory. Recent studies have revealed that properties of training trajectories can be indicative of generalization. Building on this insight stateoftheart methods have leveraged the topology of these trajectories particularly their fractal dimension to quantify generalization. Most existing works compute this quantity by assuming continuous or infinitetime training dynamics complicating the development of practical estimators capable of accurately predicting generalization without access to test data. In this paper we respect the discretetime nature of training trajectories and investigate the underlying topological quantities that can be amenable to topological data analysis tools. This leads to a new family of reliable topological complexity measures that provably bound the generalization error eliminating the need for restrictive geometric assumptions. These measures are computationally friendly enabling us to propose simple yet effective algorithms for computing generalization indices. Moreover our flexible framework can be extended to different domains tasks and architectures. Our experimental results demonstrate that our new complexity measures correlate highly with generalization error in industrystandards architectures such as transformers and deep graph networks. Our approach consistently outperforms existing topological bounds across a wide range of datasets models and optimizers highlighting the practical relevance and effectiveness of our complexity measures. Authors contributed equally. Authors contributed equally. arXiv.v cs.LG Jul Introduction Generalization a hallmark of model efficacy is one of the most fundamental attributes for certifying any machine learning model. Modern deep neural networks DNN display remarkable generalization abilities that defy the current wisdom of machine learning ML theory . The notion can be formalized through the risk minimization problem which consists of minimizing the function Rw Ezz w z where z Z X Y denotes the data distributed according to a probability distribution z on the data space Z. In practice as z is unknown ML algorithms focus on minimizing the empirical risk RbSw n Xn i w zi where S z . . . zn n z are independent samples from z. In many applications the minimization of is achieved by discrete stochastic optimization algorithms such as stochastic gradient descent SGD or the ADAM method. Such algorithms generate a sequence of iterates in Rd denoted WS wkk which depends on the data S the initialization w Rd and some additional randomness U e.g. the random batch indices in SGD. The generalization error characterizing the models performance is then defined as GSwk Rwk RbSwk. The empirical risk typically has numerous local minima which raises the question of how to characterize their generalization properties. Recently training trajectories cf. Figure a have been shown to be paramount to answer this question . Indeed these trajectories can quantify the quality of a local minimum in a compact way because they depend simultaneously on the algorithm the hyperparameters and the data which is crucial for obtaining satisfactory bounds . A wide family of trajectorydependent bounds has been developed . For instance several results on stochastic gradient Langevin dynamics continuous Langevin dynamics and SGD take into account the impact of the whole trajectory on the generalization error. Parallel to these developments several studies have brought to light the empirical links between topological properties of DNNs and their generalization performance hereby making new connections with topological data analysis TDA tools . These studies focus on the structural changes across the different layers of the network or on the final trained network and are almost exclusively empirical. This partially inspired a new class of trajectorydependent bounds focusing on topological properties of the trajectories. In particular recent studies have proposed to relate the generalization error to various kinds of intrinsic fractal dimensions that characterize the learning trajectory. Informally these bounds provide the guarantee that with probability at least we have sup wWS GSw r dimWS IT log n where dimWS denotes various equivalent fractal dimensions in particular the persistent homology dimension PHdim and the magnitude dimension . The term IT is an informationtheoretic quantity that takes different forms among different studies. Despite providing rigorous links between We use in informal statements to indicate that absolute constants andor small terms are missing. a MDS Trajectory Embeddings b Distance Matrices of Trajectories c Our Topological Generalization Measures . . . . . . dim . . . . . . . . . dim  S  S ViT  CIFAR ViT CIFAR CaiT CIFAR CaiT CIFAR Swin CIFAR Swin CIFAR GatedGCN MNIST GraphSage MNIST . . . . . . Average granulated Kendall dimPH E ours PMagn ours Figure We devise a novel class of complexity measures that capture the topological properties of discrete training trajectories. These generalization bounds correlate highly with the test performance for a variety of deep networks data domains and datasets. Figure shows different trajectories a embedded using multidimensional scaling based on the distancematrices b computed using either the Euclidean distance between weights as in or via the lossinduced pseudometric S as in . c plots the average granulated Kendall coefficients for two of our generalization measures E and PMag n in comparison to the stateoftheart persistent homology dimensions for a range of models datasets and domains revealing significant gains and practical relevance. the topology of the trajectory and generalization these bounds have major drawbacks. First and foremost as noted in fractaltrajectory bounds such as Equation do not apply to discretetime algorithms. This creates a discrepancy between these theoretical results and the TDAinspired methods to numerically evaluate them on commonly used discrete algorithms . Additionally existing bounds rely on very intricate geometric assumptions such as Ahlforsregularity or geometric stability that are not realistic in a practical discrete setting. Previous attempts were made to address this discretization issue. Specifically under the assumption that the training dynamics possess a stationary measure wS for T T is the number of iterations it was shown in that with probability over S n z and w wS GSw r dimwS IT log n where dimwS corresponds to the fractal dimension of the measure w see for formal definitions. While this was an important step this bound only becomes practically relevant when the number of iterations grows to infinity which is never attained in reallife experiments. Other attempts make use of socalled finite fractal dimensions or fine properties of the Markov transition kernels associated with the dynamics . However these studies also rely on impractical assumptions and involve intricate quantities which make them not amenable to numerical evaluation. Despite the theoretical limitations of existing topologydependent generalization bounds TDAinspired tools have been developed to numerically estimate the proposed intrinsic dimensions in practical settings. Two particular methods have emerged and successfully demonstrate correlation with the generalization error based on persistent homology PHdim and metric space magnitude magnitude dimension these two dimensions are equivalent for compact metric spaces . Because of the limitations discussed above existing theories do not account for these experiments conducted with finitetime discrete algorithms. Moreover existing empirical studies only consider very simple models and small image datasets. Because of their lack of theoretical foundations it is not clear whether they could be extended to more practical setups. Contributions. In this paper we investigate the building blocks of PH and magnitude dimensions in order to propose new topologyinspired generalization bounds that rigorously apply to widely used discretetime stochastic optimization algorithms and experimentally test our new topological complexities on practically relevant DNN architectures. Our detailed contributions are as follows We start by establishing the first theoretical links between generalization and a new kind of computationally thrifty topological complexity measure the weighted lifetime sums . We propose and elaborate on another novel topological complexity positive magnitude PMag a slightly modified version of magnitude . We rigorously link PMag with the generalization error by relying on a new proof technique. Overall our generalization bounds rooted in TDA admit the following generic form sup wWS GSw r Topological complexity IT log n . We then provide a flexible computational implementation based upon dissimilarity measures between neural nets Figure b which enables quantifying generalization across different architectures and models without the need for domain or problemspecific analysis as done in . Unlike existing trajectorybased studies operating on small models our experimental evaluation is extensive. We consider several vision transformers and graph neural networks GNN trained on multiple datasets spanning regular and irregular data domains cf. Figure c. Our findings robustly demonstrate that the novel topological generalization measures we introduce exhibit a strong correlation with test performance across diverse architectures hyperparameters and data modalities actually used in practice. All the proofs of the main results are presented in the appendix along with additional experiments. We will make our entire implementation publicly available under httpsgithub.comrorondreTDAGeneralization. Technical Background Our generalization indicators will be based upon weighted lifetime sums and magnitude capturing different topological features as we shortly dicsuss below. Let X be a finite pseudometric space. weighted lifetime sums. Persistent homology PH is an important concept in the analysis of geometric complexes . We focus on the persistent homology of degree PH . Informally it consists in tracking the connected components of a finite set at different scales. We provide in Sections A. and A. an overview of these notions. For simplicity we present here an equivalent formulation of the weighted lifetime sums based on minimum spanning trees MST . A tree over X is a connected acyclic undirected graph a set of edges whose vertices are the points in X. Given an edge e linking the points a and b we define its cost as e a b. An MST T on X is a tree minimizing the total cost P eT e. The weighted lifetime sums E are then written as E X X eT e . The celebrated persistent homology dimension PHdim of a compact pseudometric space A is then defined as dim PHA inf C Y X finite EY C. the PHdim has been proven to be related to generalization error for different pseudometrics . Our term topological complexity should not be confused with the homonym topological invariant. Magnitude. Magnitude is a recently introduced topological invariant which encodes many important invariants from geometric measure theory and integral geometry . Magnitude can be interpreted as the effective number of distinct points in a space . For s we define a weighting of the modified space X s as a map X R such that a X P bX e sabb . Given such a weighting the magnitude function of X s is defined as Mag sX X aX a. The parameter s should be interpreted as a scale through which we look at the set X . We present in Appendix A. additional properties of this function. Note that magnitude is usually defined in metric spaces we show in Appendix B. that we can seamlessly extend it to the pseudometric setting. Magnitude can also be extended to infinite compact spaces and as for PH an intrinsic dimension the magnitude dimension can be defined from magnitude as dim MagA lims log MagsA logs . It is known that dim PH and dim Mag coincide for compact metric spaces . As a result dim Mag has also been proposed as a topological generalization indicator . Total mutual information. Prior intrinsic dimensionbased studies relied on mixing assumptions Assumption H Assumption H or various mutual information terms to take into account the statistical dependence between the data and the training trajectory. Recently a new framework was proposed in to unify these approaches by proving datadependent uniform generalization bounds using simpler and smaller informationtheoretic IT terms. By leveraging these methods we derive new generalization bounds involving the same IT terms for all our introduced topological complexities. More precisely they take the form of a total mutual information between the data S and the training trajectory WS. This term is denoted IS WS and measures the dependence between S and W. We refer to Appendix A. and for exact definitions. Main Theoretical Results We now introduce our learningtheoretic setup Section . before delving into our main theoretical results in Sections . and .. . Mathematical setup Random trajectories. The primary goal of our theory is to prove uniform generalization bounds over the training trajectory wk k . We are mostly interested in the behavior near local minima of RbS. To this end we observe the trajectory between iterations and T where N is the number of iterations before reaching near a local minimum and T is the total number of iterations. Therefore we consider the set WT wi i T which we call the random trajectory. Note that WT is a set i.e. it does not contain any information about the timedependence. Moreover our setup allows the random times and T to depend on the data S through the choice of a stopping criterion as opposed to being fixed predetermined times. General Lipschitz conditions. The topological quantities described in Section as well as the intrinsic dimensions introduced in prior works require a notion of distance between parameters in Rd to be computed. In the case of fractalbased generalization bounds two cases have already been considered the Euclidean distance and the datadependent pseudometric defined in . In our work we emphasize that both examples are particular cases of a more general family of pseudometrics on the parameter space Rd . In order to fully characterize this family of pseudometrics we define the datadependent map LS Rd Rn by LSw w z . . . w zn. To fit into our framework a pseudometric must satisfy the following general Lipschitz condition. Definition . q L Lipschitz continuity. For any pseudometric on Rd and q we will say that is q L Lipschitz in w when w w Rd LSw LSw q Lnqw w . A wide variety of distances have been proposed to compare the weights of two DNNs . The above condition restricts our analysis to a family of pseudometrics containing the following examples. Example . Datadependent pseudometrics. For any p we define the pseudometrics p S w w n p LSw LSw p . The case S corresponds to the datadependent pseudometric used in we will denote it S S . Example . Euclidean distance. If w z is LLipschitz continuous in w i.e. w z w z Lw w for all z then is p L Lipschitz continuous for every p . Assumptions. Given an q L Lipschitz continuous pseudometric our approach relies only on a single assumption of a bounded loss function. For the case of the pseudometric S Example . this assumption is already made in . Assumption . We assume that the loss is bounded in B with B a constant. The boundedness of is classically assumed in the fractal TDA literature . In it is shown that the proposed theory seems to be experimentally valid even for unbounded losses. Our experimental findings suggest that this observation also applies to our work. . Persistent homology related generalization bounds In contrast to all existing fractal dimensionbased bounds we propose new generalization bounds that apply to practical discrete stochastic optimizers with a finite number of iterations. To this end our key idea involves replacing the intrinsic dimension with intermediary quantities that are used to compute them numerically. Following this points us towards the two quantities E and Mag defined in Section . We are now ready to state the first generalization bound in terms of the weighted lifetime sums where we denote E for E WT . Theorem .. Let be a pseudometric on Rd . Supposes that Assumption holds and that is q L Lipschitz for q . Then for all with probability at least we have sup iT GSwi B vuut log E log L n B n B n B r IS WT log n . The term IS WT is the total mutual information MI term that is defined in Sections and A.. It measures the statistical dependence between the random set WT and the data S n z . Such MI terms appear in previous works related to fractalbased generalization bounds . Our proof technique presented in Appendix B. makes use of a recently introduced PACBayesian framework for random sets to introduce this MI term. It is also shown in that the MI term IS WT is tighter than those appearing in the aforementioned works. We highlight the fact that Theorem . is fundamentally different from the persistent homology dimension PHdim based bounds studied in . Indeed while the growth of E for increasing finite subsets of the trajectory are used in to estimate the PHdim it does not provide any formal link between the generalization error and the value of E. Therefore the above theorem could not be cast as a corollary of these previous studies. Another important characteristic of the above theorem as well as the results of Section . is to be nonasymptotic i.e. it is true for every n N . This is an improvement over the fractal dimensionsbased bounds presented in . Mag n P M a g n Euclidean distance pseudometric S pseudometric a Comparison of Mag and PMag. Proportion of the data Relative variation E PMag n b Relative variation of E and Mag. Figure Left Comparison of Mag and PMag for s n for different pseudometrics ViT on CIFAR. Right relative variation of the quantities EWT and Mag nWT with respect to the proportion of the data used to estimated S ViT on CIFAR. . Positive magnitude PMag and related generalization bounds Recent preliminary experimental results displayed a correlation between the generalization error of DNNs and magnitude . To provide a theoretical justification for this behavior it would be tempting to mimic the proof of Theorem . and build on existing covering arguments. However while lower bounds of magnitude in terms of covering numbers have been derived in they appear to be impractical in our case. Another possibility would be to use the magnitude dimension bounds of . Yet this could not apply to our finite and discrete setting where the dimension is . Hence we identify a new quantity closely related to magnitude while being more relevant to learning theory. With the notations of Section we fix a finite metric space X and a weighting s X R of X s where s is a scale parameter. We define the positive magnitude as s PMag sX X aX sa where x maxx denotes the positive part of x. To avoid harming the readability of the paper we refer to Appendix B. for the extension of PMag to the pseudometric case. Based on a new theoretical approach we prove that the positive magnitude can be used to upper bound the generalization error see the proof in Appendix B.. This leads to the following theorem Theorem .. Let be a pseudometric such that W admits a positive magnitude according to Definition B. for every . We assume that is q L Lipschitz continuous with q . Then for any s we have with probability at least that sup iT GSwi s log PMag LsWT s B n B r IS WT log n . The IT term I in the above result is the same as in Theorem .. Given a fixed finite set W and a big enough s we establish MagsW PMagsW. Moreover we present in Figure a an empirical comparison of Mag and PMag showing a small and almost monotonic relation between both quantities. Therefore Theorem . may be seen as the first theoretical justification of the empirical relationship between magnitude and the generalization error observed in . A natural choice for the scale s would be s n ensuring a convergence rate in n . However our empirical evaluations see Section in particular Table revealed that small values of s we Table Correlation coefficients associated with the different topological complexities. Modeldataset ViTCIFAR SwinCIFAR GraphSageMNIST GatedGCNMNIST Compl.Metric lr bs lr bs lr bs lr bs dimPH S . . . . . . . . . . . . . . . . Mag n S . . . . . . . . . . . . . . . . Mag. S . . . . . . . . . . . . . . . . PMag n S . . . . . . . . . . . . . . . . PMag. S . . . . . . . . . . . . . . . . E S . . . . . . . . . . . . . . . . dimPH . . . . . . . . . . . . . . . . Mag n . . . . . . . . . . . . . . . . Mag. . . . . . . . . . . . . . . . . PMag n . . . . . . . . . . . . . . . . PMag. . . . . . . . . . . . . . . . . E . . . . . . . . . . . . . . . . dimPH . . . . . . . . . . . . . . . . Mag n . . . . . . . . . . . . . . . . Mag. . . . . . . . . . . . . . . . . PMag n . . . . . . . . . . . . . . . . PMag. . . . . . . . . . . . . . . . . E . . . . . . . . . . . . . . . . typically use s can also provide good correlation with the generalization error. This could be explained by the fact that PMagsW as s i.e. the bound may not diverge when s . For our topological complexities to be computationally efficient we focus our experiments on fixed values of s in n . We will omit the trajectory and denote Mags and PMags. Computational Considerations We now detail the numerical estimation of the topological complexities mentioned above. Computation of E. We compute E by using the giottoph library . This setup is inspired by PH frameworks used in . This technique uses the equivalent formulation of E in terms of PH see Appendix A. for details. Theorem . and its proof presented in Appendix B. suggest that the relevant value of is similar to this is what we used in our experiments. Computation of Mag and PMag. Different methods exist to evaluate magnitude . We use the Krylov approximation method which is based on preconditioned conjugate gradient iteration implemented in the Python library krypy.linsys.Cg to solve for the magnitude weights. We then sum over the weights to compute Mag and sum over the positive weights to obtain PMag. Distance matrix estimation.. Given a finite set i.e. a trajectory W Rd the calculation of our topological complexities requires computing the distance matrix D w w wwW. For large DNNs this may become challenging. Depending on we propose the following solutions. Case If is the Euclidean distance on Rd for large DNNs in our case for the transformer experiments storing the whole trajectory is challenging. Instead we use sparse random projections inspired by the JohnsonLindenstrauss lemma to map the trajectories onto a lowerdimensional subspace. We use the scikitlearn impleementation so that with high probability the relative variation of the distance matrices is at most see Appendix A. for details. Case If is of the form q S as in Example . then the computation of D requires the evaluation of the model on the entire dataset at each iteration which becomes intractable for large DNNs. In Figure the authors show that the PHdim based on the pseudometric S S is very robust to a random subsampling of a training dataset i.e. when S is replaced by B with B S and BS . Figure b shows that E and positive magnitude are also robust to this subsampling. We mainly used BS . We refer the reader to Appendix C. for details. Generalization error. Our theory like many trajectorybased studies predicts upper bounds on the worstcase generalization error over the trajectory WT . Yet experiments in previous works mainly reported the error at the last iteration. To estimate the worstcase error in a computationally feasible way we periodically evaluated the test risk between times and T with a period of iterations and reported worst test risk final train risk as the error. This is consistent as we start the trajectory WT from a weight w already in a local minimum of the empirical risk. Our main conclusions are still valid if the final generalization gap is used. This observation which is to the best of our knowledge new is briefly discussed in Appendix D.. Empirical Analysis Setup. Given a DNN and a dataset we start from a pretrained weight vector w yielding high training accuracy on classification tasks. By varying the learning rate and the batch size b we define a grid of hyperparameters. For each pair b we compute the training trajectory WT for iterations. Unless specified we use the ADAM optimizer . Based on the set WT we estimate distance matrices as described in Section . For the sake of clarity we focus on relevant pseudometrics i the Euclidean distance as in ii the datadependent pseudometric S used in and iii the loss distance. For ii S is computed based on the surrogate loss used in training e.g. the crossentropy loss while the reported generalization error is always based on accuracy gap loss which is of interest in most applications see Section . For the last one iii is defined as in Example . but with being the loss we call it pseudometric and denote it by in the tables. This last setup matches exactly our theoretical requirements. In terms of DNN architectures we focus on practically relevant models while previous studies mainly considered small networks . We examine two different families of architectures. The first family consists of vision transformers ViT CaiT Swin see Table each evaluated on both the CIFAR and CIFAR datasets. Moreover we also tested our theory on graph neural networks GNN architectures namely GatedGCN and GraphSage trained on the Superpixel MNIST dataset . To the best of our knowledge this is the first time these kinds of topological complexities have been evaluated on transformers and GNNs. We ran the experiments on NVIDIA Ti GB GPUs. Granulated Kendalls coefficients.. We assess the correlation between our complexities and the generalization error by using the granulated Kendalls coefficients GKC . While the classical Kendalls coefficients KC denoted measures the correlation between two quantities it may fail to capture their causal relationship. Instead one granulated coefficient is defined in for each hyperparameter i.e. LR for and BS for b it measures the correlation when only this hyperparameter is varying. In Table we report LR and BS and the averaged GKC LR BS for several models datasets and topological complexities. In Figures a and b we represent our topological complexities in the plane BS LR the red square indicates the region of best correlation the coefficients are in their sign is the sign of the correlation. . Analysis As explained above we focus our main experiments on the quantities E Mag n PMag n Mag and PMag each computed for the pseudometrics discussed above S . In the interest of comparison we also compute the PHdim proposed in for the and in for S which is thus tested for the first time on transformers and GNNs. Generalization gap . . . . . ph dim Batch size Learning rate Generalization gap positive magnitude Batch size Learning rate Generalization gap E alpha Batch size Learning rate Figure Sbased complexity measures vs. generalization gap for a ViT trained on CIFAR dimPH left PMag n middle and E right. Performance on vision transformers. We see in Table and Figure that our proposed topological complexities consistently outperform the PH dimensions across several vision transformer models and datasets. This suggests that PHdim previously tested only on small architectures is less scalable to industrystandards models with more parameters. Figure a including all model dataset pairs for the pseudometric S reveals important observations. First we notice that the GKC of our topological complexities are both positive and close to indicating that they are indeed good measures of generalization. We note that for most models and datasets dimPH has a small or negative BS indicating that it has less ability to explain generalization for varying batchsizes. As it was observed in for PHdim our complexities computed from the pseudometric S correlate very well with the generalization gap while this gap is based on the loss. Performance on GNNs. An important aspect of our framework is the ability to seamlessly encapsulate different data domains. In particular the possibility of using different pseudometrics can help define topological complexities that naturally take into account the internal symmetries of GNNs without any modelspecific analysis . The results of Table and Figure a confirm that our proposed topological complexities outperform PHdim and correlate strongly with the generalization error for GNNs. Additionally it may be observed that Mag n performs significantly well for GNNs and in particular better than PMag n. This points us towards the idea that further theory would be desirable to formally relate magnitude to the generalization error in that case . Comparison of the topological complexities. In Table and Figures and a it can be seen that E and PMag n perform equally well for the image and graph experiments across multiple datasets models and data domains. We see in Table that most topological complexities perform better with datadependent metrics i.e. S and than with the Euclidean distance for transformerbased experiments. This extends results obtained for PHdim in for smaller architectures. However the poor performance of Euclideanbased complexities may also be partially caused by the projections applied to the Euclidean distance matrices to make them memorywise computable see Section . This is a remaining limitation of our algorithms. On the other hand the and S datadependent pseudometrics seem to yield similar performance in all experiments. Ablations. In Figure b we reveal that changing the optimizer has little effect on the observed correlation for the same model and dataset. Interestingly we note that the PHdim computed with pseudometric S and obtained from the SGD trajectories exhibits high GKCs. This observation agrees with the results in . Figure further displays the typical behavior of several topological complexities for ViT and CIFAR. In addition to the correlation of our proposed complexities being stronger than for the PHdim we observe that E and PMag n seem to better correlate with the generalization gap for small learning rates. Finally it is consistently observed in Table and Figures a and b that using a relatively high value of the positive magnitude scale s n yields better We shall underline that while Mag with the Euclidean distance was empirically proposed as a complexity measure in a theoretical justification for Mag results in Table is still missing for moderate values of s. . . . . . . . . . bs . . . . . . . . . lr Model dataset ViT CIFAR ViT CIFAR CaiT CIFAR CaiT CIFAR Swin CIFAR Swin CIFAR GatedGCN MNIST GraphSage MNIST Topological quantities E PMag n PMag. dimPH a S pseudometric . . . . . . . . . bs . . . . . . . . . lr Optimizers metric SGD S SGDEucl. SGD ADAM S ADAMEucl. ADAM RMSprop S RMSpropEucl. RMSprop Topological quantities E PMag n PMag. dimPH b Comparison of optimizers for ViT on CIFAR. Figure Granulated Kendall coefficients for several models datasets and topological quantities. Note that our framework is directly applicable to graph networks. correlations than small values s . However both cases still provide satisfying correlation comforting the robustness of magnitude as a generalization indicator. Due to limited space we present all the correlation coefficient of one transformer model ViT for CIFAR and Swin for CIFAR in Table as illustrative examples for each dataset. The remaining results appear in the Appendix Tables and and they all follow a similar trend. Further empirical results and illustrations of this behavior are provided in Appendix D. On topological complexity generalization. Contrary to the claims in which are based on the limited experimental settings of our findings on an extensive set of evaluations on modern architectures decisively demonstrate that topological complexity measures serve as robust proxies for generalization. While fractal dimension shows some correlation with generalization our rigorous discretetime topological measures exhibit a significantly stronger correlation evidenced by the high granulated coefficients. This promising outcome underscores the need for future research in exploring deeper connections between network dynamics topology and generalization. Conclusion In this paper we proved novel generalization bounds based on several topological complexities coming from TDA namely weighted lifetime sums and a new variant of metric space magnitude which we called positive magnitude. Compared to previous studies we require fewer assumptions and operate in a discrete setting in which our proposed quantities are fully computable. Our algorithms are flexible enough to be seamlessly integrated with diverse data domains and tasks. These advantages of our framework allowed us to create a computationally cheap experimental setup as close as possible to the theoretical setup. We thus provided a comprehensive suite of experiments with several industryrelevant architectures across vision transformers and graph neural networks which have not been explored yet in this literature. We show that our proposed topological complexities correlate well with the generalization error outperforming the previously studied intrinsic dimensions. Limitations future work. The main limitation of our theory is the lack of understanding of the IT terms while they are still smaller than most prior works. Moreover a better understanding of the behavior of positive magnitude for small values of the scale factor s would be a necessary improvement. Regarding our experiments a refinement of the estimation techniques of the topological complexities would be beneficial. Despite experimenting with practically relevant architectures our future works also include scaling up our empirical analysis to include larger models and datasets in particular large language models which are still beyond the scope of this study. Broader impact. Certifying generalization is key for safe and trusted AI systems hence we believe that our study may have a positive societal impact on the general use of deep networks. Acknowledgments R.A. is supported by the United Kingdom Research and Innovation grant EPSX UKRI Centre for Doctoral Training in Biomedical AI at the University of Edinburgh School of Informatics. U.S. is partially supported by the French government under management of Agence Nationale de la Recherche as part of the Investissements davenir program reference ANRPIA PRAIRIE IA Institute. B.D. and U.S. are partially supported by the European Research Council Starting Grant DYNASTY . T.B. is partially supported by the Royal Society Research Grant RGR. References Henry Adams Manuchehr Aminian Elin Farnell Michael Kirby Joshua Mirth Rachel Neville Chris Peterson and Clayton Shonkwiler. A fractal dimension for measures via persistent homology. In Topological Data Analysis The Abel Symposium pages . Springer . Henry Adams and Michael Moy. Topology applied to machine learning From global to local. Frontiers in Artificial Intelligence . Rayna Andreeva Katharina Limbeck Bastian Rieck and Rik Sarkar. Metric space magnitude and generalisation in neural networks. In Proceedings of nd Annual Workshop on Topology Algebra and Geometry in Machine Learning TAGML volume of Proceedings of Machine Learning Research pages . PMLR . Sanjeev Arora Simon S. Du Wei Hu Zhiyuan Li and Ruosong Wang. FineGrained Analysis of Optimization and Generalization for Overparameterized TwoLayer Neural Networks May . Peter Bartlett and Shahar Mendelson. Rademacher and Gaussian Complexities Risk Bounds and Structural Results. Journal of Machine Learning Research . Ulrich Bauer. Ripser Efficient computation of VietorisRips persistence barcodes. Journal of Applied and Computational Topology September . Arash Behboodi Gabriele Cesa and Taco S Cohen. A pacbayesian generalization bound for equivariant networks. Advances in Neural Information Processing Systems . Subhrajit Bhattacharya Robert Ghrist and Vijay Kumar. Persistent homology for path planning in uncertain environments. IEEE Transactions on Robotics . Tolga Birdal Aaron Lou Leonidas J Guibas and Umut Simsekli. Intrinsic dimension persistent homology and generalization in neural networks. Advances in Neural Information Processing Systems . JeanDaniel Boissonat Frederic Chazal and Mariette Yvinec. Geometrical and Topological Inference. Cambridge Texts in Applied Mathematics. Cambridge University Press . Xavier Bresson and Thomas Laurent. Residual gated graph convnets. arXiv preprint arXiv. . Alexander Camuto George Deligiannidis Murat A Erdogdu Mert Gurbuzbalaban Umut Simsekli and Lingjiong Zhu. Fractal structure and generalization properties of stochastic optimization algorithms. Advances in neural information processing systems . Gunnar Carlsson. Topological pattern recognition for point cloud data. Acta Numerica May . Frederic Chazal and Bertrand Michel. An introduction to topological data analysis fundamental and practical aspects for data scientists. Frontiers in artificial intelligence . Thomas H. Cormen Charles E. Leiserson Ronald L. Rivest and Clifford Stein. Introduction to Algorithms. The MIT Press nd edition . Ciprian A Corneanu Meysam Madadi Sergio Escalera and Aleix M Martinez. What does it mean to learn in deep networks? and how does one detect adversarial attacks? In Proceedings of the IEEECVF Conference on Computer Vision and Pattern Recognition pages . Vin De Silva and Robert Ghrist. Coverage in sensor networks via persistent homology. Algebraic Geometric Topology . Claire Donnat and Susan Holmes. Tracking network dynamics a survey of distances and similarity metrics . Alexey Dosovitskiy Lucas Beyer Alexander Kolesnikov Dirk Weissenborn Xiaohua Zhai Thomas Unterthiner Mostafa Dehghani Matthias Minderer Georg Heigold Sylvain Gelly Jakob Uszkoreit and Neil Houlsby. An Image is Worth x Words Transformers for Image Recognition at Scale June . Benjamin Dupuis George Deligiannidis and Umut Simsekli. Generalization bounds using datadependent fractal dimensions. In International Conference on Machine Learning pages . PMLR . Benjamin Dupuis Paul Viallard George Deligiannidis and Umut Simsekli. Uniform generalization bounds on datadependent hypothesis sets via pacbayesian theory on random sets . Vijay Prakash Dwivedi Chaitanya K Joshi Anh Tuan Luu Thomas Laurent Yoshua Bengio and Xavier Bresson. Benchmarking graph neural networks. Journal of Machine Learning Research . Herbert Edelsbrunner and John Harer. Computational Topology an Introduction Semantic Scholar. American Mathematical Society . Kevin Emmett Benjamin Schweinhart and Raul Rabadan. Multiscale topology of chromatin folding. In th EAI International Conference on Bioinspired Information and Communications Technologies formerly BIONETICS pages . Kenneth Falconer. Fractal Geometry Mathematical Foundations and Applications Third Edition. Wiley . Carlos FernandezGranda. Lecture random projections . Jingwen Fu Zhizheng Zhang Dacheng Yin Yan Lu and Nanning Zheng. Learning Trajectories are Generalization Indicators October . Hanan Gani Muzammal Naseer and Mohammad Yaqub. How to train vision transformer on smallscale datasets? arXiv preprint arXiv. . Michael Gastpar Ido Nachum Jonathan Shafer and Thomas Weinberger. Fantastic generalization measures are nowhere to be found . M. Gori G. Monfardini and F. Scarselli. A new model for learning in graph domains. In Proceedings. IEEE International Joint Conference on Neural Networks . volume pages vol. . Mustafa Hajij Ghada Zamzmi Theodore Papamarkou Nina Miolane Aldo GuzmanSaenz Karthikeyan Natesan Ramamurthy Tolga Birdal Tamal K Dey Soham Mukherjee Shreyas N Samaga et al. Topological deep learning Going beyond graph data. arXiv preprint arXiv. . Will Hamilton Zhitao Ying and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems . Allen Hatcher. Algebraic Topology. Cambridge University Press . Yasuaki Hiraoka Takenobu Nakamura Akihiko Hirata Emerson G Escolar Kaname Matsue and Yasumasa Nishiura. Hierarchical structures of amorphous solids characterized by persistent homology. Proceedings of the National Academy of Sciences . Liam Hodgkinson Umut Simsekli Rajiv Khanna and Michael W. Mahoney. Generalization Bounds using Lower Tail Exponents in Stochastic Optimizers. Proceedings of the th International Conference on Machine Learning July . Ahmed Imtiaz Humayun Randall Balestriero and Richard Baraniuk. Training Dynamics of Deep Network Linear Regions. httpsarxiv.orgabs.v October . Yiding Jiang Behnam Neyshabur Hossein Mobahi Dilip Krishnan and Samy Bengio. Fantastic Generalization Measures and Where to Find Them. ICLR December . Maurice G. Kendall. A new reasure of rank correlation. Biometrika . Bobak T Kiani Thien Le Hannah Lawrence Stefanie Jegelka and Melanie Weber. On the hardness of learning under symmetries. arXiv preprint arXiv. . Diederik P. Kingma and Jimmy Ba. Adam A method for stochastic optimization . Thomas N Kipf and Max Welling. Semisupervised classification with graph convolutional networks. arXiv preprint arXiv. . Gady Kozma Zvi Lotker and Gideon Stupp. The minimal spanning tree and the upper box dimension. Proceedings of the American Mathematical Society . Alex Krizhevsky. Learning multiple layers of features from tiny images . Alex Krizhevsky Vinod Nair and Geoffrey E. Hinton. The cifar dataset . Gregory Leibon Scott Pauls Daniel Rockmore and Robert Savell. Topological structures in the equities market network. Proceedings of the National Academy of Sciences . Tom Leinster. The magnitude of metric spaces. Documenta Mathematica . Katharina Limbeck Rayna Andreeva Rik Sarkar and Bastian Rieck. Metric space magnitude for evaluating the diversity of latent representations . Ze Liu Yutong Lin Yue Cao Han Hu Yixuan Wei Zheng Zhang Stephen Lin and Baining Guo. Swin transformer Hierarchical vision transformer using shifted windows. In Proceedings of the IEEECVF international conference on computer vision pages . Xuanyuan Luo Luo Bei and Jian Li. Generalization Bounds for Gradient Methods via Discrete and Continuous Prior October . Kaifeng Lyu Zhiyuan Li and Sanjeev Arora. Understanding the Generalization Benefit of Normalization Layers Sharpness Reduction January . German Magai. Deep neural networks architectures from the perspective of manifold learning. In IEEE th International Conference on Pattern Recognition and Artificial Intelligence PRAI pages . IEEE . German Magai and Anton Ayzenberg. Topology and geometry of data manifold in deep learning. arXiv preprint arXiv. . Pertti Mattila. Geometry of Sets and Measures in Euclidean Spaces. Cambridge University Press . Pertti Mattila Manuel Moran and Josemanual Rey. Dimension of a measure. Studia mathematica . Mark W Meckes. Positive definite metric spaces. Positivity . Mark W Meckes. Magnitude diversity capacities and dimensions of metric spaces. Potential Analysis . Wenlong Mou Liwei Wang Xiyu Zhai and Kai Zheng. Generalization Bounds of SGLD for Nonconvex Learning Two Theoretical Viewpoints. In Proceedings of the st Conference On Learning Theory. arXiv July . Gregory Naitzat Andrey Zhitnikov and LekHeng Lim. Topology of deep neural networks. Journal of Machine Learning Research . Gergely Neu Gintare Karolina Dziugaite Mahdi Haghifam and Daniel M. Roy. InformationTheoretic Generalization Bounds for Stochastic Gradient Descent August . Monica Nicolau Arnold J Levine and Gunnar Carlsson. Topology based data analysis identifies a subgroup of breast cancers with a unique mutational profile and excellent survival. Proceedings of the National Academy of Sciences . Nina Otter Mason A Porter Ulrike Tillmann Peter Grindrod and Heather A Harrington. A roadmap for the computation of persistent homology. EPJ Data Science . Theodore Papamarkou Tolga Birdal Michael Bronstein Gunnar Carlsson Justin Curry Yue Gao Mustafa Hajij Roland Kwitt Pietro Lio Paolo Di Lorenzo et al. Position paper Challenges and opportunities in topological deep learning. arXiv preprint arXiv. . F. Pedregosa G. Varoquaux A. Gramfort V. Michel B. Thirion O. Grisel M. Blondel P. Prettenhofer R. Weiss V. Dubourg J. Vanderplas A. Passos D. Cournapeau M. Brucher M. Perrot and E. Duchesnay. Scikitlearn Machine learning in Python. Journal of Machine Learning Research . Ankit Pensia Varun Jog and PoLing Loh. Generalization Error Bounds for Noisy Iterative Algorithms. IEEE International Symposium on Information Theory ISIT January . Julian Burella Perez Sydney Hauke Umberto Lupo Matteo Caorsi and Alberto Dassatti. Giottoph A Python Library for HighPerformance Computation of Persistent Homology of VietorisRips Filtrations August . David PerezFernandez Asier GutierrezFandino Jordi ArmengolEstape and Marta Villegas. Characterizing and measuring the similarity of neural networks with persistent homology. arXiv preprint arXiv. . Yakov B Pesin. Dimension theory in dynamical systems contemporary views and applications. University of Chicago Press . Maithra Raghu Thomas Unterthiner Simon Kornblith Chiyuan Zhang and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in neural information processing systems . Patrick Rebeschini. Algorithmic fundations of learning . Bastian Rieck Matteo Togninalli Christian Bock Michael Moor Max Horn Thomas Gumbsch and Karsten Borgwardt. Neural persistence A complexity measure for deep neural networks using algebraic topology. arXiv preprint arXiv. . Sarav Sachs Umut Simsekli and Tim van Erven. Generalization Guarantees via Algorithmdependent Rademacher Complexity preprint. COLT . Shilan Salim. The qspread dimension and the maximum diversity of square grid metric spaces. PhD thesis University of Sheffield . Benjamin Schweinhart. Fractal dimension and the persistent homology of random geometric complexes. Advances in Mathematics . Benjamin Schweinhart. Persistent homology and the upper box dimension. Discrete Computational Geometry . Milad Sefidgaran Amin Gohari Gael Richard and Umut Simsekli. RateDistortion Theoretic Generalization Bounds for Stochastic Learning Algorithms June . Milad Sefidgaran and Abdellatif Zaidi. Datadependent Generalization Bounds via VariableSize Compressibility January . Shai ShalevSchwartz and Shai BenDavid. Understanding Machine Learning From Theory to Algorithms. Cambridge University Press . Umut Simsekli Ozan Sener George Deligiannidis and Murat A Erdogdu. Hausdorff dimension heavy tails and generalization in neural networks. Advances in Neural Information Processing Systems . Charlie Tan Ines GarcaRedondo Qiquan Wang Michael M Bronstein and Anthea Monod. On the limitations of fractal dimension as a measure of generalization. arXiv preprint arXiv. . Hugo Touvron Matthieu Cord Matthijs Douze Francisco Massa Alexandre Sablayrolles and Herve Jegou. Training dataefficient image transformers distillation through attention. In International conference on machine learning pages . PMLR . Hugo Touvron Matthieu Cord Alexandre Sablayrolles Gabriel Synnaeve and Herve Jegou. Going deeper with image transformers. In Proceedings of the IEEECVF international conference on computer vision pages . Tim van Erven and Peter Harremoes. Renyi Divergence and KullbackLeibler Divergence. IEEE Transactions on Information Theory July . Roman Vershynin. HighDimensional Probability An Introduction with Applications in Data Science. Number in Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press . Satoru Watanabe and Hayato Yamana. Topological measurement of deep neural networks using persistent homology. Annals of Mathematics and Artificial Intelligence . Jing Xu Jiaye Teng Yang Yuan and Andrew Yao. Towards DataAlgorithm Dependent Generalization A Case Study on Overparameterized Linear Regression. Advances in Neural Information Processing Systems December . Chiyuan Zhang Samy Bengio Moritz Hardt Benjamin Recht and Oriol Vinyals. Understanding deep learning requires rethinking generalization. ICLR February . Chiyuan Zhang Samy Bengio Moritz Hardt Benjamin Recht and Oriol Vinyals. Understanding deep learning still requires rethinking generalization. Communications of the ACM February . Afra Zomorodian. Topological data analysis. Advances in applied and computational topology . Appendix We now provide additional technical details and proofs that are omitted from the paper followed by experimental evidence complementing our main paper. We organize the appendix as follows Appendix A presents additional technical background related to information theory Rademacher complexity and the various topological quantities that appear in our work. In Appendix B we present the omitted proofs of all our theoretical results as well as a few additional theoretical contributions. In Appendix C we show the experimental details needed to reproduce our experiments. Finally Appendix D is dedicated to additional empirical results. A Additional technical background A. Informationtheoretic quantities The following definition is a precise definition of the total mutual information term that appears in our main theoretical results. The reader may consult for further information on this notion. Definition A. Total mutual information. Let X and Y be two random elements defined on a probability space F P note that the codomains of X and Y may be distinct. We define the total mutual information between X and Y by the following formula IX Y log sup A PXY A PX PY A . Such a term has already been used in the fractalbased generalization literature . Other works used intricate variants of this total mutual information term . We stress the fact that our proposed bounds are simpler. A. Rademacher complexity Rademacher complexity is a central tool in learning theory. As part of our theory uses this notion we now provide its definition and introduce some notation. Definition A. Rademacher complexity on a hypothesis set. Let us fix a dataset S Zn a set W Rd and . . . n some iid Rademacher random variable. Whenever it is defined we will call Rademacher complexity of over W the following quantity Rad W S n E sup wW Xn i iw zi . Rademacher complexity has already been used in Theorem . to relate the generalization error to the socalled datadependent fractal dimension. Part of our theory is based on a recent extension of such arguments in the datadependent setting . A Rademacher random variable is defined by Pi Pi . A. Persistent homology The goal of this short subsection is to present a few notions of persistent homology which is necessary for a better understanding of our contributions. Persistent homology is an important subfield of TDA capable of providing myriad of new insights for analysing data by extracting meaningful topological features. It has demonstrated its usefulness in a very diverse set of applications from biology to materials science finance robotics sensor networks and a lot more . The types of datasets which are amenable to this kind of analysis are finite metric spaces known as pointcloud datasets images networks and also levelsets of functions. More recently several studies have brought to light empirical links between persistent homology and DNNs . In particular recent studies have related the worstcase generalization error to several concept of intrinsic dimensions defined through persistent homology . As mentioned in the introduction our goal is to extend these last studies to more practical settings. In general persistent homology is defined for any degree k N denoted PHk . Intuitively PHk keeps track of the number of holes of dimension k in a set when looked at different scales. However in our work and as in we only use PH whose presentation is simpler. In this section to avoid harming the readability of the paper we only present a highlevel introduction to PH that is sufficient to understand our work. The interested reader may consult for a more indepth introduction to persistent homology. We first start by introducing briefly homology which is a classical concept in algebraic topology. We only introduce the most essential concepts for understanding persistent homology. For a more detailed introduction please consult . Definition A.. A simplicial complex is a set K of finite sets closed under the subset relation if K and then K. In the above definition is a simplex plural simplices and is a face of its coface. Definition A.. An abstract simplicial complex K is a finite collection of simplices where a face of any simplex K is also a simplex in K. Definition A.. A simplicial kchain is the formal sum of ksimplices X N i rii where each ri R where R is a fixed commutative ring with additive identity and multiplicative identity and i K. Kk is the set of simplicial kchains with addition over R which is an Rmodule. Then the set of all ksimplices of the complex K is a set of generators for Kk. For each generator the boundary of is the sum of all k faces of . Definition A.. The boundary of a ksimplex x . . . xk is the k chain k X k i i x . . . xi . . . xk where x . . . xi . . . xk is the k simplex spanned by all vertices without xi . It is common that the coefficients for homology are considered to be restricted to Z which is the field with elements and where . However the theory extends to homoogy with coefficeints in any field and since every field is a ring the definitions in terms of rings are more general. Definition A.. A chain complex is a sequence of abelian groups Ak with homomorphisms called boundary maps k Ak Ak such that k k for all k. We should note that when considering coefficients in Z a kchain can be seen as a finite collection of ksimplices. Introduce topological invariants simplicial homology groups and Betti numbers. Definition A. Simplicial Homology group. The nth simplicial homology group of a finite simplicial complex K is Hn ker nimn where ker and im are the kernel and image respectively of the boundary operator. In order to define the simplicial complexes of use in TDA we need to first understand what a nerve is. Definition A. Nerve. A simplicial complex associated to a collection of sets is called a nerve. The sets are the vertices of the complex and a simplex belongs to a complex iff its vertices have a nonempty intersection Nrv S AA . Definition A. Cech complex . The Cech complex of X for radius r is Cech rX NrvBx r x X where Bx r is the closed ball of radius r centered at x. In other words the Cech complex is the nerve of the ball neighbourhoods of a set of points X R n. The Cech complex faithfully captures the topology of the space but it is not computed in practice due to its high computational cost. Instead a different complex called VietorisRips VR is used due to ease of construction for higher dimensions. It can be shown that the VR complex is not always homotopy equivalent to the Cech complex and therefore it can be seen as an approximation. We first need to introduce the notion of a clique complex to explain what the VR is. Definition A. Clique complex. The clique complex for a graph G V E consists of all cliques of G which are all simplices V for which E contains all edges of . Now we have explicitly states all the necessary components in order to define the main complex used in TDA the VietorisRips complex. Definition A. VietorisRips complex. The VietorisRips complex of X for radius r is the clique complex of the skeleton of the Cech complex of X and r Ripsr X X u v r for all u v . Now that we have defined the most important complex in TDA we proceed to explain how we can derive important topological information at multiple scales by introducing the concept of a filtration. Definition A.. Given a simplicial complex K a filtration is a totally ordered set of subcomplexes Ki of K indexed by nonnegative integers such that for i j Ki Kj . Definition A. Filtered simplicial complex. A simplicial complex K together with a filtration function f K R such that f f whenever is a face of . The sublevel set at a value r R is f r which is a subxomplex of K. Let r r rm be the values of the simplices and Ki f ri then we call K K Km the sublevel set filtration of f. When you start with a simplicial complex K and you filter it according to a filtration f it is clear that the homology of Kr evolves as the radius r increases. For example new connected components can be formed loops can appear or disapper cavities can form. What persistent homology does and where the importance of the filtering comes in is that now we have the tools to track the topological changes associated with the different stages of the filtering process and to associate a lifetime to them track when a topological feature has first appeared and at which stage of the filtration it will disappear. This essential topological information is recorded in a set of intervals known as barcodes which can be represented as a multiset of points in R where the coordinates correspond to the birth and death points of each interval. A.. Persistent homology of degree alternative approach For the rest of the this section we only focus on homology in dimension and provide an alternative and perhaps easier to understand interpretation. Please note that the following definition is a simplified and nonstandard though equivalent definition of PH . Definition A. Persistent homology of degree PH . Let X be a finite metric space and N its cardinality. For each time t we construct an undirected graph Gt whose edges are given by x y X x y Gt x y . There exists a finite set of times t tk such that the number of connected components in Gti changes compared to Gt for t ti . Let ci be the number of connected components in Gti . By convention we set c N and t and define ni ci ci. PH is then defined as the following multiset the notation denotes multisets PH t . . . t z n times t . . . tk . . . tk z nk times . Remark A. VietorisRips filtration. The above is a simplified highlevel definition of PH . More formally the construction of the family of graphs Gt corresponds to the construction of the socalled VietorisRips filtration of X of which we only kept the simplices of dimension see for more details. We now use PH to give the definitions of the quantities of interest in our work. The following is a definition of the quantity E already mentioned in Section but seen through the lens of persistent homology. As it will be explained in Appendix A. these definitions are equivalent. We use the term time for the scalar t as it is classically done in the study of persistent homology. Note that this has nothing to do with the number of iterations appearing in the rest of the paper. Definition A. weighted lifetime sums. With the same notations as in Definition A. we define the weighted lifetime sums as E X X tPH t . Remark A. birth and death times. PHk is usually defined as a multiset of birth and death times tracking the appearance and disappearance of holes of dimension k during the construction of the VietorisRips filtration of X. In the particular case of PH all birth times are and the times that we constructed correspond to the death times. We end this section by giving the definition of the PH dimension which has been shown to be theoretically and empirically related to the generalization error of neural networks in prior works . Definition A. Persistent homology dimension of degree . Given a compact metric space X we define the PH dimension of degree by dim PHX inf C A X finite EA C . It has been shown in that for any compact metric space the PH dimension defined above is equal to the celebrated upper boxcounting dimension . A. Minimum spanning tree The persistent homology dimension used in existing generalization bounds is closely related to another notion of intrinsic dimension called minimum spanning tree MST dimension in the sense that the PH and MST dimensions of bounded metric spaces are identical. The link between persistent homology and MST is even deeper than the equality between the induced dimensions as noted by . In this section we define quantities related to MSTs which will play an important role in our proofs. In this section let us fix a finite metric space X . Let us first specify our notations for trees. A tree T on X is a connected undirected graph. We represent T by its set of edges which are denoted a b or equivalently b a as the graph is undirected. For an edge e of the form a b we define its length by e a b. Definition A. Minimum spanning tree. Let us define the cost of a tree by the sum of the length of its edges i.e. E MST T X eT e. An MST of X is defined as a tree with minimal cost. A consequence of the greedy algorithm to find such an MST is that an MST T is also minimal for any of the following costs E MST T X eT e with . Our interest in this notion comes from several results that are summed up in the following theorem. The reader can refer to for more details. Theorem A. Link between MST and persistent homology of degree . There is a bijection between the two following multisets The multiset of the lifetimes in the persistent homology of degree of the VietorisRips complex of X. The multiset of the length of the edges of an MST of X. Therefore if we fix some the weighted sum associated to the persistent homology of degree of the VietorisRips complex of X is equal to the cost E of an MST of X ie E MST T EX. In all the following we will use the notation E to denote both quantities. A. Magnitude Let us restate formally a few standard definitions of magnitude weighting and positive definite metric spaces. We refer the reader to for more details. In this section we fix a finite metric space X . Some of the presented concepts will be later extended to pseudometric spaces in Appendix B.. As before the similarity matrix of X is defined by Ma b e ab for a b X. We now define weightings and magnitude of X according to Section .. Definition A. Weighting and magnitude. A weighting of X is a function X R such that a X X bX e abb . If such a weighting exists the magnitude of X is defined by MagX X bX b. It is easily seen that this definition is independent of the choice of weighting . When a weighting exists we say that X has magnitude. Based on such a definition it is natural to inquire whether such a weighting exists. This question has been studied by several authors . This question appears to be related to the notion of positive definite space which we now define according to . Definition A. Positive definite space. X is positive definite if the similarity matrix M is positive definite. It is clear that positive definite spaces have magnitude. More interestingly we have the following result which ensures that most metric spaces considered in this study are positive definite. Theorem A. . Let p and d every finite subset of Rd p is positive definite. A. Covering and packing numbers In this section we fix a compact pseudometric space X and give definitions of covering and packing numbers. These quantities have long been of primary interest in learning theory in particular through the classical covering arguments for Rademacher complexity . More recently limits of covering arguments have been leveraged by several authors to derive uniform generalization bounds in terms of fractal dimensions which we aim to improve in this study. For x X and r we denote the closed ball centered at x and or radius r by B rx y X x y r. We can now define covering and packing. Definition A. Covering number. Let the covering number N X is the cardinality of a minimal set of points N such that X xN B x. Remark A.. There exist several conventions for the definition of such numbers all of which are equivalent up to absolute constants and in particular induce the same fractal dimensions on X see . Definition A. Packing number. Let the covering number N X is the cardinality of a maximal set of disjoint closed balls with centers in X. A. About JohnsonLindenstrauss lemma In our implementation of Euclideanbased topological quantities we use sparse random projections to project the weight vectors from Rd to a lower dimensional subspace. This is necessary because of memory constraints. Indeed storing the full trajectory WT Rd in our experiments T can become intractable for large models. Given a finite set of points W Rd and . Let N O log W JohnsonLindenstrauss lemma ensures the existence of a linear map P Rd RN such that w w W w w P w P w w w . In practice the linear maps suggested by this result can be obtained through subgaussian random projections Section .. In our work as the purpose of JohnsonLindenstrauss embeddings is mainly memory optimization we have to rely on sparse random projections. We use the implementation provided in scikitlearn . More precisely we used a relative variation of . Finally it should be noted that these projection techniques were only used for the vision transformer experiments as the GNNs that we used have a small enough number of parameters to avoid the use of random projections. A. A note on the connection to Topological Deep Learning Topological deep learning TDL is a rapidly evolving field that uses topological features to understand and design deep learning models . Our topological complexity measures can be seen as a direction towards addressing the Open Problem mentioned in concerning the discovery of topological properties of internal representations that are linked to generalization. B Omitted proofs of the theoretical results In this section we present the proofs of our main theoretical contributions. We divide our proofs into two groups of subsections Sections B.B. and B. focus on the extension in a very natural way of the quantities appearing in our bounds in pseudometric spaces. The main outcome of this analysis is the definition of positive magnitude in the pseudometric case. Note that Appendix B. is not a contribution of this paper. We placed it in this section to improve the readability of the paper. In sections B. B. B. and B. we present the proof of our main theoretical results. Before proving our main results we define the notion of metric identification which will be used in several of the following subsections. This is the same setting that was used in to naturally extend the persistent homology dimension to pseudometric spaces. Definition B. Metric identification. Let X be a pseudometric space. We can define an equivalence relation on X by a b a b . The associated quotient space which is denoted X is a metric space for the naturally induced metric which we still denote . We will also use the canonical projection X X. These notations will be used throughout the text. B. Persistent homology and MST in pseudometric spaces In this short subsection we first restate results proven in regarding persistent homology in pseudometric spaces. The main result is the following proposition which has been proven inside the proof of Lemma B.. Proposition B. . Let X be a finite pseudometric space and then we have EX E X where the pseudometric and its metric identification have been omitted from the notation. Based on Theorem A. the above result is also true when E represents the cost of a MST of X. B. Magnitude in pseudometric spaces In this section we fix X a finite pseudometric space. We denote by X its metric identification and by X X the canonical projection. Indeed if a b then we have c X a c b c. We directly extend Definition A. to the pseudometric case. In order for this definition to make sense in our context we first need to verify that it provides a wellposed definition of magnitude. This follows from the following lemma. Lemma B.. We assume that the finite pseudometric space X has magnitude. Then magnitude is independent of the choice of weighting. Proof. The proof is straightforward and identical to the metric case. Let be two weightings we have X aX a X aX X bX e ab ba X bX b X aX e aba X bX b. In the following theorem we show that magnitude is invariant through metric identification. Theorem B. Invariance of magnitude through metric identification. X has magnitude if and only if X has magnitude in which case we have MagX Mag X. Proof. We decompose X into equivalence classes as X a aX a a iI ai where denotes disjoint union and the points aiiI XI represent each equivalence class. We denote by a the equivalence class of a X. Let X R be any function. We have a X X bX e abb X iI e aai X bai b. If X has magnitude then we take to be a weighting of X we define a X a X ba b. By Equation is a weighting of X. if is a weighting of X then we define a X a a a where a denotes the cardinality of a. By Equation is a weighting of X. B. Definition of positive magnitude in the pseudometric case Let us extend our new notion of positive magnitude in finite pseudometric spaces. This is a rather complicated task. Indeed we need to ensure that the positive magnitude is independent of the choice of weighting which is not true in general. For this reason we restrict our definition to pseudometric spaces whose metric identification is positive definite and we choose one particular weighting. Definition B. Positive magnitude in finite pseudometric spaces. Let X be a finite pseudometric space whose metric identification X is positive definite. Let X R be a weighting of X then we define the positive magnitude of X denoted PMag by PMagX X xX x where x maxx denotes the positive part of x. We will say that X admits a positive magnitude if its metric identification X is positive definite. Note that X admits a unique weighting because it is positive definite. However X still admits several weightings in general. The above definition ensures that the definition of positive magnitude is independent of any choice of weighting. For the need of our proofs we will need to introduce weightings in pseudometric spaces whose sums of positive parts yield the positive magnitude. This is possible by using the following definition which corresponds to a good choice of weighting in finite pseudometric spaces. Definition B. Canonical weighting. Let X be a finite pseudometric space whose metric identification X is positive definite. Let X R be a weighting of X we define the canonical weighting X R on X by a X a a a where X X is the canonical surjection. The following lemma is then obvious but crucial to some of our theoretical results. Lemma B.. With the notation of the previous definition we have PMagX X xX x. The next proposition is a consequence of Theorem A. it shows that the pseudometrics considered in practice in our work and in our experiments admit a positive magnitude. Proposition B.. Let p and S Zn then every finite subset of Rd p S admits a positive magnitude and therefore it also has a canonical weighting. Proof. Let W w . . . wN be a finite set in Rd . We have LSw LSw p n p p S w w . Therefore if we denote by w the equivalence class of w in the metric identification it is clear that w w LSw LSw . Hence the map S n pLS naturally extends to an isometry between metric spaces W SW finite R n . By Theorem A. the finite set SW is positive definite hence it is also the case of W. Therefore W admits a positive magnitude by definition. B. Warmup covering bounds The following is deduced from the transcription of the results of to our setting. It is the starting point of our persistent homologybased analysis. Theorem B.. Let be a pseudometric on Rd . Suppose that Assumption holds and that is q L Lipschitz for q . Then for all with probability at least over n z u sup iT GSwi L B r log N WT n B r IS WT log n . The proof of this theorem will be given in the next subsection. Before discussing this proof a few remarks are in order. Covering bounds such as B. have been used in to introduce fractal dimensions more precisely through the notion of upper boxcounting dimension into the generalization bounds. This is done via the following definition of the aforementioned upper boxcounting dimension dim BX lim sup log N X log . By using a similar procedure we see that our framework could be used to introduce intrinsic dimensions associated to a wide range of pseudometrics as soon as they satisfy a q L Lipschitz continuity assumption. However arguments based on these intrinsic dimensions only make sense in the limit T which makes little sense in practical settings. To address this issue we take inspiration from two other notions that are equal to the upper boxcounting dimension and therefore lay the ground of the numerical approximation of this dimension namely the PHdimension and the magnitude dimension . Our approach is to replace the intrinsic dimensions by the intermediary quantities used to define them. This leads to the results presented in the next two subsection. B. Proof of Theorem B. Before going to the proof of Theorem B. we specify our theoretical setup which is the one introduced in . In this section we prove our results in the case T . However note that one could consider T without much technical difficulties. The setup is the following let FRd T denote the set of all finite subsets of Rd endowed with a algebra T . We consider the following probability distribution on FRd A T A Z Zn SAdn z S. As it is discussed in Section . we make the following technical measuretheoretic assumption. Assumption . The probability measure n z is a strictly positive Borel measure. Moreover for every A T the map S SA is continuous. The following example highlights the fact this is a very mild assumption. Example B.. If the data space Z is countable and the data distribution z has no null mass then the above assumption is automatically satisfied with respect to the discrete topology. Theorem B.. Let be a pseudometric on Rd . Suppose that Assumption holds and that is q L Lipschitz for q . Then for all with probability at least over n z u sup iT GSwi L B r log N WT n B r IS WT log n . Proof. Let us fix some . First note that thanks to Assumption we have that S is absolutely continuous with respect to n z almost surely. Therefore we can introduce its RadonNykodym derivative denoted by dSd. Thanks to the above notation we can apply the datadependent Rademacher complexity bound of Theorem to obtain that with probability at least we have for any sup iT Rwi RbSwi Rad WT S dS d WT log B n with Rad WT S a Rademacher complexity term defined by Rad WT S E sup wWT n Xn i iw zi where . . . n is a vector of independent centered Bernoulli random variables. By Lemma we have almost surely that dS d WT IWT S. Therefore by optimizing the choice of the parameter in the above equation we have that sup iT Rwi RbSwi Rad WT S B r IS WT log n . We now perform a covering argument very similar to classical covering arguments for Rademacher complexity. Let us fix some and introduce x . . . xN WT the centers of a minimal covering of WT for pseudometric . For any w WT there exists j such that w xj . Therefore we have sup wWT n Xn i iw zi sup jN WT n Xn i ixj zi n Xn i iw zi xj zi sup jN WT n Xn i ixj zi n Xn i w zi xj zi sup jN WT n Xn i ixj zi n q LSw LSxj q where the last line comes from Holders inequality. We can now apply Massarts lemma on the first term and the q L Lipschitz continuity of on the second term this gives us Rad WT S L B r log N WT n which concludes the proof. B. Persistent homology bounds We now present the proofs of our persistent homologybased bounds ie the results of Section .. The following lemma is a pseudometric version of a classical result of fractal geometry . Lemma B. Covering and packing in pseudometric spaces. Let X be a pseudometric space and x . . . xPX a maximal packing of X for pseudometric . Then we have N X P X. Proof. Let us fix and let x . . . xP X be centers of a maximal packing of X with closed balls. Let us assume that X iP X B xi so that we can take some x belonging to the above nonempty set. Now let us fix i . . . P X and w B xi. By the triangle inequality and the definition of w and x we have x xi z x w w xi z . Therefore we have x w and hence B xi B x so that we construct a bigger packing by adding x to x . . . xP X which is absurd. Therefore we have X S iP X B xi hence the result. The next lemma asserts that E is increasing with respect to the inclusion of sets if and only if . This is the reason why we require in Theorem .. Lemma B.. Let X be a finite pseudometric space and . Then we have E X P X Proof. In all the following we fix and . We also denote P P X. Without loss of generality we can assume P . We fix T an MST of X represented by a set of edges denoted x y with x y X note that we identify x y and y x. It is a classical result that there are X edges. For an edge e of the form a b we denote its length by e a b. For a b X with a b we denote by a b the shortest path between a and b. More precisely we represent it as a list of edges denoted a a a aK b for some K. When the context is clear we identify a b to the set of its edges a b. Let us introduce x . . . xP a maximal packing of X by closed. For every i . . . P as T is connected there exists yi X such that yi B xi and yi is the only point in the path xi yi that does not belong to the ball B xi. For each i we denote ei the only edge in xi yi to which yi belongs i.e. ei is of the form zi yi with zi B xi. By construction those edges ei are the only ones that can be shared by several paths xi yi. Let us introduce the following set of indices I i . . . P j i ei xj yj K . . . P I. Let us consider i K. Let us assume that we have j j . . . P such that ei xj yj and ei xj yj . If we denote ei as zi yi we have that zi B xi by definition of yi . Therefore by definition of yj we have zi yj because B xi B xj . We have similarly zi yj and thus yj yj . By definition of yj and yj we also have yi B xj B xj which is absurd by definition of packing. We conclude the following k K !j i ei xj yj . For k K we denote the corresponding j by k. By definition of K it is clear that k K. Moreover as yi zi B xi this implies that i i. Therefore we have constructed an involution K K such that k K k k. This implies that the cardinality of K is even and that we can write K K K with K K K K. The outcome of this construction is that we now have disjoint paths given by the xi yiiI and the xk xkkK . Therefore we get the following lower bound on EX. EX X iI X exiyi e X kK X exkxk e . As we have that EX X iI X exiyi e X kK X exkxk e . By the triangle inequality and by definition of packing we have EX X iI X kK I K P X which concludes the proof. Theorem .. Let be a pseudometric on Rd . Supposes that Assumption holds and that is q L Lipschitz for q . Then for all with probability at least we have sup iT GSwi B vuut log E log L n B n B n B r IS WT log n . Proof. For better clarity we assume T . Let us fix some and . By Theorem B. we have with probability at least sup iT Rwi RbSwi L B r log N WT n B r IS WT log n . We now bound the covering number appearing in the above equation. By Lemma B. we have E WT P WT . Moreover by Lemma B. we have E WT N WT . We now combine this with our generalization bound by choosing the value B L n and we get that with probability at least we have sup iT Rwi RbSwi B n B vuut logE WT log L n B n B r IS WT log n leading to the desired result. B. Proof of the magnitudebased generalization bounds Lemma B.. Let W Rd be a finite set and . . . n and a pseudometric such that W admits a positive magnitude according to Definition B. for every . We assume that is L q Lipschitz continuous with q . Then for any we have E exp n sup wW Xn i iw zi e B n PMag LW. where PMag is the positive magnitude see Appendix B. Proof. We first remark that by Holders inequality and the L q Lipschitz condition we have w w W Sw w n q LSw LSw q Lw w . Let us fix some . As W admits a positive magnitude we can introduce a canonical weighting W R. By definition of a weighting we have a W X bW e abb . Moreover for any n we introduce a argmaxaW Xn i ia zi. With those notations we can compute X bW e abb X bW e L S abb X bW exp Ln Xn i a zi b zi b X bW exp Ln Xn i ia zi b zi b exp Ln Xn i ia zi X bW exp Ln Xn i ib zi b. Therefore by dividing by the first term on the righthand side and using the independence of the i we deduce that E exp Ln sup wW Xn i iw zi E X bW exp Ln Xn i ib zi b X bW Yn i E h e Ln ibzi i b. By Hoeffdings lemma we have E exp Ln sup wW Xn i iw zi e B nL X bW b e B nL PMag W. The result follows by the change of variable L. Theorem .. Let be a pseudometric such that W admits a positive magnitude according to Definition B. for every . We assume that is q L Lipschitz continuous with q . Then for any s we have with probability at least that sup iT GSwi s log PMag LsWT s B n B r IS WT log n . Proof. The beginning of the proof is completely similar to the proof of B. up to Equation . More precisely we have that with probability at least sup iT Rwi RbSwi Rad WT S B r IS WT log n . By Jensens inequality we have for all Rad WT S log E exp n sup wWT Xn i iw zi . Therefore we can apply Lemma B. to write that for all s Rad WT S s B n s log PMag LsWT . We deduce that for all s we have with probability at least that sup iT Rwi RbSwi s B n s log PMag LsWT r IS WT log n . Remark B. Link between magnitude and positive magnitude. Let W RM be a finite set for some M of cardinality N and a metric on W. If we denote the similarity matrix for a given value of s by Msa b e ab then it is clear that Ms s IN . Moreover by continuity of the inverse this implies that the weighting associated to s i.e. s W R satisfy a W sa s . From this we first deduce that for s we have Mag sW N. Moreover by continuity of the inverse this means that up to a certain s the weighting saaW only has positive elements. Therefore this implies that for s big enough one has Mag sW PMag sW. Thanks to our definitions for positive magnitude in pseudometric spaces given in Appendix B. this observation extends to the pseudometric case. Remark B. Extension to infinite sets. There exist extensions of the definition of magnitude beyond finite sets . More specifically weightings are then represented by measures on the set. It is clear from the above proofs that we can extend the positive magnitude in this setting and that the proof would follow similar lines. Therefore our theory provides upper bounds of Rademacher complexity in terms of positive magnitude in more general cases than the one we use in this work. Table Architecture details for the vision transformers taken from . WS refers to Window Size. Model Dataset Depth Patch Size Token Dim Heads MLPratio WS Params ViT CIFAR ViT CIFAR Swin CIFAR Swin CIFAR CaiT CIFAR CaiT CIFAR C Additional Experimental Details In this section we give additional details regarding the models datasets and hyperparameters used in our experiments. C. Experimental setting C.. Vision Transformers Architecture and implementation details The design of the ViT has been modified to accommodate for the small datasets as per . Our implementation is based on the which is based on the timm library with the architecture parameters presented in Table . The implementation of Swin is based on the SwinTransformer libarary and the implementation of CaiT is predominantly based on the timm library with some modifications. The full version can be found in the supplementary code. Instead of training from scratch which is extremely timeconsuming we used the pretrained weights available from the GitHub repository of the paper we further fintetuned them for epochs on the dataset CIFAR or CIFAR to achieve the optimum performance reported in the paper . Then we verified that the finetuned weights achieved training performance and then they were the starting point of our computational framework. We ran the transformer experiments on NVIDIA Ti GPUs and the graph experiments on Intel Xeon Silver CPUs. C.. GNN Architecture and implementation details We will briefly talk about the details of GraphSage and GatedGCN prior works we use in our experiments. GraphSage is an improvement over the GCN Graph ConvNets model and it incorporates each nodes own features from the previous layer in an explicit way by the update equation h l i ReLUU lConcath l i MeanjNih l j where Ni is the neighbourhood of node i h l i is the feature vector and U l R dd . We use the graphpooling version of GraphSage with the following update equation h l i ReLUU lConcath l i MaxjNiReLUV lh l j where V l R dd . GatedGCN Gated Graph ConvNet uses the following update equation h l i h l i ReLUBNU lh l i X jNi e l ij V lh l i where U l V l R dd is the Hadamard product and the edge gates e l ij have the following definitions e l ij e l ij P j Ni e l ij e l ij e l ij ReLUBNA lh l i B lh l i C l e l ij where is the sigmid funciton is a small constant for numerical stability Al Bl Cl R dd and BN stands for Batch Normalization. We used the code provided by which relies on the dgl library implementation of GraphSage and GatedGCN. We trained GraphSage and GatedGCN until training accuracy following the setup in . All experiments were ran on Intel Xeon Silver CPUs. Each experiment one fixed batch size and learning rate was run on a single CPU and experiments were run on the server at any given time on different CPUs. C. Hyperparameter details Hyperparameters shared among experiments.. For the Vision Transformers experiments we varied the learning rate range and batch size in the range . For the graph experiments and batch size in the range . For all experiments we used . proportion of the training data for the computation of the pseudo matrix apart from CaiT and Swin on CIFAR where we used . proportion of the training data due to memory constraints. All experiments use a grid of hyperparameters which is specified as follows. ViT on CIFAR. We selected values for the learning rate in the range and the batch size between and data proportion for the computation of the pseudodistance S of see Section . ViT on CIFAR. We selected values for the learning rate in the range and the batch size between and data proportion for the computation of the pseudodistance S of see Section . CaiT on CIFAR. We selected values for the learning rate in the range batch size between and data proportion for the computation of the pseudodistance S of see Section . CaiT on CIFAR. We selected values for the learning rate in the range batch size between and data proportion for the computation of the pseudodistance S of see Section . Swin on CIFAR. We selected values for the learning rate in the range batch size between and data proportion for the computation of the pseudodistance S of see Section . Swin on CIFAR. We selected values for the learning rate in the range batch size between and data proportion for the computation of the pseudodistance S of see Section . GatedGCN. We selected values for the learning rate in the range the batch size between and data proportion for the computation of the pseudodistance S of see Section . We note that for due to time constraints the experiments with batch sizes of and for the Euclidean metric were not complete. GraphSage. We selected values for the learning rate in the range the batch size between and data proportion for the computation of the pseudodistance S of see Section . ViT on CIFAR Adam. We selected values for the learning rate in the range and the batch size between and data proportion for the computation of the pseudodistance S of see Section . ViT on CIFAR SGD. We selected values for the learning rate in the range and the batch size between and data proportion for the computation of the pseudodistance S of see Section . ViT on CIFAR RMSprop. We selected values for the learning rate in the range and the batch size between and data proportion for the computation of the pseudodistance S of see Section . D Additional experimental results In this section we present additional empirical results in addition to what was already presented in the main part of this document. We divide this section into three parts. First we quickly explore in Appendix D. the consequence of our choice of estimation technique of the worstcase generalization error. In Appendix D. we report additional experiments based on vision transformers and in Appendix D. we include additional illustration of the GNN experiments. D. About the final accuracy gap and the worst accuracy gap Our main theoretical results presented in Section apply to the worstcase generalization error over the trajectory i.e. on the quantity supkT Rwk RbSwk . However computing this quantity over the whole trajectory may be extremely expensive as it requires evaluating the model on the whole dataset at each iteration this is a similar problem to the one encountered for the computation of the datadependent distance matrices discussed in Section . Previous studies on worstcase TDAinspired generalization bounds circumvented this issue by reporting the final accuracy gap as the generalization error in their experiments as it is the case in our work most existing experiments consist of classification tasks. In our work we argue that the true worstcase generalization error may however have a different behavior than the final accuracy gap. In order to estimate this quantity in a computationally friendly way we used the following procedure we periodically estimated the test accuracy during the training computed its minimum value acctestworst and substracted it from the final train accuracy acctrainfinal to obtain the generalization gap GbS reported in our main experiments i.e. GbS acctrainfinal acctestworst. Note that in addition to being a good proxy to the true error appearing in our theory the above quantity could be of independent experimental interest. In order to assess that our main conclusions remain valid if the final accuracy gap is used instead of GbS we present here a few additional experiments using the final accuracy gap as a generalization measure it is denoted Accuracy gap in the figures. In the case of a ViT on CIFAR this is shown in Figure and Figure . We observe that our proposed topological complexities also correlate very well with the final accuracy gap and outperform the previously proposed PH dimensions . In addition to these findings we make two additional new observations. First the Ph dim while outperformed by our proposed metric has better granulated Kendalls coefficients when compared to the final accuracy gap than the worst generalization error goes from . to .. This may explain why we observed poor performance of PHdim in Figure a. Second we observe that the correlation seems to be slightly less good with the final accuracy gap especially for high learning rates which seems to be similar behavior to what was reported in . Accuracy gap E alpha Batch size Learning rate a E Accuracy gap positive magnitude Batch size Learning rate b PMag n Accuracy gap . . . . . . ph dim Batch size Learning rate c dimPH Figure ViT on CIFAR with Spseudometric using the final accuracy gap as a generalization measure. Accuracy gap E alpha Batch size Learning rate a E Accuracy gap positive magnitude Batch size Learning rate b PMag n Accuracy gap . . . . . . ph dim Batch size Learning rate c dimPH Figure ViT on CIFAR with pseudometric using the final accuracy gap as a generalization measure. D. Vision Transformers additional experiments We compare the performance of the different metrics by using the granulated Kendalls coefficients introduced in . The experiments presented here use different Vision Transformers ViT CaiT Swin on CIFAR and CIFAR. As a baseline we use the dimPH introduced in and the datadependent dimension with the pseudometric dimPH from . Here we present the full results on each dataset and model. They can be found in Table for CaiT and CIFAR for Swin and CIFAR for ViT and CIFAR and for CaiT and CIFAR. The plots from each experiment for every computed quantity can be found in the remaining quantities for ViT and CIFAR. Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap magnitude Batch size Generalization gap . . . . . magnitude small Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . . ph dim Batch size Learning rate Learning rate Figure ViT on CIFAR with S D. Graph Neural Networks Additional Experiments In Table we already presented the correlation coefficients for all quantities for the GNN models considered in our study GraphSage GatedGCN we have selected the models which achieve training accuracy and GraphMNIST. We can observe a nice correlation outperforing dimPH in most experiments. As it was observed for the transformerbased experiments the correlation seems to be better for the datadependentmetrics. This is an important fact as no sparse random projection was used to compute the Euclidean distance matrices in the GNN experiments it was not necessary as these models have less parameters than the tramsformers considered above. This shows that the fact the datadependent pseudometrics outperform the Euclidean distance also happens in the absence of these projections. It also shows that all quantities seem to yield better correlations in the absence of random projections at least in the GNN expsriments. The corresponding plots for GatedGCN can be seen in Figure with the pseudometric Figure for the Euclidean and for . The plots for GraphSage are reported in Figure Figure and Figure . We can observe a strong correlation on these figures outperforing dimPH in most cases. As it was observed for the transformerbased experiments the correlation seems to be better for the Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . . . magnitude small Batch size Learning rate Figure ViT on CIFAR with Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . magnitude small Batch size Learning rate Figure ViT on CIFAR with pseudometric datadependentmetrics. This is an important fact as no sparse random projection was used to compute the Euclidean distance matrices in the GNN experiments . This shows that datadependent pseudometrics outperform the Euclidean distance also in the absence of these projections. In addition all quantities seem to yield better correlations in the absence of random projections at least in the GNN expsriments. Interestingly a few failure cases can be seen on these plots. Indeed Mag. and PMag. seem to be almost constant and near . This indicates that the scale choice s . was not suited A sparse random projection was not necessary as these models have less parameters than the tramsformers considered above Table Correlation coefficients for all quantities for ViT model and CIFAR dataset. The corresponding plots are presented in Figures Figure and Figure . Metric Complexity lr bs S E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag. . . . . Mag n . . . . PMag. . . . . PMag n . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . . magnitude small Batch size Learning rate Figure ViT on CIFAR with S for these experiments this behavior was already reflected in Table through very low Kendalls coefficients indicating the absence of meaningful correlation. However Mag n and PMag n provide significantly better correlation which supports our main claims as s n has been argued in Section . to be a particulary relevant choice of scale factor. Note finally that the PHdim plots for the pseudometric failed to produce numbers in these graphs experiments this is why they are either missing or look irrelevant. As before we gave away this fact in Table by imposing our granulated Kendalls coefficients implementation to return zeros Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . . . magnitude small Batch size Learning rate Figure ViT on CIFAR with Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . magnitude small Batch size Learning rate Figure ViT on CIFAR with pseudometric in the absence of correlation hence the small numbers observed in this case. That being said this behavior should not be seen as an issue. Indeed PHdim with pseudometric consists in theory in estimating the dimension of a subset of a discrete hypercube which is always . The reason we still reported PHdim for this pseudometric is for consistence and to test the implementation of in this nonstandard setting it is however not theoretically grounded. Table Correlation coefficients for all quantities for CaiT model and CIFAR dataset. The corresponding plots can be seen in Figures and . Metric Complexity lr bs S E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . Generalization gap E alpha Batch size Generalization gap . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Learning rate Generalization gap . . . . . magnitude small Batch size Learning rate Generalization gap positive magnitude Batch size Generalization gap positive magnitude small Batch size Figure CaiT on CIFAR with Spseudometric. Generalization gap E alpha Batch size Generalization gap . . . . . . ph dim Batch size Learning rate Generalization gap positive magnitude small Batch size Generalization gap . . . . . magnitude small Batch size Learning rate Generalization gap positive magnitude Batch size Generalization gap magnitude Batch size Figure CaiT on CIFAR with distance. Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . magnitude small Batch size Learning rate Figure CaiT on CIFAR with pseudometric. Table Correlation coefficients for all quantities for CaiT model and CIFAR dataset. The corresponding plots can be seen in and Metric Complexity lr bs S E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . loss . . . . Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . . . . magnitude small Batch size Learning rate Figure CaiT on CIFAR with Spseudometric. Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . . magnitude small Batch size Learning rate Figure CaiT on CIFAR with . Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . magnitude small Batch size Learning rate Figure CaiT on CIFAR with pseudometric. Table Correlation coefficients for all quantities for Swin model and CIFAR. The corresponding plots are in Figure and . Metric Complexity lr bs S E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . . . magnitude small Batch size Learning rate Figure Swin on CIFAR with Spseudometric. Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . magnitude small Batch size Learning rate Figure Swin on CIFAR with . Generalization gap E alpha Batch size Generalization gap positive magnitude Batch size Generalization gap . . . . . . ph dim Batch size Learning rate Generalization gap magnitude Batch size Generalization gap positive magnitude small Batch size Generalization gap . . . . magnitude small Batch size Learning rate Figure Swin on CIFAR with pseudometric. Table Correlation coefficients for all quantities for Swin model and CIFAR. See Figures and for the corresponding plots. Metric Complexity lr bs S E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . E . . . . Mag n . . . . Mag. . . . . PMag n . . . . PMag. . . . . dimPH . . . . . . . . . Generalization gap E alpha Batch size . . . . . Generalization gap positive magnitude Batch size . . . . . Generalization gap . . . . . . . . . ph dim Batch size Learning rate . . . . . Generalization gap magnitude Batch size . . . . . Generalization gap positive magnitude small Batch size . . . . . Generalization gap . . . . . . . magnitude small Batch size Learning rate Figure Swin on CIFAR with Spseudometric. . . . . . Generalization gap E alpha Batch size . . . . . Generalization gap positive magnitude Batch size . . . . . Generalization gap . . . . . . . . ph dim Batch size Learning rate . . . . . Generalization gap magnitude Batch size . . . . . Generalization gap positive magnitude small Batch size . . . . . Generalization gap . . . . magnitude small Batch size Learning rate Figure Swin on CIFAR with . . . . . . Generalization gap E alpha Batch size . . . . . Generalization gap positive magnitude Batch size . . . . . Generalization gap . . . . ph dim Batch size Learning rate . . . . . Generalization gap magnitude Batch size . . . . . Generalization gap positive magnitude small Batch size . . . . . Generalization gap . . . . . . . . magnitude small Batch size Learning rate Figure Swin on CIFAR with . . . . . Generalization gap E alpha Batch size . . . . Generalization gap positive magnitude Batch size . . . . Generalization gap ph dim Batch size Learning rate . . . . Generalization gap magnitude Batch size . . . . Generalization gap positive magnitude small Batch size . . . . Generalization gap . . . . . magnitude small Batch size Learning rate Figure GraphSage on MNIST with Spseudometric. . . . . Generalization gap E alpha Batch size . . . . Generalization gap positive magnitude Batch size . . . . Generalization gap . . . . . . ph dim Batch size Learning rate . . . . Generalization gap magnitude Batch size . . . . Generalization gap positive magnitude small Batch size . . . . Generalization gap . . . . . magnitude small Batch size Learning rate Figure GraphSage on MNIST with . . . . . Generalization gap magnitude small Batch size Learning rate . . . . Generalization gap E alpha Batch size . . . . Generalization gap positive magnitude Batch size . . . Generalization gap ph dim Batch size . . . . . . . Learning rate . . . . Generalization gap magnitude Batch size . . . . Generalization gap positive magnitude small Batch size Figure GraphSage on MNIST with . . . . Generalization gap E alpha Batch size . . . Generalization gap positive magnitude Batch size . . . Generalization gap ph dim Batch size Learning rate . . . Generalization gap magnitude Batch size . . . Generalization gap positive magnitude small Batch size . . . Generalization gap . . . . . . . . magnitude small Batch size Learning rate Figure GatedGCN on MNIST with Spseudometric. . . . . Generalization gap E alpha Batch size . . . . Generalization gap positive magnitude Batch size . . . . Generalization gap . . . . . . ph dim Batch size Learning rate . . . . Generalization gap magnitude Batch size . . . . Generalization gap positive magnitude small Batch size . . . . Generalization gap . . . . . magnitude small Batch size Learning rate Figure GatedGCN on MNIST with . . . . Generalization gap E alpha Batch size . . . Generalization gap positive magnitude Batch size . . . Generalization gap magnitude Batch size Learning rate . . . Generalization gap . . . . positive magnitude small Batch size . . . Generalization gap . . . . magnitude small Batch size Learning rate Figure GatedGCN on MNIST with .

write a youtube script

I made a D printed conveyor belt all of the links are actually the same component but they have overlapping segments the links could be printed without support and I needed of them for the entire conveyor belt theyre held together with these mm toy axles this was super fun to play with when its all connected I built a little piece of crap table to test out my conveyor belt and I attached a lever to a sprocket so I could drive the entire mechanism this is just a proof of concept so now I know how to build an automatic conveyor belt for my pin Setter were getting there

write a youtube script

have you ever wondered what happens to the bowling ball after it knocks down some pins after the bowling ball collides with some pins theyre knocked onto a conveyor belt with the ball but now youre left with this problem how do you separate the ball from the bowling pins on the conveyor belt theres a wall thats high enough to let the pins through but low enough so the ball can slide through thats what Ive modeled here so theres this wall where the pins can slide underneath the divider until this motor lets this B through and then its returned to the user

write a youtube script

to move this lawnmower I need to apply a force to it but its also important to note that that force its also applied back to myself I can show this by trying to push the lumo without bracing myself like standing on one foot you may know this as Newtons third law of motion where each action has an equal and opposite reaction by applying Newtons second law of motion where force equals mass times acceleration we can work out the relative acceleration of each object if we ignore all the frictional forces like the frictional force between the wheels of the Norma and the ground if the mass of the lawnmower was the same as the mass of myself we both accelerate at the same rate in opposite directions as soon as my hands leave the handlebar of the lawnmower the force between us ends and so does the acceleration in either direction but well both have an enveloping in the direction and we can actually work out the relative velocity between us by using the law of conservation of momentum we calculate momentum by multiplying an objects mass by its velocity and by applying conservation of momentum to this example we can say that the end momentum of the loma is equal to the end momentum of myself now I know the mass of myself so I can actually calculate the velocity that I will push back by depending on whether I use a really heavy object like this Norma traveling very slowly or a very light object traveling very quickly like shooting a bullet from a gun this law of conservation of momentum can also be applied to rotating objects such as this wheel now this wheel has a very low mass in comparison to myself so I would need to rotate it to a very high velocity in order to give it enough momentum in this case angular momentum because its rotating in order to give myself enough momentum to rotate in the opposite direction however it should be enough to at least make me lose balance okay that example was quite terrible so I think I need to go and build a test rig to try to explain it a bit better now reaction wheels work in exactly the same way as that bicycle wheel did however as I mentioned because the mass of the bicycle wheel was very very low and the mass of my body weight is much higher I would need to accelerate the world very very quickly in order to move my body weight in the opposite direction so Ive made a small test truck here to try and give a better example of how reaction wheels work here you can see a heavy aluminium disc which is mounted to the shaft of an electric motor which is mounted to the backside of this blue plate now this blue plate and the whole assembly here is mounted to a pivot which its balanced about to simulate a satellite in space or a drone in midair it doesnt want to swing back to be self level in its current state now if I apply voltage to this electric motor it all causes will to spin in whatever direction the motor was wired up for and this will increase the angular momentum of this wheel and because angular momentum has to be conserved it will also apply angular momentum to this blue plate so for example if the disc were to spin clockwise the blue disc would spin in the opposite direction counterclockwise now I havent just wired this straight up to a battery with a switch instead Ive put some gyros and accelerometers on the backside of this plane which should cause the plate to try and self level itself with this white pad being the top of the plate so if I were to put it out an angle like this it should try and spin the wheel in a certain direction to bring this foam pad up to the top again so lets plug it in and see if this test ring works lets plugged in and I put a certain angle so it should be able to correct itself and Music there we go its brought the white pad up to level now the wheel is become stationary now because it doesnt need to impart any more angular momentum on to this blue pad because its where it wants to be now whats also interesting about this test rig is that if I apply an external force to this blue plate the gyros and accelerometers will detect that and therefore spin the wheel in the direction that I post trying to absorb all of the angular momentum lets give this a go you see that it spins the motor in a certain direction to try and absorb the rotation of this blue plate now whats important to know about reaction wheels that they can become whats called saturated which basically the motor wont spin any quicker and therefore it cant accelerate the wheel any more than its current rpm and this is a big issue with satellites and I think it might be an issue with this drone if I bring it back to level the motor should stop Music so now have two wheels this was the reaction wheel which was mounted onto the experimental test rig and this wheel will be mounted onto the drone now this first wheel weighs grams and this one weighs grams so its less than half the weight of this first wheel now going back to the equation for linear momentum which I mentioned at the start of this video where mass times velocity equals momentum it would suggest that this will would have a far greater momentum than this will if theyre both rotated at the same velocity but I think we need to take a look at the equation for angular momentum instead this is the equation for angular momentum instead of being mass times velocity its in fact moment of inertia times angular velocity which is the rate that the spins at but the important part is the moment of inertia if we apply conservation of momentum to this we can say that the inertia of the drone times the rotational rate of the drone is equal to the inertia of the wheel times the rotational rate of the wheel now basically what we want to do is maximize the rotational rate of the drone so that we have full attitude control of the drone and the way that we can do this is either decrease the inertia of the drone or increase the angular momentum of the wheel by either increase in the inertia of the wheel or the angular velocity now we cant increase the angular velocity anymore than the electric motor is rated for because it will only spin up to a maximum rpm so basically what were working with is the inertia of the wheel and the drone so do you have this as high as possible we need to have a really high inertia wheel and a really low inertia drone so lets talk about the inertia of the wheel to calculate moment of inertia of either a disk or a ring these are the following equations as you can see for a solid disk like this first reaction wheel the equation is M R squared which is times its mass times its radius squared whereas for a ring which is similar to this obviously in the spokes middle we can see that the moment of inertia is mass times its radius squared so for these two objects if they were the same mass and the same radius then the ring would have double the amount of moment of inertia as a disc so basically what this means is that the radius has a larger effect on the moment of inertia than the mass does because its squared so if you were to double the radius its equivalent to x in it by four whereas if you have to double the mass its just the same as x in it by two so this solid disk although it weighs well almost where weighs over double the amount of this disk they actually have equal amounts of inertia now because this is slightly more complex shape than just a simple ring what I did is I designed these two disks in my camp program and if I look at the calculations in the cam program I can prove that they both have the same amount of inertia so what does this mean in terms of this project well here I have two reaction wheels both have the same moment of inertia and when attached to the same electric motor will have the same angular velocity this means theyll both be able to impart these same angular momentum onto the drone but this will weighs half as much as this will which will not only reduce the overall weight of the drone because this obviously weighs a lot less and will hopefully increase the flight time of the drone it also allows the drone to run much smaller motors in order to lift this wheel rather than this wheel which will therefore decrease the inertia of the drone which also assists to increase the angular velocity of the drone so not only have we kept the inertia the same weve managed to decrease the inertia of the drone which is as I mentioned before is very important for increase in the angular velocity and therefore keeping the drone under control so I think its time to get building with this drone before I complete the build of this reaction wheel drone I want to run through a few of the components that are chosen to use to control the reaction wheels here you can see two brushed motors mounted perpendicular to each other degrees and on the output shafts of these motors will be mounted the reaction wheels now these will be able to control the roll and pitch axis and I wont be having a will for the yaw axis because Im going to do that through aerodynamic control of the propellers now Ive chosen to go for brushed motors over more modern brushless motors because its very difficult to control a brushless motor at very slow revolutions per minute they either need some kind of encoder or Hall effect sensor to know the exact position of the rotor its always a very difficult to change direction of a brushless motor smoothly although drone speed controllers can change the direction of a brushless motor it usually does a very hard braking maneuver and a very hard acceleration maneuver when changing direction and I reckon this will cause issues with the reaction wheel when the drone is in midair with a brush motor you can simply change the direction of the voltage and current and it will just decelerate the motor and accelerate it smoothly in the other direction the other reason is that these motors have a gearbox which means the reaction will can receive a lot of torque from the motor without spinning too quickly and causing any kind of gyroscopic issues to drive these brushed motors I have a motor driver down in the bottom of here the motor driver can drive both motors which is quite handy and its controlled via this Arduino Nano now in order for this Arduino Nano to know which way to spin the motors to control the drone I have a flight controller from a regular drone which Ive programmed to output servo controls which can be read by the Arduino Nano now this fly controller has gyros and accelerometers in it so that youll be able to detect any movement of the drone and I can also control the movement of the drone by using a receiver and a radio control so if I plug in a battery we should be able to see everything light up and it should arm in a second they were goat so that means the motors are all ready to spin and if I tilt the drone it should spin up one of the motives if I tilt it in this direction it should spin up this motor so yes and that motor spins and if I tilt it this way the other motor spins now I can also put my radio control here and control the direction using my sticks which means I can control the drone in midair so now I know that the gyros and accelerometers are working with the Arduino to control the reaction wheel motors I think its time to sword up the motors to lift this thing into the air now you might be wondering Tom why you trying to control drone using reaction wheels drones fly perfectly fine as they are honestly I have no idea well I find the physics behind the reaction wheel quite interesting with the whole conservation of momentum thing and if this drone doesnt fly its only going to be a big waste of time and money and even if this drone doesnt fly at least Ive learned something and hopefully taught you something I still really hope this thing does fly though right so I know that this still looks more like a Christmas decoration then it does a drone so Ive hung it from the ceiling to try and tune some of the parameters inside of the flight controller while hanging from the ceiling I basically free up one of the axes in this case is the roll axis reaction wheel and I can actually simulate what would happen if I apply an external force to it where the reaction will responds not only that but I can also take control with my radio control and make it spin see I can be pretty accurate s degree increments so pretty good control over that axis so the erection will drone is now complete and hopefully it looks slightly more like a drawing than it used to with the propellers on the top and the bottom in terms of the controls this drone the reaction modes will control the pitch axis and the roll axis to control the gyro axis Ill be using differential thrusts of these propellers the propellers are counter rotating to prevent any torque rotation during a hover and then to control the your axis one will speed up and one will slow down and then the differential in torque will hopefully twist the drone in midair so basically this whole thing is controlled by changing the torque of an object like a propeller or a reaction wheel so Im going to charge on the battery and then I think were ready for a test flight Ill put the dryer on the paving slab to not only keep the propeller from hitting the grass but also to making sure that it was perfectly level if it wasnt quite level with the reaction was to try to make it level and may become saturated even before it takes off Music lets try that again and again ouch time for some repairs so Ive made a few modifications aside from just swapping out the old broken propellers for some reason the pitch reaction wheel has far less torque than the roll reaction wheel which is really weird because I bought them from the same supplier and theyre the same rated motors both rated for the same RPM and also the same gear ratio but for some reason this one has a lot more resistance or torque than this one does Im not sure whether it has a different gear ratio or different magnets but what Ive done is adjusted the settings inside the flight controller to basically rotate this wheel faster than it was doing before Ive also changed the battery for a higher voltage battery which will be able to give the motors more power and also a higher end rpm which should try and delay the saturation of the reaction was a little bit longer to balance out this battery Ive added some lid weights to this leg here and my plan for the next flight I havent got that many propellers left but my plan for the next flight is just to try and get it off the ground as quick as possible and then see if the reaction wheels can take control I think if I spend too much time on the ground the reaction wheels try and fight the ground to try and roll it and pitch it and it just sort of ends up saturating them before it even takes off so lets take it out for another test flight the next few attempts didnt seem to be much better but then I took a look at the slomo of this flight and realized the drone was rotating about its your axis so I increased the your PID values and tried again do your axis seems to be holding now and I think we could be close to success or maybe not but then this happened the smoke mode clearly shows the reaction wheels trying to hold the drones orientation and they arent spending anywhere near fast enough for maximum rpm saturation to be an issue as successful as this eight second flight was the following flight took a pretty hard dive resulting in a broken top motor mount so with the death of this propellerpowered reaction wheel keeps at drone looming over the horizon it was time for one last test flight so its time to wrap up this project its been quite a long project and also quite a long video it finally ended in destruction of this top motor mount and that last crash also managed to take both reaction wheels off of the motor shafts theres a few cracks in these legs and I think the frame actually cracked around side here as well so it definitely needs some repairs done if I want to try it again as I mentioned earlier in the video Im using brushed motors to control the reaction wheels and the way to some controlling the torque output to the wheels is by varying the current supplied to the motor now the problem with this is that the torque output is n

write a youtube script

Is plantar thermography a valid digital biomarker for characterising diabetic foot ulceration risk? Akshay Jagadeesh Chanchanok Aramrat Aqsha Nur Poppy Mallinson Sanjay Kinra Faculty of Epidemiology and Population Health London School of Hygiene Tropical Medicine London WCE HT UK Abstract Background In the absence of prospective data on diabetic foot ulcers DFU crosssectional associations with causal risk factors e.g. peripheral neuropathy and peripheral arterial disease PAD could be used to establish the validity of plantar thermography for DFU risk stratification. Methods First we investigate the associations between the intrinsic clusters of plantar thermographic images with several DFU risk factors. We do this using an unsupervised deeplearning framework. We trained a Convolutional Neural Network Auto Encoder ConvAE to extract useful representations of feetsegmented thermographic images which we then hierarchically clustered. We then studied associations between obtained thermography clusters and DFU risk factors. Second to identify those associations with predictive power we used supervised learning to train Convolutional Neural Network CNN regressionclassification models that predicted the risk factor based on the thermograph and visual input. Findings Our dataset comprised thermographs from type diabetes mellitus patients aged . . years . males. On clustering we found two overlapping clusters silhouette score . indicating weak separation. There wasstrong evidence for associations between assigned clusters and several factors related to diabetic foot ulceration such as peripheral neuropathy PAD number of diabetes complications and composite DFU risk prediction scores such as MartinsMendes PODUS and SIGN. However models predicting said risk factors had poor performances AUCs . or large mean absolute errors. Interpretation The strong associations between intrinsic thermography clusters and several DFU risk factors support the validity of using thermography for characterising DFU risk. However obtained associations did not prove to be predictive likely due to spectrum bias or because thermography and classical risk factors characterise incompletely overlapping portions of the DFU risk construct. Our findings highlight the challenges in standardising ground truths when defining novel digital biomarkers. Introduction Up to a third of everyone with diabetes will develop a foot ulcer in their lifetime. These ulcers do not heal very well with nearly of them getting infected and of those infected requiring surgical amputation. Most amputations are preventable through the earlier identification and appropriate management of diabetic foot ulcers DFU. To this end it is beneficial to reliably identify diabetes patients at high risk of developing DFU. Screening such highrisk patients will facilitate the early detection of preulcerative lesions and small DFUs and also help support risk factor modification e.g. behavioural support for daily foot selfcare and offloading and limb evaluation vascular andor nerve conduction studies. Studies have used infrared thermometer foot temperature measurements at specific foot sites to help risk stratify patients DFU risks. More recently foot imaging with thermographic cameras thermography which have a higher spatial resolution temperature estimates are provided across all points of the feet and potential for automatedremote analysis is shown as a likely digital biomarker for predicting DFU risk. Literature suggests thermographs may be able to pick up on preulcerative inflammatory changes several weeks before visible skin changes thereby allowing the identification of highrisk feet. This modality could then offer an opportunity for regular screening by healthcare providers in settings where daily selfmonitoring is unreliable. However determining the validity of this approach will ideally require prospective on ulcers. Previous randomised controlled trials RCTs and prospective studies have provided inconsistent evidence that differing foot temperature distributions among diabetes participants are associated with unique DFU risks. A metaanalysis of five RCTs evaluating interventions of thermography coupled with personalised foot care and offloading for participants with hotspots contralateral feet temperature differences . C found some evidence for reduced DFU risk at any foot site relative risk . CI . .. However several studies mostly included diabetes populations with previous ulcer histories. Further only one trial considered the outcome of foot ulcer development at or adjacent to a previously identified hotspot and they found no association. We also argue that an RCT design may not be suitable for ascertaining the predictive capability of plantar thermographs in modelling DFU risk the observed reductions in DFU risk could have occurred merely due to activity reduction even if it were done unrelated to any preceding thermographic changes. Only two previous prospective studies found thermographic findings could predict DFU with a lead time of weeks. However one of them demonstrated an impractically high false positive rate of and the other included only a very small sample N patients of highrisk diabetes patients with histories of previously healed ulcers. Without prospective data it is not possible to directly validate the use of thermography for predicting absolute DFU risk. Indirectly however we can use crosssectional data to investigate whether plantar thermography is associated with classical DFU risk factors. Few previous studies have investigated such predictive associations. Characterising risk factor severity especially key ones like neuropathy and peripheral arterial disease PAD can then be used for risk stratification. Such prediction models would likely have additional clinical benefit accurate classification into neuropathic andor ischaemic foot will help tailor limb evaluation. Robust prediction model performances would then support the validity of using thermography as a digital biomarker for predicting DFU risk in diabetes populations. In this study we aimed to study the validity of plantar thermography as a digital biomarker to identify patients at high risk of diabetic foot ulceration DFU. We limited ourselves to the plantar aspect as the literature shows most DFUs develop on the weightbearing portions of the plantar aspect of the foot. First we investigate the associations between the intrinsic clusters of plantar thermographic images with DFU risk factors. We do this using an unsupervised deeplearning framework. We trained a Convolutional Neural Network Auto Encoder ConvAE model to extract useful representations of feetsegmented thermographic images. We then applied hierarchical clustering to this dataset of thermography representations and explored associations between thermography clusters and DFU risk factors. Second to identify those associations with predictive power we trained supervised learningbased Convolutional Neural Network CNN regressionclassification models that predicted the risk factor based on the thermograph and visual input. Methods Study population The details of the APCAPS population have been described elsewhere. Briefly it is a prospective intergenerational cohort based in the villages of Ranga Reddy district in the South Indian state of Telangana. For this analysis we used a subset of diabetes patients with available thermography images from the latest followup in . Image acquisition and preprocessing Thermal and corresponding registered visual images of the foot were acquired using the CAT S smartphone using a noninsulating cloth to cover nonfoot portions according to the APCAPs data acquisition protocol. The regions of interest demarcating the two feet in the thermal image were segmented using a CNN with UNet architecture models previously shown to have good segmentation performance in biomedical image processing. Two authors CA and AN created a ground truth dataset using ITKSnap version .. by independently drawing foot contours on a random sample of visual images. We created a consensus foot segmentation mask from these two masks using the STAPLE Simultaneous Truth and Performance Level Estimation algorithm. Then we trained the UNet model to segment bilateral feet from visual images achieving high performance intersection over union IoUscore of .. This trained segmentation model was used to obtain binary mask predictions from all visual images and by applying these mask predictions to the corresponding registered thermal images we extracted the thermal images of the feet for downstream analysis. Similar to previous studies our segmentation model approach used registered visual images instead of direct segmentation on thermal images as the morphological boundaries of the feet are not well demarcated in thermal compared to visual images. Extracting image representations We built a custom ConvAE model consisting of an encoder with three layers of D CNN layers used to extract thermographic image representations that are dimensionally smaller than the original segmented thermal image. The decoder consisted of transposed layers trained to reconstruct output similar to the original data using representations generated from the encoder. The ConvAE model demonstrated good reconstruction hold out set mean squared error . mean absolute error .. We then used this trained ConvAE model to extract representations useful features from all footsegmented thermographs. Hierarchical clustering We used the Ward method of hierarchical clustering to generate clusters within thermographic representations. We identified the optimal number of clusters for the dataset using the elbow method and visualised clustering results using tSNE plots. We report silhouette scores to quantify the quality of clustering. We studied associations between DFU risk factors and thermographic cluster assignment using appropriate statistical tests. Prediction models For DFU risk factors that showed strong associations with intrinsic thermography clusters we supervisedlearningbased trained several pretrained and custom CNN classificationregression models to predict those risk factors based on thermography with visual image supplementation. Theoretically it was thought that visual images would provide information regarding callosities macerations between the toes and foot deformities including hammer or claw toes that may aid in characterising DFU risk. As a sensitivity analysis we trained models to predict the thermography cluster based on thermography input. Deep UNet ConvAE and CNN prediction models were all trained using the Adam Optimiser with optimal learning rates identified through learning rate schedulers and stopped training when validation set performance metrics failed to show improvement in over epochs. Results Associations between thermographic images and DFU risk factors Our study included participants aged . . years . males with type diabetes mellitus. Participants had diabetes for years with the majority never reporting having experienced an ulcer or amputation .. ConvAE obtained latent space representation for a sample thermograph shown in Figure . Hierarchical clustering was performed to visualise the intrinsic clusters within thermographs Figure . In Panel A the dendrogram shows two primary clusters with several subclusters demonstrating the distinct clusters of the ConvAEgenerated representations of plantar thermographic images. In Panel B the elbow point indicates the optimal number of clusters to be two. At two clusters the explained variance of the clustering begins to level off and the marginal gain of adding additional clusters decreases significantly. In Panel C we visualise through tSNE the distribution of the two distinguishable yet overlapping clusters of thermographic representations. In Panel D we show that ConvAE has successfully captured hidden patterns of overlapping thermographic phenotypes finding visually discernible groups. However the silhouette score of . suggests that the obtained clusters are weakly separated indicating that while some distinct patterns are captured there is still considerable statistical similarity between the clusters. Figure Visualisation of the latent space representations for a sample feetsegmented thermography image. Conceptually the lowdimensional x x images shown can be thought of as each viewing the original thermographic highdimensional x x image through a different series of image filters convolutional operations. The ConvAE model has optimised these filters such that the original image can be successfully reconstructed by upsampling all the representations. These representations contain statistically useful thermographic image features for downstream analysis. We found compelling evidence for a higher prevalence of DFU risk factors among thermographs assigned to cluster compared to cluster Table . This is particularly evident with measures of the two underlying casual pathologies peripheral neuropathy and peripheral arterial disease. mTCNS measurements which reliably capture the clinical progression of peripheral neuropathy were much higher in cluster compared to cluster vs p . x . Similarly TBI measures which quantify vascular insufficiency in PAD in diabetes populations were much lower in cluster . . . vs . . . in cluster p . x . There was also strong evidence for higher numbers of diabetes complications and higher measures of composite DFU risk prediction scores like PODUS including measures of neuropathy PAD and history of ulcerationamputation SIGN similar measures as PODUS but additionally including visual and physical impairment and MartinsMendes original scores with measures of neuropathy PAD number of diabetes complications and physical impairment in cluster compared to cluster . Our findings indicate a convergence between thermographic clusters and variables associated with the DFU risk construct. Further we found some evidence for higher minimum foot temperatures and lower foot temperature ranges in cluster compared to cluster . . vs . . p . and . . vs . . p . respectively. These results support the trustworthiness of the ConvAE clustering as they indicate that the clusters are indeed capturing the differences in quantitative foot temperature measurements. As a sensitivity analysis to show that given a thermograph one can reliably predict cluster assignment we used supervised learning to train CNN models and found very high performance on the holdout test set AUCs .. Figure Visualisation of hierarchical clustering. Panel A The dendrogram displays the hierarchical relationships among the thermographic images with the vertical axis representing the distance or dissimilarity between clusters. Each point on the horizontal axis represents each unique thermograph in the dataset. Each of the two primary clusters appears to be subdivided into several subclusters. Panel B Elbow method for determining the optimal number of clusters denoted by elbow point in hierarchical clustering. The plot displays the explained variance as a function of the number of clusters. Panel C The tSNE plot illustrates the two conspicuously differentiated primary clusters with the presence of overlapping data points between the clusters suggesting shared characteristics and a degree of ambiguity in their segregation. Panel D shows examples of thermographic images from each of the two identified clusters. There is a noticeable thermal asymmetry in the image belonging to cluster . B B C C A A Elbow point D Table Associations between the intrinsic clusters of thermographic images and patient characteristics Patient Characteristics Diabetes Patients N p value Cluster Likely Higher Risk N Cluster Likely lower Risk N Demographics Age years . . . Sex Male . . . Comorbidity Current or Former Smoker . . . Alcohol Use Disorders Identification Test Score . Obesity Body Mass Index kgm . . . . . . . Atherosclerosis Carotid intimamedia thickness mm . . . . . . . Hypertension . . . Diabetes History Duration years . . Poor glycaemic control . . . Peripheral Neuropathy Light touch Impaired . . . Position sense Impaired . . . Pinprick test Impaired . . . x Vibration sense Impaired . . . Temperature sense Impaired . . . Modified Toronto Clinical Neuropathy Score mTCNS . x Neuropathy as per mTCNS . . . x Peripheral Arterial Disease PAD Clinical history suggestive of PAD . . . Toe Brachial Index TBI . . . . . . . x PAD as per TBI . . . . Cardiovascular Disease . . . Stroke . . . Diabetic Renal Disease . . . Diabetic Eye Disease . . . Number of DiabetesRelated Complications . x Physical Impairment . . . Visual Impairment . . . History of Diabetic Foot Previous history of foot ulcer . . . Previous history of amputations . . . Composite DFU Risk Scores PODUS . x PODUS score . . . x SIGN indicative of highrisk foot . . . MartinsMendes original . . to . . . to . . x What DFU risk factors can you predict with a thermographic image? Performances of finetuned CNNs reported on the holdout testing set revealed poor model performances in both classification several AUCs merely above . demonstrating little benefit over random classifiers and regression large mean absolute errors that render any subsequent thresholdbased clinical classification futile tasks. Based on poor model performances in predicting measures of neuropathy regressing mTCNS or classifying based on mTCNS or PAD regressing TBI or classifying based on clinical history from thermographs with or without visual image supplementation or any of the composite risk scores there is insufficient evidence to support the use of plantar thermography for predicting DFU risk factors in populations with majority previously neverulcerated diabetes patients. Discussion There is growing interest in exploring the use of plantar thermography as a digital biomarker for various prediction tasks across the spectrum of diabetic foot disease including classifying foot temperature distributions patterns or contralateral asymmetry among diabetes patient groups detection of visibly smaller DFUs predicting the risk of DFU monitoring severity and treatment of infections in DFU and monitoring for DFU recurrence. Several others report advances in technical challenges in thermal image acquisition registration andor segmentation automated temperature difference calculations and usertesting studies all of which are pertinent to establishing thermography as a useful digital biomarker. In this study we aimed to study the validity of plantar thermography as a digital biomarker to identify patients at high risk of DFU. Comparison with previous work Our analysis of crosssectional data showed strong evidence for associations between intrinsic thermographic clusters and several DFU risk factors including casually associated pathologies peripheral neuropathy and PAD and previously validated composite DFU risk prediction scores PODUS MartinsMendes and SIGN. Our findings are in line with the few previous studies not performing hotcold stimulus exposure to limbs that have found some evidence for associations between thermographic features and measures of neuropathy or PAD. Thus our findings support the construct validity of using thermography for characterising DFU risk. Interestingly none of the associations we examined predicted patients at high risk of DFU. There could probably be two reasons for this spectrum bias owing to a community setting with a relatively lower risk DFU population with short diabetes history few PAD TBI . and a majority of neverulcerated patients. Thus the whole spectrum of plantar temperature changes that occur with progressive neuropathy andor vascular compromise may not have been represented in our population thermography and classical risk factors could be characterising incompletely overlapping portions of the DFU risk construct. Theoretically these measures may relate to DFU risks at varying time frames. Peripheral neuropathy and PAD both chronic pathologies can individually or collectively contribute to the risk of DFU resulting in absolute risk estimates of . to reported at year time frames. It is plausible that plantar thermography may measure risk across a different or shorter time frame. Very few previous observational studies have developed prediction models specifically for DFU risk using thermographic images. One recent study utilized CNNbased deep learning to classify DFU risk. However their highrisk target class included diabetes participants with active ulceration which likely drove the very high classification performance accuracy . for thermal visual . for visual only on a test set of participants. This performance is probably due to the CNN models ability to distinguish participants with and without visible ulcers a capability that has been well documented. Another study demonstrated strong classification performance accuracy . in predicting PAD risk factor status using thermographic imaging in a population including PAD patients with ABI .. However their images included lowerleg thermographs in addition to plantar thermography. One study used an unsupervised thermographic clustering k approach to first subjectively assign each cluster to mild moderate or severe DFU risk based on physician review of the thermographs without detailing the criteria. The accuracy of their models predicting cluster assignments . was comparable to our sensitivity analysis results on the same. This is expected as the ground truth labels used are known to be mathematical functions of the input thermography image. Likewise models classifying thermographs into thermal change index a mathematical function of contralateral feet asymmetry adjusted for healthy control reference ranges classes have also demonstrated excellent performances. However none have previously explored the associations between such possible ground truth labels and variables related to the clinical DFU risk construct. We described strong associations between thermography clusters and DFU risk factors although explored associations did not prove to have predictive value. Taken together these findings highlight the challenges in identifying appropriate strongly predictive clinically relevant ground truth labels while defining novel digital biomarkers. Strengths and Limitations Ours was the largest thermographic study till date to demonstrate thermographic associations with DFU risk factors in the relatively lowerrisk group of diabetes patients without previous ulcer histories. We are also among very few studies to successfully demonstrate the feasibility of diabetes foot thermographic imaging using smartphones and a lowcost setup no use of black insulated material to block out background temperatures demonstrating the potential for using thermographic imaging pipelines for diabetes screening in ruralremote lowermiddleincome country settings. Implications The findings from our study justify the need for largescale prospective studies to study absolute DFU risks across intrinsic thermographic clusters inform ideal in terms of predictive efficacy and logisticaleconomic considerations thermographic screening interval to identify participants at high risk of DFU delineate thermographic image patterns and absolute DFU risks across a wider disease spectrum including both previouslyulcerated and neverulcerated populations. Conclusion The strong associations between intrinsic thermography clusters and several DFU risk factors support the validity of using thermography for characterising DFU risk. Obtained associations did not prove predictive likely due to spectrum bias or because thermography and classical risk factors characterise incompletely overlapping portions of the DFU risk construct. Our findings highlight the challenges in standardising ground truths when defining novel digital biomarkers. References McDermott K Fang M Boulton AJM Selvin E Hicks CW. Etiology Epidemiology and Disparities in the Burden of Diabetic Foot Ulcers. Diabetes Care. . Armstrong DG Tan TW Boulton AJM Bus SA. Diabetic Foot Ulcers A Review. JAMA. . Chappell FM Crawford F Horne M et al. Development and validation of a clinical prediction rule for development of diabetic foot ulceration An analysis of data from five cohort studies. BMJ Open Diabetes Res Care . DOI.bmjdrc . Faus Camarena M IzquierdoRenau M JulianRochina I Arrbola M Miralles M. Update on the Use of Infrared Thermography in the Early Detection of Diabetic Foot Complications A Bibliographic Review. Sensors. . DOI.s. Van Netten JJ Van Baal JG Liu C Van Der Heijden F Bus SA. Infrared Thermal Imaging for Automated Detection of Diabetic Foot Complications. www.jdst.org. Liu C van Netten JJ van Baal JG Bus SA van der Heijden F. Automatic detection of diabetic foot complications with infrared thermography by asymmetric analysis. J Biomed Opt . Coravos A Khozin S Mandl KD. Developing and adopting safe and effective digital biomarkers to improve patient outcomes. NPJ Digit Med. . DOI.s . Frykberg RG Gordon IL Reyzelman AM et al. Feasibility and efficacy of a smart mat technology to predict development of diabetic plantar ulcers. In Diabetes Care. American Diabetes Association Inc. . Oohashi F Ogai K Takahashi N et al. Increased temperature at the healed area detected by thermography predicts recurrent pressure ulcers. Wound Repair and Regeneration . Golledge J Fernando ME Alahakoon C et al. Efficacy of at home monitoring of foot temperature for risk reduction of diabetesrelated foot ulcer A metaanalysis. Diabetes Metab Res Rev. . DOI.dmrr.. Gatt A Cassar K Falzon O et al. The identification of higher forefoot temperatures associated with peripheral arterial disease in type diabetes mellitus as detected by thermography. Prim Care Diabetes . Landi I Glicksberg BS Lee HC et al. Deep representation learning of electronic health records to unlock patient stratification at scale. NPJ Digit Med . DOI.sz. Jang JH Kim TY Lim HS Yoon D. Unsupervised feature learning for electrocardiogram data using the convolutional variational autoencoder. PLoS One . DOI.journal.pone.. Liu X Faes L Kale AU et al. A comparison of deep learning performance against healthcare professionals in detecting diseases from medical imaging a systematic review and metaanalysis. Lancet Digit Health e. Kinra S Radha Krishna K V. Kuper H et al. Cohort profile Andhra Pradesh children and parents study APCAPS. Int J Epidemiol . Lieber J Banjara SK Mallinson PAC et al. Burden determinants consequences and care of multimorbidity in rural and urbanising Telangana India Protocol for a mixedmethods study within the APCAPS cohort. BMJ Open . DOI.bmjopen . APCAPS LSHTM. httpswww.lshtm.ac.ukresearchcentresprojectsgroupsapcapsresearch accessed July . Hernndez A ArteagaMarrero N Villa E Fabelo H Callic GM RuizAlzola J. Automatic segmentation based on deep learning techniques for diabetic foot monitoring through multimodal images. In Lecture Notes in Computer Science including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics. Springer Verlag . Yushkevich PA Piven J Hazlett HC et al. Userguided D active contour segmentation of anatomical structures Significantly improved efficiency and reliability. Neuroimage . Warfield SK Zou KH Wells WM. Simultaneous truth and performance level estimation STAPLE An algorithm for the validation of image segmentation. IEEE Trans Med Imaging . ArteagaMarrero N Hernndez A Villa E GonzlezPrez S Luque C RuizAlzola J. Segmentation approaches for diabetic foot disorders. Sensors Switzerland . Shahapure KR Nicholas C. Cluster quality analysis using silhouette score. Proceedings IEEE th International Conference on Data Science and Advanced Analytics DSAA . Khandakar A Chowdhury MEH Reaz MBI et al. A Novel Machine Learning Approach for Severity Classification of Diabetic Foot Complications Using Thermogram Images. Sensors . DOI.s. Idiaquez JF Alcantara M Bril V. Optimal cutoff value of the modified Toronto Clinical Neuropathy Score in the diagnosis of polyneuropathy. Eur J Neurol . Tehan PE Santos D Chuter VH. A systematic review of the sensitivity and specificity of the toebrachial index for detecting peripheral artery disease. Vascular Medicine United Kingdom . Kaka AS Landsteiner A Ensrud KE et al. Risk prediction models for diabetic foot ulcer development or amputation a review of reviews. J Foot Ankle Res . DOI.s. MonteiroSoares M Boyko EJ Ribeiro J Ribeiro I DinisRibeiro M. Risk stratification systems for diabetic foot ulcers A systematic review. Diabetologia . Bandalakunta Gururajarao S Venkatappa U Shivaram JM Sikkandar MY Al Amoudi A. Infrared Thermography and Soft Computing for Diabetic Foot Assessment. In Machine Learning in BioSignal Analysis and Diagnostic Imaging. Elsevier . Adam M Ng EYK Tan JH Heng ML Tong JWK Acharya UR. Computer aided diagnosis of diabetic foot using infrared thermography A review. Comput Biol Med . HernandezContreras D PeregrinaBarreto H RangelMagdaleno J GonzalezBernal JA AltamiranoRobles L. A quantitative index for classification of plantar thermal changes in the diabetic foot. Infrared Phys Technol . CruzVega I HernandezContreras D PeregrinaBarreto H RangelMagdaleno J de J RamirezCortes JM. Deep learning classification for diabetic foot thermograms. Sensors Switzerland . DOI.s. Khosa I Raza A Anjum M Ahmad W Shahab S. Automatic Diabetic Foot Ulcer Recognition Using MultiLevel Thermographic Image Data. Diagnostics . DOI.diagnostics. Khandakar A Chowdhury MEH Ibne Reaz M Bin et al. A machine learning model for early detection of diabetic foot using thermogram images. Comput Biol Med . DOI.j.compbiomed... Mori T Nagase T Takehara K et al. Morphological Pattern Classification System for Plantar Thermography of Patients with Diabetes. www.jdst.org. Nagase T Sanada H Takehara K et al. Variations of plantar thermographic patterns in normal controls and nonulcer diabetic patients Novel classification using angiosome concept. Journal of Plastic Reconstructive and Aesthetic Surgery . Hazenberg CEVB Van Netten JJ Van Baal SG Bus SA. Assessment of Signs of Foot Infection in Diabetes Patients Using Photographic Foot Imaging and Infrared Thermography. httpshome.liebertpub.comdia . Huemer K Wei Q Nallan S Jebasingh F Palaniappan L. Preventing diabetic foot ulcers in low resource settings using Pedal Elevated Temperature Risk Assessment PETRA. DOI..... Hutting KH aan de Stegge WB Kruse RR van Baal JG Bus SA van Netten JJ. Infrared thermography for monitoring severity and treatment of diabetic foot infections. Vascular Biology . Sharma N Mirza S Rastogi A Singh S Mahapatra PK. Regionwise severity analysis of diabetic plantar foot thermograms. Biomedizinische Technik . DOI.bmt. GonzlezPrez S Strm DP ArteagaMarrero N et al. Assessment of Registration Methods for Thermal Infrared and Visible Images for Diabetic Foot Monitoring. Sensors Vol Page . Scholten HJ Shih CD Ma R Malhotra K Reyzelman AM. Utilization of a Smart Sock for the Remote Monitoring of Patients With Peripheral Neuropathy Crosssectional Study of a Realworld Registry. JMIR Form Res . DOI.. Zhou Q Qian Z Wu J Liu J Ren L Ren L. Early diagnosis of diabetic peripheral neuropathy based on infrared thermal imaging technology. Diabetes Metab Res Rev . DOI.DMRR.. Djajakusumah TM Candrawinata VS Ho JP Herman H Lukman K Lesmana R. The predictive value of infrared thermal imaging IRT for peripheral artery disease A systematic review. Medicine United States. E. AstasioPicado Martnez EE GmezMartn B. Comparison of Thermal Foot Maps between Diabetic Patients with Neuropathic Vascular Neurovascular and No Complications. Curr Diabetes Rev . Varoquaux G Cheplygina V. Machine learning for medical imaging methodological failures and recommendations for the future. NPJ Digit Med. . DOI.sy. Bharara M Schoess J Armstrong DG. Coming events cast their shadows before detecting inflammation in the acute diabetic foot and the foot in remission. Diabetes Metab Res Rev Suppl . Zhang J Qiu Y Peng L Zhou Q Wang Z Qi M. A comprehensive review of methods based on deep learning for diabetesrelated foot ulcers. Front Endocrinol Lausanne . DOI.FENDO..FULL.

write a youtube script

this is an atmosphere powered car it works by using a syringe as a pneumatic spring to drive the rear wheels using these pulleys in this video Im going to go through the fundamental principles behind designing such a car such as the energy stored the amount of torque produced at the rear wheel and the theoretical range at this car to travel and then Im gonna take it to the next level lets start by addressing how this car is powered by the atmosphere this is a syringe which Ive sealed off the nozzle using an end cap which when I pull the piston outwards it expands a very small volume of air into a much large volume therefore creating a vacuum now this might seem similar to stretching a spring or rubber band but the potential energy is actually stored in the atmosphere rather than the material its made from because gases flow from volumes of high pressure to low pressure its not actually the vacuum inside of the syringe that is forcing the syringe to contract but instead the atmospheric pressure pushing on the rear side of the piston you might be wondering what I chose to expand the syringe using the atmospheric pressure to do the work rather than contracting the volume of air inside the piston well aside from the constant atmospheric pressure of roughly bar putting far less stress on the syringe and the infinitely increasing pressure of a shrinking volume it also helps of calculating the stored energy because there is a constant force exerted by the piston which allows us to calculate the work carried out by the syringe the unit for work is joules which also happens to be the unit for energy and work is calculated by multiplying a force by the distance and it supplied by expanding this syringe Im applying a force over a measurable distance I could measure this force but it could cause inaccuracies due to friction and the theoretical method of calculating it is way more interresting pressure is the measurement of force per area and in the case of atmospheric air pressure at sea level there is newtons of force per square meter so to work out the force exerted on the piston all we need to do is rearrange the equation for the force I wind up my multiply in the atmospheric air pressure by the area of the piston which can be calculated by measuring the diameter of the cylinder dividing by two together radius squaring that and then multiplying it by PI which gives us an area of . to meters squared and a force of point Newtons which is equivalent to hang a . kilogram weight from the syringe handle taking into account that my current location is slightly higher than sea level and then most likely isnt a perfect vacuum inside the syringe this . kilogram weight should be a good test to see if our calculations are close not bad so we now know the distance and the force required to expand the piston which when multiplied together gives us a work done in joules now I bet . joules that means nothing to you so to put it into perspective this tiny a battery which is about same size as a syringe can store over times its amount of energy and for those of you who might believe that you can run your cars and airplanes on just in there alone if I were to fill this milliliter syringe with petrol it will contain thousand times that energy now that weve calculated how useless vacuum powered vehicles are lets go through how we get the most from the small amount of energy weve stored as youll probably know to get the most efficient fuel consumption from your car youre best off accelerating slowly and traveling as a slow constant speed to keep drag and frictional forces to a minimum so this is where the pulleys come into play instead of the syringe producing a point Newton force over a distance of . centimeters we can use these pulleys to convert some of that force into increasing the distance for example if the string was wrapped around a pulley a half the diameter of the rear wheel the force at the ground to rear wheel contact would now be halved but the distance traveled would be double this type of gear or pulley ratio is used in all sorts of applications but probably most visible on a bicycle when you ride a bicycle you push down with your feet with a far greater force than you do when you walk however you travel much further per push and in the case of this car it has a ratio of to in other words the point Newton force from the syringe is now converted to . Newtons or equivalent to lifting . grams but in return it should travel times further therefore theoretically traveling point seven meters now I hope you found that interesting in all but I bought a multi pack of syringes from Amazon and Im not willing to settle for just the milliliter syringe Music Music so here it is my atmosphere powered radiocontrolled car it has a milliliter syringe at the frontier which is times the capacity of the milliliter syringe and interestingly enough it can actually store times the amount of energy in the atmosphere which means that the with these two syringe examples the capacity of the syringe is directly proportional to the amount of energy it can store in the atmosphere now as youll probably expect from the start of this video the large diameter piston also means that theres a larger pushing force from the atmosphere in fact with this piston there is a force of Newtons pushing it inwards towards the nozzle which is equivalent to hanging an eight kilogram weight from the piston and itll just about stay there now I decided to make a custom d printed piston so that I can attach a nylon string directly to it which means that when its fully extended to the outer position theres not a huge handle sticking out the back of the car now this nylon string is wrapped around an initial shaft of five point six millimeters in diameter which drives a pulley thats millimeters in diameter this in droves a string which goes to a shaft which is three millimeters in diameter which drives the rear wheel which is millimeters in diameter this whole pulley arrangement essentially gives the distance travelled by the piston to the distance traveled by the wheel a ratio of to so basically this car should theoretically be able to travel about meters now I dont actually think this car will travel meters for the simple fact that when you wrap the string around this shaft as many times as you need to in fact to pull this person all the way back the rear wheel needs to spin times the actual diameter of the shaft will be increased by quite a lot because it will basically be wrapping over the previous rows of string so that will actually reduce the pulley ratio the more you turn the wheel also because I dont know how long this cards actually going to travel and I dont have a really long stretch of really flat perfectly smooth ground Ive decided to fit this cart with radiocontrolled steering so that way I should be able to test it in in the house and be able to steer it round hopefully so lets go and give it a quick test after Ive won this rear wheel nine times right so Ive only wound it up about twenty milliliters on the syringe gauge for now so about one fifth fully expanded just because I want to give it a quick test the string has already increase the diameter of the shop are quite a bit but this is going to be really smallest diameter tests obviously when I run it up to milliliter gauge its going to be really thick theres gonna have a lot more talk so I want to just test that it can run with this low talk amount where the amount of string wrapped lets try this okay alright first test in three two one slow start slow start speaking up speed doesnt have much grip on those front wheels thats working really well starting okay and thats the end right so did what three four laps now Im going to wind it up to the full amount and well see how long this runs for and we are fully wound back check how much string is wrapped around that thing so as you can see the damage of the shaft has now increased by a lot so I imagine its going to be very fast at the start of the run but lets give it a go three two one right the front we always had no grip and it just went straight into the wall tipped over and then pretty much all wound it unwound itself round all of this wheel which is an absolute nightmare to rewind again because all these string is now tangled up so Im going to untangle everything put some rubber bands on the front two wheels and well give it another test in a minute or in a few right so Ive dude some rubber bands onto the front wheels just to add some rubber grip hopefully this will allow it to steer properly and not just slam into them all at the end Ive rewound it up to the milliliter mark on the syringe and I think were all good for another test drive I hope in the way they come out Ill move out the way in three two one oh its wheel spinning its moving it will spins as soon as it goes on the corner see there juner still got half the syringe to go and now its reducing the amount of talking can produce so its not doing it as often its still got about percent left maybe its really slowing down with the amount of talking from the juice why is it not going I think whats happening is the string is actually getting tangled because its wrapped up so much and its making it harder to release its still going come on come on its still working just right Ive just made a few minor modifications to it Im unable to change the pulley ratio currently so what Ive done is added grams of lead to the back to increase grip and also added another rubber band to the rear wheel so we now have double the amount of rubber touching the ground and some downforce Ive also spent a bit more time winding the string around to try and get it as wide as possible on that shaft problem is if I have a too wide on the shaft it may come off of this pulley when its been wound up so am I think this is probably the best shot Im gonna get from this car so yeah lets give it a final test get my remote and in three two one oh yes were spinning again thats a really weird shudder it makes trying to turn as slowly as possible so that doesnt do it all that was smoother as soon as it picked up speed still going really well come on now its getting into its rhythm if I just keep it turning and it doesnt have time to pick up speed then grip doesnt seem to be an issue come on dont stop yet you still got a bit left to go okay now Im gonna let it go a bit straight so it can pick up some speed because its running out of energy come on the last little bit you know you can do it come on there we go so yeah it still has some wines left on the string but its a not enough torque left to just still quite a bit see what I mean theres still quite a few turns left in the wheel which would obviously get it to that theoretical meter distance it just runs out of a bit of talk so Im lets go back to the shed and conclude this project so to conclude this project I can confirm that the atmospheric pressure surrounding us is enough to power a small radiocontrolled car like this but I really dont believe that its going to be replacing our fossil fueled cars anytime soon if ever however this was quite a fun project to do I think there can be some improvements with this car obviously to get rid of the grip issues if I were to increase the gear rate at the pulley ratio then that would reduce the amount of torque at the rear wheel and probably therefore reduce the amount and wheelspin that would have to be designed it into a version car because these pulleys are the largest sort of pulley ratio I can put in this current chassis as I said earlier its a to ratio throughout these two police and therefore to get larger pulleys on this I would need to either extend the chassis or somehow move a pulley to the side of the syringe and basically it would require a full redesign to reduce the Wow spinning the main purpose behind this project was not to prove that we can replace our fossil fueled cars with atmospheric powered cars because well I think Ive proved that its a bit ridiculous the main reason behind it was to try and inspire some people to maybe try a similar thing you dont need a d printer to build one of these because I dont know if youve ever seen mousetrap cars you can search them into Google and theres all sorts of all sorts of examples but its quite popular in American high schools where they build a car powered by a mousetrap spring many people use stuff such as CDs as wheels and also bolster wood as the chassis so you can build a similar car out of all sorts of different materials I just thought the atmospheric thing with the syringe added a new physics element to the whole equation so there now you can calculate the amount of torque it produces by the atmospheric pressure before I completely in this video I just want to say something quickly its currently getting near to the end of its late December and this is going to be my last project video of this year I will be working on new projects for next year however at a slightly slower pace due to the Christmas holidays I want to say a massive thank you to everyone thats been supporting my channel over the past year I believe at the start of this year there were have you subscribed and it looks like were going to reach thousand subscribers by the start of which is just insane that a hundred thousand of you have subscribed in the past year I need to say a massive thank you to all of my patreon supporters I know I say this in every video however they are the reason that Im able to put the time and effort and also funding these things into making these project videos possible for the rest of you to watch so huge thanks to you guys I seriously couldnt fund all these projects without your help now theres one last thing before I shut off this video for the end of this year and that is that Ive made a complete video of all of the projects Ive done this year because uploading videos on nearly a weekly basis for a whole year his yeah its brought about a lot of projects that I completely forgot Id even done so yeah if youre new to my channel and you havent seen any of my videos then if you watch this compilation video and you see something go check out my other videos and if youve been subscribed for the whole year then hopefully this will be interesting as well a reminder of what Ive been up to anyway Im gonna go and enjoy my Christmas have a good New Year and Ill see in Merry Christmas Happy New Year and good bye I was just spin it and then Music haha Applause rotate stick should till the motifs and its line Music Applause Music Music three two one Music Music Applause Music Applause Oh Applause Music Applause Music Applause Music Music Music Music

write a youtube script

Music so it turns out you know musk is the time to change the name of the SpaceX BFS or big Falcon ship to be called starship as well as mentioning that theres going to be some changes to the design now as of recording this video the new design hasnt been released yet and therefore Im going to be focusing on the current design which has three large fins at the bottom and two smaller fins at the top now if SpaceX may or may not change the design of these fins in the upcoming design however what I want to do in this project is see how these fins control the attitude of the aircraft on its approach for landing in last weeks video I went through the design of my radiocontrolled starship and briefly cut out some of these phone parts this week Im gonna hopefully finish the build on the starship and we can get some test flying done I started by covering all the foam board parts with packing tape to add some strength as well as some appropriate black and white coloring to match the real starship the parts then slotted together using the d printed motor mount as a jig I had originally planned to glue these three fins together in the middle however tape seemed to do the job and would also allow me to remove the thin for repair when it crashes Ive then attached to three degree pi section pieces which will not only hold the foam board sheets at exactly degrees apart but will also act as a lip from which the fuses are to ball mount it was in time to start mounting the electronics starting with the servos which will control the thrust vectoring vanes the fly controller which contains the gyros and accelerometers for stabilizing this craft was then mounted to the top of the three foam board fins using a d printed part also these servos I bought came with overly long whys so although this looks like a wiring nightmare its actually far less complex than you might think so lets plug in the battery and check everything is working I first moved it around to check the flight control it was correcting for my movements and then check that my control inputs are working notice how the pitch control moves both the thrust vectoring things and the large rear fins also notice that the thrust vectoring fins can tilt to produce a spiral vortex of thrust this will be used to contract the torque produced by the motor and therefore prevent the craft from spinning out of control also I set up a switch which would shift the neutral position of the large rear fins this way it should allow me to test more of a range of angles for pitch control during descent I think this is probably a good time to test that the thrustvectoring stability control is working or perhaps not I mounted the foam fuse last tube to the bottom fins which allowed me to mount the battery higher up this will increase the distance between the center of mass and the point at which the thrust vectoring force is produced essentially increasing the amount of control Authority Music at this point I was trying to tune the PID values in fly controller but later realize that the issue was actually related to the large fins at the bottom as soon as the craft starts moving sideways or there is a gust of wind the fins produce more sideways drag than the fuselage tube therefore tipping the craft into the wind and because the thrust vectoring isnt strong enough to counteract this it keeps tilting until it crashes because the previous fuse lost she wasnt very round I decided to remake it with half of it covered in black packing tape to match the actual starship color scheme and try bending it into a tube slowly with no success so making many folds info seemed to be the only way to get it to bend this can then be slotted over the fins and be taped into position I then cut out the small front vents cover them in black packing tape and mounted into the pre painted d printed nose cone then because my workbench is relatively close to the roof of my shed I took it outside for a quick assembly and it was finally ready for some proper test flights now I was quite nervous about testing this thing because its not the strongest thing Ive ever built due to I need him to be quite lightweight plus this was the only day this week where it was a really windy and raining so repairs werent really an option all cameras rolling and liftoff Music what happened here was the same as what was happening when I was testing without the nose cone however it was slightly less violent fortunately the nose cone just popped off and with a bit of tape it was good to go again so on the second flight I increase the throttle law which would therefore increase the amounts of control I had from the thrust vectoring fins I also folded the rear fins back to reduce the amount of drag produced at the bottom of the craft and this seemed to help massively with the fins folded back they acted like dihedral on a plane and helped to stabilize the roll of the craft this allowed me to stay in control and hover it around pointing bellyfirst into the oncoming air I increased the throttle to gain some altitude and attempted an approach descent lets see that again as soon as I switched up the throttle the cross spins rapidly to the left and I lose complete control until I increase the thrust again now what I think is happening here is that if he fins on at the exact same angle it essentially becomes a falling propeller and wants to spin around so for example if you were to raise the front right fin and the rear left fin itll basically caused the whole starship to you around towards the left and this could possibly be the way that SpaceX are aiming to control the your axis on the starship on descent for landing so now instead of testing it with a zero throttle Glide Im going to descend it with a small amount of thrust the real starship wont use any thrust until the absolute final landing stage but it also has a much higher horizontal momentum due to its orbital velocity and this model just drops vertically when I stop the throttle due to its high drag and low mass this method seemed to be far more controllable right up until the final part before landing where I deployed the fins to the landing position this again produces a lot of sideways drag at the bottom of the starship and therefore causes it to tip into the oncoming wind so I went back around for another approach now that was a close call but what happened when the starship is descending Billy first the fins can be hinged back and forth to change their angle of attack to the oncoming air but the issue arises when the starship is angled directly nose first into the oncoming air to control the craft the fins must now rotate in this axis to change the angle of attack but the starship is unable to do this and will therefore have no aerodynamic control Music the next couple of flights werent too different as I was just practicing the approach at landing although the final flight did make me realize how important these fins are to the control of the craft so yeah the pushrod control horn has just come loose if you can see right there the flap can move and this is just hanging loose I didnt have any glue with me so Ive just taped it and that obviously wasnt strong enough this side seam still seems good so just losing control of one of these flaps pretty much ruins all control of the aircraft or the rocket so I say okay so take it back in check all the servers are ok and then if the weathers still good the winners are picked up too much okay so well try again whilst the glue is drying on the rocket I reviewed the simulation at SpaceX have made of these starships atmospheric entry and landing phase and what so noticed is that it enters the atmosphere at a really shallow angle to slowly bleed off its velocity like the majority of spacecrafts do and it keeps facing bellyfirst until it has no more horizontal velocity at this point it is traveling at about mak . and folds the large rear fins back to reduce drag at the rear this pitches the craft and it begins to accelerate to Mach . before firing up the rockets to decelerate and finally moving the rear fins back to the degree spacing for landing so what I want to try and do is replicate the final landing phase which is falling at a steep angle and folds the rear fins just before landing to bring it in the vertical position then unfold the fins and perform a landing the craft was a bit unstable with a slightly steeper approach but the issue was mostly with the final rotation just before landing now I have no idea where the center of mass will be on the actual starship and this is probably the issue here for it yourself right itself to the vertical position just before landing it might need to center of mass to be further towards the rear therefore the front fins can act like fins on a dart and point the craft Rockets first however on this model because the front things arent adjustable having a further rear center of mass will make the craft unflyable on descent so what I think basics does is they have the centerofmass really far back and had the small front fins folded back for the majority of the descent this way the center of pressure is balanced with the center of mass and when the front fins are unfolded it suddenly becomes a backwards facing dart now I can get some more close to simulating this with my model but it really needs some help from the vectoring of the thrust so itll be interesting to see where the space hits can perform this rotation using only aerodynamic control of the fins or whether theyll need to use two gas thrusters to assist it so there we have it my radiocontrolled version of the spacex starship now obviously i couldnt replicate the simulation that spacex showed of it entering the atmosphere and landing due to various budgeting constraints however this would have performed pretty well with the pivoting of the rear fins it was a bit of a shame that I wasnt able to testing pivoting of these small front fins because I reckon this could increase in control Authority a bit and also assist with the final rotation just before landing now there is one other thing that I noticed with the design of these small fins on the actual Spacek starship and that is that the axis of which they pivot about isnt perfectly in line with the rear fins theyre actually angled slightly inwards which means that when these pivot up and down they ever so slightly change the angle of attack to the oncoming it now maybe this is a way that SpaceX are going to aerodynamically controlled the role of the starship on entry or maybe even control the pitch of the starship if it ever gets into a situation like minded where its facing directly into the oncoming air now when the space suits are going to stick to this fin design or move on from it I dont really know however I learned a lot from this project and I hope you guys did too if youre wondering where I got this spacex starship model from its actually some files that i found on the thing of us ill post a link to in the description below as you can d print your own model spacex starship i lets say a huge thanks to d prints for sponsoring the filament for this project they gave me all the roles of pla and PE TG d printer filament to make this project possible and id like to thank you guys for watching if you enjoy this video we appreciate it if you can leave the thumbs up if youre new to my channel please click Subscribe and a huge huge thanks to all of my patreon supporters you guys make these videos possible and I honestly couldnt go through the effort to make such a project and such a video for this subject so huge use thanks to you guys and yeah Ill see you in the next project goodbye

write a youtube script

a couple of months ago Elon Musk announced some new information on the SpaceX bfr rocket which was live streamed on the SpaceX YouTube channel Ill post a link to it in the description below if you want to check it out but what caught my attention was the new control method of the BFS the big falcon ship now the current Falcon rocket and also the Falcon heavy rocket reenter the atmosphere backwards with the rocket nozzles pointing first and it controls itself using aerodynamic grid fins which are mounted on the top of the rocket and they sort of act like ailerons or elevators on a plane and also direct the rocket down to the ground before a lastminute rocket burn just before landing now this new big falcon ship is designed to descend belly first rather than rocket nozzles first and the purpose of this is to try and bleed off as much velocity as possible using the aerodynamic drag considering there is an atmosphere where its landing this will mean that when it comes in to do the rotation to do the last landing burn its traveling a lot slower than if it were to just travel backwards through the atmosphere now whats interesting about this is the BF s has very large fins on the rear of the rocket and also on the nose but these dont work like regular fins on a rocket where they would maneuver the air similar to an aileron or elevator on a plane instead what they do is they rotate about the parallel axis of the rocket and actually I used to induce drag on the nose or the tail as its descending that belly first Eagle must described it as similar to the way a skydiver would fall where they manipulate the drag now this concept is really interesting to me because its almost identical to how my radiocontrolled wingsuit guy would fly he would maneuver the angle of his arms to produce more or less drag at the front and when therefore caused them to pitch up and pitch down but we all know how that went Music yeah the head snapped up anyway so what I want to do is try to replicate this method of descent by building a radiocontrolled model of the big Falcon ship so lets jump into fusion and Ill run you through my current design so heres my design of the model BFS its going to be built out of mostly foam board just to keep it as lightweight as possible with obviously some d printed parts the main power source is going to be a motor and propeller which is mounted at the bottom here and the thrust from the propeller is going to be vectored using these three fins now thats mainly just to get the rocket to altitude and also hopefully we want to perform a slow landing the main thing Im interested in here is the adjustment of these two fins here now the real Falcon chip the big Falcon ship will adjust the angle of these fins up here but because this rocket is going on this model is going to be fairly small to add the weight in complexity to adjust the top fins Im not sure thats necessarily relevant right now the main thing is to test whether these bottom fins will adjust the center of pressure of the rocket and because the center of gravity or mass will always be same when the center of pressure moves either side of the center of mass the rocket should pitch up and down so thats what I want to test and also you may be wondering why theres these odd sticks sticking out the side of the rocket and thats mainly theyre basically the landing legs on the real rocket theyll be using these fins as the landing legs however to keep this as lightweight its possible to make that hinge as strong as possible and also have these servo withstand any kind of forces from hard landing just isnt feasible so I thought Id make my own sort of landing legs wish to take out the side and will be a lot stronger now this model is quite complex with the whole thrust vectoring unit and therefore on the heavy side so Im not sure how fast its going to descend it might fall quite quickly and hit the ground quite fast and I did it originally I actually have an idea just to build a glider version which would just vary the fins at the bottom and only be powered by a very small lightweight battery and not have a motor on board and this would have been dropped from a drone and I hope you just be able to perform a controlled glide down however I thought Id get some opinions from you guys on patreon and it seemed like the majority of you wanted me to make this fully thrust effected version so yeah I think we should go ahead with this so what Im going to do is start the printer and get on with cutting out some foam because of the curve shape at the nose cone its quite difficult to construct from a sheet of foam so I decided to d print it using a single outer wall layer method whilst the printer was going in circles I started cutting out the bottom profile sections of the rocket from the foam sheet once they were cut out I mounted the motor to the d printed motor mount which also acts as a structural jig to hold the three fins together the next step was to construct me main fuselage tube which isnt easy using this foam board as it tends to snap if you bend it too far so I laminated one side using packing tape to add some tensional strength to what will be the outer side of the tube then after bending rubbing scoring or whatever youd call it the foam board started to curve and I could join the two ends together however the curve was pretty terrible and I almost slightly add some disk bulkheads inside to maintain the round shape the nose cone eventually finished printing and I take the main parts together to give an idea of what itll be like when its finished unfortunately I was unable to finish it in this video so prepare yourself for some lame excuses take it away Tom so unfortunately Im going to have to cut the part one of this build video slightly short due to a number of reasons the main issue being the electronics of the hold for this project which include the electronic servos and also ordered some lightweight batteries for powering it for some reason when I ordered them they werent shipped for two to three days so they havent arrived yet and I dont really want to be gluing stuff together without having the electronics available to obviously mount the other issue I ran into was that the nose cone is printed with a single outer wall layer to make it as lightweight as possible however for some reason my printer decided to miss a few layers under extruded a few layers so that was a well it was about a tower print just wasted so thats the end of part of this rig controlled muddle big Falcon ship I hope youre interested in this project and if you are then it would be appreciated if you could leave a thumbs up if youre new to my channel and you want to follow this project then please click Subscribe down below and a huge thanks to all of my patrons for supporting me you guys make these weekly videos possible and also thanks for suggesting or at least voting on the way I should configure this because I really think that the thrust vector design will be the most impressive if it works I suppose well have to find out thanks very much for watching and hopefully Ill see you with part two next week goodbye

write a youtube script

three two one a trebuchet stores energy by raising a counterweight which when released converts its gravitational potential energy into kinetic energy but what if there was a way to store this energy as kinetic energy before release in something like a flywheel lets make a flywheel trebuchet unfortunately my old trebuchet hasnt aged very well and the plywood sections are falling apart so ive decided to make this new trebuchet completely from aluminium the frame will be built using these extrusions which have a track built into them allowing these nuts to slot in and clamp down giving lots of flexibility for mounting positions the frame design will be an equilateral triangle for maximum strength which means cutting custom degree angle plates to join the sections together once the two frame halves were assembled the cross joint extrusions could be attached with the degree bracket and its already starting to look like a trebuchet its probably a bit overkill for this project but with the extrusions being so adjustable it can be reused for future trebuchet experiments the arms are made from by millimeter rectangular tube section as these need to be strong to survive the huge amount of force accelerating a tennis ball so fast the official worlds fastest tennis serve was miles per hour and to put that into perspective this is my old trebuchet topping out at miles per hour so to beat the record speed the gram ball will undergo a centripetal acceleration of at least gs which is equivalent of hanging a kilogram weight from the end of the arm which according to my stress analysis can bend this arm by four millimeters at the tip i then added bearings to the arms which allowed the whole assembly to be mounted to the axle with a small push everything seemed to run smooth but it needs some balancing instead of mounting another long arm to the opposite end i decided to use a counterweight because at high speed i expect the arm to produce quite a lot of drag so having only one arm will reduce the amount of energy required to spin it the balance isnt perfect but its far better and ill balance it properly once the flywheel is mounted theres just one problem i need a method of spinning it faster so i d printed a pulley which mounts to the arm then a timing belt can be used to drive it from an external source and i mounted a secondary axle and larger pulley to the frame and using the slot nuts the belt could be correctly tensioned i can then attach a crank handle to the end of the secondary shaft and wind it up to speed but first lets build the flywheel the flywheel was cut from six millimeter thick aluminium and it will have a diameter of millimeters but because thats too large to fit on my cnc router ive decided to cut it in four pieces each section can then be bolted to the arms to produce a large flywheel ring and the unsuspecting opposition has yet to recognize the potential of this machine so tomorrow we can attach the projectile slinging components and get on with the testing unlike traditional trebuchets the projectile will be constantly moving with the arm so it needs to be held in position until a trigger is pulled this nylon strap is sprung loaded to clamp the ball in position and is attached to a small hook near the axle to trigger this latch to release i built this contraption that moves into the path of the hook as the flywheel rotates so the trigger can be pulled at any point during the flywheels rotation and the ball will always release at just the right time and this release contraption can also be advanced or delayed by sliding it up and down on the frame to get the most optimal release angle i then place the ball in the sling and attach the sling strings to the end of the arm a trebuchet sling has two attachment points one is fixed to the end of the main arm and the other is loose but is hooked over this pin then as a projectile swings around the loop slides off the end of the pin and releases the projectile time for a first test albeit a slow test as i havent made the crank handle yet and oh i should probably pin this to the ground it seems like all the weight of the flywheel pushed onto this pulley and has destroyed it but nothing else seems to be broken so we could probably give this another test by hand in three two one that actually went pretty far after the first test being so slow right so i now have the crank handle on here and also ive fixed the pulley which is now half aluminium half d printed so we should be able to get up to some pretty high rpm this slowmo was filmed at frames per second and has been slowed down times using the length of the arm as a reference we can measure that the ball traveled . meters in milliseconds so by dividing the distance by the time we get a calculated speed of . meters per second or miles per hour which is miles per hour faster than my old trebuchet but its still miles per hour slower than the worlds fastest tennis set the fact that a human can hit a tennis ball faster than this is just insane so today we are going to beat the world record tennis serve i reckon the most optimal release point is when the sling is in line with the arm because then its at the largest radius from the main axle so what ive done is ive set the trigger to release the ball earlier dropped from the arm earlier and then the sling to open up later so should have the same trajectory but just hopefully slightly faster Applause i didnt even see where that went managed to snap the sling though i hope that went forwards wow lets check the high speed footage then i need to rebuild that sling right so according to the high speed footage the ball traveled . meters in frames so thats a hundredth of a second that means it went meters per second in miles per hour thats miles i think we need to redo that test and this time im going to make sure that my hand is completely off of the crank handle before the tennis ball is released just to prove that its purely the inertia of the flywheel at the point my hand stops applying power the flywheel is spinning at rpm and i know from my cad model that its moment of inertia is . kilogram meters squared from this we can calculate the flywheel has a stored energy of joules thats equivalent to raising a kilogram counterweight to the height of the axle but this flywheel only weighs about percent of that at seven kilograms once the ball is released it leaves a sling with a kinetic energy of joules meaning the usable energy efficiency is just percent however the flywheel is still spinning with an energy of joules and i reckon if we increase the mass of the projectile we can extract that last remaining amount of energy that went high doesnt quite stop it so we need to increase the weight a bit more but its getting close right so this is now gram tennis ball which is about three times the weight of a standard tennis ball thats so weird to watch at the point of trigger release the flywheel slows to almost a dead stop in a sixth of a second losing . percent of its energy whats interesting about this is the flywheel on the left has twice the stored energy to begin with but the projectiles leave the sling with almost the same kinetic energy so the heavier projectile on the right is traveling quite a bit slower but still launches about further due to it being affected less by air resistance seeing as its flywheel trebuchet likes projectiles a little heavier than a tennis ball maybe i need to get some baseballs and stick a motor on this thing but for now im very happy with the performance if you enjoyed the video itd be great if you could leave a thumbs up down below if youre new to my channel and want to see other projects similar to this then please click subscribe down below and a massive thank you to all of my supporters over on patreon.com for making these projects possible i honestly couldnt build them without your support so thanks once again thanks again for watching and ill see you in the next video goodbye

write a youtube script

Ive been working with World of Tanks to recreate a s allterrain vehicle concept called the Rhino with its unique features of half hemispherical wheels but the deeper they sink into the mud the more grip they got so in the previous video we got a freight wonton dumper move the seat in position and tilted all the axles at degrees all we need to do now is make the wheels and give it a body right so weve got my little Freights dumper its a massive amount of work weve stretched the axles out weve tilted them at degrees and weve dragged all the Hydraulics for the steering right the way through the articulation to put the seat at the front now arguably its in a worse position when it started weve got this horrible looking vehicle right really bad positive camera it looks a bit weird but what we need to do now we need to make these huge hemispherical Wheels this is the part of the Rhino which Im most interested in how were going to do that we need a plan and the sponsor World of Tanks lets not forget about them theyre making all this possible now then if you like huge tank Arsenals massive battles and of course a game which is free to play World of Tanks is for you yes now if you use my code first Army in the link in the description Ive got a little gift for you yeah youll get three free tanks youll get credits seven days premium access all for you its all worth about . pounds and its free yes free right to the barn how are we gonna make these wheels Colin Music whats the plan then first this is the plan so thinking well make some ribs a bit like a skeleton and then in between that have some plates slightly thinner a bit like a bicycle how like you know the individual components wont be very strong but once its all welded together hopefully itll be okay now weve got some flat bits top and bottom thats for when its on like Smooth ground like this it can roll properly and we need to put some tread into that a lot going on here and then afterwards weve got the massive question mark is because we have increased the diameter of our Wheels considerably weve kind of sapped a load of torque from the engine by doing that but kilos lighter than what it was and it was designed to take an extra one turn of rubble around up a Hill so it must have had quite a bit of power in hand so Im hoping because were not taking a ton of rubble anywhere were gonna be all right right lets sort of make a skeleton thing I dont know Im just gonna tried making it see where it goes oh yes were off and look I love it when the plasma car goes in time with the music editing Colin oh look its in time again no lovely little tip here if you want to grind anything or brush anything metal it slips around stick it on a bit of carpet dont go anywhere its quite a big one its a circle thats a bit worried it was gonna be too small now then I have made a few errors these ribs were gonna weld on here weld on there those go all the way around theres going to be eight of them so Paul Colin has only done six spokes on his other wheel now that wouldnt matter too much because you know you weld these and where they need to be but I was going to put some strengthening bars inside maybe down onto these Spokes and of course if they dont all line up thats not very good here we have Mark II with eight spokes the right amount of Spokes and then the center hub which has the holes which bolt straight onto the axle thatll sit there now Ive got a little bit trick Ive gone all notches slotty basically at the end of each spoke theres a little slit and then at the end of the ribs Ive got like these little catches they sit in there very nicely and then when I come to weld it all together you know Ill get them all in the right place also on the other end theres a little slit because this is going to be slip rolled into my access ring and theyre now slot together like that which go around on there thatll be a circle obviously and Ive done the same thing on the middle tube as well this is going to get slit rolled and this has got all these little tabs that stick out the top of it and theyll go in these little notches here right I better do some slip rolling in some welding then um Music around Music right weve got our skeleton Im quite happy with this everything seems to be where it should be and it seems symmetrical seems reasonably true now the difficult bit now is filling in these little bits and I thought my original idea I will draw it on the computer I will then kind of section it off and then I put it in this other program which kind of unfolds it and turns it into a flat pattern but this was what it come up with something aint quite right so I thought all right then well go semiold school semimanuals Ive got a bit of a bit of wood laid it up against it drew around the back of it and then that kind of gave me this and then I basically measured this and did my best as I could to kind of put this into the CAD software myself but then for a weather a minute Colin youve got a lot of these measurements already you know because all because I drew already find out the length length how wide that bit is across there so I kind of put them all into the computer cut it out and then we ended up with this now this is pretty good it fits in here reasonably well probably do have a little bit of adjustment but now I was looking at thinking oh wait a minute heading down the wrong path here Colin because Ive kind of cut it across at an angle here because you know to try and bend it in both directions thats English Wheeling thats going to be a right pain and I thought well this part of the wheel here is like the tread when its on concrete ground this Wants To Be A Perfect Circle it doesnt want to have like a pence piece you know like a hexagon so I kind of thought well that bit needs to be slightly different than that so we have its like a little cone of slit rolled that will fit in there like that of this little bit here which will sit on top of it and weld in there and then within that we can have our like tire tread which will then hit the floor then of course when it goes into mud it will go up this bit it aint gonna matter so Im going to weld all them in there put that on there and then Ill adjust this little bit so it sits in here and does the final part of it yes I know what youre thinking this is getting heavy Colin it is Im a little bit worried about that what about lots of things the diameter the weight of it in fact just please download the link and download the game at that link because if this doesnt work Im gonna be very happy with me are they right itd be fine Colin come on Jeopardy we like a bit of Jeopardy Music Applause Music Applause Music its not bad Applause Music conversations Music Applause Music how can that be uh lets go right do you wanna right no no no no thats just you its just yeah okay is it gonna be all warped and wobbly or is it going to be as straight and as true as it was it oh thats all right Ill take that Music right thats got that all sorted finish welding that up seems weve made my hair go a bit shorter now its all done got a little cover plate on the front we just need to put some tread blocks or like some tire tread on it and see all these holes to put the back but Im not going to do that yeah Im gonna make another one fit it to the back and just check that when the thing articulates they dont touch because it looks a lot closer than what I was hoping it was going to be right lets make another ah weve got another one Colin what are you gonna put on the back to see if they touch when they steered yes I was but Ive noticed I made another two like this for the rear with how Ive mounted them theyd end up touching the diesel tank and the hydraulic tank so Ive got to kind of step back ones out slightly now theres a bit of wobble so Im going to make the inner spokes on the rear ones a little bit thicker try and strengthen that up a little bit Ive had a little drive around all seems okay right two for the back three two one boom foreign Applause s and they do not touch when it stairs that Im very happy about now Ive welded all the holes in all the back Ive put thinner metal in there the idea was to save a bit of weight but I think basically it just made a lot more work but hopefully every little helps now before I drive this on the concrete obviously because its still on concrete its going to make horrific noise and possibly damage the floor Ive got some rubber tracks and rubber feet coming for it but while Im waiting for them this thing needs some sort of Bodywork it needs a cover cant just use it like this doesnt look very good does it now thats a lot of bending and a lot of metal shaping Ive been lent a new tool okay so we have this Morgan Rush with hydraulic press brake now this is overkill because originally I was just going to get like a manual bundle a bit like the one Ive got at home but there is advantages to having this the accuracy you can decide the radius of what you want and the repeatability of it you can bend Bend Bend Bend they will all be dead on exactly the same I shall demonstrate these jaw doors come down with a pressure of up to tons into this V block now the clever thing is this vblock can be tilted and has many different Vs on it each side has a slightly different depth V and a different profiles basically meaning you can do really tight bends like this or you can do slightly shallower more curved bends like this it also allows you to do different thicknesses of material like you can bend up to mil plate on this thing but youre not going to shove a bit of mil plate into a tiny V so you just choose a bigger V now also weve got these stops at the back here you can set the distance then theres a little computer screen up here so say if you want a degree Bend mil in for the edge you set them for mil theyll move back you push a bit of metal push your foot on the foot plate bang itll come down and apply the correct pressure and the correct there to bend a degree Bend and you can just keep doing that over and over again if you want to change the angle you can do a degree one you can do a degree Bend repeatability awesome and that is what we need for our tank okay here my plan which on the face of it will seem very complicated but its just what I want to do so Im going to get loads of panels plasma cut out press break all the sides on drilled holes like tabs that will sit in there and then theyll have individual panels all coming off at funny angles itll have a lot a lot of sharp Corners well sharp sharp but a lot of angles and Corners to it so its going to be where history meets modern game makers are going to love this in there because you can drive old tanks and new ones so Mount that there different panels all bolted together lets have a crack Applause beautiful today Music Music Applause okay this looks blooming awesome futuristic but its got that retro touch with the nuts and bolts now then this was quite simple all the bottom bits but when it comes to the middle bits I kept making mistakes I was going through bits of metal bits of cardboard on the laser cutter and then I made a discovery me angle finder put it up against the bench the blooming notches are three degrees out aint they everything Ive been measuring has been three degrees out hence why a lot of mistakes now then I might put a little uh extra roll cage across the top give me a little bit more protection I dont know yet but now were going to do this to the rear and because weve got the engine to consider we need some cooling Vents and some louvers oh Music Applause oh yes it is looking fantastic but we got one more job before we can take this thing for a spin we need some rubber Treads now my laser cutter will cut thin rubber but its not going to cut through ml slabs of rubber but Swift cut who make my plasma cutter also make huge water jet Cutters so basically one of their customers is not that far away and theyre very kindly cut these out for me now these are fantastic were going to bend these around in here put some studs on with a stud welder and fix them to it but we dont want all bolt heads sticking out of these holes so we need to machine back like a washer and bolt recess we cant do that obviously with lasers and plasmas because they only work in two Dimensions but at home I have the perfect tool for the job Music foreign here we are in the bunker shed with a one finity CNC router now Ive actually had this a long time but obviously because Ive been tunneling and doing the barn Ive not had a chance to use it very much I saw it at maker Central in I thought these big fat linear rails big lead screws that is a bit of me that is nice and solid now its a fantastic machine but can its machine rubber because rubber of course is a flexible material that moves around a bit the software it uses is by carve coat very good software very easy to use now obviously programming routers is a lot more complicated than laser cutters because its not just working in two Dimensions youve got the three dimensions youve got different tools you got the speeds youve got the way at which it comes down theres all sorts of things but dont worry about it its very simple so what Im going to do Ill put this little screw thing which sticks out here Im going to push it over the top of that we will tell that that is our Center and then Ive programmed it to come down down a bit more wear it down a bit more where it was so we should have an eight mil recess in our rubber Music hey hey Music guys work to treat look you put the washer straight in it no on it squeeze it down Im impressed that I thought its gonna make a right mess of it so to fix these to the wheel gonna use the stud welder with some internal threaded studs weve got this nice little template Jiggy thing bend that over there hold it on with magnets studs stud fix it over bolt it on and repeat times Music and thats it fitted and after of them it looks like this Music okay my s concept wheel vehicle is ready for testing now I think this thing looks absolutely brilliant but is it gonna do what its supposed to do is it going to laugh at the muddiest of mud we shall see Ive no idea how this is going to perform I look at it and think yeah its going to be really good and then Ill look at it from another angle and think Hmm not so sure anyway subscribe that is the next video and of course dont forget the sponsor World of Tanks they have made all of this possible of course if you use your link in the description and your code for his army you get your credits seven days premium access only free free tanks All Above for about and its all free to you go check it out right whats gonna happen Colin I dont know Applause

write a youtube script

this is microsoft flight simulator the latest and possibly greatest flight sim out today with amazing graphics for our eyes to enjoy and a d generated version of our world to explore it feels like the makers of the sim wanted us to fly around in small bush planes to soak up the low altitude scenery but i feel my old joystick needs an upgrade after looking at a few new joysticks they all seem similar in terms of overall design with the short pivot length table mounted style for sure the quality of parts will be different but i want something a bit more this video is sponsored by kiwiko more about them later i set about d printing some bearing mounts which attach to some aluminium extrusions i could then use millimeter shoulder bolts as the axles to remove any potential play and the same could be done with the other axis to produce a two axis gimbal that will be the main mechanical pivot of the joystick but how will this control a virtual aircraft in the simulator first we need to measure the angle of the joystick pivots and one possibility is to use a potentiometer potentiometers work by connecting a positive and negative wire to these pins which in this case supplies five volts then the third pin outputs a voltage between zero and five volts depending on the angle of the potentiometer shaft but there is a problem with this because i want this joystick to hinge at the ground the total rotation of the pivot isnt much before the joystick will hit my leg in fact its only degrees in either direction and if i rotate this potentiometer by degrees the voltage change is less than half a volt resulting in a very low measurement resolution it seems the expensive joysticks use a magnetic technology which uses something called a hall effect sensor now ive used hall effect sensors before to measure the exact angle of a motor however the cost of the two sensors plus another for the rudder pedals will be more than what i paid for my old joystick but then i realized i could use these tiny hall effect sensors which are a fifth of the cost they work in a similar way to the potentiometer where there are three pins one for positive one for negative and one that outputs a voltage depending on the nearby magnetic field so if you place it by a north pole of a magnet it outputs volts and if you place it by a south pole it outputs volts the only issue with this is the north and south poles of a magnet are degrees apart so wed only be using about six percent of the potential resolution so instead im going to use these cube magnets by mounting them slightly off center from the pivot with the north and south poles facing either side of the sensor a small degree rotation achieves the full north to south pole magnetic field change resulting in a to volt output resolution once the other axis was mounted i soldered the wires to an arduino micro to read the analog signals then using an arduino joystick library and a few lines of code the board can emulate a games controller and be set up to fly a plane the next step was to add springs to the joystick so it will selfcenter and these springs can be changed to achieve a desired resistive feel i could then add an extension to the joystick to raise it up to my chair height which has to be angled to avoid colliding with my chair when holding full up elevator i decided to go with a basic cylindrical grip to resemble a classic bush playing control but being d printed it can be swapped out for any other design and the large red button can be assigned to a desired control output which will most likely control the brake as microsoft flight simulator doesnt currently support missiles i then attached another aluminium extrusion to stand up vertically which would be the mount for the throttle control as well as a few other switches i decided to use a potentiometer for the throttle as the rotation range is much larger so the resolution doesnt need to be as precise this also meant it was very simple and compact to build which fit nicely inside a small case it might not be the most realistic looking throttle control but its very comfortable to use with just one hand and i can toggle the switches easily too the rudder pedals can then be attached to the same aluminium extrusion as the joystick which prevents the controls from moving around on the floor as well as making the distances between the controls adjustable the two pedals were linked using these ball joint rods that attach to another d printed part which will eventually house the hall effect sensor this means the rudder pedals move in opposite directions and have limited end stops to prevent over rotation and i can add some springs so they always selfcenter like the joystick the foot pedals are again a basic cylindrical design with a long five millimeter bolt down the center for added strength but this can always be modified if needed theres just one issue the pivot which is supposed to house the hall effect sensor rotates more than the degrees of the joystick pivot so it needs a slight modification instead of buying a different magnet or using a potentiometer i realized i can use a north and south pole of two separate magnets and as long as theyre not too far apart the hall effect sensor will still output a linear value all the switches and hall effect sensors can then be wired up to an arduino board and mounted inside of a small d printed case to keep the setup neat so i now have a custom homemade flight simulator joystick complete with external throttle and rudder pedals its a slightly different design to a conventional joystick but how does it perform in terms of precision latency adjustability and also the cost because im using an arduino micro to read the analog signals from the sensors the precision is ultimately limited by its input resolution which is different values this may seem like a high resolution for a degree pivot but some of the expensive joysticks boasts resolutions over also when comparing it to my old joystick which apparently has a resolution of different values its quite a bit more smooth and because the magnetic field is linear throughout the joystick rotation it essentially has zero deadband in the center which i cant say the same for my old joystick i then set up my high speed camera at frames per second to film both the joystick and the hz monitor to measure the joysticks latency this way i could measure the time from when i started to move the joystick to the moment the monitor started to update the position of this slider after averaging the eight measurements the measured latency was milliseconds however after carrying the same test with my old joystick its latency was milliseconds which both seemed pretty slow so i also tried my mouse which in the flight simulator menu had a latency of milliseconds so its milliseconds slower than my mouse which i guess isnt bad for a joystick and ive honestly never noticed any delay in the controls even when flying the most aerobatic planes Music i ultimately designed this setup for you guys to build at home which is why everything uses these adjustable extrusions to make it customizable but before building one of these yourself im sure you want to know the cost this whole setup cost me pounds which is expensive if youre just looking for a joystick and throttle however including rudder pedals it makes it a lot cheaper as even some of the budget runner pedals can cost you an extra pounds and sometimes its fun to build your own stuff as you might learn something new but if youre not going to build this joystick how about building a crate from kiwico im sure some of you were like me as a kid where my favorite part of the science museum was browsing through the gift shop dreaming of what project i could build next which spurred my curiosity and ultimately led me to become an engineer and now that kiwico provides that and more is an amazing opportunity for kids to learn something new their super cool handson projects are designed to expose kids to concepts in science technology engineering art and math and are an excellent resource for learning at home im always impressed by how well the crates are put together as they require no extra supplies everything needed is already included as well as a detailed magazine about the project so once a project is built its easy to understand how it works the crates are also an excellent gift option for the holidays and the kiwico subscription will keep the fun and learning lasting all year long so check out kiwikos excellent crates at kiwico.com forward slash tom for off your first month of any crate the link will be in the description down below thanks to kiwiko for sponsoring this video and thank you very much for watching this video also if you want to build this flight simulator setup i have a video over on my second channel which goes through all the electronics and all the code as well as the d printer file so that you can build your own one at home id like to thank you very much for watching this video if you enjoyed the video be great if you could leave a thumbs up down below if youre new to my channel and want to see other projects similar to this then please click subscribe down below and a massive thank you to all of my supporters over on patreon.com for making these projects possible i honestly couldnt do without your support so thanks once again thanks once again for watching and ill see you in the next video goodbye

write a youtube script

Music Applause Music do you remember when I made a bicycle with an omni wheel on the front wheel it baled like a Segway sideways but you could still ride it like a normal bike because the front wheel had lots of skateboard wheels all around it so it would kind of slide along sideways a while later I moved on and tried to make a bike with two Omni Wheels instead of building the electronics myself though I took the guts out of a hoverboard including the motors I tried to Velocity match the original Wheels with the new larger Wheels using belt drives which kind of worked okay but then the only way to propel it was with edfs blowing it along which wasnt very successful as soon as Id assembled that version I realized it would have been better to have made mechum wheels which have the little wheels slanted all around them these are typically used on the four corners of a robot so it can move in any direction but I wanted to put them all in a line so the bike still balances sideways but it can screw itself along I actually found a YouTube video from Matthew Watson where this method was used to drive a small robot I believe this was a PhD project and it works very well indeed I cant find any other versions of this type of locomotion and I certainly cant find any versions that are B enough to ride on so yeah you can buy mechum wheels off the shelf but the biggest I can find is mm in diameter which is really small and theyre really expensive like and I already made two massive Omni Wheels so I guess Im going to have to make four massive mechum Wheels yes of course Im going to D print most of the parts Im building a scaled up version of my mechum wheels from the smaller robot I showed earlier each wheel is going to be around mm in diameter and itll have little wheels around its circumference so I need little wheels in total just a quick add from my D printing sponsor thanks to lsot for supporting my Channel with D printers Im printing the tires in a TPU from D Fuel and the wheel hubs in D fuel proa plus you can now get off D fuel products with my special code and Ill get a small commission Music too Music theres also wedges to be made to put on each end of the axles to hold the wheels at a slant but Ill save you the footage of printing every single one of them I printed the TPU Wheels with a . mm nozzle on LSB printers and theyre incredibly tough and that means its very very hard to push the hubs in so its time for the hydraulic press Music Applause Music thanks to Simply bearings for the bearings for this project Ive got skate sized bearings for each end of the Wheels as well as of course some bigger bearings for the actual big wheels themselves so we need to make some wheel hubs to mount all this on so as usual Im going to use some plywood CN seed out so weve got some holes here being drilled in a massive Circle and thats so we can attach all of those wedges around and we also need some more holes to go and align all of the parts together with a hub in the middle and make sure those wheels are in the right place so weve cut out some circles and Ive also got some htd eight profile pulley and one of those is going to drive each wheel so Ive painted all of those black with some automotive spray paint and you notice theres some very shiny sections on there and thats actually masking tape which Ive put over where the four holes are to mount each of the wedges and thats so I can glue those on and theres something to stick to other than paint so Ive got some solventbased adhesive to stick each of those on and Im also going to screw those on as well so basically those never ever come off because Im going to be sitting on this and that leaves us with a piece of round Wood with wedges glued and screwed to it on the other side weve got the ht pulley and Ive also made a thin plywood retainer so the Belt doesnt fall off and thats going to be used to drive the wheel this Hub fits onto a D printed part which is printed with a . mm nozzle so thats a really tough D printed Hub and Ive got bolts that go all the way through which are a pretty tight fit well just knock all of those in and eventually that will hold the other side on as well for now we need to get all the wheels in so weve got these mm pieces of stainless steel rod which go into those wedges all the way around to get the spacing right Ive got a couple of nylon washers and then we can put all of the little wheels on with their bearings all the way around followed by another two nylon washers on the top of every one that just makes it perfectly spaced so they dont move weve got another set of wedges which go all the way around there and then you can probably guessed that with those bolts sticking out were going to stick another plywood Hub on top of that and then were going to glue and screw all of those wedges to it on the other side and that makes the complete wheel I did turn the bolts around those so that the piece is sticking out with the nuts on her in the nice recess that I left we need to fiz a bearing in each end of the wheel but the Hub was printed with a massive nozzle so I printed another piece with a finer nozzle a half mm nozzle that makes it much easier to tolerance and we can reprint that part if we need to without reprinting the whole Hub so that we can fit the bearings in nice and tightly and those have got mm internal diameters Ive made four of those in total and youll notice I remember to make opposite pairs so that the wheels go in opposing directions so they can screw against each other to travel in the same direction otherwise balancing is going to be really hard so what we need to do now is put them all on a chassis and put some Motors on its chassis Music time Music Applause Music so thats my main chassis and these four pieces sticking up as I hold the four Motors so actually the wheels go below this and this weird saddle thing here is where were going to put foot pegs so we need to put the wheels on this piece of bright steel bar which is hung below that and we need to make it all Music adjustable so Ive got some mm Rose joints with the mm thread to match the studing and mm holes in to match my axle and all of these are going to go through that chassis crossbar and be bolted on so that we can adjust the height of them and we can get that axle through perfectly all through all of them threading the eye of the needle every time and then we can get all the wheels aligned and spaced correctly so now its time to put the wheels on which is a bit of an ordeal that mm bright steel bars quite tightly toleranced for the bearings I can do it but its going to take me quite some time to push it through and push all the wheels on not forgetting of course to put the drive belts on first otherwise Ill never be able to get them Music on theres a color clamp on the end and therell be more of those along the axle to brace the axle against the rose joints throughout the length of the bike so now all my wheels are all perfectly in line and they all run fine and Ive done up all the nuts on that studing so that we get everything at the right height so that the belts can mesh with the motor drives so we do need some spacers um at the moment those can slide up and down but we need to align them perfectly with the motors you might have noticed I actually put the wheels on in the wrong order so I had to take them all off and do it again but even without spaces there if I slide this up and down lengthways we can see those wheels turning against each other just like my little animation so Im very happy with how the alignments gone so far its time for some motors do you remember the video I did with Colin Furs where I built a version of his rhino tank and he helped me test it yeah thats what we was made for go through the mud well Im using the same Motors and drivers from that project these are o drive KV brushless motors with an O Drive S on the back of them theres a magnet on the motor and an encoder on the board so the position and velocity or Torque can be accurately controlled this is a really neat unit and I built a twostage belt reducer for the tank this time were going to need a bit more velocity though so Ive doubled the size of the output pulley that drives the wheels the motor and reduced assembly fit onto the risers I made on top of the main chassis crossbar and the belts attenion with blocks underneath each assembly I made a quick stand out of steel for it so I can lean it up and work on it upright and Ive also fitted the top of the chassis on and thats going to hold the handlebars the saddle and the electronics and that is a bolted on piece so that I can take it off to get to those motor assemblies if I need to in the last domy bike I made a steering assembly which basically would lean sideways as well as rotating and thats what I used to control all of the axes however there was nothing solid to hang on to so that was a bit of a mistake in the end and resulted in me crashing so this time Ive just made the handlebars which are completely fixed and dont turn at all I found this saddle on Amazon which fits just right so Im pretty happy of the way its looking its quite retro right its going to be one of the best things Ive ever made if it works but we need to put some Electronics in and get the motors powered up and see what happens but before we carry on assembling that its time for a quick add from the video sponsor which is ODU Odo is an allinone management tool that provides entrepreneurs with a wide range of webbased applications to simplify the management of their business the first ODU application is free for life with unlimited hosting and support they also offer a free personalized domain name for one year once you have two or more applications youll switch to a paid plan the OD website builder application is a powerful tool for creating a website quickly easily efficiently free of charge the website configurator guides you through defining your goals choosing the color palette and inserting your logo adding pages and features and choosing your theme once the structure is in place you can start customizing simply drag and drop blocks to customize your website you can customize each block adding content including images text video documents and functionalities like animations on blocks and images changing typography colors and image shapes the grid lets you easily lay out and organize your content blocks AI generated text is available in ODU to populate any sections I quickly put together a website about this project and you can try out ODU Now by visiting the link in the video description right lets get on with the screw bike mechum wheel Omni directional Co linear drive bike project so Ive made an Electronics enclosure and in there weve got a Tec . weve got a spark fun b IMU and weve got the canbus transceiver to control all the O drives on top Ive got a battery monitor an LED strip which is actually at the moment visualizing the IMU data Ive also got various knobs on there so weve got a trim control which is for the inertial measurement unit data and Ive also got a switch to initialize the O drives and Ive got a volume control well talk about later Ive got connectors coming out including the canbas USB power the volt power so it can monitor the battery and some connector to go to the handlebars for all of the switches Ive made a battery box which holds four s lipo and were going to put those in series and then pairs in parallel so we basically get nearly volts out of that theres a amp contactor on the top of there and thats going to be linked to the e stop on The Handlebar so that I can actually cut the power off if I need to the electronic sit right in front of the saddle there so theyre nice and accessible to control yes thats a USB boost pank powering them nice and professional the gray cable is can bus thats daisy chain between all of the O drives to control them and all those batteries sit at the back and yes I do have a cover there to cover the dodgy bolts which has got volts on Ive built quite a few balancing robots in the past theres a playlist in my channel just for balancing projects that use various means to make unstable things stable the main thing that makes balancing robots work is a p ID controller this takes the angle data from the inertial measurement unit it controls the wheel velocities appropriately the proportional term simply means that the wheels go faster if the angle is greater but this isnt sufficient to make the device balance properly the most important part is the integrator the output of the integrator is the area under the graph so if the angle is constant as time goes on then the output increases as this area increases over time if the angle is increasing then the output gets bigger exponentially so the integrator makes the wheel velocity accelerate to catch the robot as it falls over which happens constantly the derivative term is the tangent of the curve of the angle data so this makes a short sharp velocity movement to damp any overshoot all three terms need to be tuned for the mass and power of the robot which is a bit like focusing a camera with three lenses but typically the highest value is the integral term and the smallest is the derivative however if the device is left at an angle without the motors powered on then the integral term tends to accumulate a large value and when its finally time to balance then it makes the robot suddenly kick until its stable this was an issue with the first Omni bike so this time Ive got a control so I can slowly turn up the output which makes it less likely to break my ankles when I power it up the button initializes the O drives and now I can slowly ramp up the output of the P ID controller to make things balance without it kicking so much it was pretty easy to tune up but bigger heavier robots tend to thrash around less because theyve got more inertia so its much easier to tune the P ID controller to make them balance right lets put the saddle on I have in fact tuned this up to have a person on rather than no person so this should be working pretty well and Im pretty happy of how well its working is almost as good as a Segway oh yeah this is really good this is so much better than the last one made of a hacked hoverboard mainly cuz I couldnt retune those P parameters on the hoverboard so even just a slight bit of uh mass shift causes this thing to travel which is good its almost as good as a Segway Ive definitely got enough velocity for balancing there but what we need to do now is try and screw it along by moving those Wheels in opposite directions and if that works then this is going to be pretty cool so Ive got a Twist grip on the right hand side which moves in one Direction and Ive also got a toggle switch so that I can drive in forward or reverse and all thats going to do is make the wheels turn in opposite directions to screw itself along and transpose that value over the balancing value right lets try and uh go forwards oh it works oh yeah this is brilliant this is brilliant oh yes way I think we might be slipping slightly on the smooth floor but this feels amazing rotating is controlled by another twist grip ideally I want a Twist grip that rotates forwards and backwards so I can turn left and right but I couldnt find one that does that so Ive just got another Direction switch and well see how that goes I did think about having foot switches for steering but I thought probably this is going to fall over at some point and theyll get buckled or broken if theyre attached to the foot pegs so I decided handlebar controls were best now of course the wheels are different distance is from the center and that means that they need to rotate at different velocities so that we can rotate around the center point basically one circumference is bigger than the other so the outside Wheels need to go faster we need to take that into account right lets see where rotating gets us oh oh oh oh I think I want my mass in the middle since thats where the rotation point is yeah that seems to be that seems to be

write a youtube script

AnyTaskTune Advanced DomainSpecific Solutions through TaskFineTuning DataTager Abstract The pervasive deployment of Large Language ModelsLLMs in various sectors often neglects the nuanced requirements of individuals and small organizations who benefit more from models precisely tailored to their specific business contexts rather than those with broadly superior general capabilities. This work introduces AnyTaskTune a novel finetuning methodology coined as TaskFineTune specifically developed to elevate model performance on a diverse array of domainspecific tasks. This method involves a meticulous process to identify and define targeted subtasks within a domain followed by the creation of specialized enhancement datasets for finetuning thereby optimizing taskspecific model performance. We conducted comprehensive finetuning experiments not only in the legal domain for tasks such as keyword extraction and sentence prediction but across over twenty different subtasks derived from the domains of finance healthcare law psychology consumer services and human resources. To substantiate our approach and facilitate community engagement we will opensource these bilingual task datasets. Our findings demonstrate that models finetuned using theTaskFineTune methodology not only achieve superior performance on these specific tasks but also significantly outperform models with higher general capabilities in their respective domains. Our work is publicly available at httpsgithub.comPandaVTDataTager. Introduction Recently the rapid advancement and deployment of Large Language Models LLMs have transformed various sectors by providing unprecedented natural language processing capabilities . Traditionally these models have been developed with a focus on enhancing their general abilities aiming to create universally powerful tools that excel across a broad spectrum of tasks and domains. However this generalized approach often fails to meet the specific requirements of individual users and small organizations whose needs are intricately tied to their unique operational contexts. To address this issue one very simple way is to finetune LLMs for different tasks. In the medical field researchers have optimized LLMs using medical datasets to support functionalities such as medical document interpretation and diagnostic consultations. For example ChatMed DISCMedLLM HyKGE IvyGPT and HuatuoGPT are notable examples demonstrating significant advancements within their specialized medical domains compared to generic LLMs. Similarly in the legal field researchers have introduced finetuned LLMs trained on legal corpora to support activities such as legal research contract analysis and legal document summarization. Models like ChatLaw LawGPT and DISCLawLLM exemplify this trend demonstrating a deeper grasp of legal language and principles compared to their generic counterparts. Within finance researchers are leveraging diverse financial datasets including conversational data and market reports to finetune LLMs for applications such as financial forecasting risk assessment and fraud detection. Examples include FinGPT DISCFinLLM and PIXIU . Additionally in other fields models like Please cite this work as DataTager. Full authorship contribution statements appear at the end of the document. Correspondence regarding this technical report can be sent to reportdatatager.com Preprint. Under review. arXiv.v cs.CL Jul FaiMA designed for Multidomain applications and Kuaiji tailored for accounting tasks illustrate the versatility and potential of finetuned LLMs in specialized domains. To address this discrepancy we introduce a novel finetuning paradigm specifically designed for Explicit Data Sets which we term task fine tune. This method diverges from conventional training techniques by emphasizing precision and specificity over general performance. By focusing on Explicit Data Sets that contain clear directive inputoutput pairs with specific instructions this approach allows for precise tailoring of models to perform designated tasks effectively. Such finetuning not only enhances the models ability to execute particular functions but also significantly improves its applicability and efficiency in realworld scenarios. The cornerstone of this work AnyTaskTune is to validate and elaborate on the task fine tune method utilizing Explicit Data Sets across various domains. This method involves identifying the spectrum of tasks required in different fields and scenarios and meticulously developing numerous Explicit Data Sets to conduct taskspecific finetuning. The aim is to optimize the models performance for individual and corporate use ensuring that it meets the nuanced demands of users in fields such as finance healthcare law psychology consumer services and human resources. To showcase the effectiveness and versatility of this approach we conducted experiments using over twenty different Explicit Data Sets that we will opensource. These datasets are bilingual and designed to cater to a diverse global audience providing a valuable resource for the community to engage in further research and enhance applicationspecific model improvements. By concentrating on specific clearly defined tasks rather than general capabilities AnyTaskTune significantly outperforms traditional models that prioritize breadth over depth. This strategic shift towards more specialized taskoriented model training represents a critical evolution in how we develop and deploy LLMs making them more relevant and valuable in industryspecific applications. The primary contributions of this paper can be summarized as follows Introduction of the task fine tune paradigm specifically designed for Explicit Data Sets emphasizing precision and specificity over general performance. Development and utilization of over twenty bilingual Explicit Data Sets across various domains to validate the effectiveness of the task fine tune method. Evaluation of multiple models and datasets across various domains demonstrating that taskspecific finetuning yields better results. Opensourcing of the developed Explicit Data Sets providing a valuable resource for the community to engage in further research and enhance applicationspecific model improvements. Implementation of our data processing model DataTagerLLM as an online accessible website httpsdatatager.com. AnyTaskTune . Model Requirements in Business Applications In practical applications simply pursuing the performance of a general model or deploying a domainspecific model alone is often insufficient to address realworld problems. While scientific research outputs need to be diverse and comprehensive businesses require standardized and controllable solutions. For most enterprises and organizations their needs are highly specific and contextualized which cannot be fully met by a single general language model. In such cases general models may underperform on specific tasks or fail to achieve the expected efficiency and accuracy. Moreover while some domainspecific models may provide good performance within their fields they lack flexibility and scalability making it difficult to adapt to everchanging business needs. Therefore there is a need for a new method to finetune and optimize models ensuring they not only possess the broad applicability of general models but also meet the specific needs of particular domains and tasks. AnyTaskTune addresses this issue by using specially designed explicit datasets for task finetuning ensuring the models precision and efficiency in specific tasks thereby enhancing its value in realworld business applications. Fields DataSets DataTager LLM Traditional Interaction Mode action action action action a. First xxx b. Second xxx c. query summary Field actiona actionb actionc actiona Task Pool a b c a b c d action action DataSet actionaLM actionbLM actioncLM actionaLM actionLM actionLM Mixed LLM action action action action a. TaskLM b. TaskLM c. TaskLM AIpowered Interaction Mode TaskLM Analysis Figure AnyTaskTune Framework . DataTager First leveraging our prior experience and expertise in vertical domains we collected raw data from multiple fields including finance healthcare law psychology consumer services and human resources. By surveying industry professionals to understand their expectations and requirements for models we synthesized multiple branch datasets. These datasets cover core tasks across various domains and include detailed and variant data from realworld scenarios. Through the training of these branch datasets we developed DataTagerLLM a largescale data synthesis model. DataTagerLLM forms the foundation of our product and serves as a core tool for further research and application. Currently we have no plans to opensource DataTagerLLM to better protect our intellectual property and commercial interests. . Explicit and Implicit DataSets We defined what constitutes a good dataset for businesses and distinguished between these datasets based on the presence of explicit instructions. Specifically datasets with clear instructions such as instruction Please summarize this news article and extract the key points input xxx are designed for models to perform specific tasks. Explicit datasets provide models with clear guidelines enhancing their ability to understand and execute specific tasks accurately and efficiently in practical applications. On the other hand datasets without explicit instructions such as instruction input xxx involve random queries without fixed task types thus considered implicit datasets. Implicit datasets are better suited for handling openended questions and unstructured data but are more challenging to train and apply. For businesses looking to enhance their operations with LLMs the first step is to identify the various task types within their current operational context. This helps determine the number of explicit datasets needed. Each dataset is then synthesized individually and a general LLM that aligns with the businesss usual interaction patterns is finetuned across multiple tasks to create a model that best fits the specific business scenarios. . AnyTaskTune Different fields have multiple interaction action modes. Take the medical field for example if triage is considered an action the actual tasks it involves include a. Patients describing their issues in nonstandard language b. Triage staff translating this into medical standard language c. Staff analyzing the completeness of the information and asking followup questions if necessary d. Extracting medical key points and directing the patient to the appropriate department. This triage action thus includes four subtasks a b c d. We can use DataTager to generate datasets for each Instruction Please extract three to six key medical information from the users description. Input Ive also been caught up in the flu wave these past few days. Now Ive been coughing for almost three months without getting better. I want to ask what should I do if Ive been coughing for nearly three months? Output Symptom Description Longterm cough lasting for three months Duration of Symptoms Three months Possible Cause Initially may have been caused by a cold infection Help Sought Seeking a solution for a persistent cough Explicit DataSet Implicit DataSet Instruction Input Ive also been caught up in the flu wave these past few days. Now Ive been coughing for almost three months without getting better. I want to ask what should I do if Ive been coughing for nearly three months? Output Im sorry to hear about your persistent cough. A cough lasting three months is considered chronic and may require medical evaluation to determine the underlying cause. Here are some general steps you can take .Consult a Doctor Its essential to see a healthcare professional. They can perform a thorough examination and may order tests such as chest Xrays spirometry or blood tests to Figure Explicit and Implicit DataSets subtask. These datasets collectively represent the triage action process. Then we train models based on these datasets with multiple approaches available If we train a separate model for each subtask the model can address that specific subtask effectively. For instance a model specifically trained for subtask a can handle the translation of patient descriptions while another model trained for subtask b can convert nonstandard language into medical standard language. If we combine these subtask datasets to train a large mixed model different instructions will trigger different task types. This method can handle multiple tasks within one model but requires a more complex training process and more computational resources. It is important to note that if we choose to train multiple small models based on our experience a . billion or billion model can handle less complex tasks well without data drift. If the performance of the mixed model declines adding some general datasets to the training process can help. This way we can use highperforming models to replace traditional subtasks in action interactions creating a new interaction model and reducing costs. Experiments To validate the effectiveness and versatility of the AnyTaskTune paradigm we undertook extensive experiments across multiple domains finance healthcare law psychology and roleplay. We compared the performance of AnyTaskTune against various models including closedsource large language models LLMs opensource LLMs and domainspecific models. Critically our experiments maintained a strict separation between training and testing datasets to ensure unbiased evaluation and reproducibility of results. . Experimental Setup Our experiments were structured as follows Model Base We utilized QwenB as the base model for AnyTaskTune training. This model was finetuned on specific subtasks within each domain and then tested to obtain the experimental results. Model Categories Closed Source LLMs GPT GPT. LLaMAb and LLaMAb. Open Source LLMs Qwen Baichuan ChatGLM and internalLM . Domain Models ChatMed DISCMedLLM FinGPT ChatLaw DISCLawLLM MindChat SoulChat and MachineMindset MBTI . Evaluated Domains and Tasks Medical Subtasks M M M. Finance Subtasks F F F. Law Subtasks L L L L. Psychology Subtasks P P. RolePlay Subtask MBTI. . Results and DomainSpecific Analysis The results of the experiments are summarized in Table . Our AnyTaskTune models finetuned on specific domain data demonstrated significant performance improvements even sometimes surpassing wellregarded models like GPT and LLaMAb. The following points detail crossdomain influences and taskspecific adaptabilities observed during testing Table Performance Comparison Across Various Domains and Models Med Finance Law Psychology RolePlay M M M F F F L L L L P P MBTI Closed Source LLMs GPT . . . . . . . . . . . . . GPT. . . . . . . . . . . . . . LLaMAb . . . . . . . . . . . . . LLaMAb . . . . . . . . . . . . . Open Source LLMs QwenB . . . . . . . . . . . . . BaichuanB . . . . . . . . . . . . . ChatGLMB . . . . . . . . . . . . . internLMB . . . . . . . . . . . . . Domain Model ChatMed . . . DISCMedLLM . . . FinGPT . . . ChatLaw . . . . DISCLawLLM . . . . MindChat . . SoulChat . . MachineMindsetMBTIGPT . Task Fine Tuned Model AnyTaskTuneQwenBMed . . . . . . . . . . . . . AnyTaskTuneQwenBFinance . . . . . . . . . . . . . AnyTaskTuneQwenBLaw . . . . . . . . . . . . . AnyTaskTuneQwenBPsychology . . . . . . . . . . . . . AnyTaskTuneQwenBRolePlay . . . . . . . . . . . . . . Detailed Comparative Analysis .. Impact Across Domains Medical Domain The AnyTaskTuneQwenBMed model showcased exceptional performance in medical subtasks notably achieving a F score of . in M which is a substantial improvement over the base QwenB models score of .. This precision underscores its reliability in handling medical data. However when this model was tested in finance and law domains its performance dropped to . and . respectively illustrating a significant decline due to the models parameters being highly specialized to medical data. Financial Domain AnyTaskTuneQwenBFinance performed excellently in financial tasks especially F and F with scores of . and . outperforming the base models scores of . and . in these tasks. Applying this model to legal tasks resulted in moderate performance drops from . in legal tasks down from . in finance tasks likely due to some overlap in structured data and terminological precision between the finance and legal fields. However performance in the medical domain was notably poorer confirming the domainspecific nature of the finetuning. Legal Domain The AnyTaskTuneQwenBLaw excelled in legal tasks with scores of . . . and . across L L L and L. Its performance in the finance domain was competitive with scores slightly lower but still respectable. However when applied to the medical and psychological domains the performance was limited further highlighting the challenges of crossdomain applicability due to different knowledge bases and data characteristics. Psychology and RolePlay Domains The AnyTaskTuneQwenBPsychology model showed good results in psychology particularly in the P task with a score of .. The nature of psychological tasks which often involve complex language and emotional analysis allows this model to perform reasonably well in roleplay scenarios as well evidenced by a score of . in the MBTI task. RolePlay Domain Similar to the psychology model AnyTaskTuneQwenBRolePlay demonstrated good performance due to overlapping requirements in emotional and language processing between the two fields. .. Task Adaptability and Generalizability While domainspecific finetuning led to exceptional performance within respective fields it generally resulted in reduced effectiveness when models were applied across unrelated domains. For instance while AnyTaskTuneQwenBMed excelled in medical tasks its application in nonmedical fields showed significant performance reductions a clear indication of the tradeoffs involved in specialized finetuning. Comparison with Base Model The original QwenB model without finetuning generally showed lower performance across all specific tasks compared to the finetuned AnyTaskTune models. This highlights the effectiveness of taskspecific finetuning in enhancing model capabilities although it restricts the models flexibility across various domains. . Subtask Details The specific types of subtasks evaluated in this experiment will be detailed in the accompanying GitHub repository. This repository will provide comprehensive descriptions and examples of each subtask offering valuable insights for further research and application development. Explicit Implicit Input Can you recommend some movies similar to Inception? Output If you enjoyed Inception you might like Interstellar The Matrix Memento and Shutter Island. These films also explore complex narratives and mindbending concepts. Input How can I reduce my home energy consumption during winter? Output To reduce energy consumption in winter consider insulating your home using programmable thermostats to reduce heating when not needed and sealing leaks around windows and doors. Additionally upgrading to energyefficient heating systems can significantly lower your energy usage. Instruction Please extract three to six key medical information from the users description. Input Ive also been caught up in the flu wave these past few days. Now Ive been coughing for almost three months without getting better. I want to ask what should I do if Ive been coughing for nearly three months? Output Symptom Description Longterm cough lasting for three months Duration of Symptoms Three months Possible Cause Initially may have been caused by a cold infection Help Sought Seeking a solution for a persistent cough Instruction You are an experienced lawyer. Please help me break down a complex case into several subcases. The subcases should be independent of each other. Input A woman had a sterilization procedure after giving birth to a son in during the era of the onechild policy and now at she wants to assert her rights. How should she proceed? Output . Legality and effectiveness of the sterilization procedure Consider the legality and outcome of the sterilization procedure conducted in to determine if there is a basis for a claim. . Medical malpractice and disputes If there were issues with the sterilization it might involve medical malpractice and disputes requiring consideration of legal avenues and evidence. . Inheritance rights and property division The child has rights to inheritance which necessitates consideration of how to approach rights assertion and issues of property division. Instruction Identify key information within the consultation assess the possible psychological state of the counselee and then provide targeted support and advice. Input How can I relieve my nervousness? I get nervous whenever I go on stage and even when I am called to answer questions in class I start shaking all over and my voice trembles. I dont know how to relieve this. Output Risk level Moderate. nThe individual describes significant nervousness and physical reactions when speaking in public which may indicate social anxiety. It is recommended to practice relaxation techniques such as deep breathing mindfulness meditation or progressive muscle relaxation. Additionally seeking help from a psychologist is advised to explore the root causes of anxiety and learn more effective coping strategies. Instruction Convert lengthy or complex consultation texts in the finance field into concise structured questions. Input How long does it take for the transfer to arrive and can I track the progress? Output How long does it usually take for the transferred funds to be credited to the account? Is there a service to track the progress of the transfer? Instruction Please extract the emotional characteristic descriptive words from users ecommerce reviews. Input Good Output Good recommendable great value for money Input This is the second bottle I doubt if its a fake because the packaging has changed the new packaging feels low quality greasy after use the first bottle was refreshing so I repurchased but the second bottle is completely different causes acne after use greasy! Regardless of authenticity I wont repurchase Output Fake lowquality packaging greasy causes acne will not repurchase Instruction Please rate different aspects of the hotel based on the following reviews and summarize the users intended meanings. Input A peaceful place amidst the hustle you can see nice scenery from the window. The hotels price is indeed a bit high Output Peacefulness Score the review mentions the hotel as a peaceful haven amidst the hustle indicating a very positive attribute that can attract guests seeking tranquility Scenery Score the view from the window is praised showing that the hotels location and design offer beautiful views enhancing the attractiveness of the stay High price Score the review notes that the hotel is pricey which could be a barrier for budgetsensitive guests despite potentially highquality services or facilities offered price remains a crucial factor for guests. Figure Task examples . Analysis This analysis confirms that while AnyTaskTune can dramatically improve performance on specific tasks its crossdomain applicability is limited. This highlights the importance of taskspecific datasets in finetuning processes ensuring that models are not only highly effective but also tailored to the particular characteristics and requirements of their intended operational domains. Conclusion In this work we introduced AnyTaskTune a novel finetuning paradigm specifically designed for Explicit Data Sets. Our approach addresses the limitations of traditional general models and domainspecific models by emphasizing precision and specificity which are critical for realworld applications. By leveraging our extensive experience in various vertical domains we developed and utilized over twenty bilingual Explicit Data Sets enabling us to finetune models for specific tasks with high accuracy and efficiency. Our experiments demonstrated that AnyTaskTune significantly outperforms both closedsource and opensource LLMs as well as existing domainspecific models across a range of tasks in finance healthcare law psychology and roleplay. The results showcased the superior performance of AnyTaskTune in handling complex and nuanced tasks thereby validating the effectiveness of taskspecific finetuning. Moreover we highlighted the practical implications of AnyTaskTune for businesses and organizations. By defining clear and directive inputoutput pairs through Explicit Data Sets we enable precise model tailoring which not only enhances task performance but also improves the models applicability and efficiency in specific operational contexts. Our methodology ensures that models are not just broadly capable but are finely tuned to meet the exact needs of various business scenarios. We also introduced DataTager our foundational data synthesis model which forms the basis for generating highquality datasets used in AnyTaskTune. While we have not opensourced DataTagerLLM we have provided comprehensive documentation and resources in our accompanying GitHub repository detailing the subtasks and datasets used in our experiments. This transparency aims to foster further research and development in the community. In conclusion AnyTaskTune represents a significant advancement in the development and deployment of Large Language Models. By shifting the focus from general capabilities to specialized taskoriented model training AnyTaskTune offers a robust and efficient solution for industryspecific applications. This strategic evolution in model finetuning not only enhances performance but also ensures that LLMs are more relevant and valuable in realworld business contexts. We believe that AnyTaskTune will pave the way for more precise and effective applications of AI in various domains ultimately contributing to the broader adoption and integration of AI technologies in everyday business operations. Authorship Credit Attribution and Acknowledgements Please cite this work as DataTager. Authorship and Contributions Paper Writing Jiaxi Cui Founder Provided the main ideas data construction model training paper writing. Wentao Zhang Organized ideas and outline and contributed to paper writing. Engineering Xudong Tong Software engineering at DataTager. Zhenwei Zhang Tencent Software engineering at DataTager. Other Contributions Jing Tang Huazhong University of Science and Technology Responsible for external presentations and resource acquisition. Amie PublicAI Beihang University Web and external resource acquisition Jing Wen Provided industry insights. Rongsheng Wang Qiyuan.Tech Provided assistance in the medical field. Pengfei Wu Peking University Tencent Offered partial assistance. Acknowledgements We would like to express our sincere gratitude to the following individuals and organizations for their invaluable support and contributions to this project Modelscopes Chen Cheng and HuggingFaces Tiezhen Wang for their tremendous support. PublicAI our Web partner for providing data support. SciLearning our channel partner for reaching student groups. YuanGroup for supporting the free exploration of early ideas. Fudan NLP team for their related research contributions. The opensource community and contributors to the various software libraries used in this project. Your dedication and hard work are deeply appreciated. Special Thanks A special thanks to the opensource community and contributors to the various software libraries used in this project. Your dedication and hard work are deeply appreciated. Contact Information For further information or inquiries please contact us at reportdatatager.com. Thank you for your interest and support in our work. The DataTager Team July References J. Bai S. Bai Y. Chu Z. Cui K. Dang X. Deng Y. Fan W. Ge Y. Han F. Huang B. Hui L. Ji M. Li J. Lin R. Lin D. Liu G. Liu C. Lu K. Lu J. Ma R. Men X. Ren X. Ren C. Tan S. Tan J. Tu P. Wang S. Wang W. Wang S. Wu B. Xu J. Xu A. Yang H. Yang J. Yang S. Yang Y. Yao B. Yu H. Yuan Z. Yuan J. Zhang X. Zhang Y. Zhang Z. Zhang C. Zhou J. Zhou X. Zhou and T. Zhu. Qwen technical report . Z. Bao W. Chen S. Xiao K. Ren J. Wu C. Zhong J. Peng X. Huang and Z. Wei. Discmedllm Bridging general large language models and realworld medical consultation . Z. Cai M. Cao H. Chen K. Chen K. Chen X. Chen X. Chen Z. Chen Z. Chen P. Chu X. Dong H. Duan Q. Fan Z. Fei Y. Gao J. Ge C. Gu Y. Gu T. Gui A. Guo Q. Guo C. He Y. Hu T. Huang T. Jiang P. Jiao Z. Jin Z. Lei J. Li J. Li L. Li S. Li W. Li Y. Li H. Liu J. Liu J. Hong K. Liu K. Liu X. Liu C. Lv H. Lv K. Lv L. Ma R. Ma Z. Ma W. Ning L. Ouyang J. Qiu Y. Qu F. Shang Y. Shao D. Song Z. Song Z. Sui P. Sun Y. Sun H. Tang B. Wang G. Wang J. Wang J. Wang R. Wang Y. Wang Z. Wang X. Wei Q. Weng F. Wu Y. Xiong C. Xu R. Xu H. Yan Y. Yan X. Yang H. Ye H. Ying J. Yu J. Yu Y. Zang C. Zhang L. Zhang P. Zhang P. Zhang R. Zhang S. Zhang S. Zhang W. Zhang W. Zhang X. Zhang X. Zhang H. Zhao Q. Zhao X. Zhao F. Zhou Z. Zhou J. Zhuo Y. Zou X. Qiu Y. Qiao and D. Lin. Internlm technical report . W. Chen Q. Wang Z. Long X. Zhang Z. Lu B. Li S. Wang J. Xu X. Bai X. Huang et al. Discfinllm A chinese financial large language model based on multiple experts finetuning. arXiv preprint arXiv. . Y. Chen X. Xing J. Lin H. Zheng Z. Wang Q. Liu and X. Xu. Soulchat Improving llms empathy listening and comfort abilities through finetuning with multiturn empathy conversations . J. Cui L. Lv J. Wen R. Wang J. Tang Y. Tian and L. Yuan. Machine mindset An mbti exploration of large language models . J. Cui M. Ning Z. Li B. Chen Y. Yan H. Li B. Ling Y. Tian and L. Yuan. Chatlaw A multiagent collaborative legal assistant with knowledge graph enhanced mixtureofexperts large language model . T. GLM A. Zeng B. Xu B. Wang C. Zhang D. Yin D. Rojas G. Feng H. Zhao H. Lai H. Yu H. Wang J. Sun J. Zhang J. Cheng J. Gui J. Tang J. Zhang J. Li L. Zhao L. Wu L. Zhong M. Liu M. Huang P. Zhang Q. Zheng R. Lu S. Duan S. Zhang S. Cao S. Yang W. L. Tam W. Zhao X. Liu X. Xia X. Zhang X. Gu X. Lv X. Liu X. Liu X. Yang X. Song X. Zhang Y. An Y. Xu Y. Niu Y. Yang Y. Li Y. Bai Y. Dong Z. Qi Z. Wang Z. Yang Z. Du Z. Hou and Z. Wang. Chatglm A family of large language models from glmb to glm all tools . X. Jiang R. Zhang Y. Xu R. Qiu Y. Fang Z. Wang J. Tang H. Ding X. Chu J. Zhao and Y. Wang. Hykge A hypothesis knowledge graph enhanced framework for accurate and reliable medical llms responses . J. Luo S. Yang X. Qiu P. Chen Y. Nai W. Zeng W. Zhang and X. Jiang. Kuaiji the first chinese accounting large language model . OpenAI J. Achiam S. Adler S. Agarwal L. Ahmad I. Akkaya F. L. Aleman D. Almeida J. Altenschmidt S. Altman S. Anadkat R. Avila I. Babuschkin S. Balaji V. Balcom P. Baltescu H. Bao M. Bavarian J. Belgum I. Bello J. Berdine G. BernadettShapiro C. Berner L. Bogdonoff O. Boiko M. Boyd A.L. Brakman G. Brockman T. Brooks M. Brundage K. Button T. Cai R. Campbell A. Cann B. Carey C. Carlson R. Carmichael B. Chan C. Chang F. Chantzis D. Chen S. Chen R. Chen J. Chen M. Chen B. Chess C. Cho C. Chu H. W. Chung D. Cummings J. Currier Y. Dai C. Decareaux T. Degry N. Deutsch D. Deville A. Dhar D. Dohan S. Dowling S. Dunning A. Ecoffet A. Eleti T. Eloundou D. Farhi L. Fedus N. Felix S. P. Fishman J. Forte I. Fulford L. Gao E. Georges C. Gibson V. Goel T. Gogineni G. Goh R. GontijoLopes J. Gordon M. Grafstein S. Gray R. Greene J. Gross S. S. Gu Y. Guo C. Hallacy J. Han J. Harris Y. He M. Heaton J. Heidecke C. Hesse A. Hickey W. Hickey P. Hoeschele B. Houghton K. Hsu S. Hu X. Hu J. Huizinga S. Jain S. Jain J. Jang A. Jiang R. Jiang H. Jin D. Jin S. Jomoto B. Jonn H. Jun T. Kaftan ukasz Kaiser A. Kamali I. Kanitscheider N. S. Keskar T. Khan L. Kilpatrick J. W. Kim C. Kim Y. Kim J. H. Kirchner J. Kiros M. Knight D. Kokotajlo ukasz Kondraciuk A. Kondrich A. Konstantinidis K. Kosic G. Krueger V. Kuo M. Lampe I. Lan T. Lee J. Leike J. Leung D. Levy C. M. Li R. Lim M. Lin S. Lin M. Litwin T. Lopez R. Lowe P. Lue A. Makanju K. Malfacini S. Manning T. Markov Y. Markovski B. Martin K. Mayer A. Mayne B. McGrew S. M. McKinney C. McLeavey P. McMillan J. McNeil D. Medina A. Mehta J. Menick L. Metz A. Mishchenko P. Mishkin V. Monaco E. Morikawa D. Mossing T. Mu M. Murati O. Murk D. Mly A. Nair R. Nakano R. Nayak A. Neelakantan R. Ngo H. Noh L. Ouyang C. OKeefe J. Pachocki A. Paino J. Palermo A. Pantuliano G. Parascandolo J. Parish E. Parparita A. Passos M. Pavlov A. Peng A. Perelman F. de Avila Belbute Peres M. Petrov H. P. de Oliveira Pinto Michael Pokorny M. Pokrass V. H. Pong T. Powell A. Power B. Power E. Proehl R. Puri A. Radford J. Rae A. Ramesh C. Raymond F. Real K. Rimbach C. Ross B. Rotsted H. Roussez N. Ryder M. Saltarelli T. Sanders S. Santurkar G. Sastry H. Schmidt D. Schnurr J. Schulman D. Selsam K. Sheppard T. Sherbakov J. Shieh S. Shoker P. Shyam S. Sidor E. Sigler M. Simens J. Sitkin K. Slama I. Sohl B. Sokolowsky Y. Song N. Staudacher F. P. Such N. Summers I. Sutskever J. Tang N. Tezak M. B. Thompson P. Tillet A. Tootoonchian E. Tseng P. Tuggle N. Turley J. Tworek J. F. C. Uribe A. Vallone A. Vijayvergiya C. Voss C. Wainwright J. J. Wang A. Wang B. Wang J. Ward J. Wei C. Weinmann A. Welihinda P. Welinder J. Weng L. Weng M. Wiethoff D. Willner C. Winter S. Wolrich H. Wong L. Workman S. Wu J. Wu M. Wu K. Xiao T. Xu S. Yoo K. Yu Q. Yuan W. Zaremba R. Zellers C. Zhang M. Zhang S. Zhao T. Zheng J. Zhuang W. Zhuk and B. Zoph. Gpt technical report . R. Wang Y. Duan C. Lam J. Chen J. Xu H. Chen X. Liu P. C.I. Pang and T. Tan. Ivygpt Interactive chinese pathway language model in medical domain. In CAAI International Conference on Artificial Intelligence pages . Springer . Q. Xie W. Han X. Zhang Y. Lai M. Peng A. LopezLira and J. Huang. Pixiu A large language model instruction data and evaluation benchmark for finance. arXiv preprint arXiv. . D. X. Xin Yan. Mindchat Psychological large language model. httpsgithub.com XDLabMindChat . A. Yang B. Xiao B. Wang B. Zhang C. Bian C. Yin C. Lv D. Pan D. Wang D. Yan F. Yang F. Deng F. Wang F. Liu G. Ai G. Dong H. Zhao H. Xu H. Sun H. Zhang H. Liu J. Ji J. Xie J. Dai K. Fang L. Su L. Song L. Liu L. Ru L. Ma M. Wang M. Liu M. Lin N. Nie P. Guo R. Sun T. Zhang T. Li T. Li W. Cheng W. Chen X. Zeng X. Wang X. Chen X. Men X. Yu X. Pan Y. Shen Y. Wang Y. Li Y. Jiang Y. Gao Y. Zhang Z. Zhou and Z. Wu. Baichuan Open largescale language models . H. Yang X.Y. Liu and C. D. Wang. Fingpt Opensource financial large language models . S. Yang X. Jiang H. Zhao W. Zeng H. Liu and Y. Jia. Faima Featureaware incontext learning for multidomain aspectbased sentiment analysis . S. Yue W. Chen S. Wang B. Li C. Shen S. Liu Y. Zhou Y. Xiao S. Yun X. Huang and Z. Wei. Disclawllm Finetuning large language models for intelligent legal services . S. Yue W. Chen S. Wang B. Li C. Shen S. Liu Y. Zhou Y. Xiao S. Yun W. Lin et al. Disclawllm Finetuning large language models for intelligent legal services. arXiv preprint arXiv. . H. Zhang J. Chen F. Jiang F. Yu Z. Chen J. Li G. Chen X. Wu Z. Zhang Q. Xiao et al. Huatuogpt towards taming language model to be a doctor. arXiv preprint arXiv. . W. X. Zhao K. Zhou J. Li T. Tang X. Wang Y. Hou Y. Min B. Zhang J. Zhang Z. Dong et al. A survey of large language models. arXiv preprint arXiv. . Z. Zhou J.X. Shi P.X. Song X.W. Yang Y.X. Jin L.Z. Guo and Y.F. Li. Lawgpt A chinese legal knowledgeenhanced large language model. arXiv preprint arXiv. . W. Zhu and X. Wang. Chatmed A chinese medical large language model. httpsgithub.commichaelwzhuChatMed .

write a youtube script

Mitigating Catastrophic Forgetting in Language Transfer via Model Merging Anton Alexandrov Veselin Raychev Mark Niklas Mller Ce Zhang Martin Vechev Kristina Toutanova INSAIT Sofia University St. Kliment Ohridski LogicStar.ai ETH Zurich University of Chicago Together AI Abstract As openweight large language models LLMs achieve ever more impressive performances across a wide range of tasks in English practitioners aim to adapt these models to different languages. However such language adaptation is often accompanied by catastrophic forgetting of the base models capabilities severely limiting the usefulness of the resulting model. We address this issue by proposing BranchandMerge BAM a new adaptation method based on iteratively merging multiple models finetuned on a subset of the available training data. BAM is based on the insight that this yields lower magnitude but higher quality weight changes reducing forgetting of the source domain while maintaining learning on the target domain. We demonstrate in an extensive empirical study on Bulgarian and German that BAM can significantly reduce forgetting while matching or even improving target domain performance compared to both standard continued pretraining and instruction finetuning across different model architectures. Introduction Large language models have shown remarkable capabilities particularly in English. However for less prevalent languages performance can be significantly lower making additional adaptation paramount Zhao et al. Cui and Yao . Catastrophic Forgetting Unfortunately most adaptation techniques come at the cost of catastrophic forgetting of the base models capabilities Zhai et al. Shi et al. Li and Lee Gogoulou et al. . At the same time retaining these capabilities is often crucial for solving downstream tasks in a new language. For example math and coding skills learned in English can be extremely helpful for general problemsolving or reasoning tasks in other languages. Correspondence author anton.alexandrovinsait.ai BranchandMerge Base Model Train on Xi Train on Xi Merge Final Model N Xi Xi X X Training Data XN Figure Illustration of BranchandMerge BAM. We first split the training data into N slices blue . We then iteratively finetune the current base model on two of these slices green and merge the resulting models to obtain the base model for the next iteration purple . We repeat this until all N data slices have been used. Experience Replay To mitigate such catastrophic forgetting mixing in source language data in the target language training set socalled experience replay has proven effective for both continued pretraining Ibrahim et al. and instruction tuning Scialom et al. Zhang et al. . However experience replay alone can not fully mitigate forgetting. Especially when the exact source data is unknown e.g. for stateoftheart language models experience replay can only be implemented approximately reducing its effectiveness and necessitating further regularization. This Work Mitigating Catastrophic Forgetting with BranchandMerge We build on ideas from continual learning and introduce BranchandMerge BAM illustrated in Fig. a novel method for adapting pretrained language models to new languages underrepresented in their unknown training data while minimizing the loss of previously learned capabilities. Concretely BAM splits the training data into N slices blue in Fig. before iteratively training the current base model on K here two such data slices in parallel green arXiv.v cs.LG Jul and finally merging them purple to obtain the initial model for the next iteration. This significantly reduces the total weight change and as a result forgetting while preserving most of the learning from the parallel training steps. In particular while target language perplexity is slightly increased compared to standard continued training the retained base model skills lead to higher downstream performance on target language tasks. Results We apply BAM to adapt MISTRALB Jiang et al. and LLAMAB AIMeta b from predominantly English to an alphabetsharing German and a nonalphabetsharing Bulgarian language considering both continued pretraining and instruction tuning. We show that BAM consistently improves benchmark performance in both the target and source language compared to standard training while not incurring additional computational or data costs. For example when applied to instruction tuning BAM significantly improves performance allowing our LLAMAB BAMtrained for Bulgarian to outperform LLAMABInstruct not only in Bulgarian by . but also in English by . by inducing smaller magnitude but more efficient weight changes. In particular we show that BAM induces more favorable tradeoffs between learning and forgetting than prior techniques such as reduced learning rates Winata et al. and LORA Biderman et al. . Key Contributions Our main contributions are We propose BranchandMerge BAM a training technique for language adaptation improving learning while mitigating forgetting Section . We develop a highquality data mix for approximate experience replay significantly improving language transfer Section . We conduct an extensive empirical investigation demonstrating the effectiveness of BAM across two target languages Section . Model Merging A wide range of model merging methods have been proposed Matena and Raffel Yadav et al. Stoica et al. Yu et al. Wortsman et al. . We experiment with LINEAR Wortsman et al. SLERP Goddard et al. Shoemake and MODEL STOCK Jang et al. merging focusing on the first two explained below. Let us consider the pretrained base model f parameterized by which was finetuned on two different datasets X and X yielding f and f respectively. We call the changes in weight due to this finetuning the task vectors i i . To obtain a single model combining the learning from both datasets we now merge these models. LINEAR model merging interpolates task vectors or equivalently parameterizations linearly so to obtain LINEAR c c c . SLERP first represents task vectors in polar coordinates before interpolating to obtain the new parameterization arccos sin c sin sinc sin where is the angle between the two parameterizations and c is the interpolation coefficient. By slight abuse of notation we write SLERP c for both the resulting parameters and the corresponding model f . BranchandMerge for Mitigating Forgetting in Language Transfer To adapt a model f pretrained on a typically unknown data distribution Xpre to a new task language without suffering from catastrophic forgetting we propose the BranchandMerge BAM method visualized in Fig. . BAM is based on first splitting the available training data into N slices blue in Fig. and then iteratively training K models in parallel on one slice each green before merging the resulting models to obtain the base model for the next training iteration purple . We first provide the intuition behind BAM before describing it in more detail. Intuition There are two key ideas underlying BAM. First lower magnitude weight changes i called task vectors lead to less forgetting but also less learning. Second the randomness in finetuning leads to task vectors i i with an unbiased error i around the locally optimal task vector Jang et al. . We can thus reduce forgetting by reducing the task vector magnitude while offsetting the reduced learning by increasing task vector quality i.e. reducing the error . If this error is unbiased and empirically Gaussian i N Train on X Train on X Merge and Figure Illustration of BAM in the loss surface over parameter space. Both and land in poor local minima but their merge lies in the valley of a better minimum. Training from there and land at the boundary of that minimum due to noise in the training process and limited data. Their merge cancels these errors and lies in the better minimum. Jang et al. merging i.e. averaging K noisy task vectors to obtain K PK i i reduces the corresponding expected error magnitude with K as N K . At the same time increasing K in BAM reduces the number of consecutive training iterations as more data slices are used per iteration and thus the expected total weight magnitude which in turn reduces both learning and forgetting. This allows BAM to trade off learning and forgetting. We visualize this in Fig. for K . There the first two task vectors and land in the basins of poor local minima with their merge falling into the basin of a better minimum highlighting the importance of BAMs iterative merging approach. Training on two more data slices yields noisy task vectors and at the edge of this loss basin with their merge falling right in the middle. In contrast simply reducing the learning rate can also reduce task vector magnitude but does not improve task vector quality. We note that the same intuitions apply to SLERP and MODEL STOCK merging. Implementation In more detail we first partition the training data Xtrain into N not necessarily i.i.d. or equalsized data slices Xi . Then we choose a parallelism factor K K for most experiments and the visualizations in Figs. and and train our current base model f independently on K of these data slices yielding fi to fiK . We merge the resulting models to obtain the base model for the next iteration f MERGEfj iK ji c. We typically choose the merging coefficient c . but note that we can easily perform a d line search over the resulting models. We then set the merged model f to be the base model f f for the Algorithm BranchandMerge BAM Require K parallelism factor f base model Xi N i data slices c merging coefficient for i N do fi trainf Xi i if i mod K i N then MERGE c return f finetuned model next training iteration and repeat this process until we have used all data slices. We formalize this approach in Algorithm . Data Mixtures for Mitigating Forgetting in Language Transfer Here we describe the data we use for continued pretraining of predominantly English base language models in order to adapt them to other languages. Outside of training methodology we find in agreement with prior work that highquality dataset mixtures are paramount for both effective language adaptation and reducing forgetting. We distinguish between experience replay of source language data and target language training data. . Approximate Experience Replay of Source Domain Data While experience replay is crucial to alleviate forgetting Rolnick et al. Ibrahim et al. the training data of most stateoftheart models remains undisclosed. We therefore rely on approximate experience replay constructing our approximate source data based on prior work Penedo et al. Together.ai Touvron et al. Groeneveld et al. . In more detail we create a dataset consisting of OpenWebText Gokaslan et al. an opensource recreation of WebText Radford et al. English Wikipedia GitHub repositories and a range of instruction finetuning datasets with a total of .B unique tokens see Table . We repeat the smaller IFT datasets times to obtain an effective dataset size of .B tokens. We note that while pretraining datasets commonly contain some instructionresponse pairs for example from Reddit our experience replay mix most likely contains a higher portion of instruction data than the unknown source distribution. Table Composition of the approximate experience replay dataset. We report the number of unique tokens how often a dataset is repeated Rep. and the resulting sampling probability Prob.. Dataset Domain Tokens Rep. Prob. OpenWebText Web .B . WikipediaEN Wiki .B . GitHub repos Code .B . OpenHermes. IFT M . SlimOrca IFT M . MetaMathQA IFT M . CodeInstructions IFT M . . Minimal Experience Replay of Source Domain Data To explore the significance of highquality data for experience replay we contrast the aforementioned approximate experience replay with what we call minimal experience replay. In minimal experience replay we exclusively utilize samples from OpenWebText Gokaslan et al. instead of a carefully curated data distribution. While minimal experience replay still incorporates source domain data during continuous pertaining we anticipate it to cause a greater distribution shift than approximate experience replay. We chose the minimal experience replay to comprise roughly oneeighth of the training data B tokens for German and B for Bulgarian. . Constructing Target Language Data While designing an optimal training data mix is still an open research problem Xie et al. Tirumala et al. Shen et al. some key components have been identified that we adhere to for our target domain data. In particular it has been shown that a small portion of code can notably improve the resulting reasoning capabilities and should thus be included Liang et al. Ma et al. Fu and Khot . Furthermore the importance of reasoning and instructionfollowing capabilities for end tasks suggests that instruction data would benefit the continued pretraining data mix. This agrees well with Jiang et al. suggesting a preinstruction tuning phase to improve learning from new documents in continued pretraining. We discuss the exact data mixes we use in Section .. Bulgarian Training Data We adapt the RedPajama v pipeline Together.ai for BulgarianCyrillic to annotate Common Crawl snap httpscommoncrawl.org Table Composition of the Bulgarian target domain dataset. We report the number of unique tokens how often a dataset is repeated Rep. and the resulting sampling probability Prob.. Dataset Domain Tokens Rep. Prob. RPvBG Web B . Legal docs Legal .B . Books Literature .B . EurLexBG Legal M . WikipediaBG Wiki M . OrcaMathBG IFT M . Bulgarian Law Legal M . ParlamintBG Transcripts M . Curlicat Mixed M . SlimOrcaBG IFT M . CodeInstructionsBG IFT M . EuroparlBG Transcripts M . MetaMathBG IFT M . OpenPlatypusBG IFT M . shots with a total of T tokens. After aggressive quality filtering and neardeduplication we obtain a dataset of to B Bulgarian tokens depending on tokenization. We augment this dataset using the Bulgarian split of publicly available multilingual highquality datasets such as Wikipedia Eurlex Baisa et al. Europarl Koehn Parlamint Erjavec et al. books and a selection of private datasets containing news articles legal texts and literature. We further include selected machinetranslated instruction data. See Table for a full list of datasets. This yields a total of .B unique tokens using the original LLAMA tokenizer which we boost to .B tokens by repeating the smaller and particularly highquality datasets between and times. German Training Data German is significantly more abundant than Bulgarian in the quantity of text available from public datasets. We thus subsample roughly of the German subset from the curated web text dataset CulturaX Nguyen et al. equal to B LLAMA tokens and include three German IFT datasets. For more details see Table . Experimental Setup We now describe the experimental setup used to validate BAMs effectiveness for language adaptation. In particular we discuss the target languages Bulgarian and German evaluation benchmarks Section . training data Section . and training setup Section .. We experiment with both continued pretraining of base models and instruction tuning of the resulting models. . Target Languages and Benchmarks To evaluate the effectiveness of BAM we conduct experiments on the transfer from general purpose predominantly English models to an alphabetsharing German and a nonalphabetsharing Bulgarian language evaluating the resulting models on both the target and source languages. While there is a large and growing number of highquality datasets for evaluating LLMs in English and to a lesser extent German these are much sparser for lowresource languages such as Bulgarian. We therefore first provide a brief overview of the English and German benchmarks we use before discussing the construction of a holistic evaluation suite for Bulgarian. English Benchmarks We consider the following domains and benchmarks in English commonsense reasoning HellaSwag Zellers et al. Winogrande Sakaguchi et al. ARCEasy ARCChallenge Clark et al. multitask capabilities MMLU Hendrycks et al. math GSMK Cobbe et al. MathQA Amini et al. and reading comprehension Belebele English Bandarkar et al. TriviaQA Joshi et al. . We provide a detailed description of these benchmarks in Appendix B.. German Benchmarks We use the German benchmarks available in the Language Model Evaluation Harness Gao et al. . Some of these benchmarks are translated from English using GPT . Plster a TruthfulQADE HellaSwagDE MMLUDE ARCDE. We further consider human curated or translated benchmarks for math MGSMDE Shi et al. paraphrasing PAWSX Yang et al. and reading comprehension BeleBele German Bandarkar et al. . A detailed description of these benchmarks can be found in Appendix B.. Bulgarian Benchmarks As the number of publicly available Bulgarian benchmarks is limited we translate all of the above English benchmarks using a combination of machine translation and over hours of professional translators work. We denote the translated benchmarks by appending BG to their name and make them publicly available. In addition we use the following Bulgarian benchmarks natural language inference XNLI Conneau et al. and highschool exams EXAMS Hardalov et al. MONTests. From these XNLI was constructed through a professional translation of English examples by Conneau et al. and the other two are natively Bulgarian. We provide more details on the construction of the translation process and the novel MONTests benchmark in Appendix B.. Evaluation Metrics We aim to measure both learning i.e. language adaptation and forgetting. To this end we consider benchmark scores and perplexity in the source and target language. Since our approximate experience replay data contains instruction tuning examples which can lead to improved English benchmark scores compared to the base model we focus on heldout English document perplexity as a measure of forgetting. We use both benchmark performance normalized accuracy and heldout document perplexity as a measure of learning in the target language see Appendix C for more details. For both English and Bulgarian we evaluate MMLU TriviaQA and EXAMS in a shot GSMK in an shot and all other benchmarks in a zeroshot setting. All German benchmarks are run in a shot setting. . Training Data Below we discuss the training data used for language adaptation. Continued Pretraining Data For German we subsample the training data including the approximate B tokens and minimal experience replay B tokens to B and B tokens respectively and divide it into N i.i.d. slices of .B and B tokens each. For Bulgarian we split the full B tokens of Bulgarian data plus B tokens of approximate or B tokens of minimal experience replay into N slices either i.i.d. or via a curriculum where the evennumbered slices contain significantly more experience replay data than the odd ones see Table in Appendix C. Instruction Finetuning Data We investigate the effectiveness of BAM for instruction finetuning after continued pretraining. We collect K samples of English finetuning data and mix it with German or Bulgarian data. For Bulgarian we generate K samples by using a mix of machine translation and professional translators to translate English samples to Bulgarian. For German we use a mix of available translated German IFT datasets. Please see Tables to as well as Appendix C for details on the resulting dataset. . Training Setup Base models We chose MISTRALB Jiang et al. and LLAMAB AIMeta b as base models due to their exceptional performance for their size and permissive licenses. Details We implement BAM in PyTorch Paszke et al. using HuggingFaces transformers library Wolf et al. and DeepSpeed Rasley et al. Rajbhandari et al. a. We train each model on NVIDIA Hs. Based on prior work and initial experiments we find that is the best maximum learning rate for continued pretraining on the models that we are using together with a batch size of for continued pretraining and for supervised finetuning. We use cosine decay to . maxlr with max . totalsteps linear warmup. Experimental Evaluation We now evaluate BAM empirically for both continued pretraining CPT and instruction finetuning IFT before conducting extensive ablations and providing further results in Appendix A. . BAM for Continued Pretraining Bulgarian CPT We use our data mix of Bulgarian data and English experience replay to adapt both LLAMAB and MISTRALB to Bulgarian comparing standard CPT and BAM in Table . We first demonstrate on MISTRAL that BAM with i.i.d. data slices matches standard CPT in Bulgarian . average score difference while reducing forgetting significantly less English NLL increase achieving a . higher average score on English benchmarks and even outperforming the base model. Using our curriculum slices only called BAM we outperform standard CPT in out of Bulgarian benchmarks while retaining the reduced forgetting. Similarly BAM achieves both a slightly higher average Bulgarian . better and a notably higher English score . better for LLAMA. We observed consistently across all of these settings that while standard CPT achieves a lower negative loglikelihood NLL in Bulgarian indicating it fits the Bulgarian language modeling task better the increased forgetting of base model capabilities higher English NLL leads to worse or equal benchmark performance. German CPT We observe very similar trends adapting LLAMAB to German see Table with BAM outperforming standard CPT both in terms of German . and English . scores with minimal experience replay. Using our approximate experience replay and injecting German IFT data further improves performance . in German and in English surpassing the base LLAMAB model now in both German and English benchmarks. . BAM for Instruction FineTuning We investigate the effectiveness of BAM for instruction finetuning reporting results in Tables and . We observe that BAM slightly improves learning of both German and Bulgarian while significantly reducing forgetting. Considering a wider range of settings in Table we observe that BAM with N K and i.i.d. data slices not only strictly outperforms standard IFT on the combined data IFT full and an equal mix of Bulgarian and English data IFT but also IFT on just English data IFT EN. Slicing the data by language BAM BG EN results in even greater improvements and outperforms LLAMABInstruct AIMeta a in both Bulgarian . and English .. We hypothesize that merging the task vectors of IFT on multiple languages removes a lot of languagespecific errors leaving a higher quality instruction following task vector. . Ablations Below we investigate various components and design decisions underlying BAM using the domain adaptation to Bulgarian. Training steps . . . Avg BG Score Minimal RP Appx. RP . . . EN NLL Figure Comparing minimal and our approximate experience replay on MISTRAL with respect to average Bulgarian benchmark scores and Negative LogLikelihood NLL on the English validation set . Approximate Experience Replay We compare our approximate experience replay described in Section . to minimal experience replay described in Section . for continued pretraining in Fig. . We observe that using minimal replay solid lines in Fig. target language performance Avg BG Score blue first increases before dropping Table Effect of BAM with N and K for the language transfer to Bulgarian. We report normalized accuracies and their averages with full results on English benchmarks deferred to Table . Model CPT IFT Bulgarian Benchmarks Avg BG Avg EN BG NLL EN NLL WG HS ARCc ARCe MMLU MathQA GSMK TrQA MON Bele XNLI EXAMS LB LLAMAB Base . . . . . . . . . . . . . . . . Standard . . . . . . . . . . . . . . . . Half LR . . . . . . . . . . . . . . . . BAM . . . . . . . . . . . . . . . . LLAMABInstruct . . . . . . . . . . . . . . . . BAM Standard . . . . . . . . . . . . . . . . BAM BAM . . . . . . . . . . . . . . . . MB MISTRALB Base . . . . . . . . . . . . . . . . Standard . . . . . . . . . . . . . . . . BAM i.i.d. . . . . . . . . . . . . . . . . BAM . . . . . . . . . . . . . . . . Table Effect of BAM with N and K for the language transfer of LLAMAB to German. We report normalized accuracies and their averages with full results on English benchmarks deferred to Table . CPT IFT German Benchmarks Avg DE Avg EN ARCc HellaSwag MMLU TruthfulQA MGSMDE PAWSDE BeleBele LLAMAB Base . . . . . . . . . Standard min. replay . . . . . . . . . BAM min. replay . . . . . . . . . BAM appx. replay . . . . . . . . . BAM appx. replay Standard . . . . . . . . . BAM appx. replay BAM . . . . . . . . . Table BAM for Bulgarian instruction tuning of our BAM trained LLAMAB. Method Avg BG Avg EN Base BAM trained . . IFT full . . IFT . . IFT EN . . IFT BG . . BAM i.i.d. . . BAM BG EN . . LLAMAInstruct . . Table Effect of approximate and minimal replay on source and target domain performanc. BAM is with i.i.d. data slices. Model Language CPT Replay Avg BG Avg DE Avg EN LLAMAB DE min Standard . . BAM . . appx BAM . . MISTRALB BG min Standard . . BAM . . appx Standard . . BAM . . off as capabilities of the base model are forgotten increasing negative loglikelihood green. In contrast using our approximate experience replay dashed line we see a much stronger increase in target domain performance and reduced forgetting of the source domain. We confirm these findings in German see Table and thus use approximate experience replay for all other experiments. BAM and Experience Replay We compare the effectiveness of BAM in the presence of minimal and approximate experience replay in Table on Bulgarian German and English benchmarks. We observe that BAM is even more effective in the minimal replay setting where the larger distribution shift induces more forgetting. There BAM can e.g. improve the performance in Bulgarian and English by . and . respectively compared to . and . respectively in the approximate replay setting. Weight Change Magnitude . . . AVG BG Score CPT BaM . . . EN NLL Figure Average Bulgarian benchmark score and English NLL over L norm of weight change depending on training method for LLAMA Training steps Weight Change Magnitude CPT CPT half lr BAM Figure L norm of weight change depending on training method for LLAMA Forgetting and Weight Change Magnitude We plot the average BG score as a measure of learning and English NLL as a measure of forgetting over weight change magnitude in Fig. . We observe that both forgetting and learning strongly correlated with weight change magnitude and that BAM is more efficient i.e. yields more learning and less forgetting at the same weight change confirming our intuition discussed in Section . Comparing BAM to standard CPT with a halved learning rate we observe almost identical weight change magnitudes see Fig. corresponding to of the standard CPT weight change. While the reduced learning rate CPT also reduces forgetting although slightly less than BAM it comes at the cost of severely reduced learning see Table . We observe a similar but stronger effect for LORA which only shows minimal learning see Table . Training Steps . . . BG Validation Loss CPT CPT Half LR BaM Odd BaM Even Figure Bulgarian validation loss over training steps for LLAMA depending on training method. BAM Odd green is trained on more Bulgarian and BAM Even red on more approximate experience replay. We show their merges as green dots. Training Dynamics with BAM We compare the training dynamics of BAM and standard CPT at full and half learning rate in Fig. . We observe that training on data slices with larger portions of experience replay even red cannot decrease Bulgarian validation loss further after a short period. However after a merge training on the Bulgarianfocused slices odd green converges significantly faster than for CPT at a similar validation loss highlighting the potential of merging to escape local minima or flatter portions of the loss landscape. Table Effect of of slice count N parallelism factor K and merging method on continued pertaining CPT of LLAMA on a reduced Bulgarian dataset. Merging Method N K Avg BG Avg EN BG NLL EN NLL base . . . . . . . . SLERP . . . . . . . . . . . . LINEAR . . . . MODEL STOCK . . . . Effect of the Parallelism Factor We investigate the effect of the parallelism factor K on dataset of B tokens obtained by combining the first two data slices X and X reporting results in Table where all settings use the same data and compute. We observe that training on all data jointly N K reduces Bulgarian NLL the most but at the cost of increased forgetting highest English NLL leading to worse benchmark performance than BAM. Comparing BAM hyperparameters we observe that increasing K reduces both learning and forgetting as we reduce weight change magnitude and improve task vector quality N K . The best tradeoff leading to the highest English and Bulgarian scores is attained with a parallelism factor of K and data slices of roughly B tokens N K . We thus choose these settings for all other experiments leading to N for the full data. Effect of the Merging Method We compare SLERP LINEAR and MODEL STOCK merging in Table and observe that SLERP and LINEAR merging achieve almost identical results with MODEL STOCK reducing forgetting at the cost of reduced learning. As SLERP merging achieves slightly better scores we use it for all other experiments. Related work Catastrophic Forgetting Neural networks trained on a specific task are known to catastrophically forget the previous task when adapted to a new one French Goodfellow et al. Kemker et al. . While this becomes less pronounced as model and pertaining data size grow Ramasesh et al. it remains a severe issue even for modern LLMs Zhai et al. Shi et al. Li and Lee Gogoulou et al. . Mitigating Catastrophic Forgetting As LLMs are frequently finetuned or continually pretrained on new tasks mitigating catastrophic forgetting has become essential and a wide range of methods has been proposed. Lee et al. suggest to randomly reset weights to their pretrained state. Biderman et al. show that LORA reduces forgetting at the cost of reduced learning. Huang et al. suggest experience replay with synthetic and Ibrahim et al. with original source domain samples. Winata et al. propose to exponentially reduce the learning rate when learning new tasks. Similar to us Lin et al. suggest to linearly merge the adapted with the original model using blockwise parameters focusing on alignment tuning instead of language transfer. Model Merging Model merging was originally proposed in federated learning McMahan et al. to lower communication costs and was successfully deployed Stoica et al. Matena and Raffel . As a way to combine multiple models without training it has recently gained popularity in the LLM community Goddard et al. . Popular methods include LINEAR or TASK ARITHMETIC Ilharco et al. which perform linear interpolation of task vectors its extension MODEL BREADCRUMBS Davari and Belilovsky which discards large weight changes TIES Yadav et al. which uses heuristics favoring large weight changes DARES Yu et al. which randomly drops weight changes before merging MODEL STOCK Jang et al. which merge weights layerwise to in expectation minimize distance to the center of the task vector distribution and SLERP Shoemake which averages weights in polar coordinates. Multiple works have shown that merging during continued pretraining or finetuning especially on nonIID data can match or improve the performance of compound training. Wortsman et al. average models finetuned without any communication. Li et al. propose a scheme for iteratively branching and merging models during training however they assume the full data distribution is available for pertaining and focus on building ensembles rather than a single model. COLD FUSION DonYehiya et al. is methodologically most similar to our work but focuses on training a base model which can then be easily adapted to a new task rather than this adaptation itself. This objective is shared by Choshen et al. which only consider a single iteration of merging. Conclusion We proposed BranchandMerge BAM training to mitigate forgetting while boosting learning in language transfer by generating lower magnitude but higher quality weight changes. We showed that combining BAM with an effective approximate experience replay data mix significantly reduces forgetting. Finally we demonstrated that our approach can benefit both continuous pertaining and instruction tuning in both alphabetsharing German and nonsharing Bulgarian languages. For instance we outperform LLAMABInstruct with the same base model in both source English . and target Bulgarian . languages. Limitations Our study focuses on language transfer to two languages with different characteristics and considers two models of up to billion parameters. However to establish the general applicability of our approach potentially even to general domain adaptation further experiments across a broader set of languages and tasks as well as model architectures will be necessary. We considered specific data mixes for the continued pretraining in both considered languages which we observe to yield good performance it is possible that the success of BranchandMerge depends on the composition of these datasets. While infeasible when adapting stateoftheart pretrained models with unknown training set distribution an evaluation of our method with exact experience replay would provide further understanding of its performance relative to the stateoftheart in continuous learning including joint training on all data. While we consider a broad range of up to benchmarks per language they are still limited in their domain coverage. As BAM does not outperform standard training across all benchmarks this benchmark composition can affect the resulting conclusions. While we originally optimized hyperparameters for standard training and carried them over to BAM it is possible although unlikely that a more extensive hyperparameter search would benefit standard training more than BAM. Ethical Considerations We believe our work empowers practitioners to more efficiently adapt strong pretrained models to other potentially lowresource languages thus contributing to the democratization of large language models. However such models can of course also be abused and in particular if our approach generalizes beyond language to general domain adaptation by malicious practitioners who could more efficiently adapt the models for nefarious tasks. Acknowledgments This research was partially funded by the Ministry of Education and Science of Bulgaria support for INSAIT part of the Bulgarian National Roadmap for Research Infrastructure. This work has been done as part of the EU grant ELSA European Lighthouse on Secure and Safe AI grant agreement no. . Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or European Commission. Neither the European Union nor the European Commission can be held responsible for them. The work has received funding from the Swiss State Secretariat for Education Research and Innovation SERI. We would like to thank Dr. Maurice Weber for helping adapt the RPv pipeline to Bulgarian and create the Bulgarian dataset. We would also like to thank Hristo Venev for his help with system related issues. References AIMeta. a. Llama instruct details. AIMeta. b. Llama model card. Aida Amini Saadia Gabriel Shanchuan Lin Rik KoncelKedziorski Yejin Choi and Hannaneh Hajishirzi. . MathQA Towards interpretable math word problem solving with operationbased formalisms. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics Human Language Technologies Volume Long and Short Papers pages Minneapolis Minnesota. Association for Computational Linguistics. Vt Baisa Jan Michelfeit Marek Medved and Milo Jakubcek. . European Union language resources in Sketch Engine. In Proceedings of the Tenth International Conference on Language Resources and Evaluation LREC pages Portoro Slovenia. European Language Resources Association ELRA. Lucas Bandarkar Davis Liang Benjamin Muller Mikel Artetxe Satya Narayan Shukla Donald Husa Naman Goyal Abhinandan Krishnan Luke Zettlemoyer and Madian Khabsa. . The belebele benchmark a parallel reading comprehension dataset in language variants. Preprint arXiv.. Dan Biderman Jose Gonzalez Ortiz Jacob Portes Mansheej Paul Philip Greengard Connor Jennings Daniel King Sam Havens Vitaliy Chiley Jonathan Frankle et al. . Lora learns less and forgets less. arXiv preprint arXiv.. S. Chaudhary. . Code alpaca An instructionfollowing llama model for code generation. Zhihong Chen Shuo Yan Juhao Liang Feng Jiang Xiangbo Wu Fei Yu Guiming Hardy Chen Junying Chen Hongbo Zhang Li Jianquan Wan Xiang and Benyou Wang. . MultilingualSIFT Multilingual Supervised Instruction Finetuning. Leshem Choshen Elad Venezian Noam Slonim and Yoav Katz. . Fusing finetuned models for better pretraining. Preprint arXiv.. Peter Clark Isaac Cowhey Oren Etzioni Tushar Khot Ashish Sabharwal Carissa Schoenick and Oyvind Tafjord. . Think you have solved question answering? try arc the ai reasoning challenge. ArXiv abs.. Karl Cobbe Vineet Kosaraju Mohammad Bavarian Mark Chen Heewoo Jun Lukasz Kaiser Matthias Plappert Jerry Tworek Jacob Hilton Reiichiro Nakano Christopher Hesse and John Schulman. . Training verifiers to solve math word problems. ArXiv abs.. Alexis Conneau Ruty Rinott Guillaume Lample Adina Williams Samuel Bowman Holger Schwenk and Veselin Stoyanov. . XNLI Evaluating crosslingual sentence representations. In Proceedings of the Conference on Empirical Methods in Natural Language Processing pages Brussels Belgium. Association for Computational Linguistics. Yiming Cui and Xin Yao. . Rethinking LLM language adaptation A case study on chinese mixtral. CoRR abs.. Luigi Daniele and Suphavadeeprasit. . Amplifyinstruct Synthetically generated diverse multiturn conversations for efficient llm training. arXiv preprint arXivcoming soon. Tri Dao. . Flashattention Faster attention with better parallelism and work partitioning. In The Twelfth International Conference on Learning Representations. MohammadReza Davari and Eugene Belilovsky. . Model breadcrumbs Scaling multitask model merging with sparse masks. CoRR abs.. Shachar DonYehiya Elad Venezian Colin Raffel Noam Slonim and Leshem Choshen. . ColD fusion Collaborative descent for distributed multitask finetuning. In Proceedings of the st Annual Meeting of the Association for Computational Linguistics Volume Long Papers pages Toronto Canada. Association for Computational Linguistics. Toma Erjavec Maciej Ogrodniczuk Petya Osenova et al. . The parlamint corpora of parliamentary proceedings. Language Resources and Evaluation . Wikimedia Foundation. Wikimedia downloads. Robert M French. . Catastrophic forgetting in connectionist networks. Trends in cognitive sciences . Hao Fu Yao Peng and Tushar Khot. . How does gpt obtain its ability? tracing emergent abilities of language models to their sources. Yao Fus Notion. Leo Gao Jonathan Tow Baber Abbasi Stella Biderman Sid Black Anthony DiPofi Charles Foster Laurence Golding Jeffrey Hsu Alain Le Noach Haonan Li Kyle McDonell Niklas Muennighoff Chris Ociepa Jason Phang Laria Reynolds Hailey Schoelkopf Aviya Skowron Lintang Sutawika Eric Tang Anish Thite Ben Wang Kevin Wang and Andy Zou. . A framework for fewshot language model evaluation. Leonidas Gee Andrea Zugarini Leonardo Rigutini and Paolo Torroni. . Fast vocabulary transfer for language model compression. In Proceedings of the Conference on Empirical Methods in Natural Language Processing Industry Track pages Abu Dhabi UAE. Association for Computational Linguistics. Charles Goddard Shamane Siriwardhana Malikeh Ehghaghi Luke Meyers Vlad Karpukhin Brian Benedict Mark McQuade and Jacob Solawetz. . Arcees mergekit A toolkit for merging large language models. arXiv preprint arXiv.. Evangelia Gogoulou Timothe Lesort Magnus Boman and Joakim Nivre. . A study of continual learning under language shift. CoRR abs.. Aaron Gokaslan Vanya Cohen Ellie Pavlick and Stefanie Tellex. . Openwebtext corpus. http Skylion.github.ioOpenWebTextCorpus. Ian J. Goodfellow Mehdi Mirza Xia Da Aaron C. Courville and Yoshua Bengio. . An empirical investigation of catastrophic forgeting in gradientbased neural networks. In nd International Conference on Learning Representations ICLR Banff AB Canada April Conference Track Proceedings. Dirk Groeneveld Iz Beltagy Pete Walsh Akshita Bhagia Rodney Kinney Oyvind Tafjord Ananya Harsh Jha Hamish Ivison Ian Magnusson Yizhong Wang Shane Arora David Atkinson Russell Authur Khyathi Raghavi Chandu Arman Cohan Jennifer Dumas Yanai Elazar Yuling Gu Jack Hessel Tushar Khot William Merrill Jacob Morrison Niklas Muennighoff Aakanksha Naik Crystal Nam Matthew E. Peters Valentina Pyatkin Abhilasha Ravichander Dustin Schwenk Saurabh Shah Will Smith Emma Strubell Nishant Subramani Mitchell Wortsman Pradeep Dasigi Nathan Lambert Kyle Richardson Luke Zettlemoyer Jesse Dodge Kyle Lo Luca Soldaini Noah A. Smith and Hannaneh Hajishirzi. . Olmo Accelerating the science of language models. CoRR abs.. Momchil Hardalov Todor Mihaylov Dimitrina Zlatkova Yoan Dinkov Ivan Koychev and Preslav Nakov. . EXAMS A multisubject high school examinations dataset for crosslingual and multilingual question answering. In Proceedings of the Conference on Empirical Methods in Natural Language Processing EMNLP pages Online. Association for Computational Linguistics. Dan Hendrycks Collin Burns Steven Basart Andy Zou Mantas Mazeika Dawn Song and Jacob Steinhardt. . Measuring massive multitask language understanding. In International Conference on Learning Representations. Edward J Hu yelong shen Phillip Wallis Zeyuan AllenZhu Yuanzhi Li Shean Wang Lu Wang and Weizhu Chen. . LoRA Lowrank adaptation of large language models. In International Conference on Learning Representations. Jianheng Huang Leyang Cui Ante Wang Chengyi Yang Xinting Liao Linfeng Song Junfeng Yao and Jinsong Su. . Mitigating catastrophic forgetting in large language models with selfsynthesized rehearsal. CoRR abs.. Adam Ibrahim Benjamin Thrien Kshitij Gupta Mats L. Richter Quentin Anthony Timothe Lesort Eugene Belilovsky and Irina Rish. . Simple and scalable strategies to continually pretrain large language models. Submitted to Transactions on Machine Learning Research. Under review. Gabriel Ilharco Marco Tlio Ribeiro Mitchell Wortsman Ludwig Schmidt Hannaneh Hajishirzi and Ali Farhadi. . Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations ICLR Kigali Rwanda May . OpenReview.net. Neel Jain Ping yeh Chiang Yuxin Wen John Kirchenbauer HongMin Chu Gowthami Somepalli Brian R. Bartoldson Bhavya Kailkhura Avi Schwarzschild Aniruddha Saha Micah Goldblum Jonas Geiping and Tom Goldstein. . NEFTune Noisy embeddings improve instruction finetuning. In The Twelfth International Conference on Learning Representations. DongHwan Jang Sangdoo Yun and Dongyoon Han. . Model stock All we need is just a few finetuned models. CoRR abs.. Albert Q. Jiang Alexandre Sablayrolles Arthur Mensch Chris Bamford Devendra Singh Chaplot Diego de Las Casas Florian Bressand Gianna Lengyel Guillaume Lample Lucile Saulnier Llio Renard Lavaud MarieAnne Lachaux Pierre Stock Teven Le Scao Thibaut Lavril Thomas Wang Timothe Lacroix and William El Sayed. . Mistral b. CoRR abs.. Zhengbao Jiang Zhiqing Sun Weijia Shi Pedro Rodriguez Chunting Zhou Graham Neubig Xi Victoria Lin Wentau Yih and Srinivasan Iyer. . Instructiontuned language models are better knowledge learners. CoRR abs.. Mandar Joshi Eunsol Choi Daniel Weld and Luke Zettlemoyer. . TriviaQA A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the th Annual Meeting of the Association for Computational Linguistics Volume Long Papers pages Vancouver Canada. Association for Computational Linguistics. Ronald Kemker Marc McClure Angelina Abitino Tyler L. Hayes and Christopher Kanan. . Measuring catastrophic forgetting in neural networks. In Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence AAAI the th innovative Applications of Artificial Intelligence IAAI and the th AAAI Symposium on Educational Advances in Artificial Intelligence EAAI New Orleans Louisiana USA February pages . AAAI Press. Philipp Koehn. . Europarl A parallel corpus for statistical machine translation. In Proceedings of Machine Translation Summit X Papers pages Phuket Thailand. Taku Kudo and John Richardson. . SentencePiece A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the Conference on Empirical Methods in Natural Language Processing System Demonstrations pages Brussels Belgium. Association for Computational Linguistics. Ariel N. Lee Cole J. Hunter and Nataniel Ruiz. . Platypus Quick cheap and powerful refinement of llms. Cheolhyoung Lee Kyunghyun Cho and Wanmo Kang. . Mixout Effective regularization to finetune largescale pretrained language models. In th International Conference on Learning Representations ICLR Addis Ababa Ethiopia April . OpenReview.net. ChenAn Li and HungYi Lee. . Examining forgetting in continual pretraining of aligned large language models. CoRR abs.. Margaret Li Suchin Gururangan Tim Dettmers Mike Lewis Tim Althoff Noah A. Smith and Luke Zettlemoyer. . Branchtrainmerge Embarrassingly parallel training of expert language models. In First Workshop on Interpolation Regularizers and Beyond at NeurIPS . Wing Lian Guan Wang Bleys Goodson Eugene Pentland Austin Cook Chanvichet Vong and Teknium. . Slimorca An open dataset of gpt augmented flan reasoning traces with verification. Percy Liang Rishi Bommasani Tony Lee Dimitris Tsipras Dilara Soylu Michihiro Yasunaga Yian Zhang Deepak Narayanan Yuhuai Wu Ananya Kumar Benjamin Newman Binhang Yuan Bobby Yan Ce Zhang Christian Cosgrove Christopher D. Manning Christopher R Diana AcostaNavas Drew A. Hudson Eric Zelikman Esin Durmus Faisal Ladhak Frieda Rong Hongyu Ren Huaxiu Yao Jue Wang Keshav Santhanam Laurel J. Orr Lucia Zheng Mert Yksekgnl Mirac Suzgun Nathan Kim Neel Guha Niladri S. Chatterji Omar Khattab Peter Henderson Qian Huang Ryan Chi Sang Michael Xie Shibani Santurkar Surya Ganguli Tatsunori Hashimoto Thomas Icard Tianyi Zhang Vishrav Chaudhary William Wang Xuechen Li Yifan Mai Yuhui Zhang and Yuta Koreeda. . Holistic evaluation of language models. CoRR abs.. Yong Lin Hangyu Lin Wei Xiong Shizhe Diao Jianmeng Liu Jipeng Zhang Rui Pan Haoxiang Wang Wenbin Hu Hanning Zhang Hanze Dong Renjie Pi Han Zhao Nan Jiang Heng Ji Yuan Yao and Tong Zhang. . Mitigating the alignment tax of rlhf. Preprint arXiv.. Shayne Longpre Le Hou Tu Vu Albert Webson Hyung Won Chung Yi Tay Denny Zhou Quoc V Le Barret Zoph Jason Wei and Adam Roberts. . The flan collection Designing data and methods for effective instruction tuning. In Proceedings of the th International Conference on Machine Learning volume of Proceedings of Machine Learning Research pages . PMLR. Anton Lozhkov Loubna Ben Allal Leandro von Werra and Thomas Wolf. . Finewebedu. Yingwei Ma Yue Liu Yue Yu Yuanliang Zhang Yu Jiang Changjian Wang and Shanshan Li. . At which training stage does code data help llms reasoning? CoRR abs.. Michael Matena and Colin Raffel. . Merging models with fisherweighted averaging. In Advances in Neural Information Processing Systems Annual Conference on Neural Information Processing Systems NeurIPS New Orleans LA USA November December . Brendan McMahan Eider Moore Daniel Ramage Seth Hampson and Blaise Agera y Arcas. . Communicationefficient learning of deep networks from decentralized data. In Proceedings of the th International Conference on Artificial Intelligence and Statistics AISTATS April Fort Lauderdale FL USA volume of Proceedings of Machine Learning Research pages . PMLR. Arindam Mitra Hamed Khanpour Corby Rosset and Ahmed Awadallah. . Orcamath Unlocking the potential of slms in grade school math. Preprint arXiv.. Vladislav Mosin Igor Samenko Borislav Kozlovskii Alexey Tikhonov and Ivan P. Yamshchikov. . Finetuning transformers Vocabulary transfer. Artif. Intell. C. Subhabrata Mukherjee Arindam Mitra Ganesh Jawahar Sahaj Agarwal Hamid Palangi and Ahmed Awadallah. . Orca Progressive learning from complex explanation traces of gpt. Preprint arXiv.. Galileo Mark Namata Ben London Lise Getoor and Bert Huang. . Querydriven active surveying for collective classification. In International Workshop on Mining and Learning with Graphs Edinburgh Scotland. Thuat Nguyen Chien Van Nguyen Viet Dac Lai Hieu Man Nghia Trung Ngo Franck Dernoncourt Ryan A. Rossi and Thien Huu Nguyen. . CulturaX A cleaned enormous and multilingual dataset for large language models in languages. In Proceedings of the Joint International Conference on Computational Linguistics Language Resources and Evaluation LRECCOLING pages Torino Italia. ELRA and ICCL. Adam Paszke Sam Gross Francisco Massa Adam Lerer James Bradbury Gregory Chanan Trevor Killeen Zeming Lin Natalia Gimelshein Luca Antiga Alban Desmaison Andreas Kpf Edward Z. Yang Zachary DeVito Martin Raison Alykhan Tejani Sasank Chilamkurthy Benoit Steiner Lu Fang Junjie Bai and Soumith Chintala. . Pytorch An imperative style highperformance deep learning library. In Advances in Neural Information Processing Systems Annual Conference on Neural Information Processing Systems NeurIPS December Vancouver BC Canada pages . Guilherme Penedo Quentin Malartic Daniel Hesslow Ruxandra Cojocaru Hamza Alobeidli Alessandro Cappelli Baptiste Pannier Ebtesam Almazrouei and Julien Launay. . The refinedweb dataset for falcon LLM outperforming curated corpora with web data only. In Advances in Neural Information Processing Systems Annual Conference on Neural Information Processing Systems NeurIPS New Orleans LA USA December . Bjrn Plster. a. German Benchmark Datasets. Bjrn Plster. b. Leolmopenschnabeltier dataset. Alec Radford Jeff Wu Rewon Child David Luan Dario Amodei and Ilya Sutskever. . Language models are unsupervised multitask learners. Samyam Rajbhandari Jeff Rasley Olatunji Ruwase and Yuxiong He. a. Zero Memory optimizations toward training trillion parameter models. In SC International Conference for High Performance Computing Networking Storage and Analysis pages . Samyam Rajbhandari Jeff Rasley Olatunji Ruwase and Yuxiong He. b. Zero memory optimizations toward training trillion parameter models. In Proceedings of the International Conference for High Performance Computing Networking Storage and Analysis SC . IEEE Press. Vinay Venkatesh Ramasesh Aitor Lewkowycz and Ethan Dyer. . Effect of scale on catastrophic forgetting in neural networks. In The Tenth International Conference on Learning Representations ICLR Virtual Event April . OpenReview.net. Jeff Rasley Samyam Rajbhandari Olatunji Ruwase and Yuxiong He. . Deepspeed System optimizations enable training deep learning models with over billion parameters. In Proceedings of the th ACM SIGKDD International Conference on Knowledge Discovery Data Mining KDD page New York NY USA. Association for Computing Machinery. David Rolnick Arun Ahuja Jonathan Schwarz Timothy P. Lillicrap and Gregory Wayne. . Experience replay for continual learning. In Advances in Neural Information Processing Systems Annual Conference on Neural Information Processing Systems NeurIPS December Vancouver BC Canada pages . Keisuke Sakaguchi Ronan Le Bras Chandra Bhagavatula and Yejin Choi. . Winogrande an adversarial winograd schema challenge at scale. Commun. ACM . Thomas Scialom Tuhin Chakrabarty and Smaranda Muresan. . Finetuned language models are continual learners. In Proceedings of the Conference on Empirical Methods in Natural Language Processing EMNLP Abu Dhabi United Arab Emirates December pages . Association for Computational Linguistics. Zhiqiang Shen Tianhua Tao Liqun Ma Willie Neiswanger Zhengzhong Liu Hongyi Wang Bowen Tan Joel Hestness Natalia Vassilieva Daria Soboleva and Eric P. Xing. . Slimpajamadc Understanding data combinations for LLM training. CoRR abs.. Freda Shi Mirac Suzgun Markus Freitag Xuezhi Wang Suraj Srivats Soroush Vosoughi Hyung Won Chung Yi Tay Sebastian Ruder Denny Zhou Dipanjan Das and Jason Wei. . Language models are multilingual chainofthought reasoners. In The Eleventh International Conference on Learning Representations. Haizhou Shi Zihao Xu Hengyi Wang Weiyi Qin Wenyuan Wang Yibin Wang and Hao Wang. . Continual learning of large language models A comprehensive survey. CoRR abs.. Ken Shoemake. . Animating rotation with quaternion curves. In Proceedings of the th Annual Conference on Computer Graphics and Interactive Techniques SIGGRAPH San Francisco California USA July pages . ACM. George Stoica Daniel Bolya Jakob Bjorner Taylor Hearn and Judy Hoffman. . Zipit! merging models from different tasks without training. CoRR abs.. Teknium. . Openhermes . An open dataset of synthetic data for generalist llm assistants. Kushal Tirumala Daniel Simig Armen Aghajanyan and Ari Morcos. . D improving LLM pretraining via document deduplication and diversification. In Advances in Neural Information Processing Systems Annual Conference on Neural Information Processing Systems NeurIPS New Orleans LA USA December . Together.ai. . Redpajama an open dataset for training large language models. Hugo Touvron Thibaut Lavril Gautier Izacard Xavier Martinet MarieAnne Lachaux Timothe Lacroix Baptiste Rozire Naman Goyal Eric Hambro Faisal Azhar Aurlien Rodriguez Armand Joulin Edouard Grave and Guillaume Lample. . Llama Open and efficient foundation language models. CoRR abs.. Tams Vradi Bence Nyki Svetla Koeva Marko Tadic Vanja tefanec Maciej Ogrodniczuk Bartomiej Niton Piotr Pezik Verginica Barbu Mi titelu Elena Irimia Maria Mitrofan Dan Tufis Radovan Garabk Simon Krek and Andra Repar. . Introducing the CURLICAT corpora Sevenlanguage domain specific annotated corpora from curated sources. In Proceedings of the Thirteenth Language Resources and Evaluation Conference pages Marseille France. European Language Resources Association. Genta Indra Winata Lingjue Xie Karthik Radhakrishnan Shijie Wu Xisen Jin Pengxiang Cheng Mayank Kulkarni and Daniel PreotiucPietro. . Overcoming catastrophic forgetting in massively multilingual continual learning. In Findings of the Association for Computational Linguistics ACL Toronto Canada July pages . Association for Computational Linguistics. Thomas Wolf Lysandre Debut Victor Sanh Julien Chaumond Clement Delangue Anthony Moi Pierric Cistac Tim Rault Rmi Louf Morgan Funtowicz and Jamie Brew. . Huggingfaces transformers Stateoftheart natural language processing. CoRR abs.. Mitchell Wortsman Suchin Gururangan Shen Li Ali Farhadi Ludwig Schmidt Michael Rabbat and Ari S. Morcos. . lofi distributed finetuning without communication. Transactions on Machine Learning Research. Mitchell Wortsman Gabriel Ilharco Samir Yitzhak Gadre Rebecca Roelofs Raphael Gontijo Lopes Ari S. Morcos Hongseok Namkoong Ali Farhadi Yair Carmon Simon Kornblith and Ludwig Schmidt. . Model soups averaging weights of multiple finetuned models improves accuracy without increasing inference time. In International Conference on Machine Learning ICML July Baltimore Maryland USA volume of Proceedings of Machine Learning Research pages . PMLR. Sang Michael Xie Hieu Pham Xuanyi Dong Nan Du Hanxiao Liu Yifeng Lu Percy Liang Quoc V. Le Tengyu Ma and Adams Wei Yu. . Doremi Optimizing data mixtures speeds up language model pretraining. In Advances in Neural Information Processing Systems Annual Conference on Neural Information Processing Systems NeurIPS New Orleans LA USA December . Linting Xue Noah Constant Adam Roberts Mihir Kale Rami AlRfou Aditya Siddhant Aditya Barua and Colin Raffel. . mT A massively multilingual pretrained texttotext transformer. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics Human Language Technologies pages Online. Association for Computational Linguistics. Prateek Yadav Derek Tam Leshem Choshen Colin A. Raffel and Mohit Bansal. . Tiesmerging Resolving interference when merging models. In Advances in Neural Information Processing Systems Annual Conference on Neural Information Processing Systems NeurIPS New Orleans LA USA December . Yinfei Yang Yuan Zhang Chris Tar and Jason Baldridge. . PAWSX A crosslingual adversarial dataset for paraphrase identification. In Proceedings of the Conference on Empirical Methods in Natural Language Processing and the th International Joint Conference on Natural Language Processing EMNLPIJCNLP pages Hong Kong China. Association for Computational Linguistics. Le Yu Bowen Yu Haiyang Yu Fei Huang and Yongbin Li. . Language models are super mario Absorbing abilities from homologous models as a free lunch. CoRR abs.. Longhui Yu Weisen Jiang Han Shi Jincheng YU Zhengying Liu Yu Zhang James Kwok Zhenguo Li Adrian Weller and Weiyang Liu. . Metamath Bootstrap your own mathematical questions for large language models. In The Twelfth International Conference on Learning Representations. Rowan Zellers Ari Holtzman Yonatan Bisk Ali Farhadi and Yejin Choi. . HellaSwag Can a machine really finish your sentence? In Proceedings of the th Annual Meeting of the Association for Computational Linguistics pages Florence Italy. Association for Computational Linguistics. Yuexiang Zhai Shengbang Tong Xiao Li Mu Cai Qing Qu Yong Jae Lee and Yi Ma. . Investigating the catastrophic forgetting in multimodal large language models. CoRR abs.. Zihan Zhang Meng Fang Ling Chen and MohammadReza NamaziRad. . CITB A benchmark for continual instruction tuning. In Findings of the Association for Computational Linguistics EMNLP Singapore December pages . Association for Computational Linguistics. Jun Zhao Zhihao Zhang Luhui Gao Qi Zhang Tao Gui and Xuanjing Huang. . Llama beyond english An empirical study on language capability transfer. CoRR abs.. A Extended Evaluation Data Slice Order We evaluate the effect of data slice order on LLAMA in Table . We observe that reversing the order of data slices for BAM training has only a minimal effect on both forgetting and domain adaptation. Interestingly merging the two models obtained via these different orders is strictly better than either although at twice the computational cost. This highlights again the effectiveness of BAM at finding optimal task vectors by merging out the error component. Tokenizer Extension A common challenge with LLM domain adaptation is that the LLMs tokenizer may not be well suited for the target domain expressed in a higher fertility. This entails longer training slower inference shorter effective context length as well as potential performance degradation. In our language transfer setting from English to Bulgarian we also make use of tokenizervocabulary expansion for some of our experiments to reduce the computational cost. In the case of MISTRALB we find that Bulgarian tokenization is subpar. To this end we train a SentencePieceBPE Kudo and Richardson tokenizer with a vocabulary of k tokens on highquality Bulgarian text. We find that a mix of RPv BG and Wikipedia where the whole Bulgarian Wikipedia comprises these gave the lowest fertility on a sample from mC Xue et al. . After removing all tokens that do not include at least one Cyrillic character or are already in the original tokenizer we are left with exactly new tokens which are then appended to the original Mistral tokenizer with their respective SentencePiece scores. This whole procedure ensures that the English tokenization remains practically unchanged which is important to reduce Catastrophic Forgetting. We initialize the new input and output embeddings with their mean tokenization using the original tokenizer and add them to the models vocabulary in the style of VIPI Mosin et al. and FVT Gee et al. . We report results for MISTRALB in Table and use an k effectively k tokenizer extension for all further MISTRALB experiments due to the greatly increased training and inference efficiency at very similar performance and retain the original LLAMAB tokenizer due to its already huge vocabulary and lower fertility. Note Reducing or increasing the amount of Web data in that tokenizer training mix resulted in higher fertility on the mC sample. The reason for this is not fully clear and we intend on investigating this in future work. LowRank Adaptation LoRA has become widely popular as a method for cheaper finetuning of LLMs Hu et al. . Taking into consideration the contribution of Biderman et al. which puts LoRA in the context of learning less but also forgetting less we also show how LoRA fairs in our Language Transfer setting. Due to limited compute resources we do not perform an extensive hyperparameter sweep and instead copy what we can from the Code CPT experiment in Biderman et al. . As far as we know the batch sizes are not mentioned there and we decide to stick to while deducting that the original may have used . We also proportionately increase the learning rate and find that e converges the fastest. The comparison in Table is in the reduced Btoken setting same as in Table . We indeed observe a better preservation of the English Negative Log Likelihood but also a significant reduction in learned Bulgarian capabilities. It may be the case that the Language Transfer adaptation is not as lowrank as it is for Code and the referred LoRA rank parameter should be set higher than . B Benchmark Details B. Benchmark Descriptions Below we provide short descriptions of all datasets and note the license they are published under. German language benchmarks is run in a shot setting. For the other evaluations we specify the number of shots below or use shots when not specified. HellaSwag MIT License Zellers et al. is a common sense reasoning benchmark asking an LLM to select a logical continuation of a sentence. Evaluated on the sample validation set. Winogrande Appache . LicenseSakaguchi et al. is a common sense reasoning benchmark asking an LLM to fill in a blank from a choice of two entities to logically complete a sentence. Evaluated on the sample validation set of winograndexl. ARCEasy and Challenge CC BYSA License Clark et al. is a dataset of science exam questions. Evaluated on the hard Table English Benchmark performance of LLAMAB continuously pretrained on Bulgarian Training WG HS ARCc ARCe MMLU Bele MathQA GSMK TrQA AVG Base . . . . . . . . . . CPT . . . . . . . . . . BAM . . . . . . . . . . Table English Benchmark performance of LLAMAB continuously pretrained on German Model WG HS ARCc ARCe MMLU Bele MathQA GSMK TrQA AVG Base . . . . . . . . . . CPT . . . . . . . . . . BAM . . . . . . . . . . Table Effect of data slice order on BAM. Data Order Avg EN Avg BG BG pplx. EN pplx. Base Model . . . . Standard . . . . Reversed . . . . Merged . . . . Table Effect of tokenizer extension on performance before and after continuous pertaining CPT of MISTRALB. Training Extension Fertility Avg BG Avg EN Base None . . . k . . . CPT None . . . k . . . sample ARCChallenge and easy samples ARCEasy. MMLU MIT License Hendrycks et al. is a multitask language understanding benchmark covering a wide range of different tasks. Evaluated on test set samples. We evaluate MMLU using shots. GSMK MIT License Cobbe et al. is a mathematical reasoning benchmark consisting of gradeschool math questions for which free text answers must be provided. Evaluated on .k test set samples. We run GSMk with shot chainofthought generation. MathQA Apache . License Amini et al. is a multiple choice mathematical reasoning benchmark. Evaluated on validation set samples. Table Effect of LORA regularization compared to BAM on LLAMA. Training Avg BG Avg EN BG nll EN nll Base . . . . CPT . . . . BAM . . . . LORA . . . . Belebele CCBYNC . License Bandarkar et al. is a multiple choice reading comprehension dataset. Evaluated on samples per language. TriviaQA Apache . License Joshi et al. is trivia question dataset. Evaluated on .k validation set samples. We use shot evaluation. XNLI CC BYNC . License Conneau et al. is a language understanding dataset where the task is to decide whether two statements contradict oneanother are neutral or one entails the other. Evaluated on .k validation samples. EXAMS CC BYSA . License Hardalov et al. is a high school exam question dataset covering a range of subjects. Evaluated test set samples in Bulgarian. We use shot evaluation. PAWS Special License permitting free use for any purpose Yang et al. is a reading comprehension dataset where the task is to decide whether two benchmarks are paraphrases. Evaluated on test set samples. MGSM CC BY . License Shi et al. is mathematical reasoning benchmark manually translated from GSMk. Evaluated on test set samples. B. Bulgarian Benchmarks Translation We use Google Translate to machine translate the text of the benchmark problems and answers. Additional we identified a set of heuristics for cases where the machine translation is of low quality such as inconsistent translations of the same word and not following exact format in both source and target sentences. In all such cases we gave the tasks to human translators with additional instructions on possible problems we identified in each benchmark. Overall human translators manually translated test set samples. A notable example of a benchmark with significant problems that we expect to repeat in many other languages is Winogrande challenge Sakaguchi et al. . In this case one of two words have to be chosen based on world knowledge and reasoning. However with machine translation or nave human translation to nonEnglish the actual answer can be revealed in a much easier way by the means of having only one answer that is in gender agreement with other words in the sentence. We performed manual translations that used synonyms that do not exhibit such behavior and as a result the translated benchmark is not easier than the original. The translated versions of the benchmarks with these fixes are made publicly available. MON The MON dataset is obtained as private data from the Bulgarian Ministry of Education. This contains exam questions with possible choices only one of which is correct spanning topics from th to th grade tests previously given for external tests to schools in Bulgaria. The questions span all subjects tested by the official Bulgarian curriculum but exclude problems such as geometry tasks that include images in their problem definition or answers. The dataset is not publicly available and as a result we expect it to be less likely to be in any of the training data in any form. C Dataset Details C. IFT Set Composition We make note of the good performance and instruction following capabilities of the Intel NeuralChat models and decide to include SlimOrca Lian et al. Mukherjee et al. Longpre et al. and MetaMathQA Yu et al. in our English Table Composition of the Bulgarian IFT dataset. Dataset Domain Examples Repetitions Prob OpenHermes.BG Mixed Conversations . CapybaraBG Mixed Conversations . MetaMathBG Math . CodeAlpacaBG Code . Table Composition of the English IFT dataset. Dataset Domain Examples Repetitions Prob SlimOrca Mixed Conversations . MetaMathQA Math . Capybara Mixed Conversations . IFT data mix. To fill in the gap of multiturn conversation data we additionally include the Capybara dataset Daniele and Suphavadeeprasit which we have observed from our experience boost the models chattiness and overall response quality. The fact that there are no publicly available general Bulgarian IFT datasets lead us to the translation of already existing ones. We use machine translation to produce K Bulgarian translated samples from the OpenHermes. Teknium dataset K samples from MetaMathQA Yu et al. and K samples of code with Bulgarian instructions from CodeAlpaca Chaudhary . We take special care in the translation of the Capybara Daniele and Suphavadeeprasit and OpenHermes datasets. Through a combination of classification and manual inspection we identify examples where the machine translation is not good enough to make a sensible training example e.g. instructions that require rhyming as the words that rhyme in English will most likely not rhyme in Bulgarian. The identified of the Capybara dataset is then manually translatedadjusted to fit the Bulgarian language. See Table for full details and licenses. C. Validation Set Composition Constructing validation datasets for language model training especially when such are trained on webcrawl data is a challenging task with respect to avoiding data contamination. Our Bulgarian validation set consists of a total of K examples K of which are a heldout set of news articles from a specific media outlet and the other K is a mix of dialogs questions and answers literary works and legal documents. The English validation dataset is comprised of K random samples from the FineWebEdu dataset Lozhkov et al. K samples from arXiv scientific papers K from Table Composition of the German IFT dataset. Dataset Domain Examples Repetitions Prob evolinstructdeutsch Mixed Conversations . alpacagptdeutsch Mixed Conversations . OpenSchnabeltier Mixed Singleturn . Table Sources and licenses of used datasets Dataset Source License RPv pipeline Together.ai Apache . OpenWebText Gokaslan et al. CC. CulturaX Nguyen et al. CC. FineWebEdu Lozhkov et al. ODCBY PubMed Namata et al. Unknown EurLex Baisa et al. CCBYNCSA Wikipedia Foundation CCBYSA. OrcaMath Mitra et al. MIT Parlamint Erjavec et al. CCBY OpenHermes. Teknium Unknown Capybara Daniele and Suphavadeeprasit Apache . Curlicat Vradi et al. CCBYSA. SlimOrca Lian et al. Mukherjee et al. MIT CodeAlpaca Chaudhary CCBY. Europarl Koehn Unknown MetaMath Yu et al. MIT OpenPlatypus Lee et al. Apache . alpacagptdeutsch Chen et al. Apache . OpenSchnabeltier Plster b Apache . evolinstructdeutsch Chen et al. Apache . the PubMed dataset Namata et al. and K books from the Project Gutenberg . D Experimental Setup and Evaliation Details D. Training parameters We use the same exact training hyperparameters for both MISTRALB and LLAMAB based models. We stick to the size context lengths and train with sequence packing without truncation. Based on prior work and initial experiments we find that e is the best maximum learning rate for continued pretraining in our settings with a batch size of for continued pretraining and for supervised finetuning effectively training for M and M tokens respectively. The optimizer in use is AdamW with . and . and . weight decay rate. We use a cosine decay learning rate scheduler that decays the LR to . maxlr with max . totalsteps of linear warmup. For finetuning we have found that training for more than epochs on a given IFT dataset with the aforementioned hyperparameters is not beneficial and exaggerates catastrophic forgetting. Additionally we add embedding vector noise during training through NEFTune Jain et al. with a noise . In this stage we train only on the IFT completions and not on the prompts. This is important to prevent unwanted selftalking behavior in httpswww.gutenberg.org live usage. Since we train on GPUs at once we exploit DeepSpeed ZeRO Rasley et al. Rajbhandari et al. b stage with mixed precision training in bf. Combining this with activation checkpointing and FlashAttention Dao allows us to use a batch size of during training and evaluation. For reference our setup allows the models to train with up to tokens per second per GPU. D. Computational Budget All model training and evaluations were conducted on a cluster of NVIDIA H GPUS nodes x GPUs with InfiniBand and available CPU cores per node. The total computational cost of the experiments included in this paper including exploratory ones not mentioned here is around NVIDIA H GPU hours. The tokenizer extension we perform on MISTRALB Base helps reduce the training and inference cost of our Mistralbased models by roughly . Table Composition of the Bulgarian curriculum splits. Split Total BG Total Replay Dataset Repetitions Replay X .B M WikipediaBG OpenWebText . Bulgarian Law EurLexBG IFTBG RPvBG . X .B .B WikipediaEN . OpenWebText . GitHub repos . IFTEN RPvBG . X .B M WikipediaBG OpenWebText . Bulgarian Law EurLexBG IFTBG RPvBG . ParlamintBG EuroparlBG Legal docs . X .B .B WikipediaEN . OpenWebText . GitHub repos . IFTEN RPvBG . X .B M WikipediaBG OpenWebText . Bulgarian Law Books IFTBG RPvBG . ParlamintBG EuroparlBG Legal docs . X .B .B WikipediaEN . OpenWebText . GitHub repos . IFTEN RPvBG . X .B M WikipediaBG OpenWebText . Bulgarian Law Books IFTBG RPvBG . ParlamintBG EuroparlBG Legal docs . X .B .B WikipediaEN . OpenWebText . GitHub repos . IFTEN RPvBG .

write a youtube script

MetaUrban A Simulation Platform for Embodied AI in Urban Spaces Wayne Wu Honglin He Yiran Wang Chenda Duan Jack He Zhizheng Liu Quanyi Li Bolei Zhou University of California Los Angeles httpsmetadriverse.github.iometaurban Delivery Bot Electric Wheelchair Robot Dog Humanoid Robot Sidewalk Crosswalk Street Block Plaza Figure Top Humans and mobile machines start sharing public urban spaces. Bottom MetaUrban simulator facilitates embodied AI research in urban settings by composing interactive urban environments with diverse street blocks objects and movements of pedestrians vulnerable road users and other agents. Abstract Public urban spaces like streetscapes and plazas serve residents and accommodate social life in all its vibrant variations. Recent advances in Robotics and Embodied AI make public urban spaces no longer exclusive to humans. Food delivery bots and electric wheelchairs have started sharing sidewalks with pedestrians while diverse robot dogs and humanoids have recently emerged in the street. Ensuring the generalizability and safety of these forthcoming mobile machines is crucial when navigating through the bustling streets in urban spaces. In this work we present MetaUrban a compositional simulation platform for Embodied AI research in urban spaces. MetaUrban can construct an infinite number of interactive urban scenes from compositional elements covering a vast array of ground plans object placements pedestrians vulnerable road users and other mobile agents appearances and dynamics. We design point navigation and social navigation tasks as the pilot study using MetaUrban for embodied AI research and establish various baselines of Reinforcement Learning and Imitation Learning. Experiments demonstrate that the compositional nature of the simulated environments can substantially improve the generalizability and safety of the trained mobile agents. MetaUrban will be made publicly available to provide more research opportunities and foster safe and trustworthy embodied AI in urban spaces. Preprint. Under review. arXiv.v cs.CV Jul Introduction Public urban spaces vary widely in type form and size encompassing streetscapes plazas and parks. They are crucial spaces for transit and transport as well as providing opportunities to stage various social events. Since the early th century the study of public urban spaces has long been a cornerstone of urban sociology and planning . For example William H. Whyte in his seminal work City Rediscovering the Center revealed that the complexity and vibrant interaction in public spaces profoundly determine humans social life underscoring the critical role these environments play in urban safety and vitality. Recent development of Robotics and Embodied AI makes the urban space no longer exclusive to humans. Various mobile machines have started emerging. For example elders and physically disabled people maneuver electronic wheelchairs on the street while food delivery bots navigate on the sidewalk to accomplish the lastmile food delivery task. Various mobile legged robots like robot dog Spot from Boston Dynamics and humanoid robot Optimus from Tesla are also forthcoming. We can thus imagine a future of public urban spaces that will be shared and cohabitated by humans and mobile machines driven by Embodied AI. Ensuring the generalizability and safety of these mobile machines becomes essential. Simulation platforms have played a crucial role in enabling systematic and scalable training of the embodied AI agents and the safety evaluation before realworld deployment. However most of the existing simulators focus either on indoor household environments or outdoor driving environments . For example platforms like AITHOR Habitat and iGibson are designed for household assistant robots in which the environments are mainly apartments or houses with furniture and appliances platforms like SUMO CARLA and MetaDrive are designed for research on autonomous driving and transportation. Yet simulating urban spaces with diverse layouts and objects complex dynamics of pedestrians is much less explored. Distinct from the indoor household and driving environments the urban space has unique characteristics. Lets follow the adventure of a lastmile delivery bot who aims to deliver a lunch order from a nearby pizzeria to the campus. First it faces a longhorizon journey across several street blocks at a onemile distance with multifarious road hazards such as fragmented curbs and rugged ground caused by tree roots on sidewalks. Then it must safely navigate the cluttered street full of obstacles like trash bins parked scooters and potted plants. In addition it needs to handle pedestrians and crowds properly to avoid collisions. It should also take special care around disabled people in wheelchairs. Thus the layout diversity object distribution and dynamic complexity bring unique challenges to the design of simulation environments and the study of the generalizability and safety of Embodied AI agents operating in urban spaces. We present MetaUrban a compositional simulation platform for Embodied AI research in urban spaces. First we introduce Hierarchical Layout Generation a procedural generation approach that can generate infinite layouts hierarchically from street blocks to sidewalks functional zones and object locations. It can generate scenes at an arbitrary scale with various connections and divisions of street blocks object locations and terrains which are critical for improving the generalizability of trained agents. Then we design the Scalable Object Retrieval an automatic pipeline that can obtain an arbitrary number of highquality objects with realworld distribution. We first compute the object category distribution from broad realworld data to form a description pool. Then with the sampled descriptions from the pool we effectively retrieve objects from largescale D asset repositories with a VLMbased openvocabulary searching schema. Finally we propose the Cohabitant Populating method to generate complex dynamics in urban spaces. We first tailor recent D human and motion datasets to get rigged pedestrian models each with movements. Then to form safetycritical scenarios we integrate Vulnerable Road Users VRUs like bikers skateboarders and scooter riders. To broaden the category of mobile machines in urban scenes we include delivery bots electric wheelchairs robot dogs and humanoid robots. Then based on path planning algorithms we can get complex trajectories among hundreds of environmental agents simultaneously with collision and deadlock avoidance. It is critical for enhancing the social conformity and safety of the mobile agents. Based on MetaUrban we construct a largescale dataset MetaUrbanK that includes training scenes and test scenes. We further create an unseen test set with manually designed scenes to evaluate trained models generalizability. Besides we provide steps of highquality expert demonstrations by humans and welltrained agents to enable offline reinforcement learning and imitation learning research. We design point navigation and social navigation tasks as a pilot study using MetaUrban for embodied AI research and establish extensive baselines for Reinforcement Learning Safe Reinforcement Learning Offline Reinforcement Learning and Imitation Learning. Experiments demonstrate that the generalizability and safety of the trained mobile agents can be substantially improved by the compositional nature of the simulated environments. We will make the simulator available to enable more research opportunities for the community and foster safe and trustworthy embodied AI in urban spaces. Related Work Many simulation platforms have been developed for Embodied AI research depending on the target environments such as indoor homes and offices driving freeways and roadways and crowds in warehouses and squares. We compare representative ones with the proposed MetaUrban simulator. Indoor Environments. Platforms for indoor environments are mainly designed for household assistant robots emphasizing the affordance realism and diversity of objects as well as the interactivity of environments. VirtualHome pivots towards simulating routine human activities at home. AITHOR and its extensions such as ManipulaTHOR RoboTHOR and ProcTHOR focus on detailed agentobject interactions dynamic object state changes and procedural scene generation alongside robust physics simulations. Habitat offers environments reconstructed from D scans of realworld interiors. Its subsequent iterations Habitat . and Habitat . introduce interactable objects and deformable humanoid agents respectively. iGibson provides photorealistic environments. Its upgrades Gibson . and OmniGibson focus on household tasks with object state changes and a realistic physics simulation of everyday activities respectively. ThreeDWorld targets realworld physics by integrating highfidelity simulations of liquids and deformable objects. However unlike MetaUrban these simulators are focused on indoor environments with particular tasks like object rearrangement and manipulation. Driving Environments. Platforms for driving environments are mainly designed for autonomous vehicle research and development. Simulators like GTA V SimCV AIRSIM CARLA and its extension SUMMIT offer realistic environments that mimic the physical worlds detailed visuals weather conditions and daytonight transitions. Other simulators enhance efficiency and extensibility at the expense of visual realism such as Udacity DeepDrive Highwayenv and DriverGym . MetaDrive trades off between visual quality and efficiency offering a lightweight driving simulator that can support the research of generalizable RL algorithms for vehicles. Although some of the simulators involve traffic participants other than vehicles such as pedestrians and cyclists all of them focus on vehiclecentric driving scenarios and neglect environments and things happening in public urban spaces like sidewalks and plazas. Social Navigation Environments. Other than indoor and driving environments social navigation platforms emphasize the social compatibility of robots. Simulators like CrowdNav GymCollisionAvoidance and SocialGym . model scenes and agents in D maps focusing more on the development of path planning algorithms. Other simulators such as HuNavSim SEAN . and SocNavBench upgrade the environment to D space and introduce human pedestrians to support the development of more complex algorithms. However social navigation platforms focus on crowd navigation with oversimplified objects and surrounding environmental structures in the scenes. This proposed platform addresses the gap between the existing social navigation platforms and the realworld urban spaces regarding environmental diversity and complexity. We will compare MetaUrban with other simulators below through the scale sensor and feature dimensions. For the scale MetaUrban can generate infinite scenes with a procedural generation pipeline. It also provides the largest number of humans and movements among all simulation environments. For the sensor MetaUrban provides RGBD semantic and lidar while acoustic is our next step to better support multimodel tasks. For the feature different from other simulators MetaUrban provides realworld distribution of the objects categories and uses a more sophisticated path plan algorithm to get natural human trajectories. It also provides flexible user interfaces mouse keyboard and joystick and racing wheel which vastly ease the collection of human expert demonstration data. MetaUrban uses PyBullet as its physical engine which is opensource and highly accurate in physics simulation providing a costeffective and flexible solution for future developments. MetaUrban uses PandaD for rendering which is a lightweight opensource game engine with seamless Python integration providing a flexible and accessible development environment. A detailed comparison table is included in the Appendix. In summary none of the recent simulation platforms have been constructed for urban spaces and the proposed simulator differs from them significantly in terms of diverse layouts objects human dynamics and different types of mobile agents like delivery robots electric wheelchairs robot dogs humanoid robots and vehicles and their intricate interactions. We believe MetaUrban can provide a lot of new research opportunities for Embodied AI in urban settings. MetaUrban Simulator MetaUrban is a compositional simulation platform that can generate infinite training and evaluation environments for Embodied AI in urban spaces. Figure depicts the procedural generation pipeline. MetaUrban uses a structured description script to create urban scenes. Based on the provided information about street blocks sidewalks objects agents and more it starts with the street block map then plans the ground layout by dividing different function zones then places static objects and finally populates dynamic agents. Figure Procedural generation. MetaUrban can automatically generate complex urban scenes with its compositional nature. From the second to the fourth column the top row shows the D road maps and the bottom row shows the birdeye view of D scenes. This section highlights three key designs in the MetaUrban simulator to support exhibiting three unique characteristics of urban spaces diverse layouts particular object distribution and complex dynamics. Section . introduces Hierarchical Layout Generation which can infinitely generate diverse layouts with different functional zone divisions and object locations that are critical for the generalizability of agents. Section . introduces Scalable Object Retrieval which harnesses worldwide urban scene data to obtain realworld object distributions in different places and then builds largescale highquality static objects set with VLMenabled openvocabulary searching. It is useful for training agents specialized for urban scenes. Section . introduces Cohabitant Populating in which we leverage the advancements in digital humans to enrich the appearances movements and trajectories of pedestrians and vulnerable road users as well as incorporate other agents to form a vivid cohabiting environment. It is critical for improving the social conformity and safety of the mobile agents. . Hierarchical Layout Generation The diversity of scene layout i.e. the connection and categories of blocks the specifications of sidewalks and crosswalks as well as the placement of objects is crucial for enhancing the generalizability of trained agents maneuvering in public spaces. In the hierarchical layout generation framework we start by sampling the categories of street blocks and dividing sidewalks and crosswalks and then allocate various objects with which we can get infinite urban scene layouts with arbitrary sizes and specifications of maps. Ground plan. We design typical street block categories i.e. straight intersection roundabout circle and Tjunction. In the simulator to form a large map with several blocks we can sample the category number and order of blocks as well as the number and width of lanes in one block to get different maps. Then each block can simulate its own walkable areas sidewalks and crosswalks which are key areas for urban spaces with plenty of interactions. Figure Ground plan. Left Sidewalk is divided into four functional zones building frontage clear and furnishing zone. Right Seven typical sidewalk templates from a to g. As shown in Figure left according to the Global Street Design Guide provided by the Global Designing Cities Initiative we divide the sidewalk into four functional zones building zone frontage zone clear zone and furnishing zone. Based on their different combinations of functional zones we further construct typical templates for sidewalks Figure right. To form a sidewalk we can first sample the layout from the templates and then assign proportions for different function zones. For crosswalks we provide candidates at the start and the end of each roadway which support specifying the needed crosswalks or sampling them by a density parameter. Finally roadways sidewalks and crosswalks can take a terrain map as substrate to form different ground situations. Object placement. After determining the ground layout we can place objects on the ground. We divide objects into three classes. Standard infrastructure such as poles trees and signs are placed periodically along the road. Nonstandard infrastructure such as buildings bonsai and trash bins are placed randomly in the designated function zones. Clutter such as drink cans bags and bicycles are placed randomly across all functional zones. We can get different street styles by specifying an object pool while getting different compactness by specifying a density parameter. Figure shows different objects placed with a sampled ground plan and objects location. Figure Object placement. Placing Different objects with the same ground plan and objects location. . Scalable Object Retrieval Hierarchical layout generation decides the scenes layout and where to place the objects. However to make the trained agents generalizable when navigating through scenes composed of various objects in the real world what objects to place is another crucial question. In this section we propose the Scalable Assets Retrieval pipeline in which we first get realworld object distributions from web data and then retrieve objects from D asset repositories through an openvocabulary search schema based on VLMs. This pipeline is flexible and extensible the retrieved objects can be scaled to arbitrary sizes as we continue to exploit more web data for scene descriptions and include more D assets as the candidate objects. Realworld object distribution extraction. Urban spaces have unique structures and object distributions such as the infrastructure built by the urban planning administration and clutters placed by people. Thus we design a realworld distribution extraction method to get a description pool depicting the frequent objects in urban spaces. As illustrated in Figure a we first leverage offtheshelf academic datasets for scene understanding CityScape and Mapillary Vistas to get a list of objects that are with high frequency to be put in the urban space. However the number of objects is limited because of the closedset definitions in the image datasets. We introduce two openset sources to get broader object distribution from the real world. Google Street data. We first collect urban space images from countries across six continents. Then we harness GPTo and openset segmentation model GroundedSAM to get descriptions of objects in the urban public space. Urban planning description data. We further get a list of essential objects in public urban spaces through a thorough survey of urban design handbooks. Finally by combining these three data sources we can get an object description pool with items of descriptions that form the realworld object category distribution. Figure Scalable assets retrieval. a Realworld distribution extraction. We get object distribution for urban spaces from three sources academic datasets Google Street data and text description data. b Openvocabulary search. We use the VLM to get image and text embedding respectively. Then based on the relevant scores we can get the objects with high rankings. Openvocabulary search. The recent development of largescale D object repositories enables efficiently constructing a dataset for a specific scene. However these large repositories have three intrinsic issues to harness these repositories most of the data is unrelated to the urban scene the data quality in large repositories is uneven and the data has no reliable attribute annotations. To this end we introduce an openvocabulary search method to tackle these issues. As shown in Figure b the whole pipeline is based on an imagetext retrieval architecture. We first sample objects from Objaverse and ObjaverseXL repositories to get projected multiview images. Here a naive uniform view sampling will bring lowquality harmful images. Following we select and prioritize informative viewpoints which significantly enhance retrieval effectiveness. Then we leverage the encoder of a Vision Language Model BLIP to extract features from projected images and sampled descriptions from the object description pool respectively to calculate relevant scores. Then we can get target objects with relevant scores up to a threshold. This method lets us get an urbanspecific dataset with highquality objects in realworld category distributions. In addition we provide an interface for customizing training objects in the scene by providing images or text descriptions taking advantage of recent advances in D object reconstruction and generation . . Cohabitant Populating In this section we will describe how to populate these static urban scenes with varied agents regarding appearances movements and trajectories through Cohabitant Populating. Following BEDLAM and AGORA we represent humans as parametric human model SMPLX in which the D human body is controlled by a set of parameters for pose shape and facial expression respectively. Then built upon SynBody s asset repository D rigged human models are procedually generated by sampling from garments hairs beards accessories and cloth and skin textures. To form safetycritical scenarios we also include vulnerable road users like bikers skateboarders and scooter riders. For the other agents we incorporate the D assets of COCO Robotics and Starships delivery robots Drive Medicals electric wheelchair Boston Dynamics robot dog and Agility Robotics humanoid robot. We provide two kinds of human movements in the simulator daily movements and unique movements. Daily movements provide the basic human dynamics in daily life i.e. idle walking and running. Unique movements are the complicated dynamics that appear randomly in public spaces such as dancing and exercising. We harness the BEDLAM dataset to obtain unique movements. For humans and other agents with daily movements we simulate their trajectories using the ORCA social forces model and Push and Rotate PR algorithm . ORCA uses a joint optimization and a centralized controller that guarantees that agents will not collide with each other or any other objects identified as obstacles. Push and Rotate PR is a multiagent pathfinding algorithm that can resolve any potential deadlock by local coordination. In the future an interesting direction is to endow personal traits like job personality and purpose to humans and harness the advantages of LLMs and LVMs to enable social and interactive behaviors of humans in urban scenes. MetaUrbanK Dataset Dataset Construction. Based on the MetaUrban simulator we construct the MetaUrbanK dataset including interactive urban scenes for training MetaUrbantrain and scenes for testing MetaUrbantest. For the train and test sets we sample randomly from the templates af of sidewalks shown in Figure right with the same distributions of objects and dynamics. We further construct an unseen test set MetaUrbanunseen with scenes for zeroshot experiments in which we sample from the unseen template g Wide Commercial Sidewalk unseen objects trajectories of agents with further designers manual adjustments according to realworld scenes. In addition to enable the finetuning experiments we construct a training set of scenes with the same distribution of MetaUrbanunseen termed MetaUrbanfinetune. K scenes can be generated in hours on a local workstation. Notably our MetaUrban platform can easily extend the scale of urban scenes from a multiblock level to a whole city level. To enable the Offline RL and IL training we collect expert demonstration data from a welltrained RL agent and human operators forming steps of highquality demonstration data. The success rate of the demonstration data is which can be taken as a reference for the experiments of Offline RL and IL. The Appendix provides detailed visualizations of the MetaUrbanK dataset. Figure MetaUrbanK statistics. Statistics. Scenes in this dataset are connected by one to three street blocks covering average m areas. There are average . static objects per m and dynamic agents per street block. The average distance between each two objects is .m. Figure shows distributions of the number of objects left areas of objects occupying middle and episode length right. As shown in the distribution of object numbers there are lots of objects in each scenario with a minimal value of . As shown in the distribution of objects areas objects in the dataset occupy large areas which complies with a normal distribution centered at m . As shown in the distribution of episode length the average episode is m and more than of them are more than steps m. From these distributions we can observe that scenes are significantly challenging in MetaUrbanK for agents to navigate through which are crowded and with long horizons. Experiments Experimental settings. Tasks. We design two common tasks in urban scenes Point Navigation PointNav and Social Navigation SocialNav. In PointNav the agents goal is to navigate to the target coordinates in static environments without access to a prebuilt environment map. In SocialNav the agent is required to reach a point goal in dynamic environments that contain moving environmental agents. The agent shall avoid collisions or proximity to environmental agents beyond thresholds to avoid penalization distance . meters. The agents action space in the experiments consists of acceleration brake and steering. The observations contain a vector denoting the LiDAR signal a vector summarizing the agents state and the navigation information that guides the agent toward the destination. Methods. We evaluate typical baseline models to build comprehensive benchmarks on MetaUrban across Reinforcement Learning PPO Safe Reinforcement Learning PPOLag and PPOET Offline Reinforcement Learning IQL and TDBC and Imitation Learning BC and GAIL . Evaluation metrics. The agent is evaluated using the Success Rate SR and Success weighted by Path Length SPL metrics which measure the success and efficiency of the path taken by the agent. For SocialNav except Success Rate SR the Social Navigation Score SNS is also used to evaluate the social complicity of the agent. For both tasks we further report the Cumulative Cost CC to evaluate the safety properties of the agent. It records the crash frequency to obstacles or environmental agents. Please refer to the Appendix for details of models rewards and hyperparameters. Table Benchmarks. The benchmark of PointNav and SocialNav tasks on the MetaUrbanK dataset. Seven representative methods of RL safe RL offline RL and IL are evaluated for each benchmark. indicate the best performance among online methods RL and Safe RL and offline methods offline RL and IL respectively. Category Method PointNav SocialNav Test Zeroshot Test Zeroshot SR SPL Cost SR SPL Cost SR SNS Cost SR SNS Cost RL PPO . . . . . . . . . . . . Safe RL PPOLag . . . . . . . . . . . . PPOET . . . . . . . . . . . . Offline RL IQL . . . . . . . . . . . . TDBC . . . . . . . . . . . . IL BC . . . . . . . . . . . . GAIL . . . . . . . . . . . . . Benchmarks We build two benchmarks on the MetaUrbanK dataset for PointNav and SocialNav tasks. We train typical baselines on the MetaUrbantrain dataset and then evaluate them on the MetaUrbantest set. We use the demonstration data provided in MetaUrbanK for offline RL and IL training. We further make zeroshot evaluations on the MetaUrbanunseen set to demonstrate the generalizability of models trained on the MetaUrbanK dataset while directly tested on unseen environments. Table shows the results in the PointNav and SocialNav benchmarks. From the results we can draw key observations. The tasks are far from being solved. The highest success rates are only and for PointNav and SocialNav tasks achieved by the baselines indicating the difficulty of these tasks in the urban environments composed by MetaUrban. Note that these benchmarks are built on a medium level of object and dynamic density increasing the density will further degrade the performances shown in ablation studies. Models trained on MetaUrbanK have strong generalizability in unseen environments. With zeroshot testing models can still achieve and success rates on average for PointNav and SocialNav tasks. These results are strong since the models generalize to not only unseen objects and layouts but also unseen dynamics of environmental agents. It demonstrates that the compositional nature of MetaUrban supporting the coverage of a large spectrum of complex urban scenes can successfully empower generalization ability to the trained models. SocialNav is much harder than PointNav due to the dynamics of the mobile environmental agents. On average the success rate decreases by from PointNav to SocialNav indicating that dynamic agents such as pedestrians vulnerable road users and other agents common in urban scenes present significant challenges to the trained agent. Safe RL remarkably improves the safety property at the expense of effectiveness. Among all tasks and settings the safe RL models achieve the best performance in the Cumulative Cost indicating that these models are successful at avoiding collision with pedestrians and objects. However the success rate and SPL or SNS decrease accordingly indicating future efforts to balance the safety and effectiveness of agents in complex urban scenes. . Ablation Study In this section we evaluate the generalizability scaling ability and effects of the density of static objects and dynamic agents. For unified evaluations we use PPO for all ablation studies. Except for the results on dynamic density we use the PointNav task. Observations and hyperparameters remain the same for model training across different evaluations. Figure Ablation study. a Evaluation of generalizability. b Evaluation of scaling ability. c Evaluation of the density of static objects. d Evaluation of the density of dynamic agents. Evaluation of generalizability. To evaluate the generalizable ability of agents trained on data generated by MetaUrban we compare the success rate of four settings in Figure a. Setting and Setting are the results of training on MetaUrbantrain while testing on MetaUrbantest and MetaUrbanunseen respectively. We can observe a performance drop on MetaUrbanunseen. However the zeroshot results still achieve success rate facing various outofdistribution scenes demonstrating the strong generalizability of models trained on largescale data created by MetaUrban. Setting and Setting are the results of direct training on MetaUrbanfinetune and finetuning on MetaUrbanfinetune from the pretrained model on MetaUrbantrain. Compared between Setting and Setting we can observe an obvious performance drop which is caused by an underfitting of the insufficient and complex finetuning data. Setting outperforms Setting by a large margin demonstrating that the model trained on the MetaUrbanK dataset can provide informative priors as good initializations for quick tuning. Evaluation of scaling ability. To evaluate the scaling ability of MetaUrbans compositional architecture we train models on a different number of generated scenes from to . As shown in Figure b the performance improves remarkably from to as we include more scenes for training demonstrating the strong scaling ability of MetaUrban. MetaUrbans compositional nature has the potential to extend more diverse scenes with a larger element repository in the future which could further boost the agents performance. Evaluation of static and dynamic density. To evaluate the influence of static object density and dynamic environmental agents we evaluate the different proportions of them on the PointNav and SocialNav tasks respectively from to . Note that we keep the number of training scenes unchanged when sampling different densities. As shown in Figure c and d with the increasing density of both static objects and dynamic agents the success rates of both train and test experience dramatic degradations demonstrating the challenges for embodied agents when facing crowded streets in urban scenes. In our experiments we observe many interesting failure cases that can indicate promising future directions to improve AIs performance in MetaUrban and ultimately in realworld urban scenes. We make a detailed discussion in the Appendix. Conclusion We propose a new compositional simulator MetaUrban to facilitate embodied AI and robotics research in urban scenes. MetaUrban can generate infinite urban environments with complex scene structures and diverse movements of pedestrians and other mobile agents. These environments used as training data can significantly improve the generalizability and safety of the embodied AI underlying different mobile machines from food delivery bots to humanoids. We commit ourselves to developing the opensource simulator and fostering the community effort to turn it into a sustainable community infrastructure. References Josh Achiam Steven Adler Sandhini Agarwal Lama Ahmad Ilge Akkaya Florencia Leoni Aleman Diogo Almeida Janko Altenschmidt Sam Altman Shyamal Anadkat et al. Gpt technical report. arXiv preprint arXiv. . Peter Anderson Angel X. Chang Devendra Singh Chaplot Alexey Dosovitskiy Saurabh Gupta Vladlen Koltun Jana Kosecka Jitendra Malik Roozbeh Mottaghi Manolis Savva and Amir R. Zamir. On evaluation of embodied navigation agents. arXiv preprint arXiv. . Michael Bain and Claude Sammut. A framework for behavioural cloning. In Koichi Furukawa Donald Michie and Stephen H. Muggleton editors MI . Dhruv Batra Aaron Gokaslan Aniruddha Kembhavi Oleksandr Maksymets Roozbeh Mottaghi Manolis Savva Alexander Toshev and Erik Wijmans. Objectnav revisited On evaluation of embodied agents navigating to objects. arXiv preprint arXiv. . Abhijat Biswas Allan Wang Gustavo Silvera Aaron Steinfeld and Henny Admoni. Socnavbench A grounded simulation testing framework for evaluating social navigation. THRI . Michael J Black Priyanka Patel Joachim Tesch and Jinlong Yang. Bedlam A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In CVPR . Panpan Cai Yiyuan Lee Yuanfu Luo and David Hsu. Summit A simulator for urban driving in massive mixed traffic. In ICRA . Changan Chen Yuejiang Liu Sven Kreiss and Alexandre Alahi. Crowdrobot interaction Crowdaware robot navigation with attentionbased deep reinforcement learning. In ICRA . Zilong Chen Feng Wang and Huaping Liu. Texttod using gaussian splatting. arXiv preprint arXiv. . Marius Cordts Mohamed Omran Sebastian Ramos Timo Rehfeld Markus Enzweiler Rodrigo Benenson Uwe Franke Stefan Roth and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR . Boris De Wilde Adriaan W Ter Mors and Cees Witteveen. Push and rotate a complete multiagent pathfinding algorithm. JAIR . Matt Deitke Dhruv Batra Yonatan Bisk Tommaso Campari Angel X. Chang Devendra Singh Chaplot Changan Chen Claudia PrezDArpino Kiana Ehsani Ali Farhadi Li FeiFei Anthony G. Francis Chuang Gan Kristen Grauman David Hall Winson Han Unnat Jain Aniruddha Kembhavi Jacob Krantz Stefan Lee Chengshu Li Sagnik Majumder Oleksandr Maksymets Roberto MartnMartn Roozbeh Mottaghi Sonia Raychaudhuri Mike Roberts Silvio Savarese Manolis Savva Mohit Shridhar Niko Snderhauf Andrew Szot Ben Talbot Joshua B. Tenenbaum Jesse Thomason Alexander Toshev Joanne Truong Luca Weihs and Jiajun Wu. Retrospectives on the embodied ai workshop. arXiv preprint arXiv. . Matt Deitke Winson Han Alvaro Herrasti Aniruddha Kembhavi Eric Kolve Roozbeh Mottaghi Jordi Salvador Dustin Schwenk Eli VanderBilt Matthew Wallingford Luca Weihs Mark Yatskar and Ali Farhadi. Robothor An open simulationtoreal embodied AI platform. In CVPR . Matt Deitke Ruoshi Liu Matthew Wallingford Huong Ngo Oscar Michel Aditya Kusupati Alan Fan Christian Laforte Vikram Voleti Samir Yitzhak Gadre et al. Objaversexl A universe of m d objects. NeuIPS . Matt Deitke Ruoshi Liu Matthew Wallingford Huong Ngo Oscar Michel Aditya Kusupati Alan Fan Christian Laforte Vikram Voleti Samir Yitzhak Gadre Eli VanderBilt Aniruddha Kembhavi Carl Vondrick Georgia Gkioxari Kiana Ehsani Ludwig Schmidt and Ali Farhadi. Objaversexl A universe of m d objects. arXiv preprint arXiv. . Matt Deitke Dustin Schwenk Jordi Salvador Luca Weihs Oscar Michel Eli VanderBilt Ludwig Schmidt Kiana Ehsani Aniruddha Kembhavi and Ali Farhadi. Objaverse A universe of annotated d objects. In CVPR . Matt Deitke Dustin Schwenk Jordi Salvador Luca Weihs Oscar Michel Eli VanderBilt Ludwig Schmidt Kiana Ehsani Aniruddha Kembhavi and Ali Farhadi. Objaverse A universe of annotated d objects. In Proceedings of the IEEECVF Conference on Computer Vision and Pattern Recognition pages . Matt Deitke Eli VanderBilt Alvaro Herrasti Luca Weihs Kiana Ehsani Jordi Salvador Winson Han Eric Kolve Aniruddha Kembhavi and Roozbeh Mottaghi. Procthor Largescale embodied ai using procedural generation. NeuIPS . Alexey Dosovitskiy German Ros Felipe Codevilla Antonio Lopez and Vladlen Koltun. Carla An open urban driving simulator. In CoRL . Kiana Ehsani Winson Han Alvaro Herrasti Eli VanderBilt Luca Weihs Eric Kolve Aniruddha Kembhavi and Roozbeh Mottaghi. Manipulathor A framework for visual object manipulation. In CVPR . Martin Ester HansPeter Kriegel Jorg Sander Xiaowei Xu et al. A densitybased algorithm for discovering clusters in large spatial databases with noise. In kdd volume pages . Michael Everett Yu Fan Chen and Jonathan P. How. Motion planning among dynamic decisionmaking agents with deep reinforcement learning. In IROS . Scott Fujimoto and Shixiang Shane Gu. A minimalist approach to offline reinforcement learning. NeuIPS . Chuang Gan Jeremy Schwartz Seth Alter Damian Mrowca Martin Schrimpf James Traer Julian De Freitas Jonas Kubilius Abhishek Bhandwaldar Nick Haber Megumi Sano Kuno Kim Elias Wang Michael Lingelbach Aidan Curtis Kevin T. Feigelis Daniel Bear Dan Gutfreund David D. Cox Antonio Torralba James J. DiCarlo Josh Tenenbaum Josh H. McDermott and Dan Yamins. Threedworld A platform for interactive multimodal physical simulation. In NeurIPS Datasets and Benchmarks . Timnit Gebru Jamie H. Morgenstern Briana Vecchione Jennifer Wortman Vaughan Hanna M. Wallach Hal Daum and Kate Crawford. Datasheets for datasets. Comm. of the ACM . Patrick Geddes. Cities in Evolution. . Mike Goslin and Mark R Mine. The pandad graphics engine. Computer . Peter Hall and Mark TewdwrJones. Urban and regional planning. Routledge . Gans Herbert. The Urban Villagers Group and Class in the Life of ItalianAmericans. . Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. NeuIPS . Global Designing Cities Initiative and National Association of City Transportation Officials. Global street design guide. Island Press . Jane Jacobs. The Death and Life of Great American Cities. . Jiaming Ji Jiayi Zhou Borong Zhang Juntao Dai Xuehai Pan Ruiyang Sun Weidong Huang Yiran Geng Mickel Liu and Yaodong Yang. Omnisafe An infrastructure for accelerating safe reinforcement learning research. arXiv preprint arXiv. . Bernhard Kerbl Georgios Kopanas Thomas Leimkhler and George Drettakis. d gaussian splatting for realtime radiance field rendering. TOG . Eric Kolve Roozbeh Mottaghi Winson Han Eli VanderBilt Luca Weihs Alvaro Herrasti Matt Deitke Kiana Ehsani Daniel Gordon Yuke Zhu Kembhavi Aniruddha Gupta Abhinav and Farhadi Ali. Aithor An interactive d environment for visual ai. arXiv preprint arXiv. . Ilya Kostrikov Ashvin Nair and Sergey Levine. Offline reinforcement learning with implicit qlearning. arXiv preprint arXiv. . Parth Kothari Christian Perone Luca Bergamini Alexandre Alahi and Peter Ondruska. Drivergym Democratising reinforcement learning for autonomous driving. arXiv preprint arXiv. . Daniel Krajzewicz Georg Hertkorn Christian Rssel and Peter Wagner. Sumo simulation of urban mobilityan opensource traffic simulation. In MESM . Edouard Leurent. An environment for autonomous driving decisionmaking. httpsgithub.com eleurenthighwayenv . Chengshu Li Fei Xia Roberto MartnMartn Michael Lingelbach Sanjana Srivastava Bokui Shen Kent Elliott Vainio Cem Gokmen Gokul Dharan Tanish Jain Andrey Kurenkov C. Karen Liu Hyowon Gweon Jiajun Wu Li FeiFei and Silvio Savarese. igibson . Objectcentric simulation for robot learning of everyday household tasks. In CoRL . Chengshu Li Ruohan Zhang Josiah Wong Cem Gokmen Sanjana Srivastava Roberto MartnMartn Chen Wang Gabrael Levine Wensi Ai Benjamin Martinez Hang Yin Michael Lingelbach Minjune Hwang Ayano Hiranaka Sujay Garlanka Arman Aydin Sharon Lee Jiankai Sun Mona Anvari Manasi Sharma Dhruva Bansal Samuel Hunter KyuYoung Kim Alan Lou Caleb R. Matthews Ivan VillaRenteria Jerry Huayang Tang Claire Tang Fei Xia Yunzhu Li Silvio Savarese Hyowon Gweon C. Karen Liu Jiajun Wu and Li FeiFei. Behaviork A humancentered embodied ai benchmark with everyday activities and realistic simulation. CoRL . Junnan Li Dongxu Li Silvio Savarese and Steven Hoi. Blip Bootstrapping languageimage pretraining with frozen image encoders and large language models . Junnan Li Dongxu Li Caiming Xiong and Steven Hoi. Blip Bootstrapping languageimage pretraining for unified visionlanguage understanding and generation. In ICML . Quanyi Li Zhenghao Peng Lan Feng Qihang Zhang Zhenghai Xue and Bolei Zhou. Metadrive Composing diverse driving scenarios for generalizable reinforcement learning. TPAMI . Haotian Liu Chunyuan Li Qingyang Wu and Yong Jae Lee. Visual instruction tuning. NeuIPS . Ruoshi Liu Rundi Wu Basile Van Hoorick Pavel Tokmakov Sergey Zakharov and Carl Vondrick. Zeroto Zeroshot one image to d object. In Proceedings of the IEEECVF International Conference on Computer Vision . Shilong Liu Zhaoyang Zeng Tianhe Ren Feng Li Hao Zhang Jie Yang Chunyuan Li Jianwei Yang Hang Su Jun Zhu et al. Grounding dino Marrying dino with grounded pretraining for openset object detection. arXiv preprint arXiv. . Tiange Luo Justin Johnson and Honglak Lee. View selection for d captioning via diffusion ranking. arXiv preprint arXiv. . Tiange Luo Chris Rockwell Honglak Lee and Justin Johnson. Scalable d captioning with pretrained models. NeuIPS . Mark Martinez Chawin Sitawarin Kevin Finch Lennart Meincke Alex Yablonski and Alain Kornhauser. Beyond grand theft auto v for training testing and enhancing deep learning in self driving cars. arXiv preprint arXiv. . Matthias Mller Vincent Casser Jean Lahoud Neil Smith and Bernard Ghanem. Simcv A photorealistic simulator for computer vision applications. IJCV . Gerhard Neuhold Tobias Ollmann Samuel Rota Bulo and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV . OpenAI. Gpto. httpsopenai.comindexhellogpto . OpenStreetMap contributors. Planet dump retrieved from httpsplanet.osm.org . httpswww. openstreetmap.org . Joon Sung Park Joseph OBrien Carrie Jun Cai Meredith Ringel Morris Percy Liang and Michael S Bernstein. Generative agents Interactive simulacra of human behavior. In UIST . Robert Ezra Park Ernest Watson Burgess Roderick Duncan McKenzie and Louis Wirth. The City. . Priyanka Patel ChunHao P Huang Joachim Tesch David T Hoffmann Shashank Tripathi and Michael J Black. Agora Avatars in geography optimized for regression analysis. In CVPR . Georgios Pavlakos Vasileios Choutas Nima Ghorbani Timo Bolkart Ahmed A. A. Osman Dimitrios Tzionas and Michael J. Black. Expressive body capture D hands face and body from a single image. In CVPR . No PrezHigueras Roberto Otero Fernando Caballero and Luis Merino. Hunavsim A ros human navigation simulator for benchmarking humanaware robot navigation. arXiv preprint arXiv. . Ben Poole Ajay Jain Jonathan T Barron and Ben Mildenhall. Dreamfusion Texttod using d diffusion. In ICLR . Xavier Puig Kevin Ra Marko Boben Jiaman Li Tingwu Wang Sanja Fidler and Antonio Torralba. Virtualhome Simulating household activities via programs. In CVPR . Xavier Puig Tianmin Shu Joshua B Tenenbaum and Antonio Torralba. Nopa Neurallyguided online probabilistic assistance for building socially intelligent home assistants. In ICRA . Xavier Puig Eric Undersander Andrew Szot Mikael Dallaire Cote TsungYen Yang Ruslan Partsey Ruta Desai Alexander William Clegg Michal Hlavac So Yeon Min Vladimir Vondrus Thophile Gervet VincentPierre Berges John M. Turner Oleksandr Maksymets Zsolt Kira Mrinal Kalakrishnan Jitendra Malik Devendra Singh Chaplot Unnat Jain Dhruv Batra Akshara Rai and Roozbeh Mottaghi. Habitat . A cohabitat for humans avatars and robots. In ICLR . Alex Ray Joshua Achiam and Dario Amodei. Benchmarking safe exploration in deep reinforcement learning. arXiv preprint arXiv. . Nils Reimers and Iryna Gurevych. Sentencebert Sentence embeddings using siamese bertnetworks. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics . Tianhe Ren Shilong Liu Ailing Zeng Jing Lin Kunchang Li He Cao Jiayu Chen Xinyu Huang Yukang Chen Feng Yan et al. Grounded sam Assembling openworld models for diverse visual tasks. arXiv preprint arXiv. . Manolis Savva Jitendra Malik Devi Parikh Dhruv Batra Abhishek Kadian Oleksandr Maksymets Yili Zhao Erik Wijmans Bhavana Jain Julian Straub Jia Liu and Vladlen Koltun. Habitat A platform for embodied AI research. In ICCV . John Schulman Filip Wolski Prafulla Dhariwal Alec Radford and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv. . Shital Shah Debadeepta Dey Chris Lovett and Ashish Kapoor. Airsim Highfidelity visual and physical simulation for autonomous vehicles. In FSR . Bokui Shen Fei Xia Chengshu Li Roberto MartnMartn Linxi Fan Guanzhi Wang Claudia PrezDArpino Shyamal Buch Sanjana Srivastava Lyne Tchapmi Tchapmi Micael Vainio Kent Wong Josiah FeiFei Li and Savarese Silvio. igibson . a simulation environment for interactive tasks in large realistic scenes. In IROS . Zayne Sprague Rohan Chandra Jarrett Holtz and Joydeep Biswas. Socialgym . Simulator for multiagent social robot navigation in shared human spaces. arXiv preprint arXiv. . Hao Sun Ziping Xu Meng Fang Zhenghao Peng Jiadong Guo Bo Dai and Bolei Zhou. Safe exploration by solving early terminated mdp. arXiv preprint arXiv. . Andrew Szot Alexander Clegg Eric Undersander Erik Wijmans Yili Zhao John M. Turner Noah Maestre Mustafa Mukadam Devendra Singh Chaplot Oleksandr Maksymets Aaron Gokaslan Vladimir Vondrus Sameer Dharur Franziska Meier Wojciech Galuba Angel X. Chang Zsolt Kira Vladlen Koltun Jitendra Malik Manolis Savva and Dhruv Batra. Habitat . Training home assistants to rearrange their habitat. In NeuIPS . Deepdrive Team. Deepdrive a simulator that allows anyone with a pc to push the stateoftheart in selfdriving. httpsgithub.comdeepdrivedeepdrive. Udacity Team. Udacitys selfdriving car simulator A selfdriving car simulator built with unity. https github.comudacityselfdrivingcarsim. Nathan Tsoi Alec Xiang Peter Yu Samuel S Sohn Greg Schwartz Subashri Ramesh Mohamed Hussein Anjali W Gupta Mubbasir Kapadia and Marynel Vzquez. Sean . Formalizing and generating social situations for robot navigation. RAL . Jur Van Den Berg Stephen J Guy Ming Lin and Dinesh Manocha. Reciprocal nbody collision avoidance. In ISRR . Stephen Victor Ward. Planning the twentiethcentury city the advanced capitalist world. . William H Whyte. City Rediscovering the center. University of Pennsylvania Press . Tong Wu Jiarui Zhang Xiao Fu Yuxin Wang Jiawei Ren Liang Pan Wayne Wu Lei Yang Jiaqi Wang Chen Qian Dahua Lin and Ziwei Liu. Omniobjectd Largevocabulary d object dataset for realistic perception reconstruction and generation. In CVPR . Zhitao Yang Zhongang Cai Haiyi Mei Shuai Liu Zhaoxi Chen Weiye Xiao Yukun Wei Zhongfei Qing Chen Wei Bo Dai Wayne Wu Chen Qian Dahua Lin Ziwei Liu and Lei Yang. Synbody Synthetic dataset with layered human models for d human perception and modeling. In ICCV . Appendix A MetaUrban Visualization A. Static Scene Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Dynamic Scene Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Static Asset Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Dynamic Asset Samples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B MetaUrban Simulator B. Layout Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Object Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Cohabitant Populating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Scene Customization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B. Simulator Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C Experiment Details C. PointNav Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C. SocialNav Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D Datasheet E Performance F Discussion A MetaUrban Visualization A. Static Scene Samples Street blocks. We design five typical street block categories straight curve intersection Tjunction and roundabout. In the simulator to form a large map with several blocks we can sample the category number order lane number and other related parameters of the blocks. We use the algorithm Block Incremental Generation BIG proposed in MetaDrive to generate the target road network defined by users. Figure provides demonstrations of generated street maps composed of different numbers of blocks. Figure Examples of block maps. Generated block maps with a different number of street blocks. Ground layouts. We construct seven typical templates for sidewalks more details about the design and the generation process are given in the Section B.. As shown in Figure different types of sidewalks can be sampled on the same street block each type has its unique division and specification of functional zones. Figure further shows several block maps with a different type of sidewalks. Figure Examples of sidewalks. Generated sidewalks with seven templates a to g. Figure Examples of block maps with sidewalks. Generated block maps with a different type of sidewalks. Static objects. To generate static objects we build the object placement distribution conditioned on geometric zones of sidewalks which will be discussed in Section B.. To better distinguish between the difficulty of scenes on the same road network we use the object density s to control the crowding level on the sidewalk. This indicates the ratio of the minimum distance to the default distance between objects. Figure shows block maps with different object densities. We can observe that when the density increases the walkable region will become more and more crowded. Figure further gives egoview results by randomly sampling viewpoints on block maps. Figure Examples of block maps with different object densities. Each row is randomly sampled block maps with one object density from to . Figure Examples of egoview results in static scenes. Each row is a different object placement with the same object density . For each row we sample viewpoints to show egoview results. A. Dynamic Scene Samples Dynamic agents such as pedestrians vulnerable road users like bikers skateboarders scooter riders mobile machines delivery bots electric wheelchairs robot dogs and humanoid robots and vehicles will be present in the environment. The density of dynamic agents can be controlled with dynamic density ratio d. Figure shows egoview results by randomly sampling viewpoints on block maps. The urban spaces are well populated with different agents. Figure Examples of egoview results in dynamic scenes. Each row is a different specification of dynamics appearances movements and trajectories with the same dynamic density . For each row we sample viewpoints to show egoview results. A. Static Asset Samples We provide highquality static object assets. The roadside infrastructure is divided into three categories Standard infrastructure including poles trees and signs is placed at regular intervals along the road. Nonstandard infrastructure such as buildings bonsai and trash bins is placed randomly within designated zones. Clutter such as drink cans bags and bicycles is scattered randomly across all functional zones. Figure and show examples of these three categories respectively. Figure Examples of static objects standard infrastructure. Figure Examples of static objects nonstandard infrastructure. Figure Examples of static objects clutter. A. Dynamic Asset Samples Human assets. MetaUrban provides rigged D human models sampled from garments hairs beards accessories and cloth and skin textures from SynBody dataset. Figure shows randomly sampled humans which have large variations. Figure Examples of dynamics rigged humans. Vulnerable road user assets. MetaUrban provides kinds of vulnerable road users to form safecritical scenarios. They are bikers skateboarders scooter riders and electric wheelchair users as shown in the first row of Figure . Note that electric wheelchairs as a humanAI shared control system can also be seen as mobile machines not only vulnerable road users. Figure Examples of dynamics vulnerable road users. Mobile machine assets. MetaUrban provides kinds of mobile machines Starship Yandex Rover and COCO Robotics delivery bots Boston Dynamics robot dog Agility Robotics humanoid robot and Drive Medicals electric wheelchair. Figure shows the first assets while the electric wheelchair as a crosscategory asset vulnerable road user and mobile machine is shown in Figure . Figure Examples of dynamics mobile machines. Vehicle assets. MetaUrban provides kinds of vehicles covering different body types sizes and appearances. Figure shows sampled vehicles. B MetaUrban Simulator B. Layout Generation This section gives details about the process we developed to procedurally generate scenes with sidewalks and crosswalks as well as sample and place static objects on the sidewalk. Ground plan. As shown in Figure Top we define functional zones and geometric zones for sampling the type of sidewalks and choosing the distribution of parameters for each sidewalk component. As shown in Figure Bottom we construct typical templates for sidewalks each type of them has its unique distribution of geometric zones. To match the distribution with the real Figure Examples of dynamics vehicles. world we set the distribution of the zone width to a uniform distribution for each geometric zone the maximum and minimum values of the uniform distribution are set according to the Global Street Design Guide provided by the Global Designing Cities Initiative. Figure Architecture of ground layouts. Top The mapping from functional zones to geometric zones. Bottom Specifications of geometric zones for sidewalk templates. To generate a scene we will first sample the template of the sidewalk z from its distribution z ZT ZT z z ... z followed by the sampling of widths of each geometric zone wi fwz i i ... where fwz i is the width distribution of the ith geometric zone under the sidewalk template z. Crosswalks are crucial for the connectivity of scenes. MetaUrban provides candidates at the beginning and end of each roadway of a block. Then locations of the crosswalk can be controlled by a crosswalk density parameter or be specified by users directly. Object placement. Figure illustrates the iterative process of placing objects in the scene. First we convert the polygon of the geometric zone of the sidewalk into rectangles. We will place objects on each geometric zone independently. At each step of placing on the specific zone we can obtain rectangles that are not occupied. Then we check from the starting region to the ending region for the current retrieved object class. We place it if possible or we start to place the next class. In the simulator we use rectangle bounding boxes to represent all objects physically to adopt this object placement method. B. Object Retrieval Distribution extraction. Distinguished from the recent indoor simulation platform there are no readytouse highquality asset datasets for urban spaces. Urban spaces have their unique data Figure Iterative object placement. Given a ground plan we first convert geometric zones into rectangles. Then we place objects on zones one by one. In the placing of each zone we attempt to put sampled classes of objects in the rectangles one by one. distribution such as the infrastructure built by the urban planning administration fire hydrants and bus stops and clutters placed by people scooters and advertising boards. Thus we design a realworld distribution extraction method to get a description pool depicting what objects are frequently shown in urban spaces. We first leverage offtheshelf scene understanding datasets Mapillary Vistas and CityScape . Using the provided annotation polygon we find the overlapping object with the sidewalk and get a list of objects that are with high frequency to be put in the urban space such as tree and bench. However the number of objects is limited because of the closedset definitions in the image datasets. To get broader object distribution from the real world we introduce two openset sources worldwide Google Street data and urban planning description data. For the Google Street data we collect urban space images from countries across six continents. The selection of image locations was performed by randomly sampling points along the major roads of cities using OpenStreetMaps road network. Image orientation was determined based on road gradient to enhance the relevance of captured scenes. For object detection in these images we initially employed GPTo to generate a list of candidate objects. This was followed by the application of GroundedDino to obtain bounding boxes for these objects. We refined these boxes using nonmaximum suppression NMS to ensure the accuracy of object identification. Further refinement was achieved through the use of the GroundedSAM model an openset segmentation approach which filtered the bounding boxes to identify objects specifically located in public urban spaces. A key part of our method involves determining overlaps between identified objects and sidewalks. For each object detected we calculate its spatial intersection with sidewalk regions derived from the datasets. This overlap analysis helps in curating a list of objects that are relevant to urban public spaces. To address the diverse descriptions generated by GPTo and ensure semantic uniformity we cluster the embeddings of descriptions using DBSCAN which result in distinct object clusters with unique descriptors such as a gray trash bin and potted cactus. We use allmpnetbasev model from SentenceTransformers to embed each description. For the urban planning description data we get a list of essential objects in public urban spaces such as drinking fountains and bike racks through a thorough survey of urban design handbooks. Finally by combining these three data sources we can get an object description pool with items of descriptions that can form the realworld object category distribution. Figure illustrates the distribution of objects in urban space extracted from all of the worldwide collected data. Houses gates and trees emerge as universal elements dominating the urban landscapes across all depicted countries reflecting their fundamental role in both urban and rural settings. Figure illustrates the object distribution of example countries from continents showcasing distinct environmental and cultural characteristics through object prevalence. The data also highlights notable regional distinctions Japan for instance features a higher incidence of poles and road cones hinting at unique aspects of its urban infrastructure. In contrast Brazils considerable frequency of gates and metal gates suggests prominent architectural and security preferences. Such variances not only reveal the diverse urban aesthetics and functional priorities across different regions but also enhance our understanding of how specific objects can define the character and utility of public spaces Figure Distribution of objects in urban spaces for all collected data worldwide. globally. This comparative analysis of object distributions contributes significantly to constructing regionspecific sidewalks simulation environments. Figure Distribution of objects in urban spaces across different countries. Two example images are shown together with each distribution figure demonstrating large variations among different countries. Openvocabulary search. To effectively retrieve digital assets corresponding to the object description pool we developed a robust pipeline utilizing the Objaverse and ObjaverseXL repositories known for their extensive digital assets. The process begins with the extraction of digital assets using a multithreaded approach for further processing. Each downloaded asset is then rendered into distinct images capturing various angles to provide a comprehensive visual representation. Following viewpoints with higher quality are used for the calculation of visual feature embedding. For the matching process we leverage the BLIP model a pretrained feature extractor to align visual data with our textual descriptions. This involves processing the images to extract visual features and concurrently transforming textual descriptions into embeddings. These embeddings are compared using cosine similarity to determine the semantic correspondence between text and images allowing us to identify and collect the digital assets that best match the descriptions. Once the assets are collected a meticulous review process is initiated for each category. We manually inspect each asset filtering out those that are of low resolution lack realism or do not meet our quality standards. The selected assets are then uploaded into MetaUrban to adjust asset characteristics such as size position and orientation. This meticulous curation ensures that only highquality digital assets are incorporated into our static object dataset. Object repository extension. MetaDrive provides an interface for including objects enabled by recent advances in D content generation such as D object reconstruction and generation . Thus one can easily further extend the object repository with generated contents. Also this function can work together with scene customization Section B. to get customized scenes with specific objects. B. Cohabitant Populating Appearances. We include D human models kinds of vulnerable road users bikers skateboarders scooter riders and electric wheelchair users and kinds of mobile machines as cohabitants in the MetaUrban simulator. The number of dynamic agents in a scene can be set by the parameters respectively. The environment initialization time and RAM usage are only proportional to the number of individual agents. For example same agents will take the same initialization time and RAM usage as one. This schema can be used to significantly increase the maximum number of spawned agents for a specific hardware. Movements. We include daily movements idle walking and running as well as unique movements from the BEDLAM dataset. All of the motion sequences are trimmed and checked by designers one by one to ensure their quality. With the same skeletal binding all of the unique movements can be transferred to all of the D human models directly. Thus we can get numbers of humanmotion pairs. Trajectories. We harness ORCA and Push and Rotate PR algorithm to get the trajectories of all dynamic agents. First we build the mask that indicates whether the grid is a walkable region or not. Then we sample the start and ending points for each agent randomly followed by generating their D trajectories by using the model of ORCA and PR . The trajectory plan process is efficient running within s for agents on a Core i CPU processor. Vehicles will also be added in dynamic scenes. All traffic vehicles will follow IDM policies as MetaDrive does. B. Scene Customization MetaUrban supplies various compositional elements such as street blocks objects pedestrians vulnerable road users and other mobile agents appearances and dynamics. With just a few simple lines of specification it is easy to create customized urban spaces of interest such as street corners plazas and parks. B. User Interface MetaUrban provides user interfaces for two purposes Demonstration data collection for Offline RL and IL. Object labeling and scene customization. For demonstration data collection MetaUrban provides interfaces for mouse keyboard joystick and racing wheel. We can easily collect human expert demonstrations as shown in Figure . In addition MetaUrban provides tools for object labeling size orientation and attributes and scene customization assigning the locations of the selected objects. Figure Demonstration data collection with the user interface. B. Simulator Comparison We will compare MetaUrban with other simulators below in Table using the scale sensor and feature dimensions. For the scale MetaUrban can generate infinite scenes with a procedural generation pipeline. It provides the largest number of humans and movements among all simulation environments. For objects so far we have provided . Compared to other simulators all of the objects from MetaUrban are urbanspecific. Also we provide an interface to extend object data to any size easily with recent advances in D content generation Section B.. For the sensor MetaUrban provides RGBD semantic and lidar. For the feature different from other simulators MetaUrban provides realworld distribution of the objects categories and uses a more sophisticated path plan algorithm to get the natural agents trajectories. It also provides flexible user interfaces mouse keyboard joystick and racing wheel which vastly ease the collection of human expert demonstration data. MetaUrban uses PyBullet as its physical engine and PandaD for rendering. Table Comparison of Embodied AI simulators. We compare MetaUrban to simulators specialized for three environments indoor driving and social navigation environments. Scale Sensor Feature Simulator of Scenes of Objects of Rigged Humans of Human Motions RGBD Semantic LiDAR Acoustic Object Category Distribution Env. Agent Trajectory User Interface Physics Engine Scenario HuNavSim Social Force Gazebo Social SEAN . Manual Social Force Unity Social SocNavBench Social SUMO Driving CARLA Manual Rulebased Keyboard Joystick Unreal Driving MetaDrive Manual Rulebased PyBullet Driving AITHOR Manual Mouse Unity Indoor ThreeDWorld Manual VR Flex Indoor iGibson . Manual Mouse VR PyBullet Indoor ProcTHOR Manual Unity Indoor OmniGibson Manual PhysX Indoor Habitat . Manual Rulebased Mouse Keyboard VR Bullet Indoor MetaUrban Realworld ORCA Mouse Keyboard PyBullet Urban PR Joystick Racing Wheel C Experiment Details This section discusses the settings of environments action spaces observation spaces evaluation metrics training details for methods as well as the reward and cost in the benchmarks of Point Navigation PointNav and Social Navigation SocialNav respectively. C. PointNav Experiments Environments. For PointNav experiments there are only static objects besides the ego agent in the environment. To evaluate the trained policy we split seven types of sidewalks into six types for training and validation with one for test. The one used for the test is the Wide Commercial Sidewalk in which the frontage zone buffer will be as well as some unseen objects. We use delivery robots as the ego agent in our experiments. The task of agents in PointNav experiments is following the trajectory in the environment that navigates from start points to ending points ensuring that it does not collide with other objects. To generate such a task we harness ORCA and Push and Rotate PR algorithm to get the trajectory of the ego agent after placing objects. The process is the same as discussed in Section B.. Notably there may be some trajectories with small moving distances we set a threshold of m to filter out scenarios with small moving distances for testing to evaluate different methods more effectively. Action spaces. We use the same continuous action space as MetaDrive which is a dimensional vector that normalized to . . indicating the acceleration and steering rate of the agent. Considering that the dynamics of a delivery robot is different from a vehicle we change some core parameters like maximum velocity maximum acceleration maximum steering rate and so on. Observation spaces. Multimodal observations are provided by MetaUrban including RGB Depth Semantic Map and LidAR. We use LidAR in all of our experiments for its D information of the surrounding environment which provides distance and direction of the nearest object within a m maximum detecting distance centering at the ego. Evaluation metrics. For PointNav an episode is considered successful if the agent issues the DONE action defined as completing of the set route within maximum steps. The agent is evaluated using the Success Rate SR and Success weighted by Path Length SPL metrics which measure the effectiveness and efficiency of the path taken by the agent. Additionally to measure the safety performance of the trained policy we define the cost function by two events i.e. crashing with objects on the sidewalk or buildings in the building zone. cost is given once those events occur. Methods. In our study we employ a diverse set of baseline models to establish comprehensive benchmarks on MetaUrban. These models span various domains including Reinforcement Learning Safe Reinforcement Learning Offline Reinforcement Learning and Imitation Learning. Reinforcement learning. In the realm of Reinforcement Learning we use the Proximal Policy Optimization PPO for evaluation. PPO is a widely adopted and effective method that strikes a balance between sample complexity and ease of tuning and it is easy to scale as it adopts parallel and distributed training well. The agent in this setting is trained to maximize the reward which we carefully design to encapsulate the desired behavior of the agent in the MetaUrban environment. The specifics of the reward structure will be discussed in the subsequent paragraph. We train the PPO using the same set of hyperparameters with parallel environments which occupy processes. The total training time is hours and M environment steps for PointNav on a single Nvidia A GPU. The detailed hyperparameters are provided in Table . Table Hyperparameters of RL and SafeRL for PointNav. PPOPPOLagPPOET Hyperparameters Value Environmental horizon T Learning rate e Discount factor . GAE parameter . Clip parameter . Train batch size SGD minibatch size Value loss coefficient . Entropy loss coefficient . Cost limit Safe reinforcement learning. As driving in urban spaces is a safetycritical application it is important to evaluate Safe Reinforcement Learning SafeRL algorithms. In the domain of SafeRL we utilize two approaches PPO with a Lagrangian constraint PPOLag and PPO with modeling of Early Terminated Markov Decision Processes PPOET . Both methods aim to ensure that the learned policies adhere to specific safety constraints while optimizing the reward. PPOLag incorporates a Lagrangian term into the objective function to enforce the constraints while PPOET changes the modeling of the Constrained Markov Decision Process CMDP to a new unconstrained MDP the optimal policy that coincidences with the original CMDP. For PPOLag it considers the learning objectivate as Equation rather than adding negative cost as rewards. max min E R C d where R C and d are episodic reward episodic cost parameters of the policy and given cost threshold respectively. The rule for PPOET is to stop when the constraint cost exceeds a given value which can be easily implemented in practice. We implement both of these SafeRL methods based on OmniSafe . We train both of them with parallel environments and the training takes hours for PointNav on a single Nvidia A GPU. The detailed hyperparameters are provided in Table . Offline reinforcement learning. For Offline Reinforcement Learning we employ two prominent methods Implicit QLearning IQL and Twin Delayed Deep Deterministic Policy Gradient with Behavior Cloning TDBC . We create the dataset for PointNav by combining human demonstrations with demonstrations from a welltrained PPO policy consisting of samples with approximately success rate. The training is purely offline and takes around hours on a single Nvidia A GPU for epochs. The detailed hyperparameters for IQL and TDBC are provided in Table and respectively. Table Hyperparameters of IQL. IQL Hyperparameters Value Learning rate e Discount factor . Target critic update ratio e Inverse temperature . Log std range . . Expectile . Table Hyperparameters of TDBC. TDBC Hyperparameters Value Learning rate e Discount factor . Target critic update ratio e Actor update delay BC loss coefficient . Imitation learning. For Imitation Learning algorithms we use the same highquality mixed demonstration used in Offline Reinforcement Learning. In the Imitation Learning setting the agent learns to mimic the behavior shown in the expert demonstration and it is differentiated from Offline Reinforcement Learning in the sense that the agent does not have access to the rewards. We employ two wellestablished methods Behavior Cloning BC and Generative Adversarial Imitation Learning GAIL . BC is a straightforward approach that trains the agent to directly match the actions of the expert given the observed states. GAIL on the other hand formulates the imitation learning problem as a twoplayer game between the agent and a discriminator which tries to distinguish between the agents behavior and the experts demonstrations. The detailed hyperparameters for IQL and TDBC are provided in table and respectively. Table Hyperparameters of BC. BC Hyperparameters Value Dataset size Learning rate e SGD batch size SGD epoch Table Hyperparameters of GAIL. GAIL Hyperparameters Value Dataset size SGD batch size Sample batch size Generator Learning rate e Discriminator Learning rate e Generator optimization epoch Discriminator optimization epoch Clip parameter . Reward and cost. The reward function is composed as follows R Rterm cRdisp cRlateral cRsteering cRcrash Specifically Terminal reward Rterm a sparse reward set to if the vehicle reaches the destination and for out of route. If given Rterm at any time step t the episode will be terminated at t immediately. Displacement reward Rdisp a dense reward defined as Rdisp dt dt wherein the dt and d denote the longitudinal position of the ego agent in Frenet coordinates of current lane at time t and t respectively. We set the weight of Rdisp as c .. Lateral reward Rlateral a dense reward defined as Rlateral lt wherein the lt denotes the lateral offset of the ego agent in Frenet coordinates of current lane at time t which is designed to prevent agent driving on non walkable areas. We set the weight of Rlateral as c .. Steering smoothness reward Rsteering a dense reward defined as Rsteering st st vt wherein the st and st denotes the steering of the agent at t and t respectively. And vt denotes the speed of the agent at time t. This reward term is designed as a regularization to prevent the agent changing the steering too frequently. We set the weight of Rsteering as c .. Crash reward Rcrash a dense negative reward defined as ct wherein the ct denotes the collision between agents and any other objects at time t and is the indicator function. Its notable we do not use the termination strategy for collision as in MetaDrive . We set the weight of Rcrash as c .. And for benchmarking Safe RL algorithms collision to any objects raises a cost at each time step. C. SocialNav Experiments Environments. For SocialNav experiments most settings are the same as the ones in PointNav. The most important difference is that dynamic agents will also be present in the environment. The trajectories of environmental agents are generated together by using the model of ORCA with PR . Since vehicles are inherited from MetaDrive we use the same parameter to control its density i.e. traffic density . in our experiments. Evaluation metrics. For SocialNav an episode is considered successful if the agent issues the DONE action defined as completing of the set route within maximum steps. The agent is evaluated using the Success Rate SR and Social Navigation Score SNS which is the average of Success weighted by Time Length STL and Personal Space Compliance PSC. SNS measures the agent in terms of safety and efficiency. Methods. We benchmark the same methods as in PointNav experiments with the same hyperparameters. However due to the involvement of lots of dynamic agents the training speed of SocialNav is about approximately of PointNav on online methods. The cost scheme is defined as raising a cost of at each time step if the ego agent crashes with any agents vehicles or objects. D Datasheet Motivation For what purpose was the dataset created? The dataset was created to enable agents training on diverse scenes and facilitate Embodied AI research in urban spaces. Who created and funded the dataset? This work was created and funded by the MetaUrban team at the University of California Los Angeles. Composition What do the instances that comprise the dataset represent? Each instance is a JSON file including the configuration of our MetaUrban environment and a specific seed. How many instances are there in total of each type if appropriate? There are urban scenes released in the MetaUrbanK dataset along with the code to sample substantially more. Does the dataset contain all possible instances or is it a sample not necessarily random of instances from a larger set? We offer urban scenes with the ability to generate more using procedural generation scripts. What data does each instance consist of? Each scene is specified as a JSON file including the configuration of our MetaUrban environment and a specific seed. Is there a label or target associated with each instance? No. Is any information missing from individual instances? No. Are relationships between individual instances made explicit e.g. users movie ratings social network links? Each urban scene is created independently so there are no connections between the scenes. Are there recommended data splits? Yes. See Section in the main paper. Are there any errors sources of noise or redundancies in the dataset? No. Is the dataset selfcontained or does it link to or otherwise rely on external resources e.g. websites tweets other datasets? The dataset is selfcontained. Does the dataset contain data that might be considered confidential? No. Does the dataset contain data that if viewed directly might be offensive insulting threatening or might otherwise cause anxiety? No. Collection Process How was the data associated with each instance acquired? Each scene was procedurally generated. If the dataset is a sample from a larger set what was the sampling strategy? The dataset consists of scenes each by sampling the parameters of its composed elements. Who was involved in the data collection process? The authors were the sole individuals responsible for creating the dataset. Over what timeframe was the data collected? Data was collected in May . Were any ethical review processes conducted? No. PreprocessingCleaningLabeling Was any preprocessingcleaninglabeling of the data done? We label each objects location area and pivots to make them spawn in target functional zones and face a natural direction. We use VLMs to automatically label D images of cities worldwide which enables the extraction of realworld category distribution of objects in urban spaces. Was the raw data saved in addition to the preprocessedcleanedlabeled data? There is no raw data. Is the software that was used to preprocesscleanlabel the data available? The code related to preprocessing cleaning and labeling the data will be made available. Uses Has the dataset been used for any tasks already? Yes. See Section of the main paper. What other tasks could the dataset be used for? The scenes can be used in a wide variety of tasks in embodied AI computer vision and urban planning. Is there anything about the composition of the dataset or the way it was collected and preprocessedcleanedlabeled that might impact future uses? No. Are there tasks for which the dataset should not be used? Our dataset can be used for both commercial and noncommercial purposes. Distribution Will the dataset be distributed to third parties outside of the entity on behalf of which the dataset was created? Yes. We plan to make the entirety of the work opensource including the code used to generate scenes and train agents the scripts to get the MetaUrbanK dataset and the asset repositories. How will the dataset be distributed? The scene files will be distributed with a custom Python package. The code asset and repositories will be distributed on GitHub. Will the dataset be distributed under a copyright or other intellectual property IP license andor under applicable terms of use ToU? The scene dataset D asset repository and code will be released under the Apache . license. Have any third parties imposed IPbased or other restrictions on the data associated with the instances? For D human assets we use Synbody . Its license is CC BYNCSA .. For movement sequences we use BEDLAM . See httpsbedlam.is.tue.mpg.de license.html for its license. Do any export controls or other regulatory restrictions apply to the dataset or to individual instances? No. Maintenance Who will be supportinghostingmaintaining the dataset? The authors will be providing support hosting and maintaining the dataset. How can the ownercuratormanager of the dataset be contacted? For inquiries email metaurbanteamgmail.com. Is there an erratum? We will use GitHub issues to track issues with the dataset. Will the dataset be updated? We will continue adding support for new features to make the urban scenes more diverse and realistic. We also intend to support new tasks in the future. If the dataset relates to people are there applicable limits on the retention of the data associated with the instances e.g. were the individuals in question told that their data would be retained for a fixed period of time and then deleted? The dataset does not relate to people. Will older versions of the dataset continue to be supportedhostedmaintained? Yes. Revision history will be available for older versions of the dataset. If others want to extendaugmentbuild oncontribute to the dataset is there a mechanism for them to do so? Yes. The work will be opensourced and we intend to offer support to assist others in using and building upon the dataset. Table A datasheet for MetaUrban and MetaUrbanK. E Performance We measure the performance of MetaUrban under varying street blocks different densities of static objects and dynamic agents in the scene. All experiments are conducted on a single Nvidia V GPU and in a single process. For the environment there are approximately objects covering m on average. We sample steps for actions and run times to report the average and standard error results of FPS. For the RGB and depth image we use the resolution. On average for the RGB Depth and LiDAR observation we achieve and FPS in training respectively. F Discussion Impact. As the first urban space simulator MetaUrban could benefit broad areas across Embodied AI Economy and Society. Embodied AI. MetaUrban contributes to advancing areas such as robot navigation social robotics and interactive systems. It could facilitate the development of robust AI systems capable of understanding and navigating complex urban environments. Economy. MetaUrban could be used in businesses and services operating in urban environments such as lastmile food delivery assistive wheelchairs and trashcleaning robots. It could also drive innovation in urban planning and infrastructure development by providing simulation tools and insights into how spaces are utilized thereby enhancing the economic and societal efficiency of public urban spaces like sidewalks and parks. Society. By enabling the safe integration of robots and AI systems in public spaces MetaUrban could support the development of assistive technologies that can aid in accessibility and public services. Using AI in public spaces might foster new forms of social interaction and community services making urban spaces more livable and joyful. Potential negative societal impacts. The integration of AI and robots in urban environments while beneficial raises several concerns. Increased surveillance could infringe on privacy while automation may lead to job displacement and exacerbate economic inequalities. Societal dependency on technology poses risks of dysfunction during failures and the presence of robots might alter social norms and interactions. Thus the environmental impact of manufacturing and operating urban simulators must be carefully managed. Addressing these issues is crucial for ensuring that the benefits of such technologies are realized without detrimental societal consequences. Limitations. Realworld scene distribution. In this work we extract object category distribution from realworld data of urban spaces. Other than the realworld distribution of object categories the distribution of object location and scene layout is also important for constructing specialized scenes for agent training. Extraction of such distribution relies on an accurate reconstruction of D scenes from realworld videos or even images and thus is extremely challenging. An interesting direction is extracting realworld scene distribution from inthewild videos including object category object location and scene layout. Then we can build a digital twin of a target scene for the agents training. It could help to develop scenespecific agents. Interactive agent behaviors. In this work we construct the environmental agents dynamic with deterministic methods determining their movements and trajectories with rules. However in the real world all environmental agents are interactive their behaviors are affected by each other and the surrounding environments. An interesting research direction is to endow personal traits like job personality and purpose to agents and harness the advances of LLMs and LVMs to form social and interactive behaviors of agents in urban scenes spontaneously. Robotspecific tasks. In this work we design point navigation and social navigation tasks as the pilot study using MetaUrban for embodied AI research. However different robots delivery bots and humanoid robots have significantly different kinematics and functions. Thus more tasks such as locomotion and manipulation can be explored. An important direction is to design different tasks based on different robots which could enable various complex but crucial services in urban environments. Efficiency. In this work different from indoor scenes and driving simulators MetaUrban supports generating complex interactive urban scenes with arbitrary scales. However with the increase in scale the number of objects and dynamic agents will surge dramatically which will bring the degradation of the efficiency of physical simulation and rendering. A promising direction is to integrate more sophisticated physical engines and renders such as NVIDIA Omniverse. Future work. Foundation model. MetaUrban can easily generate infinite urban scenes with a large quantity of semantics and complex interactions which could facilitate the pretraining of foundation models like LLMs and LVMs that can be used for downstream agent learning tasks. Humanrobot cohabitate. Mobile machines have started emerging in the urban space which makes it no longer exclusive to humans. We plan to work with urban sociologists to study the influence of robots on human urban life through both simulation and field experiments. Improve limitations. The directions discussed in limitations realworld scene distribution extraction interactive agent behavior simulation robotspecific task design and efficiency optimization are also meaningful directions. In summary MetaUrban as a new urban environment simulator will bring a lot of new interesting research directions. We are dedicated to maintaining MetaUrban in the long term and supporting the communitys efforts to develop it into a sustainable infrastructure.

write a youtube script

Video Diffusion Alignment via Reward Gradients Mihir Prabhudesai Russell Mendonca Zheyang Qin Katerina Fragkiadaki Deepak Pathak Carnegie Mellon University A child painting in an art class using watercolors and a brush on paper. A joyful dog playing in the snow leaving paw prints and trying to catch snowflakes on its nose. A shark playing chess. A fairy tends to enchanted glowing flowers. A raccoon drumming on bongos under a starry night sky. A snow princess stands on the balcony of her ice castle her hair adorned with delicate snowflakes overlooking her serene realm. Figure Generations from video diffusion models after adaptation with VADER using reward functions for aesthetics and textimage alignment. More visualization results are available at https vadervid.github.io Abstract We have made significant progress towards building foundational video diffusion models. As these models are trained using largescale unsupervised data it has become crucial to adapt these models to specific downstream tasks. Adapting these models via supervised finetuning requires collecting target datasets of videos which is challenging and tedious. In this work we utilize pretrained reward models that are learned via preferences on top of powerful vision discriminative models to adapt video diffusion models. These models contain dense gradient information with respect to generated RGB pixels which is critical to efficient learning in complex search spaces such as videos. We show that backpropagating gradients from these reward models to a video diffusion model can allow for compute and sample efficient alignment. We show results across a variety of reward models and video diffusion models demonstrating that our approach can learn much more efficiently in terms of reward queries and computation than prior gradientfree approaches. Our code model weights and more visualization are available at httpsvadervid.github.io. Equal Contribution Preprint. Under review. arXiv.v cs.CV Jul Introduction We would like to build systems capable of generating videos for a wide array of applications ranging from movie production creative storyboarding ondemand entertainment ARVR content generation and planning for robotics. The most common current approach involves training foundational video diffusion models on extensive webscale datasets. However this strategy while crucial mainly produces videos that resemble typical online content featuring dull colors suboptimal camera angles and inadequate alignment between text and video content. Contrast this with the needs of an animator who wishes to bring a storyboard to life based on a script and a few preliminary sketches. Such creators are looking for output that not only adheres closely to the provided text but also maintains temporal consistency and showcases desirable camera perspectives. Relying on generalpurpose generative models may not suffice to meet these specific requirements. This discrepancy stems from the fact that largescale diffusion models are generally trained on a broad spectrum of internet videos which does not guarantee their efficacy for particular applications. Training these models to maximize likelihood across a vast dataset does not necessarily translate to highquality performance for specialized tasks. Moreover the internet is a mixed bag when it comes to content quality and models trained to maximize likelihood might inadvertently replicate lowerquality aspects of the data. This leads us to the question How can we tailor diffusion models to produce videos that excel in taskspecific objectives ensuring they are wellaligned with the desired outcomes? The conventional approach to aligning generative models in the language and image domains begins with supervised finetuning . This involves collecting a target dataset that contains expected behaviors followed by finetuning the generative model on this dataset. Applying this strategy to video generation however presents a significantly greater challenge. It requires obtaining a dataset of target videos a task that is not only more costly and laborious than similar endeavors in language or image domains but also significantly more complex. Furthermore even if we were able to collect a video target dataset the process would have to be repeated for every new video task making it prohibitively expensive. Is there a different source of signal we can use for aligning video diffusion instead of trying to collect a target dataset of desired videos? Reward models play a crucial role in aligning image and text generations. These models are generally built on top of powerful image or text discriminative models such as CLIP or BERT . To use them as reward models people either finetune them via small amounts of human preferences data or use them directly without any finetuning for instance CLIP can be used to improve imagetext alignment or object detectors can be used to remove or add objects in the images . This begs the question how should reward models be used to adapt the generation pipeline of diffusion models? There are two broad categories of approaches those that utilize reward gradients and others that use the reward only as a scalar feedback and instead rely on estimated policy gradients . It has been previously found that utilizing the reward gradient directly to update the model can be much more efficient in terms of the number of reward queries since the reward gradient contains rich information of how the reward function is affected by the diffusion generation . However in texttoimage generation space reward gradientfree approaches are still dominant since these methods can be easily trained within hours and the efficiency gains of leveraging reward gradients are not significant. In this work we find that as we increase the dimensionality of generation i.e transition from image to video the gap between the reward gradient and policy gradient based approaches increases. This is because of the additional amount and increased specificity of feedback that is backpropagated to the model. For reward gradient based approaches the feedback gradients linearly scale with respect to the generated resolution as it yields distinct scalar feedback for each spatial and temporal dimension. In contrast policy gradient methods receive a single scalar feedback for the entire video output. We test this hypothesis in Figure where we find that the gap between reward gradient and policy gradient approaches increases as we increase the generated video resolution. We believe this makes it crucial to backpropagate reward gradient information for video diffusion alignment. We propose VADER an approach to adapt foundational video diffusion models using the gradients of reward models. VADER aligns various video diffusion models using a broad range of pretrained vision models. Specifically we show results of aligning texttovideo VideoCrafter OpenSora and ModelScope and imagetovideo Stable Video Diffusion diffusion models while using reward models that were trained on tasks such as image aesthetics imagetext alignment object detection videoactionclassification and video masked autoencoding. Further we suggest various tricks to improve memory usage which allow us to train VADER on a single GPU with GB of VRAM. We include qualitative visualizations that show VADER significantly improves upon the base model generations across various tasks. We also show that VADER achieves much higher performance than alternative alignment methods that do not utilize reward gradients such as DPO or DDPO. Finally we show that alignment using VADER can easily generalize to prompts that were not seen during training. Our code is available at httpsvadervid.github.io. Related Work Denoising diffusion models have made significant progress in generative capabilities across various modalities such as images videos and D shapes . These models are trained using largescale unsupervised or weakly supervised datasets. This form of training results in them having capabilities that are very general however most end usecases of these models have specific requirements such as highfidelity generation or better text alignment . To be suitable for these usecases models are often finetuned using likelihood or rewardbased objectives . Likelihood objectives are often difficult to scale as they require access to the preferred behaviour datasets. Reward or preference based datasets on the other hand are much easier to collect as they require a human to simply provide preference or reward for the data generated by the generative model. Further widely available pretrained vision models can also be used as reward models thus making it much easier to do reward finetuning . The standard approach for reward or preference based finetuning is to do reinforcement learning via policy gradients . For instance the work of does rewardweighted likelihood and the work of applies PPO . Recent works of find that instead of using policy gradients directly backpropagating gradients from the reward model to diffusion process helps significantly with sample efficiency. A recent method DPO does not train an explicit reward model but instead directly optimizes on the human preference data. While this makes the pipeline much simpler it doesnt solve the sample inefficiency issue of policy gradient methods as it still backpropagates a single scalar feedback for the entire video output. While we have made significant progress in aligning image diffusion models this has remained challenging for video diffusion models . In this work we take up this challenging task. We find that naively applying prior techniques of image alignment to video diffusion can result in significant memory overheads. Further we demonstrate how widely available image or video discriminative models can be used to align video diffusion models. Concurrent to our work InstructVideo also aligns video diffusion models via human preference however this method requires access to a dataset of videos. Such a dataset is difficult to obtain for each different task and becomes difficult to scale especially to large numbers of tasks. In this work we show that one can easily align video diffusion models using pretrained reward models while not assuming access to any video dataset. Background Diffusion models have emerged as a powerful paradigm in the field of generative modeling. These models operate by modeling a data distribution through a sequential process of adding and removing noise. The forward diffusion process transforms a data sample x into a completely noised state over a series of steps T. This process is defined by the following equation xt sqrt bar alpha t x sqrt bar alpha t epsilon quad epsilon sim mathcal Nmathbf mathbf where represents noise drawn from a standard Gaussian distribution. Here t Qt i i denotes the cumulative product of i i which indicates the proportion of the original datas signal retained at each timestep t. The reverse diffusion process reconstructs the original data sample from its noised version by progressively denoising it through a learned model. This model is represented by xtt and estimates the noise added at each timestep t. Diffusion models can easily be extended for conditional generation. This is achieved by adding c as an input to the denoising model mathcal Ltext difftheta mathcal D frac mathcal D sum xi ciin mathcal D epsilon theta sqrt bar alpha t xi sqrt bar alpha t epsilon ci t epsilon where D denotes a dataset consisting of imageconditiong pairs. This loss function minimizes the distance between the estimated noise and the actual noise and aligns with the variational lower bound for log pxc. To sample from the learned distribution pxc one starts with a noise sample xT N and iteratively applies the reverse diffusion process xt frac sqrt alpha t left xt frac beta tsqrt bar alpha t epsilon theta xttc right sigma t mathbf z quad mathbf z sim mathcal Nmathbf mathbf The above formulation captures the essence of diffusion models which highlights their ability to generate structured data from random noise. VADER Video Diffusion via Reward Gradients xT xt xt x Reward Model Loss Reverse Diffusion Send Gradients Loss Figure VADER aligns various pretrained video diffusion models by backpropagating gradients from the reward model to efficiently adapt to specific tasks. We present our approach for adapting video diffusion models to perform a specific task specified via a reward function R.. Given a video diffusion model p. dataset of contexts Dc and a reward function R. we seek to maximize the following objective J EcDcxpxc Rx c To learn efficiently both in terms of the number of reward queries and compute time we seek to utilize the gradient structure of the reward function with respect to the weights of the diffusion model. This is applicable to all reward functions that are differentiable in nature. We compute the gradient Rx c of these differentiable rewards and use it to update the diffusion model weights . The gradient is given by Rx c X T t Rx c xt xt . Algorithm VADER Require Diffusion Model weights Require Reward function R. Require Denoising Scheduler f eg DDIM EDM Require Gradient cutoff step K while training do for t T.. do pred xt c t if t K then pred stopgradpred end if xt f.steppred t xt end for g Rx c g end while VADER is flexible in terms of the denoising schedule we demonstrate results with DDIM and EDM solver . To prevent overoptimization we utilize truncated backpropagation where the gradient is back propagated only for K steps where K T where T is the total diffusion timesteps. Using a smaller value of K also reduces the memory burden of having to backpropagate gradients making training more feasible. We provide the pseudocode of the full training process in Algorithm . Next we discuss the type of reward functions we consider for aligning video models. Reward Models Consider a diffusion model that takes conditioning vector c as input and generates a video x of length N consisting of a series of images ik for each timestep k from to N. Then the objective function we maximize is as follows J EciN Ri i...ik...iN c We use a broad range of reward functions for aligning video diffusion models. Below we list down the distinct types of reward functions we consider. ImageText Similarity Reward The generations from the diffusion model correspond to the text provided by the user as input. To ensure that the video is aligned with the text provided we can define a reward that measures the similarity between the generated video and the provided text. To take advantage of popular largescale imagetext models such as CLIP we can take the following approach. For the entire video to be well aligned each of the individual frames of the video likely need to have high similarity with the context c. Given an imagecontext similarity model gimg we have Ri i...ik...iN c X k Rik c X k gimgik c Then we have J EkN gimgik c using linearity of expectation as in the imagealignment case. We conduct experiments using the HPS v and PickScore reward models for imagetext alignment. As the above objective only sits on individual images it could potentially result in a collapse where the predicted images are the exact same or temporally incoherent. However we dont find this to happen empirically we think the initial pretraining sufficiently regularizes the finetuning process to prevent such cases. VideoText Similarity Reward Instead of using per image similarity model gimg it could be beneficial to evaluate the similarity between the whole video and the text. This would allow the model to generate videos where certain frames deviate from the context allowing for richer more diverse expressive generations. This also allows generating videos with more motion and movement which is better captured by multiple frames. Given a videotext similarity model gvid we have J E gvidi i...ik...iN c. In our experiments we use a VideoMAE finetuned on action classification as gvid which can classify an input video into one of a set of action text descriptions. We provide the target class text as input to the texttovideo diffusion model and use the predicted probability of the ground truth class from VideoMAE as the reward. Image Generation Objective While text similarity is a strong signal to optimize some use cases might be better addressed by reward models that only sit on the generated image. There is a prevalence of powerful imagebased discriminative models such as Object Detectors and Depth Predictors. These models utilize the image as input to produce various useful metrics of the image which can be used as a reward. The generated video is likely to be better aligned with the task if the reward obtained on each of the generated frames is high. Hence we define the reward in this case to be the mean of the rewards evaluated on each of the individual frames i.e Ri i...ik...iN c P k Rik. Note that given the generated frames this is independent of the text input c. Hence we have J EkN Rik EkN Mik via linearity of expectation where M is a discriminative model that takes an image as input to produce a metric that can be used to define a reward. We use the Aesthetic Reward model and Object Detector reward model for our experiments. Video Generation Objective With access to an external model that takes in multiple image frames we can directly optimize for desired qualities of the generated video. Given a video metric model N the corresponding reward is J E Ni i ..ik...iN. VideoCrafter VADER Ours The raccoon is wearing a red coat and holding a snowball. A dog playing a slide guitar on a porch during a gentle rainstorm. A strong lion and a graceful lioness resting together in the shade of a big tree on a wide grassland. A peaceful deer eating grass in a thick forest with sunlight filtering through the trees. A dog strumming an acoustic guitar by a lakeside campfire under the stars. The fox is wearing a red hat and playing with leaves. Figure Texttovideo generation results for VideoCrafter and VADER. We show results for VideoCrafter TexttoVideo model on the left and results for VADER on the right where we use VideoCrafter as our base model. The reward models applied are a combination of HPSV. and Aesthetic model in the first three rows and PickScore in the last three rows. The videos in the third and last rows are generated based on prompts that are not encountered during training. Longhorizon consistent generation In our experiments we adopt this formulation to enable a feature that is quite challenging for many opensource video diffusion models that of generating clips that are longer in length. For this task we use Stable Video Diffusion which is an imagetovideo diffusion model. We increase the context length of Stable Video Diffusion by x by making it autoregressive. Specifically we pass the last generated frame by the model as input for generating the next video sequence. However we find this to not work well as the model was never trained over its own generations thus resulting in a distribution shift. In order to improve the generations we use a video metric model N VJEPA that given a set of frames produces a score about how predictive the frames are from one another. We apply this model on the autoregressive generations to encourage these to remain consistent with the earlier frames. Training the model in this manner allows us to make the video clips temporally and spatially coherent. x x x x x x Resolution . . . . . . Reward Reward Gradient Policy Gradient Base Figure Reward gradient vs policy gradient approaches as output resolution increases We train DDPO and VADER with increasing resolution of the generated video. In the above curve we report the reward achieved after steps of optimization we find that as the resolution of the generation increases the performance gap between both approaches significantly increases. Reducing Memory Overhead Training video diffusion models is very memory intensive as the amount of memory linearly scales with respect to the number of generated frames. While VADER significantly improves the sample efficiency of finetuning these models it comes at the cost of increased memory. This is because the differentiable reward is computed on the generated frame which is a result of sequential denoising steps. i Standard Tricks To reduce the memory usage we use LoRA that only updates a subset of the model parameters further we use mixed precision that stores nontrainable parameters in fp. To reduce memory usage during backpropagation we use gradient checkpointing and for the long horizon tasks offload the storage of the backward computation graph from the GPU memory to the CPU memory. ii Truncated Backprop Additionally In our experiments we only backpropagate through the diffusion model for one timestep instead of backpropagating through multiple timesteps and have found this approach to obtain competitive results while requiring much less memory. iii Subsampling Frames Since all the video diffusion models we consider are latent diffusion models we further reduce memory usage by not decoding all the frames to RGB pixels. Instead we randomly subsample the frames and only decode and apply loss on the subsampled ones. We conduct our experiments on A GPUS GB VRAM and our model takes an average of hours to train. However our codebase supports training on a single GPU with GB VRAM. Before VADER Ours A book and a cup of tea on a blanket in a sunflower field. A book and a cup of hot chocolate on a windowsill with a snowy view. A book and a cup of coffee on a rustic wooden table in a cabin. Figure Object removal remove book using VADER. The left column displays results from the base model VideoCrafter while the right column shows results from VADER after finetuning the base model to not display books by using an object detector as a reward model. As can be seen VADER effectively removes book and replaces it with some other object. Results In this work we focus on finetuning various conditional video diffusion models including VideoCrafter OpenSora Stable Video Diffusion and ModelScope through a comprehensive set of reward models tailored for images and videos. These include the Aesthetic model for images HPSv and PickScore for imagetext alignment YOLOS for object removal VideoMAE for action classification and VJEPA selfsupervised loss for temporal consistency. Our experiments aim to answer the following questions How does VADER compare against gradientfree techniques such as DDPO or DPO regarding sample efficiency and computational demand? To what extent can the model generalize to prompts that are not seen during training? How do the finetuned models compare against one another as judged by human evaluators? How does VADER perform across a variety of image and video reward models? This evaluation framework assesses the effectiveness of VADER in creating highquality aligned video content from a range of input conditioning. Baselines. We compare VADER against the following methods VideoCrafter OpenSora . and ModelScope are current stateoftheart publicly available texttovideo diffusion models. We serve them as base models for finetuning and comparison in our experiments in texttovideo space. Stable Video Diffusion is the current stateofart publicly available imagetovideo diffusion model. In all our experiments in imagetovideo space we use their base model for finetuning and comparison. DDPO is a recent image diffusion alignment method that uses policy gradients to adapt diffusion model weights. Specifically it applies PPO algorithm to the diffusion denoising process. We extend their code for adapting video diffusion models. DiffusionDPO extends the recent development of Direct Preference Optimization DPO in the LLM space to image diffusion models. They show that directly modeling the likelihood using the preference data can alleviate the need for a reward model. We extend their implementation to aligning video diffusion models where we use the reward model to obtain the required preference data. Reward models. We use the following reward models to finetune the video diffusion model. Aesthetic Reward Model We use the LAION aesthetic predictor V which takes an image as input and outputs its aesthetic score in the range of . The model is trained on top of CLIP image embeddings for which it uses a dataset of image ratings provided by humans ranging from to where images rated as are classified as art pieces. Human Preference Reward Models We use HPSv and PickScore which take as input an imagetext pair and predict human preference for the generated image. HPSv is trained by finetuning CLIP model with a vast dataset that includes instances of human preference rankings among image pairs while PickScore is trained by finetuning CLIP model with examples of human preferences. These datasets are among the most extensive in the field offering a solid foundation for enhancing imagetext alignment. Object Removal We design a reward model based on YOLOS a Vision Transformer based object detection model trained on annotated images. Our reward is one minus the confidence score of the target object category from which video models learns to remove the target object category from the video. Video Action Classification While the above reward models sit on individual images we employ a reward model that takes in the whole video as input. This can help with getting gradients for the temporal aspect of video generation. Specifically we consider VideoMAE which is finetuned for the task of action classification on Kinetics dataset . Our reward is the probability predicted by the action classifier for the desired behavior. Temporal Consistency via VJEPA While action classification models are limited to a fixed set of action labels here we consider a more general reward function. Specifically we use selfsupervised masked prediction objective as a reward function to improve temporal consistency. Specifically we use VJEPA as our reward model where the reward is the negative of the masked autoencoding loss in the VJEPA feature space. Note that we employ exactly the same loss objective that VJEPA uses in their training procedure. Before VADER Ours a man in a trendy suit taking a selfie in a city square surrounded by modern buildings and a fountain. A bear enjoying a slice of cake at a picnic. OpenSora A shark riding a bike. A bear playing chess. ModelScope Figure Aligning OpenSora . and ModelScope with VADER. The left column shows results from the base models while results from VADER are demonstrated on the right. The first two rows use OpenSora as the base model and the last two rows use ModelScope. The reward models applied are PickScore in the first row HPSv. in the second row HPSv in the third row and the Aesthetic reward model in the last row. Prompts. We consider the following set of prompt datasets for reward finetuning of texttovideo and imagetovideo diffusion models. Activity Prompts Text We consider the activity prompts from the DDPO . Each prompt is structured as an animal activity using a collection of familiar animals. The activity for each prompt is selected from a trio of options riding a bike playing chess and washing dishes. HPSv Action Prompts Text Here we filter out prompts from a set of prompts introduced in the HPS v dataset for textimage alignment. We filter prompts such that they contain action or motion information in them. ChatGPT Created Prompts Text We prompt ChatGPT to generate some vivid and creatively designed text descriptions for various scenarios such as books placed beside cups animals dressed in clothing and animals playing musical instruments. ImageNet Dog Category Image For imagetovideo diffusion model we consider the images in the Labrador retriever and Maltese category of ImageNet as our set of prompts. Stable Diffusion Images Image Here we consider all images from Stable Diffusion online demo webpage. . Sample and Computational Efficiency Training of largescale video diffusion models is done by a small set of entities with access to a large amount of computing however finetuning of these models is done by a large set of entities with access to a small amount of computing. Thus it becomes imperative to have finetuning approaches that boost both sample and computational efficiency. In this section we compare VADERs sample and computational efficiency with other reinforcement learning approaches such as DDPO and DPO. In Figure we visualize the reward curves during training where the xaxis in the upper half of the figure is the number of reward queries and the one in the bottom half is the GPUhours. As can be seen VADER is significantly more efficient in terms of sample and computation than DDPO or DPO. This is mainly due to the fact that we send dense gradients from the reward model to the diffusion weights while the baselines only backpropagate scalar feedback. Reward Queries . . . . . Aesthetic Score Aesthetics Reward Queries . . . . HPSv Score Text Alignment Reward Queries . . . . Class Accuracy Action Prediction Compute GPUhours . . . . . Aesthetic Score Compute GPUhours . . . . HPSv Score Compute GPUhours . . . . Class Accuracy VADER Ours DPO DDPO Figure Training efficiency comparison Top Sample efficiency comparison with DPO and DDPO. Bottom Computational efficiency comparison with DPO and DDPO. It can be seen that VADER starts converging within at most GPUhours of training while DPO or DDPO do not show much improvement. . Generalization Ability Method Aes TV HPS TV ActP Aes IV Train. Test. Train. Test. Train. Train. Test. Base . . . . . . . DDPO . . . . . NA NA DPO . . . . . NA NA Ours . . . . . . . Table Reward on Prompts in train test. We split the prompts into train and test sets such that the prompts in the test set do not have any overlap with the ones for training. We find that VADER achieves the best on both metrics. A desired property of finetuning is generalization i.e. the model finetuned on a limited set of prompts has the ability to generalize to unseen prompts. In this section we extensively evaluate this property across multiple reward models and baselines. While training texttovideo TV models we use HPSv Action Prompts in our training set whereas we use Activity Prompts in our test set. We use Labrador dog category in our training set for training imagetovideo IV models while Maltese category forms our test set. Table showcases VADERs generalization ability. . Human Evaluation Method Fidelity Text Align ModelScope . . VADER Ours . . Table Human Evaluation results for HPS reward model where the task is imagetext alignment. We carried out a study to evaluate human preferences via Amazon Mechanical Turk. The test consisted of a sidebyside comparison between VADER and ModelScope. To test how well the videos sampled from both the models aligned with their text prompts we showed participants two videos generated by both VADER and a baseline method asking them to identify which video better matched the given text. For evaluating video quality we asked participants to compare two videos generated in response to the same prompt one from VADER and one from a baseline and decide which videos quality seemed higher. We gathered responses for each comparison. The results illustrated in Table show a preference for VADER over the baseline methods. . Qualitative Visualization In this section we visualize the generated videos for VADER and the respective baseline. We conduct extensive visualizations across all the considered reward functions on various base models. HPS Reward Model In Figure we visualize the results before and after finetuning VideoCrafter using both HPSv. and Aesthetic reward function together in the top three rows. Before finetuning the raccoon does not hold a snowball and the fox wears no hat which is not aligned with the text description however the videos generated from VADER does not result in these inconsistencies. Further VADER successfully generalizes to unseen prompts as shown in the third row of Figure where the dogs paw is less like a human hand than the video on the left. Similar improvements can be observed in videos generated from OpenSora V. and ModelScope as shown in the second and third rows of Figure . Aesthetic Reward Model In Figure in the top three rows we visualize the results before and after finetuning ModelScope using a combination of Aesthetic reward function and HPSv. model. Also we finetune ModelScope via Aesthetic Reward function and demonstrate its generated video in the last row in Figure . We observe that Aesthetic finetuning makes the generated videos more artistic. PickScore Model In the bottom three rows of Figure we showcase videos generated by PickScore finetuned VideoCrafter. VADER shows improved textvideo alignment than the base model. In the last row we test both models using a prompt that is not seen during training time. Further video generated from PickScore finetuned OpenSora is displayed in the first row of Figure . Object Removal Figure displays the videos generated by VideoCrafter after finetuning using the YOLOSbased objection removal reward function. In this example books are the target objects for removal. These videos demonstrate the successful replacement of books with alternative objects like a blanket or bread. Video Action Classification In Figure we visualize the video generation of ModelScope and VADER. In this case we finetune VADER using the action classification objective for the action specified in the prompt. For the prompt A person eating donuts we find that VADER makes the human face more evident along with adding sprinkles to the donut. Earlier generations are often misclassified as baking cookies which is a different action class in the kinetics dataset. The addition of colors and sprinkles to the donut makes it more distinguishable from cookies leading to a higher reward. ModelScope VADER Ours A person playing Piano A person eating Donuts Figure Video action classifiers as reward model. We use VideoMAE action classification model as a reward function to finetune ModelScopes TexttoVideo Model. We see that after finetuning VADER generates videos that correspond better to the actions. VJEPA reward model In Figure we show results for increasing the length of the video generated by Stable Video Diffusion SVD. For generating longrange videos on SVD we use autoregressive inference where the last frame generated by SVD is given as conditioning input for generating the next set of images. We perform three steps of inference thus expanding the context length of SVD by three times. However as one can see in the images bordered in red after one step of inference SVD starts accumulating errors in its predictions. This results in deforming the teddy bear or affecting the rocket in motion. VADER uses VJEPA objective of masked encoding to enforce selfconsistency in the generated video. This manages to resolve the temporal and spatial discrepancy in the generations as shown in Figure . Stable Video Diffusion VADER Ours Stable Video Diffusion VADER Ours Figure Improving temporal and spatial consistency of Stable Video Diffusion SVD ImagetoVideo Model. Given the leftmost frame as input we use autoregressive inference to generate N frames in the future where N is the context length of SVD. However this suffers from error accumulation resulting in corrupted frames as highlighted in the red border. We find that VADER can improve the spatiotemporal consistency of SVD by using VJEPAs masked encoding loss as its reward function. Conclusion We presented VADER which is a sample and compute efficient framework for finetuning pretrained video diffusion models via reward gradients. We utilized various reward functions evaluated on images or videos to finetune the video diffusion model. We further showcased that our framework is agnostic to conditioning and can work on both texttovideo and imagetovideo diffusion models. We hope our work creates more interest towards adapting video diffusion models. References Bardes A. Garrido Q. Ponce J. Chen X. Rabbat M. LeCun Y. Assran M. Ballas N. Vjepa Latent video prediction for visual representation learning Black K. Janner M. Du Y. Kostrikov I. Levine S. Training diffusion models with reinforcement learning. arXiv preprint arXiv. Blattmann A. Dockhorn T. Kulal S. Mendelevitch D. Kilian M. Lorenz D. Levi Y. English Z. Voleti V. Letts A. et al. Stable video diffusion Scaling latent video diffusion models to large datasets. arXiv preprint arXiv. Brooks T. Holynski A. Efros A.A. Instructpixpix Learning to follow image editing instructions. In Proceedings of the IEEECVF Conference on Computer Vision and Pattern Recognition. pp. Chen H. Zhang Y. Cun X. Xia M. Wang X. Weng C. Shan Y. Videocrafter Overcoming data limitations for highquality video diffusion models Clark K. Vicol P. Swersky K. Fleet D.J. Directly finetuning diffusion models on differentiable rewards. arXiv preprint arXiv. Dong H. Xiong W. Goyal D. Zhang Y. Chow W. Pan R. Diao S. Zhang J. Shum K. Zhang T. Raft Reward ranked finetuning for generative foundation model alignment Fang Y. Liao B. Wang X. Fang J. Qi J. Wu R. Niu J. Liu W. You only look at one sequence Rethinking transformer in vision through object detection. CoRR abs. httpsarxiv.orgabs. Feng W. He X. Fu T.J. Jampani V. Akula A. Narayana P. Basu S. Wang X.E. Wang W.Y. Trainingfree structured diffusion guidance for compositional texttoimage synthesis Ho J. Chan W. Saharia C. Whang J. Gao R. Gritsenko A. Kingma D.P. Poole B. Norouzi M. Fleet D.J. Salimans T. Imagen video High definition video generation with diffusion models Ho J. Jain A. Abbeel P. Denoising diffusion probabilistic models. CoRR abs. httpsarxiv.orgabs. Ho J. Salimans T. Gritsenko A. Chan W. Norouzi M. Fleet D.J. Video diffusion models Hu E.J. Shen Y. Wallis P. AllenZhu Z. Li Y. Wang S. Wang L. Chen W. Lora Lowrank adaptation of large language models Karras T. Aittala M. Aila T. Laine S. Elucidating the design space of diffusionbased generative models. Advances in Neural Information Processing Systems Kay W. Carreira J. Simonyan K. Zhang B. Hillier C. Vijayanarasimhan S. Viola F. Green T. Back T. Natsev P. et al. The kinetics human action video dataset. arXiv preprint arXiv. Kirstain Y. Polyak A. Singer U. Matiana S. Penna J. Levy O. Pickapic An open dataset of user preferences for texttoimage generation Lambert N. Pyatkin V. Morrison J. Miranda L. Lin B.Y. Chandu K. Dziri N. Kumar S. Zick T. Choi Y. et al. Rewardbench Evaluating reward models for language modeling. arXiv preprint arXiv. Lee K. Liu H. Ryu M. Watkins O. Du Y. Boutilier C. Abbeel P. Ghavamzadeh M. Gu S.S. Aligning texttoimage models using human feedback Liu R. Wu R. Van Hoorick B. Tokmakov P. Zakharov S. Vondrick C. Zeroto Zeroshot one image to d object. In Proceedings of the IEEECVF International Conference on Computer Vision. pp. Prabhudesai M. Goyal A. Pathak D. Fragkiadaki K. Aligning texttoimage diffusion models with reward backpropagation. arXiv preprint arXiv. Radford A. Kim J.W. Hallacy C. Ramesh A. Goh G. Agarwal S. Sastry G. Askell A. Mishkin P. Clark J. et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. pp. . PMLR Rafailov R. Sharma A. Mitchell E. Manning C.D. Ermon S. Finn C. Direct preference optimization Your language model is secretly a reward model. Advances in Neural Information Processing Systems Sauer A. Boesel F. Dockhorn T. Blattmann A. Esser P. Rombach R. Fast highresolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv. Schuhmann C. Laoin aesthetic predictor httpslaion.aiblog laionaesthetics Schulman J. Wolski F. Dhariwal P. Radford A. Klimov O. Proximal policy optimization algorithms. arXiv preprint arXiv. SohlDickstein J. Weiss E.A. Maheswaranathan N. Ganguli S. Deep unsupervised learning using nonequilibrium thermodynamics Song J. Meng C. Ermon S. Denoising diffusion implicit models Tallec C. Ollivier Y. Unbiasing truncated backpropagation through time. arXiv preprint arXiv. Tong Z. Song Y. Wang J. Wang L. Videomae Masked autoencoders are dataefficient learners for selfsupervised video pretraining. Advances in neural information processing systems Wallace B. Dang M. Rafailov R. Zhou L. Lou A. Purushwalkam S. Ermon S. Xiong C. Joty S. Naik N. Diffusion model alignment using direct preference optimization. arXiv preprint arXiv. Wang J. Yuan H. Chen D. Zhang Y. Wang X. Zhang S. Modelscope texttovideo technical report. arXiv preprint arXiv. Wu X. Hao Y. Sun K. Chen Y. Zhu F. Zhao R. Li H. Human preference score v A solid benchmark for evaluating human preferences of texttoimage synthesis. arXiv preprint arXiv. Xu J. Liu X. Wu Y. Tong Y. Li Q. Ding M. Tang J. Dong Y. Imagereward Learning and evaluating human preferences for texttoimage generation Yuan H. Zhang S. Wang X. Wei Y. Feng T. Pan Y. Zhang Y. Liu Z. Albanie S. Ni D. Instructvideo Instructing video diffusion models with human feedback. arXiv preprint arXiv. Zheng Z. Peng X. Yang T. Shen C. Li S. Liu H. Zhou Y. Li T. You Y. Opensora Democratizing efficient video production for all March httpsgithub.com hpcaitechOpenSora

write a youtube script

Music Applause Music do you remember when I made a bicycle with an omni wheel on the front wheel it baled like a Segway sideways but you could still ride it like a normal bike because the front wheel had lots of skateboard wheels all around it so it would kind of slide along sideways a while later I moved on and tried to make a bike with two Omni Wheels instead of building the electronics myself though I took the guts out of a hoverboard including the motors I tried to Velocity match the original Wheels with the new larger Wheels using belt drives which kind of worked okay but then the only way to propel it was with edfs blowing it along which wasnt very successful as soon as Id assembled that version I realized it would have been better to have made mechum wheels which have the little wheels slanted all around them these are typically used on the four corners of a robot so it can move in any direction but I wanted to put them all in a line so the bike still balances sideways but it can screw itself along I actually found a YouTube video from Matthew Watson where this method was used to drive a small robot I believe this was a PhD project and it works very well indeed I cant find any other versions of this type of locomotion and I certainly cant find any versions that are B enough to ride on so yeah you can buy mechum wheels off the shelf but the biggest I can find is mm in diameter which is really small and theyre really expensive like and I already made two massive Omni Wheels so I guess Im going to have to make four massive mechum Wheels yes of course Im going to D print most of the parts Im building a scaled up version of my mechum wheels from the smaller robot I showed earlier each wheel is going to be around mm in diameter and itll have little wheels around its circumference so I need little wheels in total just a quick add from my D printing sponsor thanks to lsot for supporting my Channel with D printers Im printing the tires in a TPU from D Fuel and the wheel hubs in D fuel proa plus you can now get off D fuel products with my special code and Ill get a small commission Music too Music theres also wedges to be made to put on each end of the axles to hold the wheels at a slant but Ill save you the footage of printing every single one of them I printed the TPU Wheels with a . mm nozzle on LSB printers and theyre incredibly tough and that means its very very hard to push the hubs in so its time for the hydraulic press Music Applause Music thanks to Simply bearings for the bearings for this project Ive got skate sized bearings for each end of the Wheels as well as of course some bigger bearings for the actual big wheels themselves so we need to make some wheel hubs to mount all this on so as usual Im going to use some plywood CN seed out so weve got some holes here being drilled in a massive Circle and thats so we can attach all of those wedges around and we also need some more holes to go and align all of the parts together with a hub in the middle and make sure those wheels are in the right place so weve cut out some circles and Ive also got some htd eight profile pulley and one of those is going to drive each wheel so Ive painted all of those black with some automotive spray paint and you notice theres some very shiny sections on there and thats actually masking tape which Ive put over where the four holes are to mount each of the wedges and thats so I can glue those on and theres something to stick to other than paint so Ive got some solventbased adhesive to stick each of those on and Im also going to screw those on as well so basically those never ever come off because Im going to be sitting on this and that leaves us with a piece of round Wood with wedges glued and screwed to it on the other side weve got the ht pulley and Ive also made a thin plywood retainer so the Belt doesnt fall off and thats going to be used to drive the wheel this Hub fits onto a D printed part which is printed with a . mm nozzle so thats a really tough D printed Hub and Ive got bolts that go all the way through which are a pretty tight fit well just knock all of those in and eventually that will hold the other side on as well for now we need to get all the wheels in so weve got these mm pieces of stainless steel rod which go into those wedges all the way around to get the spacing right Ive got a couple of nylon washers and then we can put all of the little wheels on with their bearings all the way around followed by another two nylon washers on the top of every one that just makes it perfectly spaced so they dont move weve got another set of wedges which go all the way around there and then you can probably guessed that with those bolts sticking out were going to stick another plywood Hub on top of that and then were going to glue and screw all of those wedges to it on the other side and that makes the complete wheel I did turn the bolts around those so that the piece is sticking out with the nuts on her in the nice recess that I left we need to fiz a bearing in each end of the wheel but the Hub was printed with a massive nozzle so I printed another piece with a finer nozzle a half mm nozzle that makes it much easier to tolerance and we can reprint that part if we need to without reprinting the whole Hub so that we can fit the bearings in nice and tightly and those have got mm internal diameters Ive made four of those in total and youll notice I remember to make opposite pairs so that the wheels go in opposing directions so they can screw against each other to travel in the same direction otherwise balancing is going to be really hard so what we need to do now is put them all on a chassis and put some Motors on its chassis Music time Music Applause Music so thats my main chassis and these four pieces sticking up as I hold the four Motors so actually the wheels go below this and this weird saddle thing here is where were going to put foot pegs so we need to put the wheels on this piece of bright steel bar which is hung below that and we need to make it all Music adjustable so Ive got some mm Rose joints with the mm thread to match the studing and mm holes in to match my axle and all of these are going to go through that chassis crossbar and be bolted on so that we can adjust the height of them and we can get that axle through perfectly all through all of them threading the eye of the needle every time and then we can get all the wheels aligned and spaced correctly so now its time to put the wheels on which is a bit of an ordeal that mm bright steel bars quite tightly toleranced for the bearings I can do it but its going to take me quite some time to push it through and push all the wheels on not forgetting of course to put the drive belts on first otherwise Ill never be able to get them Music on theres a color clamp on the end and therell be more of those along the axle to brace the axle against the rose joints throughout the length of the bike so now all my wheels are all perfectly in line and they all run fine and Ive done up all the nuts on that studing so that we get everything at the right height so that the belts can mesh with the motor drives so we do need some spacers um at the moment those can slide up and down but we need to align them perfectly with the motors you might have noticed I actually put the wheels on in the wrong order so I had to take them all off and do it again but even without spaces there if I slide this up and down lengthways we can see those wheels turning against each other just like my little animation so Im very happy with how the alignments gone so far its time for some motors do you remember the video I did with Colin Furs where I built a version of his rhino tank and he helped me test it yeah thats what we was made for go through the mud well Im using the same Motors and drivers from that project these are o drive KV brushless motors with an O Drive S on the back of them theres a magnet on the motor and an encoder on the board so the position and velocity or Torque can be accurately controlled this is a really neat unit and I built a twostage belt reducer for the tank this time were going to need a bit more velocity though so Ive doubled the size of the output pulley that drives the wheels the motor and reduced assembly fit onto the risers I made on top of the main chassis crossbar and the belts attenion with blocks underneath each assembly I made a quick stand out of steel for it so I can lean it up and work on it upright and Ive also fitted the top of the chassis on and thats going to hold the handlebars the saddle and the electronics and that is a bolted on piece so that I can take it off to get to those motor assemblies if I need to in the last domy bike I made a steering assembly which basically would lean sideways as well as rotating and thats what I used to control all of the axes however there was nothing solid to hang on to so that was a bit of a mistake in the end and resulted in me crashing so this time Ive just made the handlebars which are completely fixed and dont turn at all I found this saddle on Amazon which fits just right so Im pretty happy of the way its looking its quite retro right its going to be one of the best things Ive ever made if it works but we need to put some Electronics in and get the motors powered up and see what happens but before we carry on assembling that its time for a quick add from the video sponsor which is ODU Odo is an allinone management tool that provides entrepreneurs with a wide range of webbased applications to simplify the management of their business the first ODU application is free for life with unlimited hosting and support they also offer a free personalized domain name for one year once you have two or more applications youll switch to a paid plan the OD website builder application is a powerful tool for creating a website quickly easily efficiently free of charge the website configurator guides you through defining your goals choosing the color palette and inserting your logo adding pages and features and choosing your theme once the structure is in place you can start customizing simply drag and drop blocks to customize your website you can customize each block adding content including images text video documents and functionalities like animations on blocks and images changing typography colors and image shapes the grid lets you easily lay out and organize your content blocks AI generated text is available in ODU to populate any sections I quickly put together a website about this project and you can try out ODU Now by visiting the link in the video description right lets get on with the screw bike mechum wheel Omni directional Co linear drive bike project so Ive made an Electronics enclosure and in there weve got a Tec . weve got a spark fun b IMU and weve got the canbus transceiver to control all the O drives on top Ive got a battery monitor an LED strip which is actually at the moment visualizing the IMU data Ive also got various knobs on there so weve got a trim control which is for the inertial measurement unit data and Ive also got a switch to initialize the O drives and Ive got a volume control well talk about later Ive got connectors coming out including the canbas USB power the volt power so it can monitor the battery and some connector to go to the handlebars for all of the switches Ive made a battery box which holds four s lipo and were going to put those in series and then pairs in parallel so we basically get nearly volts out of that theres a amp contactor on the top of there and thats going to be linked to the e stop on The Handlebar so that I can actually cut the power off if I need to the electronic sit right in front of the saddle there so theyre nice and accessible to control yes thats a USB boost pank powering them nice and professional the gray cable is can bus thats daisy chain between all of the O drives to control them and all those batteries sit at the back and yes I do have a cover there to cover the dodgy bolts which has got volts on Ive built quite a few balancing robots in the past theres a playlist in my channel just for balancing projects that use various means to make unstable things stable the main thing that makes balancing robots work is a p ID controller this takes the angle data from the inertial measurement unit it controls the wheel velocities appropriately the proportional term simply means that the wheels go faster if the angle is greater but this isnt sufficient to make the device balance properly the most important part is the integrator the output of the integrator is the area under the graph so if the angle is constant as time goes on then the output increases as this area increases over time if the angle is increasing then the output gets bigger exponentially so the integrator makes the wheel velocity accelerate to catch the robot as it falls over which happens constantly the derivative term is the tangent of the curve of the angle data so this makes a short sharp velocity movement to damp any overshoot all three terms need to be tuned for the mass and power of the robot which is a bit like focusing a camera with three lenses but typically the highest value is the integral term and the smallest is the derivative however if the device is left at an angle without the motors powered on then the integral term tends to accumulate a large value and when its finally time to balance then it makes the robot suddenly kick until its stable this was an issue with the first Omni bike so this time Ive got a control so I can slowly turn up the output which makes it less likely to break my ankles when I power it up the button initializes the O drives and now I can slowly ramp up the output of the P ID controller to make things balance without it kicking so much it was pretty easy to tune up but bigger heavier robots tend to thrash around less because theyve got more inertia so its much easier to tune the P ID controller to make them balance right lets put the saddle on I have in fact tuned this up to have a person on rather than no person so this should be working pretty well and Im pretty happy of how well its working is almost as good as a Segway oh yeah this is really good this is so much better than the last one made of a hacked hoverboard mainly cuz I couldnt retune those P parameters on the hoverboard so even just a slight bit of uh mass shift causes this thing to travel which is good its almost as good as a Segway Ive definitely got enough velocity for balancing there but what we need to do now is try and screw it along by moving those Wheels in opposite directions and if that works then this is going to be pretty cool so Ive got a Twist grip on the right hand side which moves in one Direction and Ive also got a toggle switch so that I can drive in forward or reverse and all thats going to do is make the wheels turn in opposite directions to screw itself along and transpose that value over the balancing value right lets try and uh go forwards oh it works oh yeah this is brilliant this is brilliant oh yes way I think we might be slipping slightly on the smooth floor but this feels amazing rotating is controlled by another twist grip ideally I want a Twist grip that rotates forwards and backwards so I can turn left and right but I couldnt find one that does that so Ive just got another Direction switch and well see how that goes I did think about having foot switches for steering but I thought probably this is going to fall over at some point and theyll get buckled or broken if theyre attached to the foot pegs so I decided handlebar controls were best now of course the wheels are different distance is from the center and that means that they need to rotate at different velocities so that we can rotate around the center point basically one circumference is bigger than the other so the outside Wheels need to go faster we need to take that into account right lets see where rotating gets us oh oh oh oh I think I want my mass in the middle since thats where the rotation point is yeah that seems to be that seems to be

write a youtube script

do you remember when Colin F pulled a string and it made a cup of tea and he was just lying in bed and the strings werent all through the house and then he pulled the strings and it made a cup of tea in the kitchen I was thinking maybe we can make a robot tea maker with like robot arms and everything but people like mechanics so the simple approach probably is going to do better marble machines are popular at the moment arent they have you seen Ian Mirandas marble machine clock and the other one from jbv creative so I thought Ill make a marble machine but instead of being a clock its going to make a cup of tea this is really the mic by the way do you like it it makes me feel like Im talking on the telephone its not very good mic so dont listen too hard yeah havent really thought this through that much Ive basically done some design about how its going to move the cup on with a couple of seesaws and I havent designed the rest theres a big gap in the middle but I think it needs to be really tall so that we get lots of like gravity for the marble so Im just going to do it as I go and hope for the best well I do most projects really so Im just going to print all the stuff and then well build the bits Ive designed already and then well work out what to to do as we go as we work out the run of the marble and what it needs to do just a quick ad from my D printing sponsor thanks to lot D printers for supporting my Channel with D printers Ive got quite a few D printers and it really helps big projects like this to get all the parts done in time thanks to D fuel for the filament for this project you can now get off at D fuel.com with my special code and Link and Ill get a small commission Im going to use some vslot Extrusion for this project to make the base of the thing and as usual be using some droing t nuts on M bolts which just turn into that slot there and hold it together so Ive made a frame with some D printed Corners which is going to make the base of the whole machine another accessory for vslot Extrusion is V Wheels which youll notice have a kind of V profile on and that fits into the v slot so we can make a slider that moves really really freely that of course is going to be the slider for the cup so it goes through the production line of making the te that fits right on there it moves pretty freely cuz the whole thing is just going to be driven by gravity with a ball bearing or a marble coming down so it needs to be pretty free moving yep its another company that helped me with the project thanks to Simply bearings for the bearings of this project there are loads of bearings in this because I need things to be really free moving CU I already said its just going to be driven by gravity so any friction we need to get rid of really Ive got these rotating things which have a bearing in the middle on each side just like a skateboard wheel and also one on the ends these are all cascaded so they dont touch each other and these are going to be the pushers that push the cup along as the marble or B bearing moves through the machine its going to turn each of these in order as well as do other stuff weve got about eight stages actually thats going to push the cup all the way through the machine the bearings on the end there hit this angle piece so it should roll off and decelerate the cup as I said at the beginning the whole thing is quite tall so Ive got this piece of aluminium box section fixed on a bracket and thats over a meter tall yep its quite a tall stick and its quite wobbly at the moment as well its quite tall because we need the gravity to sort of convey the force to go sideway way at the bottom and do all the other things like put the water in and stir the tea and all that stuff so uh for now I havent really planned this Ive kind of drawn some things out Ive printed some of them so were going to start with putting a couple of stages together here and see if we can convey the force to the bottom the first part is a seesaw which sort of tips over as the ball rolls down and pulls a string basically and Ive also got a piece that feeds it which sits before it the parts Ive printed are on here its quite wobbly but theres going to be more sticks going up and like more scaffolding between them and stuff So eventually it wont be so bad but now weve just got one stage that works though at the end of the Seesaw of course theres a curved piece and a hole so the ball cant just fly off into the air it has to go down the Seesaw before it can exit so that seesaw needs to pull the cup along to the first stage so Ive got this thing with multiple notches so I can put the string in different positions to experiment and get it perfectly right and that string goes all the way down and it goes round a pulley and then of course it goes to that Pusher so that when it unwinds it pushes the tea along so theres the Seesaw pushing the cup along to the first stage to make the cup of tea so lets just chop the ball bearing in and see what happens it goes quite slowly but it pulls a cup along and it seems to work all right lets try a bigger ball bearing This Is mm in diameter yeah thats much better it looks like were on the right track for making a cup of tea but now its time to put the water in so Ive made a cradle for The Jug to fit in which fits like that and it tips over with some bearings in there again so its nice and free moving and then of course thats going to tip hot water into the cup so Ive got another seesaw with another string which apart from overshooting the cup seems to work all right so the whole ball run works if I put this massive Ball Bearing in here but this is definitely going to slop boiling water all over the place cuz it moves quite quickly so we need some sort of limiter thats going to make that move down much slower and gently pour the water but Im to come back and solve that at the end along with how we make it Brew tea for minutes the next stage is slightly different I thought we try a different mechanism which is this round curved mechanism that takes the ball round in the middle of that theres a kind of paddle or a wiper thats fit onto some bearings so as the ball comes round it pushes that round and the middle of thats a pulley that can pull the string so Ive attached another string to that so as it goes around you can see it pulling that string Ive also got a magnet that holds it in place so it doesnt wander off by itself and that string goes around another pulley all the way down to the bottom round a bit of stainless steel and onto the next Pusher so as I push it round it pulls the cup on to the next stage I really like this mechanism actually because if we change the angle of this and make it more shallow and we can make the ball go more slowly and that makes the wiper go more slowly so then we probably could have used use that so it didnt slop the water instead of the previous stage but I really want to fix this one now so were going to leave it as it is now we need to put the tea bag in the next stage is a kind of elevator or lift if youre in England which has got a plunger the ball goes in and the lift goes down and its got a plunger on the outside which plunges the tea bag into the cup so that have a bit of wire attached with the tea bag on the end so as its down it plunges into the cup and when its finished it Springs back up again and takes the tea bag out and we have some sort of door on the elevator so we can Brew the tea for some amount of time but before we carry on with that its time for a quick ad from the videos sponsor which is PV case now PV case is a Next Generation AutoCAD based piece of PV software focused on Automation and accuracy it allows you to simulate the actual location of a solar plant from the earlier stages of planning incorporating D topographical data points so PV case is the ideal choice for companies undertaking large commercial and Industrial projects as well as utility scale plans the software really is intuitive and has streamlined processes to help reduce the learning curve and improve productivity features include everything from the prototyping stage electrical design stringing shading and terrain analysis and automatic generation of construction documentation so pvks really does enable engineers and designers to take the project all the way through from its initial conception to the procurement phase this really is an endtoend approach which saves time and reduces errors it streamlined so you dont have to switch between tools or other software platforms other features include slope analysis piling and collision analysis automatic topographical D cabling side byside design comparison and Rapid D building preparation try PPKs for free by following the link in the video description right lets get back to building the next stages of the tea making machine the next stage is another one of those curv pieces and again thats got a pulley and another string that pulls the cup round and after that weve got another two of those at the back of this with a shoot in the middle so those are going to do the extra two stages of pouring milk in and then finally moving the cup out of the machine so theres the first one pulling a pulley just above that milk jug thats going to put milk in that seems to be working pretty well and then the final one actually pulls the cup along so it exits the machine and you can drink the tea well thats most of the main mechanics together but lets see if the thing runs all the way through if I put this massive Ball Bearing in well it looks like something went wrong cuz the cups still over here and it should be over here so it looks like those pushes are all moving at the wrong time probably because the vibrations in the machine are causing them to creep so the cup gets stuck halfway along Yep looks very much like they just move around with that massive ball bearing smashing its way through the machine so instead of this mm ball bearing were going to use the slightly smaller mm ball bearing however this old machine was designed for the bigger one so some of the holes are the wrong size um it will mean that we dont slush water around so much cuz theres not so much violent Force pushing it but actually for that stage that tips the first jug theres not enough mass in this one so we need to make a few modifications throughout the whole machine Ive also put some foam dampers on those pushers so they dont creep which is just a loop of foam that just means that theyre not so free moving despite having put bearings in there yep that seems to work pretty well made a little thing and stuck it in the top that hole so the smaller ball bearing doesnt come through before the Seesaw is finished so it basically has to go all the way to the bottom Ive also done that on the next stage as well which is the stage that pushes The Jug that pushes the water and youll notice it goes down freely now Ive added this box on which I can put some more mass in as well if I need to to offset the weight of the water and that balances it quite well which means it doesnt move down quite so violently and slush water everywhere youll notice also that The Jug actually stays in position now so it empties completely and it locks down once its been pulled Ive done this with basically a magnet which faces another magnet attached to the bracket so it kind of locks into place there and that means that it wont violently flick back up again and throw the water in the air hopefully Ive done that to both jugs so they both got these magnetic locks that gives it a definite sort of robotic action and that means that jug is going to get completely empty before it moves on rather than flying back and flinging liquid in the air thats still getting poured so Im pretty happy with that we may need to adjust the position of the jug back and forth in its cradle so it doesnt overpour the mug as it comes along and we might also need to put some Mass on the back there so it doesnt pour as quick but we cant really do that without the water in so thats going to need testing pretty happy with the way it looks though so lets uh run the whole thing and see what Applause happens cool look at that jug Applause action and that jug action we forgot to brew the tea tea bags come in a box like this not sponsored and the tea bag is on the end of this bit of wire now on the bottom of the Elevator Shaft so whats supposed to have happened is the ball comes in pushes the teabag down and dunks it the ball exits and it comes out again but you need to brew tea for more than one second or whatever at the moment it just goes in and pops out as the machines Applause running so I was going to have an escalator that went up and then a massive hel to Skelter that took a for the ball to come back down again thats quite a big thing to build and Ive built quite a lot already in this so what Im going to do is just use an electronic timing gate so Ive got one of these infrared readymade sensor modules thats got a little adjuster on it for adjusting the proximity and thats going to sense when that gate before the lift gets triggered thats plugged into an Arduino and theres a little Servo that opens the elevator door so basically well set up a timer on there so the ball falls into the elevator it dunks the teabag at the moment it only Waits seconds but will turn that up to brew the tea and then it pops out again and the rest of the machine continues so Im pretty happy with Music that right You heard of T well this is T and were in my kitchen where we do food projects like using meat instead of Music motors its time for tea right Kettle water goes in here so we get a nice hot cup of tea hope it doesnt melt the plastic should be all right in that jug and well put a splash of milk in there as we want it in our tea thats probably enough then all we have to do to make the tea is put this B bearing in the top lets see if it works Applause its a bit splashy still oh right the tea bags in while the Teas Brewing I just want to tell you that all the designs I do are open source for most of my projects and theyre all in my GitHub and theyre not just stls theyre actual solid models you can put into cads and edit as if you drew them and also publish all the code so if youd like to support me through patreon or YouTube channel membership then patrons and YouTube channel members actually get a separate cut of the video with most of the ads stripped out which is quite good as well and it doesnt cost very much so check those things out in the description Ive also got a Discord where you can discuss some of the projects right I think it might be time for the tea milk and the tea made yeah we have spilled a bit but apart from that it works fine yeah we might need to implement a drip tray of some sort but apart from that its made the perfect cup of tea

write a youtube script

I previously built a small mono wheel which is a single wheel that runs along with one motor driving it it balances on its Edge Like a Rolling coin mono Wheels have been built large enough for people to ride in which looks like a fun project my mono wheel had some stability control which measured the angle of the internals and modified the motor speed accordingly this meant that the internals always remained at one angle instead of oscillating between driving up the inside of the wheel and falling down again which is what happens if you just run the motor at constant speed I also built a second version which used control moment Gyros to stay stable and that one could stand still on the spot but is there an easier way to make a fun Roc control vehicle which is inherently stable a ped rail is a wheel of legs and feet this type of wheel has an internal guide which controls the length of each leg so that they can form a flat surface on the bottom and reach out at the sides to climb over obstacles I built a fourwheel drive crawler version of this with added suspension which works really well but what would happen if we combine the two to make a ped rail mono wheel hopefully itll be stable because the flat profile along the bottom will prevent it from rolling out of control if we use two side by side then we can use a differential drive to steer and itll stand still by itself Im not really sure whats going to happen here though will it work at all maybe the insides will spin around and the outside will stay still or the insides will continually drive up the outside and it will do wheelies and fall flat again all the time maybe itll actually work but we wont find out unless we try it the machine Ive designed also has the ability to shift the two sides up down so it can lean to go around corners so theres quite a lot of parts to print for this project Weve Got Gear Rings which a motor is going to drive around to push the two sides around and because theres two sides of course weve got two of those there are various other ring parts and itll become apparent what those are for when I assemble it as well as the core which has got the guide on and some pieces here printed with support material which Mount Motors and servos and all the other bits and pieces on just a quick ad from my D printing sponsor thanks to LOL spot for supporting my Channel with D printers its much easier to make things quicker when youve got so many printers to use in parallel there are lots of other bits and pieces to make including some little tabs and some gears and all of those mechanical Parts Im also doing some TPU feat and I need to print of those so Im using two printers at once to get them all done in time thanks to D fuel for the filament for this project you can now get off at D fuel.com with my special code and Link and Ill get a small commission its time for some assembly so weve got of these tabs which fit onto each of the wheels so these are the little tabs which hold all of the sticks in so those fit all the way around the circumference of each of these black wheels those are screwed onto some posts which leaves a gap in between the tabs where the feet can fit all the way around and Ive toleranced this so theyre fairly loose which means that theyll flop in and out fairly easily I need to make this project run really smoothly so thanks to Simply bearings for all the bearings in this project check out Simply bearings. co.uk these bearings have an mm inside diameter and mm outside diameter Ive got a little plastic tab that fits in there and that allows me to screw it onto each of the feet so that its nice and free moving and it makes the middle the right size for the screw each side of the wheel has legs or feet so thats in total and bearings and tabs and screws that have been assembled all of those are inserted and of course those all move quite freely but we need to put some grips on the ends so Ive made of these t feet for each side of the wheel which are flexible filament so all those TPU are feet are fitted all the way around with the screw in each side and if I turn this around now then you can see the ones at the bottom drop out and the ones at the top drop in so theyre really free moving so no problems there if I run this on the ground then obviously it makes a flat line at the bottom here as it touches the ground so thats working really well and of course if we could lock these in place as we want to or drive them along then when its rolling along it stops all of these make a flat line at the bottom so that will stop it wobbling so we should be to make it quite statically stable so all we need now is the guide that goes in the middle this is the core for one side and its also the guide which guides all those little feet round so the bearings run around that contour and that makes them the right length Ive got some more lowprofile bearings which are mounted on these plastic tabs so one goes each side and that sandwiches the bearing in the middle those screw onto the core so that they run freely and Ive got three of those on each of the cores theres a yellow spacer ring that fits onto the original ring holding the feet and then the core fits into that Gap so those bearings run in The Recess on top of that is the gear ring that the motors finally going to run on and all of that screws together so that runs pretty much okay obviously those bearings are rubbing on their side when the whole things lying down on its side like that but when we put it upright theyll all run on their edges as intended so yeah thats much better if I just tilt it up slightly so if I turn the whole thing upside down you can see those feet which are now at the top are following that Contour quite nicely and its quite free moving so hopefully a motor will have no problems pushing it along and if I turn it up the other way then it follows the ground quite nicely and the whole thing runs okay even though weve got those TPU feet giving us friction against the ground so Im pretty happy with that so far so that seems to run all right if we look at all those feet then they all come to the right length and make a flat bottom if we lock the inside in place it looks like its going to be pretty stable but we need a motor to drive it so Ive got this vt Gear Head motor with a plat attached to it and Ive also got a spur gear with a captive nut and a grub screw so I can attach it to the shaft so that fits onto one of those Towers on the core where that tab on the motor is screwed on and that obviously binds with the ring gear so when to put Power on the motor it makes all of the legs spin round yeah the frame rate of the camera makes it look like its going backwards much more slowly than it is but its its actually spinning FAS in the other direction and obviously Ive been making two of those all the way along and these arent mirrors of each other theyre actually identical Parts which are basically exactly the same and thats cuz on one of them weve got this sort of Tolerance piece that I printed separately and that fits into a groove on the other one and the two of those fit back to back so that they mesh so its actually the same part just flipped round and put onto the back but before we put the rest of that together and see if it actually works at all its time for a quick from the video sponsor which is PCB way PCB way is a OneStop shop for PCB manufacturing assembly and other types of manufacturing services including contract manufacturing all under the same roof PCB way manufactures all sorts of boards including Standard fiberglass boards but also aluminium pcbs flexible pcbs and rigid Flex pcbs which are part rigid and part flexible PCB way also provides CNC services including online CNC Machining sheet metal fabric a D printing and injection molding the CNC Machining Services include a wide range of materials including aluminum stainless steel and various Plastics PCB Ways D printing services include SLS SLA DLP fdm and more in a variety of materials check out the PCB way website to browse through a variety of finishes and get a quote check out the PCB way shared project section this is a community of User submitted projects with PCB schematics and parts listing so you can reproduce the projects you see there they also have a module store which has all sorts of items for sale such as Arduino boards tool kits robot parts and kits and sensor modules and theres currently a Christmas sale with up to off find out more now at pcb.com and Ill put that link in the description to this video right lets put the rest of this together and see what happens so as well as fitting the two sides together thats also the mechanism which allows them to slide so it can steer as it goes around corners and as a result of those pieces being the same the motors are offset so they nest nicely together and its nice and compact so thats actually pretty stable now because its flat at the bottom and the insides locked pretty much by the motors we do need to do something to actuate this which well put in in a moment and obviously we need to put Electronics in to drive it along and then well see if its going to work at all to drive this sliding mechanism weve got a rack on one side and then weve got a pinion mounted on a Servo on the other side and thats fitted on both of those pieces into the Gap that I left so weve got a TNT . and an orange radio receiver so we can radio control it and weve also got two batteries . Volt s lipos one fed each side and some voltage Regulators for the servo and the electronics Ive got two bts motor drivers fitted underneath and Ive put covers on those so they dont get smashed on anything that it drives over and Im going to be using my universal remote that I built in another project some time ago which is a DSM remote which sends radio signals to that DRX receiver so if I move one of the sticks we can see that that thing slides side by side and that seems to be working pretty well Ive got some motion filtering on there as well so that it doesnt react too quickly the Motions are nice and smooth so if I let go of the stick it doesnt just spring back really fast I tried to leave the middle mechanically clear just like a mono wheel that we started with but obviously theres lots of wires and electronics in there but there is a hold straight through the middle so lets see what happens so initially that seems pretty good I can drive and it actually drives along and it doesnt just stay still with the inside spinning round and it doesnt sort of drive up the inside and do wheelies so thats good I can steer although you can see its a bit sketchy when I try and run one motor and not the other because all of those flat feet basically making that flat profile along the bottom get stuck and you can see it kind of chattering on the surface there and lifting off the ground and weird things happening if I try and turn on the spot then even weirder things happen all of those legs extend and it does a wobbly dance and everything goes out of control so thats not quite what I was expecting doesnt really steer like a tank where we can run one track forwards and One backwards and turn on the spot but its kind of fun anyway but other than that I can drive around and if Im really careful how I steer so I just run one motor slightly slower then it pretty much works but otherwise those sticks seem to lift up and the other side scrape on the ground around and it does all sorts of weird things so tipping over with the uh Servo doesnt really help me at all that just seems to cause even more problems but lets give that a proper test so if I turn here and lean over to one side and then well try and drive to the right then actually what happens is we dont get any friction on the ground with that side at all and um yeah it doesnt work at all really lets try that the other way we can kind of turn but then bad things happen so I can actually make the inside spin round if if I drive forwards and backwards quickly in quick succession but it quickly settles and everything goes back to normal again its actually quite hard to achieve the original ped rail I built was really good at climbing over things but this one not so much probably cuz it hasnt got so much mass on the ground it eventually gets there but basically what happens is all of the feet just get stuck and then the inside spin round if I drive really slowly then I can kind of get onto the item but if theres any recesses in there like this bin Le it gets stuck again I really want to make it turn better so Ive replaced all the TPU feet with rigid feet just made out of PLA and hopefully that will help it slip sideways better so it doesnt get stuck this runs much better and its much better at steering because now those feet arent gripping they can swivel on the spot and that means I can make tighter turns without so many issues its a bit more like a tank where we can run one track faster than the other and it will turn so Im pretty happy with the way this drives now and we can see theres much less issues with those feet riding out of their guides as it turns so its actually quite a fun thing to drive around with leaning over really doesnt help me much more though we still have equivalent friction on each side so its slightly better but really not much point in having that we might as well just leave it in its Central mode where both of those are aligned and turning on the spot still does some really weird things we still get basically the same effect where the legs jump out of their sockets and yeah kind of spins around but its a bit weird really I thought Id try it on grass which initially seemed really good until it got stuck in a dip and then the legs get stuck and the outside just spins around and basically I couldnt really get it out of there even if I run fast or slow I can move a little bit but pretty much theres too much friction with the ground on all of those feet and now its just basically completely stuck I can just about turn but yeah its not very good at all really but great on a smooth surface

write a youtube script

a number of people have sent me various videos of the new Star Wars droids walking around Disneys Galaxys Edge Theme Park these are smish bipedal Droid characters which are a bit like bd but kind of bigger with bigger joints because they arent fiction these are actually pretty Dynamic robots which have come out of Disney imagineering the same team most likely built the baby Groot robot which we saw a couple of years ago key features include the ability to walk on uneven terrain and stay upright if they get pushed around theyre also quite emotive characters and appear to potentially have some form of AI to make them interact although theyre also puppeteered interestingly it looks like they dont have sideways ankle joints they just pivot on the feet which are long edges that touch the ground otherwise as far as I can tell they have an axis at the ankle knee and hip and then two more axes for side stepping and rotating the leg so they can walk in a curve they also have a head which does a few things and in some footage it looks like that may be helping stability front to back but its not always apparent and probably isnt necessary since other bipedal robots dont need to throw Mass back and forth like that so a lot of people have asked if I can make a project like that but Im not a team of researchers funded by Disney Im just one person in my house making YouTube videos for clicks whoops some of you may remember my walking Star Wars gonk droid that I built some time ago I was pretty happy with the results of the time this robot had pretty big feet which it definitely needed and I was on a pretty tight budget at the time so I used motors that only cost around each in money and of course a lot of D printing it was dynamically stable though so measured its angle with an inertial measurement unit and tried to balance however unlike the robot dogs I went on to build the gong Droid had no inverse kinematics or interpolation which well talk about later in this video so I was just adding some numbers to some of the angles based on how far it Lent over and trying to coax it along Ive learned a lot since then and I also have more budget so Im going to attempt a new bipedal robot based on the Disney droids the robot Im going to build in this video will only have four Motors Im going to ignore four of the hip joints as well and also simplify the legs so we just just end up with something thats a bit like tipping a box onto its corners without the body being able to translate directly sideways I am however going to walk on feet with edges to Pivot sideways so no big feet this time around if this works okay then I can move on to build something more advanced in a future video with a head and some of the missing joints so my legs are parallelograms which means I only need one motor for each half instead of separate ankle knee and hip Motors and yeah thats all the axes it has this leg structure is similar to how I recently built the walking atap project which you can check out in my channel although that didnt have to balance sideways and also the original gonk Droid but that had really big feet so well see how it goes for me this time theres quite a few parts to make for this project because its bipedal which means we have two legs and we have to make all the parts twice plus the bit of the body but its not as bad as making robot dogs I guess with four legs just a quick ad from my D printing sponsor thanks to LOLs bot for supporting my Channel with D printers here are some of the sticks that make up the parallelograms for the leg and weve also got these feet parts to print which weve got four of those cuz theres two on each side the body doesnt actually fit on the normal L spot Taz bed so Im using the Taz long bed which has a bed twice as long as the other printers thatll enable me to do the hips and the main piece of the body which would color coding to look just like the Disney droids thanks to D fuel for the filament for this project you can now get off at D fuel.com with my special code in link and Ill get a small commission Im going to be using these Dynamics or servos for this project and these were pretty previously sent to me by robotis who are the manufacturer and Ive used these in some projects before these are XM W t servos each one of those is mounted in a housing with a bearing on the bottom so we can make a double hinged or double BRAC joint and that should be pretty sturdy for our robot this is the hip section for the robot and on the bottom of that piece Ive got another D print which is glued and screwed on and thats going to be the back pivot point for the parallelogram so Im just going to put some mm studing through there thats fixed on with some nuts either side and thats going to make the back pivot point for the legs so each of these servos actually has some M holes and Ive also left some holes in the D print and those can be screwed or bolted through into that hip section thats going to be the mounting for most of these servos I did design a plate for the bottom of these as well but Im pretty sure that Servo is held on okay now and it also restricts access to the connector where I need to plug in the data cable so Im not going to bother installing those right now the output of the servos has a nice plate with lots of M holes and of course I can attach my D print there right to it and thats going to make the first part of the leg on those servos on each side and thats basically the thighs on the other side of the servo the inside of the thighs weve got another piece and that piece has got some lumps installed that fit into the bearings I put a screw through those that means theyve got a metal core and they wont easily snap off so that should be super tough that plugs into the bearing which is on the back of that Servo Mount and then that moves quite freely and theres a spacer between those that holds them together and holds the whole thing so that the pieces dont just pull out of the bearings so those are my two fronts of my thighs now complete the legs are made of parallelograms so weve got the back of the parallelogram as well which are made of these parts which have some bearings installed so those fit onto the studding on one side and those are obviously running on bearings and then on the back theyll fit onto the knee and that will make the parallelogram for the top half of the leg the knee sections are two more servos that are going to control the bottom of the leg so weve got that bearing and the servo horn on the front and two bearings on the back so we can complete the parallelogram so the levers are attached to those that go down to the feet and then on the back of those weve got another piece of a little lumpin that fits into those bearings and that means weve got that double brace section again to go down to the feet the feet have these TPU strips on which are basically the edges Im going to walk on and Im able to reposition those to different screw holes so we can change the pivot point and again weve got these orange levers that go down to the back of the ankle to complete the parallelogram and thats on a bolt through a bearing with that alog together weve got the bottoms of the legs with the feet attached and of course each of those are parallelograms with the servo at the knee which means everything moves parallel together youll notice that orange lever is precariously attached in the middle there with a gap either side and that means that the orange lever from the top of the leg can come and overlap with it so theyre all pivot on the same pivot Points and youll also notice theres a bearing on the outside here and theres also spacer bearing on the inside and thats against so the top levers can overlap on the same pivot Points and I left those lumps in there on the top of the leg so we can now put the bottoms of the leg on and everythings upside down here with the hips onto the bench the missing piece is on the outside of the thigh which is another piece with a lump on and that fits into that bearing on the knee and then that screws on and holds everything together so now weve got basically our complete parallelogram as soon as we put that orange lever on and that as I said just fits either side of the bottom half of the lever and that bolt goes through so now weve got two parallelograms with all of the pivot Points overlapping thats going to make it much easier to solve inverse kinematics these are my lumps that I put on the bottom made of TPU for the pivot Points and Im probably going to move those around Ive put them quite close together at the moment this is the body we printed on the Taz long bed which fits on there and thats got a split in it so I can take the top off and we can look at the electronics that consists of various boards in a stack which you can see just there on the bottom of the stack is a TC . which is the main microcontroller followed by the dynamixel interface Shield which Ive connected to a Serial port on the Tec and also on the top weve got the b which is a new spark fun inertial measurement unit that should just give us the angle Im using the library and the example provided by spark fun on the i squ c interface to the teen C and we can see weve got our three axes of angle there for pitch roll and Y but for these experiments were only going to use the roll axis Ive got a s LiPo battery which is going to power everything sat under the electronics so now Ive wired in the dynamixel interface and Ive just split it off and Daisy chained the servos in the hips followed by the knees and thats all you need to do to control them but before we see what we can get out of that its time for a quick ad from the videos sponsor which is PVC case now PVC case is a Next Generation AutoCAD based piece of PV software focused on Automation and accuracy it allows you to simulate the actual location of a Solar plant from the earlier stages of planning incorporating D topographical data points so PV case is the ideal choice for companies undertaking large commercial and Industrial projects as well as utility scale plans the software really is intuitive and has streamlined processes to help reduce the learning curve and improve productivity features include everything from the prototyping stage electrical design stringing shading and terrain analysis and automatic generation of construction documentation so PB case really does enable engineers and designers to to take the project all the way through from its initial conception to the procurement phase this really is an endtoend approach which saves time and reduces errors its streamlined so you dont need to switch between tools or other software platforms other features include slope analysis piling and collision analysis automated topographical D cabling side byside design comparison and Rapid D building preparation try pvk for free by following the link in the description to this video right lets get this powered up well its standing under its own power Ive moved these TPU foot blocks slightly further apart so its a bit more stable now so kind of like you know feels like it wants to walk along really so uh lets code it up and see if we can do something Ive talked quite a bit about inverse kinematics in some other videos what we need to achieve is being able to specify the foot position in cartisian coordinates which are straight lines like the axes of a graph and then have the maths work out what the actual joint angle should be for us this isnt too hard with high school trigonometry if we divide the robot up into triangles firstly we take the side profile of the leg we know two of the lengths and we need to specify the other one which means we know all three sides of the triangle we can use Pythagoras Theorem and trigonometry to work out the angles here then once we know which angles are required to make the leg the length we need we can treat that as a straight line and draw another triangle to move the foot back and forth thats a right angle triangle so its also easy to solve the hypotenuse of that triangle gets fed into the previous bit of math to specify the new leg length Ive attached a radio receiver to my electronic Stacks which is an orange receiver for DSM remotes and I built my own DSM remote in another video that you can check out which Ill be using in this project so now if I turn one of the sticks we can move in the Zed axis up and down and it should go in a perfectly straight line we can also move backwards and forwards if Im really careful not to tip over backwards as we move the mass essentially you can see the body though moving in a straight line back and forth and all of those controls mixed together because one piece of mass is fed into the other and that means I can move anywhere within the axes of my graph even if I dont want to move in a straight line for instance I can plot a curve or I can move diagonally and all of those axes work together just specifying the two coordinates its all very well and good while Im moving the stick slowly but if I make a step sequence basically I move those legs on timers then the servos are going to move as fast as they can possibly go between the two positions and thats far too severe probably kicking the leg like this is going to tip the robot over so the next thing we need to do is use interpolation which allows me to plot all the points from A to B over a longer period of time but with the step change input Im using the Arduino ramp library for this and Ive used this several times before so now I can make those legs move much more slowly over a longer period of time and basically interpolates all the positions on the way as if Im moving the sticks really slowly instead of making that harsh step change I can also specify any coordinate for the start and end and that means I can move the legs diagonally or plot any sort of path that I want to do we can also of course make both legs move and make them move independently to do whatever motions wed like lets put it on its feet and see what happens obviously with those legs moving to fixed positions on fixed timers it just pushes itself back and forth but lets try and push it onto one foot and see if we can get it to take steps so you can see it kind of looks like it wants to work there but obviously its got no sense of balance we might get two or three steps out of it but then of course it overbalances and yep bit of a disaster so far if I take the orange body piece off that removes some of the mass so now it can thrash around even quicker which um means it probably doesnt work at all so I can kind of try and hold it in position but really what we need is for it toh work out when its tipping over itself and stop it tipping over and stop it building up momentum like this and just tipping over so it looks like those Motors are fast enough and powerful enough for it to be able to walk which is quite good but obviously unregulated with fixed timers on fixed positions we might get two or three steps in but then basically inertia builds up as it rocks backwards and forwards and eventually of course it tips over so now of course what we need to do is use that inertial measurement unit that we installed which is right on top there to measure the angle and somehow modify those steps to regulate them to try and get it balanced consistently without overbalancing and tipping over so heres what we did have which is just the feet moving up and down when one moves up the other one moves down and they move up or down and that takes half a second so a second walking cycle basically so now what Ive done is put a gap in between triggering each of those legs to take a step and Ive tied that to one of my sticks so as I turn the stick the Gap in between the steps gets bigger so you can see thats a really big gap there and as I turn the stick down again then that Gap gets shorter and shorter until theres no difference and we end up with what we had before and I plan to regulate this Gap with the inertial measurement unit data to keep it stable so some initial values there which basically Ive got the inertial measurement unit roll axes which is measuring the angle side to side and as that angle gets bigger it makes the Gap in between steps bigger thats all it does theres no fancy P controllers or anything its just a multiplication so if I just hold the programming cable out the way there so its not dragging you can see that if it tips too far then it kind of slows down and thats kills the momentum so it doesnt keep overbalancing but it doesnt look that consistent and its still leaning quite a lot so I think we can still do better than this this is exactly the same thing but Im showing the timer in real time there so the other statistics are that the leg is mm at its longest and at his shortest so only actually moving mm the whole leg cycle is milliseconds so thats milliseconds to lift it up and to put it down and then the trigger timer ranges from a minimum of milliseconds up to whatever it is and I think I just multiplied the IMU roll data by something like and then added it on and thats all its doing to vary that Gap in between steps and you can see that its quite inconsistent sometimes theres a big gap and as youd expect sometimes theres a small Gap so even though its stable it looks a bit crazy still I thought Id speed things up a bit so it doesnt have to lean so far so now my leg cycle is reduced to milliseconds so milliseconds up and milliseconds to move the leg down and my Trigger Time ranges from milliseconds up to whatever based on adding that IMU data just multiplied up till it works by trial and error its still quite inconsistent looking though because that timer is ranging by quite a large amount so the answer to that is to have two timers one is shortening the leg cycle so its actually making that up and down time shorter when it leans over more but its making the trigger between the legs longer so thats basically making the step shorter and making the gap between them longer but Al together that gives us roughly the same amount so that looks much more consistent and it looks much more stable and repeatable Yep this is exactly the same thing but were just going to have a closeup there we can see that its definitely walking on those edges and its definitely taking its feet off the ground which is pretty good so no shuffling and its not relying on those big plates to balance its definitely walking on those edges um yeah Im pretty happy with how this has gone for a two we project I wasnt really sure if Id be able to make this work at all without any hip joints to translate sideways and just tipping onto the corners like a box tipping onto its Corners basically so Im pretty happy of how this has worked out so far the only one thing I will point out though is that the tables quite wobbly and that may be causing me some issues its a similar thing to if you put lots of metronomes on a moving table they all go in sync so it may be that the robot is also reflecting the coefficient of the wobbliness of the table but um Im pretty sure that this is pretty good for something I threw together it does start by itself when I pick it up it feels a bit sketchy I probably need another little sequence to start off the first step differently if I push the feet along though it feels very much like it wants to walk along and still balances okay so I should probably try and code up something here to try and make it it takes some rudimentary steps and see if we can get it to local mode yeah stopping isnt graceful either really so I just added some extra stages to my step sequencer to move the feet back when theyre on the ground and move them forward when theyre off the ground and now it basically takes rudimentary steps if it reminds you of Huey Dey and Louie then post in the comments and this is where I probably should have grabbed the robot instead of trying to stop it with the remote because disaster struck and yeah the plastic is completely smashed there so thats the end of that well I need to rebuild the hips anyway to put the joints in so it can go like this and also so it can go like this so it can turn on the spot so thats at least another two axes per hip that I need to rebuild so thats going to happen in a future video but Im pretty happy with that for a twoe project its given me quite a lot of confidence I can actually make this work and Im also going to put a head on next time as well Im going to publish all the cads so far for version one as open source and all thats on my GitHub so if youd like to support me through patreon then have a look at that or YouTube channel membership and patrons and YouTube channel members can get all of the videos up to a week early all right thats all for now

write a youtube script

a mobus strip or Loop is a onesided object known as a nonorientable surface meaning that within it one cannot consistently distinguish clockwise from counterclockwise turns you can read more about this on Wikipedia but we can easily demonstrate the principle by getting a strip of paper putting one twisting it and joining the ends together this means that we essentially have one continuous side so if we follow it all around we travel over what would have been both sides of the strip when it was flat and return to where we started also if we cut a mobia strip in half by just cutting all the way down the middle of it all the way around we end up with one massive loop with two twists in it which is quite interesting so I got a YouTube comment on one of my other funny tank videos from one of my YouTube channel members saying why dont we make a tank with mobia strip Treads with a twis in like that and they didnt really know why that was a good idea other than it looks cool and I dont know either so were going to build one anyway and see what happens so Ive designed some track links and they have these Square kind of pivot sections in the front here and that allows two of those tracks to Pivot against each other as well as bending normally to go around sprockets and that gives us around of movement either way so its time to D print a few of those up and see if this actually works in real life and what sort of rotation we can get out of this were going to need quite a few links for the tracks so Im using some mm stainless steel in there and that fits nicely in the slot and we get that rotation at the link point so with a couple of those together we can see that it bends in both directions so it pivots like this and also bends to go around the sprockets which is just what we need for a tank so theres track links together and as you can see that almost twists all the way around with Dees on each one we should get the complete but were not quite there probably need another one to get round to thats probably because of the tolerancing on the D prints on those swiveling parts obviously it bends the other way so it can go around the sprockets Im not sure how easily its going to run when we twist it and we want to get it back on the sprockets at the end to come back to the bottom again but there we go lets put some more together and see what happens we dont do these things cuz theyre easy though we do them cuz we thought theyd be easy so theres links together which pretty much gives us quite comfortably obviously this tank is going to be quite long though just so we can get that twist in the length of it so the rest of this looks pretty much like a normal tank the only difference is weve got this massive Gap in the top here so weve got space to get the twist in the track in this is going to turn out to be quite a sizable assembly so just a quick ad from my D printing sponsor thanks to LOL spot for supporting my channel with D printers makes it much easier when youve got lots of printers to make all the parts in parallel this is part of the drivetrain which is basically a cylinder which fits between two sprockets and those go at each end to drive the two holes in the tracks theres a lot of those needed though because we need four sprockets per track and two tracks and each sprockets got two ends the whole things pretty long so Im using the L bot Taz long bed which has a bed twice as long as the other printers to do the main side pieces for the track thanks to D fuel for the filament for this project you can now get off at D fuel.com with my special code and Link and Ill get a small commission I printed those side Parts with a . mm nozzle as well so actually really big tough parts and they didnt take too long to print obviously theres one of those that goes on each side of each track so there actually four in total that fit like that here are the sprockets themselves so theres two ends youll notice one has a longer piece on thats to put a pulley on one of them to drive them and those of course go either side of the cylinder there and they just screw down thanks to Simply bearings for the bearings for this project the last time I did a tank I didnt put any bearings in it was just metal on plastic and it didnt run very well so check out Simply bearings. co.uk of course these sprockets run really well now Im just running them on studing but that studing allows me to screw the whole tank together with nuts I remember to put bogey wheels on this tank so these fit in the middle there are three of them which fit in the middle of the track at the bottom youll notice those are quite skin and well find out why that is later and those are fixed with just some nuts done up against each other either side so they stay in place and that holds the whole frame together I did remember to put the belt on and the pulley so we can drive one of the sprockets later the next parts are these which Hold the top sprockets and these screw onto the frame separately and well find out why that is very shortly also got these pieces which screw between the two sides and that just makes the whole frame a bit more rigid so it doesnt twist and its just not relying on that studing to kind of hold the whole thing rigid the top sprockets fit between these gray plates and I made those separately and that allows me to be able to reprint them really easily to tension the track up by moving the top sprockets apart without having to go and reprint the whole side with that massive printer so that just makes things a bit easier for getting things tensioned so thats one side done weve got to put the track on now and see if we can get that twist in Ive put this track together with links in and I think we need another eight or maybe in there to try and get that twist in of course which will make that section a bit longer so lets lets put together the rest and see if this actually runs fine with the twis in and if this works at all with the assembly the length is the minimum amount of track links I could get in there to get that twist in and have the track go all the way around is links so there it is it actually seems to run surprisingly well with that twisting I was expecting it to be far worse however if I run it too fast or I run it too far then some odd things happen so that track isnt quite as tight as it should be and that results in the sprockets at the bottom there popping completely out of the holes in each track link that theyre supposed to drive and sort of getting stuck in between the tracks so yeah that track is quite loose and I couldnt really make it tighter I do have these gray plates that I can just reprint though so Ive moved the holes around mm to the left on one end and thats made that track quite a lot tighter Ive used these motors quite a few times before for crawling Vehicles they are volt worm driven gear boxes and theyre pretty powerful Ive made a hub for each of those each one has an M nut and a grub screw in there and then theres a pulley piece which fits on separately so I can reprint that if I need to ret tolerance it as well that fits onto a plate which also allows me to rolance the tension by just reprinting the plate so there it is fitted in there and of course thats driving that pulley and sprocket I fitted earlier so here it is running under power Ive just put a battery on the motor there running it backwards and forwards this has actually worked out pretty well Im surprised that twist works as well without the chain coming off and also the sprockets now stay into the track Ling so Im pretty happy with that so we just need to build another one so you might have noticed these two bits of studing poking out the other side obviously thats for the other track that were going to build out there but before we put the rest of that together and see how well it works its time for a quick answer on the video sponsor which is PCB way PCB way is a OneStop shop for PCB manufacturing assembly and other types of manufacturing services including contract manufacturing and its all Under One Roof PCB way manufactures all sorts of boards including Standard fiberglass boards but also aluminium pcbs flexible pcbs and rigid Flex pcbs which are part rigid and part flexible PCB way also provides CNC services including online CNC Machining sheet metal fabrication D printing and injection molding the CNC Machining Services include a wide range of materials including aluminum stainless steel and various Plastics if you dont see the material you like you can also choose from Custom materials PCB Ways D printing services include SLS SLA DLP fdm and More in iety of materials check out the PCB way website to browse through a variety of finishes and get a quote check out the PCB way Shar project section this is a community of User submitted projects with PCB schematics and parts listing so you can reproduce the projects you see there they also have a module store which has all sorts of items for sale such as Arduino boards toolkits robot parts and kits and sensor modules find out more now at pcb.com and Ill put that link in the description to this video right lets put this together and and see what happens the other sides pretty much exactly the same so in the middle we now have some Electronics made of an Arduino Mega and an orange RC radio receiver and Ive got two s lipos wide in series driving the motors through two bts motor drivers and I think those are rated at something like volts and amps which is more than enough for those Motors as usual Im using my DSM remote which is a universal remote for all my projects and you can check that out in my channel on the whole it runs pretty well Im pretty happy how this has worked out its pretty much as good as any RC tank I think so yeah lets just drive that up and down a bit seems to work fine the twist in the track also has stayed on and it just seems to just Cascade properly off and on the sprockets at the end so Im really happy with that too but lets see what we can do with it can we drive over obstacles and all sorts of things that tank should do well a bit of aluminium extrusions no problem for it but lets try a bit of a ramp so yeah that seems to work okay even though my tracks have no grips and theyre entirely smooth lets try an even bigger ramp made of some smooth shiny acrylic and that seems to work just fine too what about an even bigger smoother ramp lets see if we can get up there well no joy to start with because those Treads wont go over the ledge which is probably about an inch and a half deep Im just going to help it and just get those front Treads up to start with uh its okay while were pushing against the ground but as soon as we leave the ground then that tables just too smooth and weve got no grips on our track so thats no good at all well that seems to work just as well as any other D printed tank I guess but the question you want the answer to Is Why have we got this mobia strip track why have we got this twist in the track so the thing I could come up with was that when we look at really big track vehicles like Megabots for instance theyve got loads and loads of friction with the ground they have real problems turning because when one track goes forward and the other one goes backwards some of that track at the ends basically has to slide sideways so they have loads of problems with friction and they get stuck so what they did was put smooth nylon plates all over the Treads to reduce that friction and make the track smooth now in my case the tracks are smooth anyway so its fine at turning and things but its also really bad at gripping ramps as weve already seen but in this case weve got this twist so we can have one side of the track with treads on thats something that grips and the other side smooth and then we get the best of both worlds um although weve only actually got one side cuz its a mobia strip so Im not quite sure how thats going to work out so I cut up some big sticky pads into lots of small pieces and put them all over the tracks making sure I avoid the holes where those sprockets go of course and also the bit in the Middle where the dolly wheels go and Ive covered the whole track and basically weve covered one length of the track so that means that we get one entire circumference of the track which is completely smooth and then eventually we get one whole circumference essentially the other side where weve got grips on and so we have one transition point where it swaps over thats actually only in one place on the track so theres all of the grips and then you can see the transition Point coming round then we get one complete circumference of a completely smooth track again and another complete circumference with the grips on which gives us quite a lot of variety so you can see the bogey Wheels run in the middle of those pads there at the bottom which is why theyre skinny and why I put the pads there and also those pads run in between the two halves of the sprocket and thats why that cylinder is so much smaller than the two half of the sprockets which are screwed onto the end I also moved the holes in the gray Parts again to tension those tracks up so theyre much much tighter so nothing weird happens now this grips lets try driving again though and obviously backwards and forwards works perfectly well no matter what the orientation of the tracks and those grips is but as soon as they get out of syn and we try and turn with the tracks turning in opposite directions then basically of course the smooth side just slips and the grippy side overwhelms it and so lots of weird things happen when we try to steer depending on what the orientation of those two tracks is and the position of the slippy part and the grippy part so there the tracks are turning in opposite directions but the one with the grips is winning we cant turn at all here they are in sync with two smooth sides facing the ground and of course that just turns fine now and we dont have too much friction either so we dont get that screeching sound from the rubber grips wiping on the floor so that would be quite good if it was a massive robot with loads of Mass on but of course as soon as we get one grippy track that just totally overwhelms the other one and itll only go in a straight line but lets try the ramp again so obviously the smooth tracks wont get us over the edge there on the tabletop as soon as the grips come round then that drives up there perfectly fine and because weve got that whole circumference of grips weve still got plenty more to go and weve reached the end of the ramp so that plan worked out quite well Im not going to drive off the end though so lets just drive back down again and now weve got two smooth surfaces on the ground we can turn around perfectly well it will of course drive on other surfaces perfectly well just like a normal RC tank so if theres a bit of grip on the ground like the grass then basically things work perfectly well it doesnt really matter what the orientation of the grips and smooth is and I can still turn perfectly well on the grass there because we just seem to get equal friction driving over uneven surfaces like piles of wood works pretty well as well not really seeing any slip when weve got the smooth parts of the tracks got one smooth and one grippy there coming around seems to work just perfectly well the grips do help of course but basically it doesnt seem to really matter and it will drive over any older loads of wood or any other obstacles perfectly well just like a normal RC tank Im going to publish all the cad and code for this as open source and you can find the link in the description to this video and dont forget you can support me through patreon if youd like to and if you like funny tank projects and check out my channel Ive got a tank that bends to steer a wave Drive tank and also a triangularshaped omnidirectional tank that can move in any direction and youll find all of those in my channel

write a youtube script

a couple of months ago Elon Musk announced some new information on the SpaceX bfr rocket which was live streamed on the SpaceX YouTube channel Ill post a link to it in the description below if you want to check it out but what caught my attention was the new control method of the BFS the big falcon ship now the current Falcon rocket and also the Falcon heavy rocket reenter the atmosphere backwards with the rocket nozzles pointing first and it controls itself using aerodynamic grid fins which are mounted on the top of the rocket and they sort of act like ailerons or elevators on a plane and also direct the rocket down to the ground before a lastminute rocket burn just before landing now this new big falcon ship is designed to descend belly first rather than rocket nozzles first and the purpose of this is to try and bleed off as much velocity as possible using the aerodynamic drag considering there is an atmosphere where its landing this will mean that when it comes in to do the rotation to do the last landing burn its traveling a lot slower than if it were to just travel backwards through the atmosphere now whats interesting about this is the BF s has very large fins on the rear of the rocket and also on the nose but these dont work like regular fins on a rocket where they would maneuver the air similar to an aileron or elevator on a plane instead what they do is they rotate about the parallel axis of the rocket and actually I used to induce drag on the nose or the tail as its descending that belly first Eagle must described it as similar to the way a skydiver would fall where they manipulate the drag now this concept is really interesting to me because its almost identical to how my radiocontrolled wingsuit guy would fly he would maneuver the angle of his arms to produce more or less drag at the front and when therefore caused them to pitch up and pitch down but we all know how that went Music yeah the head snapped up anyway so what I want to do is try to replicate this method of descent by building a radiocontrolled model of the big Falcon ship so lets jump into fusion and Ill run you through my current design so heres my design of the model BFS its going to be built out of mostly foam board just to keep it as lightweight as possible with obviously some d printed parts the main power source is going to be a motor and propeller which is mounted at the bottom here and the thrust from the propeller is going to be vectored using these three fins now thats mainly just to get the rocket to altitude and also hopefully we want to perform a slow landing the main thing Im interested in here is the adjustment of these two fins here now the real Falcon chip the big Falcon ship will adjust the angle of these fins up here but because this rocket is going on this model is going to be fairly small to add the weight in complexity to adjust the top fins Im not sure thats necessarily relevant right now the main thing is to test whether these bottom fins will adjust the center of pressure of the rocket and because the center of gravity or mass will always be same when the center of pressure moves either side of the center of mass the rocket should pitch up and down so thats what I want to test and also you may be wondering why theres these odd sticks sticking out the side of the rocket and thats mainly theyre basically the landing legs on the real rocket theyll be using these fins as the landing legs however to keep this as lightweight its possible to make that hinge as strong as possible and also have these servo withstand any kind of forces from hard landing just isnt feasible so I thought Id make my own sort of landing legs wish to take out the side and will be a lot stronger now this model is quite complex with the whole thrust vectoring unit and therefore on the heavy side so Im not sure how fast its going to descend it might fall quite quickly and hit the ground quite fast and I did it originally I actually have an idea just to build a glider version which would just vary the fins at the bottom and only be powered by a very small lightweight battery and not have a motor on board and this would have been dropped from a drone and I hope you just be able to perform a controlled glide down however I thought Id get some opinions from you guys on patreon and it seemed like the majority of you wanted me to make this fully thrust effected version so yeah I think we should go ahead with this so what Im going to do is start the printer and get on with cutting out some foam because of the curve shape at the nose cone its quite difficult to construct from a sheet of foam so I decided to d print it using a single outer wall layer method whilst the printer was going in circles I started cutting out the bottom profile sections of the rocket from the foam sheet once they were cut out I mounted the motor to the d printed motor mount which also acts as a structural jig to hold the three fins together the next step was to construct me main fuselage tube which isnt easy using this foam board as it tends to snap if you bend it too far so I laminated one side using packing tape to add some tensional strength to what will be the outer side of the tube then after bending rubbing scoring or whatever youd call it the foam board started to curve and I could join the two ends together however the curve was pretty terrible and I almost slightly add some disk bulkheads inside to maintain the round shape the nose cone eventually finished printing and I take the main parts together to give an idea of what itll be like when its finished unfortunately I was unable to finish it in this video so prepare yourself for some lame excuses take it away Tom so unfortunately Im going to have to cut the part one of this build video slightly short due to a number of reasons the main issue being the electronics of the hold for this project which include the electronic servos and also ordered some lightweight batteries for powering it for some reason when I ordered them they werent shipped for two to three days so they havent arrived yet and I dont really want to be gluing stuff together without having the electronics available to obviously mount the other issue I ran into was that the nose cone is printed with a single outer wall layer to make it as lightweight as possible however for some reason my printer decided to miss a few layers under extruded a few layers so that was a well it was about a tower print just wasted so thats the end of part of this rig controlled muddle big Falcon ship I hope youre interested in this project and if you are then it would be appreciated if you could leave a thumbs up if youre new to my channel and you want to follow this project then please click Subscribe down below and a huge thanks to all of my patrons for supporting me you guys make these weekly videos possible and also thanks for suggesting or at least voting on the way I should configure this because I really think that the thrust vector design will be the most impressive if it works I suppose well have to find out thanks very much for watching and hopefully Ill see you with part two next week goodbye

write a youtube script

worms locomote using a system of peristalsis or peristaltic motion this means that sections of their body expand and contract to push themselves along there have been quite a few research projects for robots that move like this and I previously Built My Own simple version which Ed four concertina mechanisms driven by servos to make four expanding and Contracting sections this didnt work very well because as each section expands and contracts it also gets longer and shorter which pulls or pushes the other sections in and out either side so of all sections doing this getting it to be synchronized efficiently was quite difficult worms have little bristly hairs on them which act a bit like a ratchet to help them go forwards and not slip backwards so ratchet Wheels would probably have helped on my mechanical Worm but this time were going to try and make a more efficient design without putting wheels on it the plan is to keep the concertina Parts but make them slide independently of each other so they dont push and pull the sections either side lets get the parts D printed and assembled and itll be clearer how its going to work we need to to make four sections for the worm so weve got lots of duplicate pieces to make but heres just a few of them theres some levers to make the concertina and various other parts to link the sections together just a quick add from my D printing sponsor thanks to L bot D printers for supporting my Channel with D printers it makes it much easier to have lots of printers that I can print all the parts in parallel and thanks to D fuel for the filament for this project all these parts are printed in Pro pla plus Im going to use these V wheels and vslot aluminium Extrusion to make the sliders for this project normally you need more than one wheel though both ends of my concertina are going to slide up and down and each one has four wheels so that it doesnt skew and jam you can see that weve got the wheels inside there those are supported on bolts and then my aluminium Extrusion is Extrusion and that slides in there just fine so weve got these end stops as well which are also the joints between the sections of the worm and both ends of that can slide the constantina part of this is really similar to last time its kind of trellis thing with just some normal joints that get shorter and fatter as it moves and of course that fits onto the two sliding sections which allows either end to move and the whole thing to slide up and down so unlike the previous worm which was just this concertina piece and it meant when it got shorter and fatter this segment got shorter and it pulled the other segments in as well weve now got these fixed ends that always stay the same length the other thing we can do is pull this up and then slide this piece along here bring this down again bring it up again and slide it along so it should give us a much better Locomotion so we just need to put some Motors on there now Im planning to use servos but lets hope theyre powerful enough to operate those segments Im using kg torque Metal Gear servos Ive used in a lot of projects and they come with these brackets for building robots and those are pretty powerful one of those concertin is the whole thing up and one drives it the whole length of that sliding rail so that should give us every combination of motion that we need to make it locomote Im using an Arduino Mega here and Ive got a . vol s lipo to just power those servos directly so that should all be just fine so Ive just got a very simple set of timers and Im using the servo library on the Arduino to drive those servos to the position so it does go along but the only reason that really works is because it gets overbalanced and thats cuz the servos are moving a full speed to the position that Im telling them to on fixed timers so its kind of jerking around a lot what we really want is a much slower motion and we need them to move over a much longer time scale so that we get a much smoother motion to do that were going to use the Arduino ramp Library which is really useful and that allows us to specify a step input change and it will interpolate between the positions so you can see the blue line is the step input change and the red line is our output if we use that to drive the servo instead of it moving as fast as it can to get there it will slowly change over time and we can specify the new position and the time we want to interplate over I fitted an enable switch on which starts all the timers so now we can see with that ramp Library running driving the servos weve got a lovely smooth motion so you can see the whole thing expanding Contracting on one end driving The Spine down expanding again Contracting again and repeating it but its moving much smoother than it was each action out of the three actions here takes second so its a second cycle interpolating with that library with each being a second move its not getting anywhere though and thats because it needs the other sections to push it along but luckily Ive been making four of these as weve been doing the whole thing so Ive got all four sections to stick together for the Worm but before we see how well that works all together its time for a quick ad from the video sponsor which is PCB way PCB way is a OneStop shop for PCB manufacturing assembly and other types of Manufacturing Services all under the same roof PCB way manufactures all sorts of boards including Standard fiberglass boards but also aluminium pcbs flexible pcbs and rigid Flex pcbs which are part rigid and part flexible PCB way also provides CNC services including online CNC Machining sheet metal fabrication D printing and injection molding the CNC Machining Services include a wide range of materials including aluminium stainless steel and various Plastics if you dont see the material youd like you can also choose some custom materials PCB ways D printing services include SLS SLA DLP fdm and more in a variety of materials check out the PCB way website to browse for a variet of finishes and get a quote check out the PCB way shared project section this is a community of User submitted projects with PCB schematics and parts listings so you can reproduce the projects you see there they also have a module store which has all sorts of items for sale such as Arduino boards toolkits robot parts and kits and sensor modules find out more now at pcb.com when Ill put that link in the description to this video right lets power up this robot worm these have got ball joints on each end which just basically a ball and a cup for it to fit into and Im going to bungee those together so they can move all around I just tied some bungee cord on there looped around a few times and tied a knot so obviously with two sections doing the same thing we dont get very far we do get a little bit of motion but thats just because that spine has some inertia because of the metal and it just pushes it forward as it moves a little bit but really what we need to do is adjust the sink so that one pushes forward and the other one pushes backwards as it goes so lets try two sections which are out of sync one is getting triggered milliseconds thats a second and a half after the other one and as a cycle takes seconds for each of them one is basically out of phase if you like theyre not quite doing what they should but its certainly locomoting along heres all four sections linked together theres still a second and a half behind each other as we go from the front to the back and since each section is on a second cycle that means that the first and third sections are in sync with each other and so are the second and fourth sections this seems to be quite effective because it means that its actually lifting off the ground in the middle weve got two sections pushing the spine forward while the other sections are pushing it backwards and you can actually see those sections lifting clear off the ground because theyre supported by the other sections either side when they get thinner as they get fatter they push back and since some sections are off the ground thats much more friction so they virtually dont slip at all you can see theyre pretty much stuck in place as they push backwards its quite a satisfying motion to watch and these rollers on the the bottom are actually fixed theyre not rolling around at all you can see some of them have got two screws in and they also dont have any grips on theyre just smooth D prints so Im pretty happy with the way this locomotes with virtually no slippage I tried messing around with the timing here so basically now weve decreased the sink to second that means the second section back is second behind the first one the third one is second behind the second one and the last one right on the tail is second behind the third one but this doesnt seem seem to work at all really its virtually getting nowhere so then I Tred decreasing the up to milliseconds but of course were actually getting closer to them all doing the same thing at the same time so that means theyre fighting each other instead of basically helping each other with some pushing back and some being thinner and going forward at that time so this really doesnt work at all and it really doesnt loc a out well so then I thought Id go the other way and I increased it to seconds remember we had a second and a half when it was working really well and yeah this is just as bad basically kind of is going forward but I really think the best kind of thing to do here is have two and two in sync with each other some pushing and some pulling so obviously when theyre fatter theyre pushing the whole thing back and the thinner ones can lift off the ground and go for of that time and that just seems to be the most obvious answer and the best way to sync the whole thing up so Im pretty happy with that obviously the front and back segments are dragging on the ground because theyre not supported by sections either side but they still dont have that much friction because they are getting thinner and of course weve got two sections of that point actually pushing back which are fatter so that doesnt really seem to make any difference theres hardly any load on it and the floors smooth so everything works fine Id really like to make a much bigger version that I can ride on like some of my other weird Vehicles it would be pretty cool to take to events that means building each section about as big as a bicycle though which means there would be quite a lot of friction on the front and back sections as theyre unsupported and try to push forward one option is building a kind of suspension bridge that holds the back and front sections up and that could also be used to make it steer by having cables that pull down the sides so it bends as it goes I have considered that the expanding and Contracting concertina Parts arent actually that efficient so maybe we could just have some sort of foot that moves up and down and slides back and forth or even multiple feet on sliders next to each other in each section thats not quite as much fun to look at though so Ill probably stick with something that looks like peristalsis

write a youtube script

do you remember when William Osmond did all those egg drop challenges with random people where you have to build something to protect an egg and throw it out of a window and see if it breaks and then Mark Rober did an egg drop challenge from space how are we gonna beat that yes its the human egg drop challenge Im challenging kids invent stuff to build something to protect a human which were going to drop off a crane at varying Heights only were not going to use a real human were going to use the crash test dummy that I built some time ago so it doesnt get demonetized on YouTube so Im going to build something to protect the crash test dummy and kids invent stuff are going to build something to protect the crashless dummy that kids have designed Crash Test Dummies cost a million dollars and thats because they dont make very many about a year for the Auto industry so I decided that so I could do more dangerous things in my channel and not put myself in danger which YouTube wouldnt like Id build my own crash test dummy so I made this thing out of wood and steel and various other materials including D printing its got joints that can dislocate in various places and its fully articulated his face is actually made of Lego so you can see if his face would have got broken and it does have data looking as well so Ive got two accelerometers one in the head and one in the body and those log to SD cards so we can look at the graph of the data later its got load cells in Its Ribs which work like weighing scales and its got guts made of stress balls which are full of those water balls or Orbeez which can get slashed and its guts can come out I previously tested it by throwing it off my shed which is the only thing I could really do with it to see how it worked basically the dummy survived the fall okay and then we got the data which clearly showed it going antigravity where it fell and all the accelerometers went antigravity obviously hitting the ground as a massive Spike and then laying on Its Ribs eventually which is why weve got those constant values shortly after the drop shopping with Jimmy B Magical Mystery Tour we get to go to all the best places oh okay oh right the other end right down there its wood shopping with Jimmy B as well as wood Ive also got some D printed Parts a quick ad for my D printing sponsor thanks to lulsbot for supporting my Channel with D printers this is a . millimeter nozzle so its going to make these parts really chunky and really tough and thanks to D fuel for the filament for this project and lots of other projects all these parts are printed in Pro pla plus so they should stand up I printed a lot of those because I didnt really know what I was going to build exactly until I got here but anyway were going to glue those onto the end of the wood so the most of the structure is D printing and wood and then were going to screw those on and Im using Gorilla Glue here which is the foaming type to fill all the gaps and wood screws as well so theyre nice and strong what are you doing with all these toilet plungers weve got a serious blockage roofs house is um its just really quite like quite problematic on the sewage front so um yeah just gonna do uh roof has a mansion with um block toilets and that were going to simultaneously unblock them all at the same time uh yeah and if you want to know what all these toilet plungers are for and check out lots of other crazy inventions designed by kids and built by the kids invent stuff YouTube channel then check out their Channel and Ill put the link in the description and a card in this video my plan for this build was to make some sort of tensegrity structure security is the combination of tension and rigidity and I did a video about this before essentially its a structure of sticks and cords which suspends itself in D space so none of the sticks or chords are touching each other NASA prototyped a robot like this which could Roll Along by changing its geometry they also said that it could probably be dropped on a planet and absorb some of the load so Im hoping that this sort of configuration will make quite a good protective structure for the crash test dummy I thought Id try bungee cords first so heres a bunch of bungee cords I bought off Amazon movies that hold this and then ran we live here now were part of Jimmy Bs assembly Direction when its this big its pretty hard to put together by yourself because its not something you can just support with your extra fingers obviously these bungees arent really tight enough so were gonna have to either use lots of them or think of another idea Im waiting for some new straps to arrive so Ive been helping kids invent stuff with uh putting their Dome together thats made out of bamboo and rubber were going to drop it off a crane its going really well I decided that ratchet straps would be a really good idea for this structure we need chords plus some to support a seat in the middle so I had to wait for three packs of Amazon Basics ratchet straps to arrive so here we are reassembling again but Im trying not to take too long because Ruth and Sean really need to get on with their own build so weve got most of them in apart from two straps hows that feeling solid yeah yeah theres lots of loose tails and none of them are the right length so we need to adjust it but its definitely getting there I think these ratchet straps are a good idea so that should be all of it together so now it should be totally rigid right you can let go or at least itll tip over but it should stay in one piece with the ball just roll it yeah I need to adjust the spacings and then the seats going to fit in the middle there should be . no . meters between all of those and all the straps should be . so youll bolt it onto here and then maybe youll have to touch it Im just going to suspend it in the Middle with more straps and bungees right lets uh tidy it up and see what we can get the ratchet straps of course can be adjusted to any length and they need to be . meters roughly so I just went around measuring them all with a tape measure and evening the whole structure up then I tied up all the ends so its not so messy not sure if well cut those down or I might just leave them tied up but thats looking pretty good its looking like a pretty rigid structure I can pick up in one go and Im pretty hopeful that if I suspend a seat in this and put a person in or a crash test dummy that actually if we drop them a little distance theyll probably be okay seems like quite a tough springy structure that will take quite a lot of the load and I can climb right up on it and theres no problems so you can see that none of the wood is touching any of the other wood all of its suspended in space obviously the ratchet straps me to the corners on the ends of the pieces of wood but those dont touch each other either so what weve got here is a proper tensegrity structure using six sticks and cords I need to be quite careful with the D printed Loops that Ive made to make sure they dont take too much of the load so the strap should go around the wood and that Loop should just be there to stop it slipping down the wood some of them the strap is on the actual Loop there like that one so we need to be quite careful that theyre not getting pulled too much Im using the red seat which I borrowed straight off another creation so if youve watched my channel youll remember the omnidirectional ball wheeled vehicle I bought this seat on eBay it looked like it was probably the right size but when I got it it was massive so I just went with it so were going to use that again I suspended the seat in there the best I can using another six ratchet straps and that brings the total of ratchet straps to so Ive suspended it in the best I can into this three point or six point structure you can pick it up though which is bad because when it drops its gonna just hit the top so I put some bungees in wherever I can holding it down to the bottom three points of the structure so hopefully thatll stop it going into antigravity it looks pretty springy together so Im quite hopeful about this at the moment this is a time capsule from the future but before we find out how that happened its time for a quick ad from the video sponsor which is PV case now PV case is a Next Generation AutoCAD based piece of PV software focused on Automation and accuracy it allows you to simulate the actual location of a solar plant from the earliest stages of planning incorporating D topographical data points so PV case is the ideal choice for companies undertaking large commercial and Industrial projects as well as utility scale plants the software is really intuitive and has streamlined processes to help reduce the learning curve and improve productivity features include everything from the prototyping stage electrical design stringing shading and terrain analysis and automatic generation of construction documentation so PV case really does enable the engineers and designers to take the project all the way through from its initial conception to the procurement phase this really is an endtoend approach which saves time and reduces errors its streamlined so you dont need to switch between tools other software platforms and PV cases recently acquired Anderson optimization the worlds most popular solar sighting software platform other features include slope analysis piling and collision analysis automated topographical D cabling sidebyside design comparison and Rapid D building preparation try PB case for free by following the link in the description to this video right lets get back and find out what happened to him right put the seat in which is sort of suspended in here its quite hard to put a five point harness seat into I dont know three or six Point Integrity but um I think the dummy might get its face smashed in um but I cant put it much lower because its gonna obviously get down and hit the bottom and theres not really space for its legs Id like to turn it a bit more I dont really think I know how to because it doesnt really fit but I think it seems pretty solid I mean something will happen now I can just chill out because Im finished so we shifted everything overnight and now weve reassembled it all down at the test site I swiped some thick Industrial Rubber left over from the kids invent stuff bills and I zip tied it onto the ends of my bits of wood here to make rubber feet the hope is that the wood doesnt split if one of those hits the ground really hard although were testing on grass so I think all together its going to be fine Ive also put some cushions in to make it more comfortable for the dummy Im just checking the fuel level because the last thing I want is for James to go in the air and get stuck this is what youll be using today this is the Genie Lift ss This is like a professional use bitter kit its not its not your gardening sort of stuff this is for like putting roofs on massive industrial buildings and stuff and I just happen to have one yes and you need to use it so here you are right hows it going guys are you nearly ready yeah we just um attaching ARCA mode as you do so youve got toilet plungers and a foldable portable toilet if we just sit on we like we like really wheres the toilet roll no toilet roll holder and are you feeling hopeful oh were feeling hopeful but weve been deflated by you know what do you feel about uh I think it will its got a lot of crumple zones that will absorb the load yeah so obviously its down to wherever the crash test dummy is okay rather than the structure yes I know so well see mine doesnt really Crush easily what weve done is weve designed really it reflects our engineering style but really the true test is going to be whether the mannequin that youve created survives so weve shifted the burden the benefit thats the idea the same as the egg its all on you James these have gotta bewinged in the air now so Im tying some ropes onto the three points that stick out the furthest of the sides and Ive got a carabiner there so I can winch it right up and Ive got the dummy installed in the field where were going to be testing it looking pretty comfortable on the cushions so Im pretty hopeful that hes going to have a nice ride doesnt need to be pulled down at an angle because thats how the thing goes Sean brought one of those quick release things but we couldnt get it to release with his weight on it so after a bit of messing around with different chords to pull it and some WD we finally got a good test out of it right you want it right up up thats it thats it there about there so were doing a one meter drop test just to see what happens about a meter off the ground okay A bit higher just a little bit its about two foot there tiny tiny uh yeah about there thats it foreign looking at the data in your favorite spreadsheet program we can see that weve got some constant g of gravity which is correct for the accelerometers facing upwards and then weve got the ribs had bungees on them holding into the seat which is why theyre constantly measuring something at least on one of them then weve got the antigravity part where we drop it for a short amount a big spike where it hits the ground and then weve got lots of mess but that basically is the bounce of the whole structure for a short time until everything settles back again I felt fairly confident having done a short drop so we took the crane up to meters and I measured the height by just flying the Drone up parallel with it and looking at the altimeter on the display on my handset because we didnt have a tape measure long enough basically foreign heres the data for that drop we can see weve got a longer antigravity period because the drop was longer essentially a big spike where we hit the ground a little bit of Bounce but not so much because the structure actually failed this time and then some constant measurements afterwards and it looks like we only just got to over Gs which means it would have potentially been survivable by a human so to Gs is what a fighter pilot might experience but what was it that broke well it wasnt the D print and it wasnt the straps it was actually that Timber which is CLS construction timber for doing Stud Walling it looks like there might have been a knot in one of those pieces at least and that was the weak spot which is where it broke but it looks like those D prints every single one of them is actually still intact its a bit like when I made that Caster wheel bike and the steel broke but the D printed steering columns stayed in one place so it does show that making those parts with a . millimeter nozzle makes them incredibly tough so thats pretty impressive really and it looks like the dummies mostly in one piece apart from he did lose one hand but they are any artist hands for posing and drawing pictures of hands so its not surprising really but yep his Lego face is in one piece and everything else is perfectly fine no stab wounds and all of his legs are still in one piece as well uh if you didnt think it was very high its quite high its very high hes lost a hand from the falls and Jimmy Bs but hes still good to go here we go now squeeze them in and strap them into the commode so next it was time for kids invent stuff to winch up their toilet plunger and bamboo invention and drop it so they decided to do just one drop to see what happened but if you want to see what happens with that youll have to watch it in the kids and Ben stuff channel so go over there check out their version of the build what eventually happens to their creation and all of their other projects as well alright thats all for now

write a youtube script

I was recently looking at the new Legend of Zelda tears of the Kingdom game it has loads of gadgets which you can use to complete the game including a cart sled fan a spring and shock and beam emitters theres also a stabilizer device which basically makes things go upright so Vehicles can be made more stable and you can also use it to make a catapult I started to wonder if this would work in real life and also remembered that I needed a way to make walking robots more stable like my tripod robot so it doesnt have to be as Dynamic and has more time to take steps I previously experimented with some small stabilizers which use gyroscopes to make wheeled robots balance so that seems like a good place to start a spinning Mass produces something called gyroscopic procession weve all seen spinning tops standing on their ends however if you start locking any of the axes in place then they wont stand up anymore so we need to be a bit more clever if we want to use a gyroscope to stabilize a robot and have control over it if we move one axis of a gyroscope then the procession causes a force in a perpendicular axis this is because the spinning force on the control axis Force combine resulting in a false degrees around the gyroscope in the direction its spinning when a spinning top is balancing its constantly processing constantly falling in One Direction but the resulting output force is degrees around in the direction its spinning and as a result that makes another false degrees around again so we see this motion which makes it constantly wobble as it constantly Falls however if we constrain one control axis we can use the gyroscope to exert a reliable force in a a perpendicular axis as we wish in my previous builds I measured the roll angle of the device with a solid state inertial measurement unit and built a controller from an Arduino to control the pitch of the gyroscope accordingly with a Servo this exerts a force back in the roll axis of the device to make it balance I moved on from that using two Gyros to make a twowheel inline balancing robots and also a onewheel balancing robot which balanced forwards and backwards like a Segway but balanced side to side using the Gyros note this is different from a Reaction Wheel which spins a mass in either direction around a stationary Center Point to make an opposing reaction force my gyro units were pretty small though and the Gyros didnt exert a lot of force so they only just about worked the gyros in these were D printed with some ball bearings in internal Pockets to give them Mass I want to make a much heavier Duty unit for some big walking robots so this time Im using car disc brakes each disc is about three and a half kilograms which is about times more mass than the original D printed Gyros Scopes were going to use a belt drive to drive these discs around so Im printing some pulleys and Im also printing some hubs to mount onto those discs so that weve got somewhere to put bearings theres lots of parts to make here weve got to make quite a heavy duty unit to hold these and were actually going to make a pair of them just a quick ad from my D printing sponsor thanks to lols bar for supporting my Channel with D printers and thanks to D fuel for the filament for this project all these parts are printed in Pro pla Plus the disc brakes have five big holes two holes which are tapped to M and two countersunk holes Ive made these hubs which have bearings mounted in them so we can spin the whole thing around smoothly and those also have the similar keys in them and two holes so that we can bolt them onto the discs Im just using some M studding in the disc there and of course one Hub fits either side and we can bolt that up with some lock nuts so those are my Hubs fitted with a bearing either side weve got some bright steel bar which tends to be quite well tolerant so that fits well into the middle of the two bearings and thats our axle that were going to spin everything on if we spin that by hand then I can actually already feel the gyroscopic recession if I try and move it in one axis I can feel the perpendicular Force so that seems to be working quite well I think were going to exert a lot of force I fitted a pulley onto my Hubs so we can actually drive this with a motor and Im going to be using the good old brushless KV Motors that you can get from HobbyKing the frame to build all the parts that were gonna need is gonna be made out of Extrusion Im using dropin tnuts as usual so I can attach D prints and various brackets and those just turn in and do up tight so we can attach those to the other pieces Ive also got some metal corner brackets to make sure I get my degree angles right at the bottom Ive got the degree metal brackets and Ive also got D printed brackets across all of the corners Ive only got those at one end because of the other end of the whole assembly I need to put the motor and the disc mounting brackets and that of course looks like a couple of clamps which are going to hold the millimeter bright steel bar and hold that disc and I can position those up and down to tension the belt drive so as youd imagine the disc fits in there and that spins around nicely and I remembered to put the belt on as well so that I can spin that with the motor and Id have to take it to pieces to get the belt on at the bottom theres a motor and again thats on brackets so I can move it up and down and get that belt tension and Ive just D printed the pulley on here Im using a vesque ESC or at least a clone of the vesque from flip sky and just an RC controller and an RC receiver and Ive got s lipo so around volts which should make it go really fast but before we power that up its time for a quick ad from the video sponsor which is Opera one lots of things are browserbased these days arent they so I use a browser a lot and you probably do too Opera one is a browser but its also more than that I really love how it puts everything in one place and that makes my working life creating content mostly more streamlined obviously I produce content online but its not just on YouTube Opera one integrates my other social media including Twitter and Instagram and also other apps like WhatsApp Messenger or telegram if you use them right there on the sidebar and theres also a great feature called tab Islands which you can use to group together similar browser tabs to make your browser organization much easier Opera one has an integrated ad blocker and VPN and those are right there built into the browser and theres also integrated AI called Aria Operas ais integrated seamlessly into the browser and thats completely free and unlimited you can ask Aria questions have it explain explore further uneven translate text for you and I found itll also write code which of course I find quite useful Opera runs smoothly as well the webs really active and animated these days operon handles all of these features with a separate compositor thread that paints the active elements in the browser rather than using the UI thread so animations run smoothly all the time and that makes a much better browsing experience so check it out and install Opera one today and the links in the description to this video right lets power up a massive gyro Music so yep thats a little bit terrifying there is a bit of off center wobble in it which makes it walk a little bit but its actually not too bad for something Ive just cobbled together so Im pretty happy with that so if I pick it up and move it in one axis which in this case is actually twisting it in its your axis which is the vertical axis so basically Im just rotating it between my two hands we can see we do get a perpendicular force that causes the bottom to swing out and the faster I move it the more Force we get so if I try and turn it as fast as I can like a steering wheel and basically we can see that bottom kicking out quite a lot and doing it much quicker as well this exerts quite a lot of force its really hard to even hold it upright when Im turning it so this is going to be pretty good so that does exactly what we wanted to so as I turn it this way its making a force of this way the motor helps as well because thats quite heavy and the Bell Drive means the motors turning in the same direction if it was a gear reduction it would be turning in the opposite direction and that would actually cancel out the gyroscopic effect with the mass of the motor and its own gyroscopic effect so thats pretty good but in order to get rid of gyroscopic recession this wobble that we see in Gyros standing still or spinning tops and basically cancel that out were going to use two of these units so yes Ive built another one exactly the same the mass is building up on these quite a bit now weve probably got kilograms here so those are going to spin in opposite directions and then theyre going to move in their control axis in the other direction so now we need to mount these on something and have a motor thats going to control the actual active control axis so weve got a base which is D prints and in the D prints weve got some thrust bearings which take force from the top and weve also got a normal bearing in the middle to hold that in the middle and stop it slipping out the top of thats just got a peg goes in the bearing in the middle and the other half of the thrust bearing so if I put that on it takes quite a lot of load on the top there which is what we need so of course those gyro units fit on there and those get bolted onto the the same as the other parts and the tops again have pivot Points on which will have a top on and of course all thats fitted together with dropping tnuts and screws on top of that we need to keep them in sync and we need to move them in opposite directions so weve got these gear track pieces which fit on top there and youll notice one is much longer than the other on top of that weve got another piece that looks like the bottom but without the thrust bearings because its not holding any Force were only really holding the force at the bottom and all together it looks like this which is a top and bottom held on with some Extrusion so its all nice and rigid and those two Gyros move in opposite directions they will spin in opposite directions and the control axis is in opposite directions one track is longer and that is so we can put the motor on to actually control it which is quite a powerful volt motor which has the other spur gear on and that is going to fit onto the frame with a very simple bracket so that I can position it and get that gear mesh tension correct so hopefully thats fast enough and its responsive enough to make this whole thing balanced but that remains to be seen we do need to know what position this is at though so I left one of those stubs a little bit longer and theres a grub screw in there which a pot can fit into and that black bracket is going to be glued on and that means I can get feedback to turn that motor into a Servo which is currently controlled with a BTS motor driver and an Arduino Mega so Ive got another pot which I can use to drive that so now when I turn it we can turn that control axis and we can set the position using the feedback pot and that works pretty well so its just like a big Servo which is turning that control axis and we use an actual Servo in the previous versions I now have two escs one each side one for each of those brushless motors and I put this thing on a pivot point so we can see if we can actually make it exert some Force so sitting it there spinning up the Gyros until we go full speed and then turning that pot we can move the control axis and we can see immediately that it exerts quite a lot of force so that clearly can exert a fall sideways when those Gyros move in their control axis so thats basically working but in order to make this balance and make it balance other things like robots we need to read the angle its at and then go make the gyro control axis respond accordingly so I had an mpu inertial measurement unit which has a little Arduino attached to it and then it reads that data from it and sends the data over serial to the mega so you can see heres the trace of the angle as I move it in either direction we get a positive and negative swing to make it balanced Im going to be using a PID controller and theres an Arduino library to make it easier the balancing set point is the angle we wanted to balance at which in this case is zero the PID controller then gives us an output which we give to the control axis to drive the position the PID controller has three terms proportional derivative and integral proportional will give us a bigger output value as the input angle increases which in this case drives the output control axis to a larger angle but this isnt sufficient on its own to make it balanced the integral term is the most important term for balancing devices and Ive previously featured a video from Jack Monaco in which he builds a mechanical integrator and explains what integration is ultimately the integral term is the area under the graph the output will therefore continue to increase even if the input is constant because time is progressing so the area gets bigger so the integral term makes the output accelerate towards its Target as the integrator tries to fix the error between input and set point over time this can also cause overshoot though so we need some derivative gain and that is the tangent of the curve of the graph so this tries to fix shortterm errors and makes the output respond more sharply since the gyroscopes only exert a force when the control axis moves and a larger Force the faster it moves we should probably be driving the control axis velocity from the PID output controller rather than the position in this case though were driving the position and the control axis motor velocity does actually increase the bigger the error between the set point angle and where it actually is its hacky but it works and it also means that the gyro control axis is always sent when the device is upright and it never creeps like it would do if we just control its velocity so Im pretty happy that that can just about balance but what can we put it on thats a bit more fun yes its a scooter so I fixed that to the handlebars with some more brackets and its just sat on the kick plate there and that seems to be balancing all right I put the battery off center there to help balance up the other side where the control axis motor is so balance is pretty much in the middle by itself but obviously it needs the Gyros to stay upright now I can slide this along and you can see its clearly bouncing on two inline Wheels if Id actually motorized the back wheel and had a Servo on the steering I could probably Drive the whole thing along radio control there are some oscillations in this and it will need retuning based on whatever I put it on wherever that pivot point is well need to retune that PID controller but if I leave it alone for long enough it seems to settle but if I give it a sharp and mess about with it you can see those gyros in the control axis compensating and it kind of uncannily pushing itself upright again so yep seems to work all right it can exert quite a lot of force so Im pretty confident we could use this on a walking robot so its not quite as good as the original which can pick itself up from lying down and seems to work in all axes which would require another pair of Gyros but it does look like its physically possible so obviously were going to need a much bigger tripod robot than this to carry something like this but the plan would be that it would be much much bigger then wed have the gyro unit rotating around on the top so it would rotate around over two legs and then it could pick up the third leg and balance using the joy Rose and then it would rotate around again over another two legs and pick up a different third leg and thats how it walk along but thats a whole other project

write a youtube script

day of the tripods was originally in H.G Wells novel written in around . there have been many adaptations since including a film in and another remake featuring Tom Cruise in and many other TV series and other media generally the consistent theme across most but not all adaptations is that the aliens are tripods H.G Wells was an English author so if you ever visit woking in Surrey you can see a lifesize tripod in the Town Center if you like Alien Invasion movies then you should also check out War of the Worlds which is the one where plants come alive and eat people but can tripods really walk they only have three legs which means that they have to stand on two legs while taking a step with the other leg in the movie they appear to have legs which look a bit like tentacles but we dont see how the ends touch the ground that much I found various other tripod robot projects on YouTube but none of them really take proper steps even the ones with big flat feeds so could we really make a tripod robot that could take actual steps and walk along maybe it could walk by pumping itself up and down and taking a step with one foot while its inertia carries it upwards well maybe it will be able to lean over onto two legs and take a step before it quickly falls back again Music in any case Im going to try and make my tripod walk on points rather than flat feet with ankle actuators mostly because I suspect that this will make calculating the inverse kinematics easier because I dont have to deal with the ankle joint so its time to D print all of the parts a quick ad for my D printing sponsor thanks to lull spot D printers for supporting my Channel with D printers it makes it much easier to get these projects done every couple of weeks when Ive got lots of printers all working at once to make the parts in parallel and thanks to D fuel for the filament for this project and lots of other projects all these parts printed in Pro pla plus Im using some of these kilogram torque servos for these which have a metal bracket on both sides to attach them and of course theres three legs so everything has to fit in a triangle formation with degrees between the three legs essentially so weve got three servos and those are all fitted together with these D prints that just plug into the bracket there so with three of those at degrees apart were ready to attach some more things to them there is a plate thats going to go top and bottom to hold those prints together though foreign fitted and the three brackets fitted to the output of the servo we can put the next stage on which is the smaller bracket to hold the next Servo and one of those fits onto each of those and these are just screwed through so I can access the screw holes into a block on the inside otherwise its impossible to put the horn on and of course fitted on those are three more servos which are basically going to be the shoulder joints of each of the legs fitted onto those of the upper leg which of course gives me the shoulder motion and the leg swing and fitted onto the next brackets are of course the knee Motors so there weve got a fairly basic leg there which can move in three axes at the shoulder knee and swing around and all of those are degrees apart I put little rubber feet on made of TPU so its nice and grippy and thats pretty much the whole structure we need something to control it though which fits onto the top of those original blocks and thats going to be an Arduino Mega Ive got an RC receiver there so I can control it from my universal remote and theres lots of wires there to break out for all of the servos with power distribution theres a kind of cage which makes up the robots body there and as well as protecting the electronics if it falls over it means I can stick the whole thing upside down to calibrate and test the legs the structure for this isnt that straightforward though mainly because weve got these three legs that all face out at degrees for each other so in order to get them to take steps in the same direction we need to work out some inverse kinematics that takes that angle into account so that means being able to position the feet in a known coordinate system preferably in straight lines in Cartesian coordinates and bitter will have some maths basically running in the code on the Arduino that works out the joint angles for all of these including the leg swing to be able to get the foot to the required position I recently built an attack which could walk along the main difference there though is that all the legs facing the same direction so once I solve the inverse kinematics for one leg they could all follow the same motion and make a walking pattern called a gate but I still have a shoulder and knee joint on my tripod robot so lets start with that but before we find out how the mass for that works its time for a quick ad from the video sponsor which is fan home fan homes mission is to inspire fans with high quality build up models and collections from your favorite Brands like Marvel and Star Wars all the products are original designs you wont find elsewhere each shipment includes fully Illustrated magazines full of inspiring content and Pat with information like rare behind the scenes images just choose your collection or build up model and every month youll receive exclusive products along with their magazine build the iconic Optimus Prime replica standing nearly centimeters tall and composed of meticulously detailed metal and ABS plastic parts Optimus Prime will be the centerpiece of a collection with features such as light up effects and full articulation for maximum posability Builders will get to know Optimus Prime and the world of Transformers throughout the subscription with a special full color monthly magazine learn about his Origins his home planet Cybertron and the interstellar struggle between the autobots and the Decepticons subscribers will also receive the Transformers backpack Optimus Prime tshirt Transformers keyring and mug and Transformers posters click the link in the description to start your collection today firstly Im going to specify the leg length and have the code work out what the knee and shoulder angle should be thats easy because I know the upper and lower leg lengths and I know the other side of the triangle because thats the overall leg length I want to specify this is just High School mathematics so I went on mathisfun.com and you can find out how to work it out using the law of cosines there Ive tied one of my control sticks to the leg length input so now if I turn the stick the feet should move up and down in a straight line as I scroll through all the positions next we can solve the next axis which is to make the feet move in and out as weve already solved the inverse kinematics for the leg length we can treat the leg as a straight line and then we just solve the next axis as a right angle triangle passing the hypotenuse of the triangle to the previous bit of code which controls the leg length this is just some simple trigonometry because its a right angle triangle so now I can use another controller input to move all the feet in and out in a straight line thats not much good for walking though because what we actually want is all those legs to move in a straight linear axis so we can take steps in the same direction action rather than them all moving in and out from the center degrees which is what they do now the lake facing upwards in the diagram already moves in a straight line in the direction we want to take steps but the other legs do not we still have the swing axes though but in order to get the angle we need to do another calculation we can treat the other two legs as right angle triangles where we specify the two sides in Cartesian coordinates to place the foots and then work out the swing angle using trigonometry the hypotenuse of the triangle is again passed on to the previous code which specifies how far the foot Moves In and Out so now all my legs can move in a straight line using that swing axis on two of them to flip the leg around when its needed and of course I can still mix in the zaxis which Now controls the height of the robot from the ground while keeping the other axes consistently placed and yes of course if I turn it the other way up and put it on its feet as it should be then yes it can shift sideways and thats fine and it can do it with the legs at multiple Heights as well which means we can pick those feet up and move them in a linear axis to take steps I should of course be able to move in the other axis to and from the camera but I havent done the math for that so so far were stuck with this the next stage would normally be some sort of interpolation which Scrolls through all the foot positions from A to B for you this is usually needed because if we make a step change in foot position then various joints of the leg will need to move at different rotational distances and if we use the default survey velocities then that means some of the joints will get there before others and our foot wont move in a straight line anymore you can check out more about that in the atat build and also any of my other robot dog projects but in this build I need the servos to move as fast as they possibly go so for now Im not bothering with interpolation you can see the effect of this if I move the legs up and down on time as though the feet appear to flick out as one joint catches up later than the other but when the feet are on the ground it looks fine if I suddenly move the robot in the direction we want to take steps then you can see it almost over balances and one foot lifts off the ground so it seems quite hopeful theyll be able to take steps just moving in a sharp motion like this I then made it actually take a step by raising the foot up in the zaxis and putting it down and that seems pretty hopeful so I thought I should probably move on and solve the other axis of inverse kinematics as well as being able to move the legs forward and backwards we can now move them all sideways in a straight line as well just by solving one more triangle and with his feet on the floor we can now move it forwards and backwards and sideways and of course we can move the legs independently up and down in a straight line sideways or front and back so that means we should be able to put a gate together to have it take steps blue vaccines mixed together so I can move in any direction with an infinite number of angles so I finished programming the gate so it leans in the direction of two legs and puts all his Mass over there and then when it almost over balances it takes a step with the third leg which is almost lifting off the ground anyway one of the things that does happen is theres a bit of drag on one side because as its moved two feet forward it hasnt moved the other one forward its not symmetrical anymore and doing it like this means that one leg seems to drag or bounce and that causes it to turn in a circle all the time so if we speed this up to now you can see that in fact it is walking it is taking its feet off the ground quite convincingly but it is going around in a circle the white battery pack at the top is powering the electronics but thats the back of the robot and its walking in the opposite direction to that but youll see that its constantly turning to the left and walking in an anticlockwise Circle back to normal speed and we can see that it is definitely taking its feet off the ground and taking steps so Im pretty happy that it actually works even though it doesnt have big flat feet supporting it it isnt dynamically stable though because it has no sense of balance so theres no inertial measurement unit in here telling it how to balance so the limbs can react to over balancing and thats why theres quite a big delay in between the steps because the whole motion has to settle before it can take the next step if it didnt inertia would build up as it wobbled side to side and eventually it would tip over so for now thats As Good As It Gets but how could we make it work better well do you remember those balancing robots I built there are quite a few of them in my channel some of them are just two wheel balancing robots but I also built a number of devices that use mechanisms to balance on a single weird or Edge reaction Wheels spin a large mass in either direction around a stationary Center Point to cause a reaction force and push the robot the other way if we measure the angle and dynamically control the Reaction Wheel we can make a device that balances I also built a number of projects with physical spinning gyroscopes a gyroscope will cause a force in a direction perpendicular to the angle you move it in so if we put a Servo on this axis and tune its response to the measured angle of the robot then we can make it balance on a single edge either of these methods could be used to balance a tripod robot on two legs while it takes steps so thats probably what Ill try next theres already a project on YouTube you can check out called gyrubot which basically is a twolegged walking robot which uses four gyroscopes to balance using gyroscopic precession the legs are fixed to walk in a fixed gate pattern and then its got these gyroscopes you can see moving around on top in reaction to it falling over which youre pushing it back the other way so here we can see it walking over a pile of stuff and it works perfectly well so I guess my version would be like this but with three legs instead if I was going to try that what I probably do is have a pair of gyroscopes that rotate around on the top here so they can rotate around above two legs that are going to stay on the ground and that would let one leg pick up and it would also counterbalance it a bit by putting the mass over that side and I think that should work quite well I probably want a bigger structure with bigger leg Motors something a bit bigger like one of the open dog projects so thats going to be another project but for now Im going to put all this on GitHub its all open source all the cadencoders I usually do so if youd like to support me through patreon or YouTube channel membership those links are in the description below and patrons and YouTube channel members can get access to all the videos up to a week early and sneak peeks and pages of whats coming up and Discord benefits

write a youtube script

have you seen these drone shows of up to a thousand drones draw pictures in the sky the drones used here look like they cost around two thousand dollars each during one drone show in Australia drones fell from the display into the river that was only of the drones though putting the total cost of drones in the display at around million dollars these drone displays use a super accurate satellite positioning system called gnss rtk which stands for realtime kinematics and the receivers are available for just a few hundred dollars each then all you need is a radio link to communicate with each drone so you can tell it where to go and some software to coordinate them but what about if we want to fly drones autonomously indoors there are lots of demos on YouTube of multiple drones flying around performing tasks these systems tend to use a motion capture system like those manufactured by Vicon which have multiple cameras and ir illuminators and some software that solves the skeleton of an actor based on markers they wear a system like this has been used frequently by stuff made here for tracking objects in motion these systems tend to cost around upwards though but whats a cheaper solution I found a system from Marvel mined robotics which uses Ultrasonics to work out the position of a mobile robot there are four base stations each with what looks like four ultrasonic transmitters on and then you can use a number of receivers which can work out where they are relative to the transmitters all of the units are linked together with a radio link presumably for sync the demos look good although I havent tested them personally apparently they work on drones okay and the basic set costs nearly euros but what could I make that costs a lot less well youve probably seen these cheap ultrasonic transceivers which have an ultrasonic transmitter and receiver some pins on and some Electronics on the back and these cost about three pounds each Ive linked one to an Arduino here so we can see what they do and basically theyll send an ultrasonic paying look for the reflection and work out the time of flight so you can work out the distance these work pretty reliably and I think the maximum range is about four meters so obviously it only works with the item in front theres about a three field of view so if I pick the Box up or move it too far to the side then we just measure the maximum distance to the next object in the background so yeah you can see me moving to the side there and you can see that big sort of Spike we get where we just miss the edge of the object and we just get the wall behind me if I wanted to take something much further away Im probably going to need a bigger object to make sure its reliable like this big board so I can actually get up to the four meter maximum distance Im only doing about two meters here you can still see theres some noise on the trace that were getting there on the Arduino plotter which is basically just measuring the distance which is given to you by the unit so theres not too much coding here so we could use these ultrasonic sensors to work out the distance to a solid object and we could position a robot or a drone provided there were like solid walls in the space that were uninterrupted so it could always get a reading so what we really want though is something more like a GPS system where weve got satellites in the sky and then weve got a receiver on the robot or the Drone or the car or whatever it is and it works anywhere without having to have that sort of solid surface to actually send and reflect the ultrasonic ping from so itd be better if we had one of these transmitting that was a beacon and then another one receiving that could move around and that could receive the Ping and the two could talk to each other to work out the distance the problem with using those modules like that though is that they deal with sending and receiving the ultrasonic signal themselves using the electronics on the back of the board you send a microsecond pulse which makes the unit emit eight cycles of ultrasonic bursts then the echo pin gives you back a pulse depending on how far the sound takes to reflect back the unit also wont listen until its been triggered so if you want to send with one unit and receive one another then they both need to be triggered at once unless both units can be wired together then that sync system needs to be Wireless so to solve that Im using this massive infrared Illuminator for CCTV with lots of infrared LEDs on and Im going to use that as a beacon and receive the infrared with this little module its actually a proximity module but Ive taken the LED off so it only receives and doesnt see itself so now Ive got a mosfet attached to the Arduino and you can see its light on there flashing on and off as Im pulsing it and thats going to switch the current to switch on the infrared Illuminator if you use an old camcorder without a very good IR filter you can actually see the LEDs flashing otherwise theyre impossible to see with the human eye so theres my module receiving and if I take the infrared Beacon away you can see the light stops flashing and if I put it back it works again so lets see what sort of range we get out of this Ive got the receiver in my hand here Im just going to walk backwards to the other end of the room into the dark and we can see that works at quite some distance and it can still see that infrared Beacon because of that massive amount of infrared light and I can turn that thing to about degrees so pretty much horizontally in either direction before it stops working because infrared lights reflecting off everything so now the setup looks like an infrared Beacon alongside the ultrasonic which is transmitting and those are both getting triggered at the same time on the other side Ive got the infrared receiver which is going to receive that sync pulse and Trigger the receiving ultrasonic so it goes into listening mode and Ive taken off the transmitter from that so it doesnt hear itself so now we should be able to transmit with one and receive on the other so now we get a ranger which works just as well as it did with one module reflecting off a flat surface only we dont need the flat surface anymore weve just got that one tiny Beacon and were not reliant on there being a large flat wall there so I can move around quite a bit and we still get Fairly reliable results and I think Ive probably gone bigger than the degree field of view that we would normally get with one module so pretty happy with that so far but how far will it work so Ive put that on the edge of the table and I can walk all the way back here to about four to five meters which is the maximum range for these devices and we still get Fairly reliable results and I can move sideways as well a fair amount and it still works now theres a big Square Spike there I think that was the infrared cutting out where it couldnt see the infrared Beacon anymore so obviously everything stops working but otherwise thats pretty impressive we get quite good results and were not reliant on a large object to reflect the sound off just that tiny module there sat on some breadboard acting as a beacon well that worked better than expected I was thinking about these Ultrasonics though and basically theyre not supposed to actually see themselves like this theyre supposed to see the reflection of the sound that they send so that might explain why its not totally reliable and theres still a few dropouts so Im just going to try something which is basically turning around the transmitter and bouncing it off another surface and then I thought about that surface and decided we can probably extend the field of view if we make that a specific shape Music yep its a dome so its basically an inverted satellite dish which reflects the ultrasonic off a dome to try and spread it some more Ive got the infrared Illuminator on the front along with its mosfet at the back and the ultrasonic module at the back and then the ultrasonic transmitter is actually stuck in the front here so it reflects back off the Dome surface so lets do exactly the same thing with that now youll notice theres two lines on my plot the blue line is the raw data the green one is a cubic spline interpolation that Smooths out all the problems I actually got chat GPT to write the code for this and Ill publish it as usual there is some infrared drop out so if I go to certain regions the infrared receiver just there cant see the infrared Beacon anymore so again we get some rubbish data but the green line averages it out so that wont be catastrophic I get quite a wide range of motions side to side here its not worse than it was anyway Im not sure if it is better its really hard to tell and everything works as long as the infrared receiver can still see the beacon and if it doesnt we get those massive Square spikes but otherwise the green line is pretty smooth but before we carry on with that its time for a quick ad from the video sponsor which is Paypal honey honey is a number one shopping tool in America its just a little button that sits at the top of your browser its an online saving sidekick honey automatically searches for promo codes so you dont have to when youre at the checkout for whatever youre buying honey tells you if there are coupons available for the purchase it works out what gives you the best savings and applies them to your cards I use honey all the time to save money on batteries tools and other stuff I buy for projects it works for things youre already buying on lots of sites where youre already shopping many people dont understand how easy it is its fun to use honey because its exciting to see how much you can save on something you are planning to buy anyway when theres a coupon honey finds discounts of on average and honey has over a hundred thousand fivestar Google reviews you can add it to your browser for free at joinhoney.com jamesbrussen but make sure you use my special link with Slash jamesbroughton on the end because thats how you can support the channel right lets get back to this positioning system Ive built two of those one of them is over in this corner and the other ones over in this corner each one has is infrared beacon on and of course each one is an ultrasonic transmitter and I also have another infrared beacon say in the middle here so we cant possibly miss the infrared my receiver is now two ultrasonic receivers both with the transmitter removed so they dont hear themselves and an Arduino Mega in the box and Ive angled those roughly so they point at the two transmitters and then to keep those Ultrasonics in sync were using the infrared when it turns on and off so on the rising Edge when all those infrared beacons turn on it triggers the ultrasonic on this side and the receiver on this side when it receives that infrared pulse and when the infrared turns off then it triggers the other side of the ultrasonic and the same on the receiver and then we can differentiate between the signal coming from each side so Im just showing the interpolated filtered data on the plot here basically for both of those sensors if I move in and out then of course they move up and down together as we get closer and further and if I move side to side they diverge from each other as we get closer to one transmitter and further away from the other so on the whole that seems pretty reliable with those three infrared beacons we dont get any infrared drop out now so thats working pretty well but waving my hands in the air is not very exciting so what were going to try and do is fly a drone and use that ultrasonic positioning system to hold it in place and see how that works for us just a quick ad from my D printing sponsor thanks to LOL spot for supporting my Channel with D printers and thanks to D fuel for the filament for this project all these parts are printed in Pro pla plus which will hopefully be strong enough so weve got a basic drone structure here with four arms and a body in the middle so theres two more pieces here which fits onto the two corners there which are the arms to attach the motors to and weve got the cheapest Motors off Amazon which are the KV motors which are pretty cheap this is a bit of a funny size its roughly a frame so Ive made it pretty small to fly indoors in contrast to the cheap Motors Im using a pixel C from hollybro which was about pounds including the power distribution board its overkill for this project but Im planning some more things in the future and my transmitter for now is the good old Fly Sky set the lower shelf of my drone supports the power distribution board and on top of that theres another shelf which holds the Pix Hawk and yes youve guessed it on top of that is my ultrasonic receiver system theres quite a lot of stuff attached to this drone so hopefully it will fly Im using six inch diameter rotors with three props or whatever you want to call those on so we get more thrust well it does fly I can fly manually its not too hard to hold in position although theres quite a lot of turbulence in this small room with those rotors going crackers it does have altitude hole but Im not using it so Im trying to hold that manually as well and I can still sort of fly its a bit chaotic though so hopefully the positioning system will be able to hold it in position I think Ill put it down for now before I crash I thought I should test the positioning system here with the rotors actually running to check it still works and thats because basically theyre quite loud and were trying to listen for Ultrasonic sounds also sound is just basically a vibration in the air and weve got this massive downdraft which is probably going to blow the air away but yeah you can see as I power up the rotors there on the Drone that our data becomes completely rubbish and it just Spikes all over the place if I power off the rotors again then yeah we get lovely smooth beta again so this is a massive problem now the pixel does actually support an official ultrasonic module which connects straight to it and it says that basically it can cut through the rotor downdraft to avoid obstacles similarly the Marvel mine system supports drones although the Drone pictured here is much smaller with much less downdraft and perhaps both of these ultrasonic systems are tuned to specific frequency that arent interfered with by the downdraft from the rotors so thats probably not going to work for a drone very easily I do have another solution which Im going to try in the future which is this which you might recognize if youve watched some of my older videos and I already own it and its Optical and its going to be pretty good so Im going to come back to drones but for now were going to try another solution for the Ultrasonics I took the whole setup down to my workshop where theres a lot more space for testing and we dont have slopy ceilings and things like that so yes were going to be putting this on the omnidirectional robot and this is something I built in my channel a while ago its got ball wheels and it can move in any direction and that means we can move around keeping that ultrasonic facing forward which is quite important now I can steer the robot as well if I want to but for now were just going to drive backwards and forwards and left and right and you can check the original build out in my channel thats my little LED flashing so it looks like the infrared is working and thats quite a big space so it looks like weve got range for that at least as before weve got the two ultrasonic transmitters and the three infrared transmitters all in a line and those are around three meters from where the robot is currently standing and theres a bit of space behind it Ive laid down two bits of wood on the ground either side here just by the two ultrasonic transmitters and these are roughly what seems to be the usable area and again those are about three meters apart so weve basically got a three by three of usable space so Im just manually controlling the robot and looking at the data the green line is the left hand ultrasonic and the yellow line is the right h

write a youtube script

hi Im Sam and Im Ella and this is using meat instead of Motors how big is your bicep oh its pretty big thats meat what muscle is meat thats right and mussels are activated by electrical impulses from the nervous system so maybe we can build a robot that uses meat instead of motors yeah well have sausages honey roasted yeah that looks alright so we went on a field trip to waitrose which is a supermarket in England and we got some Meats what have we got in the bag a bit of ham some ham honey roast sausages sausages and some chickens so mainly Meats right so what were going to do is take this battery thats a drone lipo and were going to put Power on sausages lets see what happens so I dont know how Im gonna oh thats very soft its kind of attached its just out put this on here and then Im going to shuffle this in this battery and well see if it moves nothing Im not doing anything lets just try the other one this is coming off in clothes why is this meat so weird just just try again I think its making a sound but its not really moving I dont think we can use this as a motor whats that thats vegan thats not meat yeah thats not me thats not meat its not meat thats not meat not meat thats not meat thats not meat its vegan thats not me thats not meat thats not meat thats not meat its not meat its not meat not meat right are you ready for ham I think Im ready this is this is Hams turn lets uh try and connect so its the clips stick a bit better the ham maybe this will work right there we go come on no nothing no ham why this is meat yeah oh hang on its because its cooked is it cooked meat oh this is cooked so thats why maybe that doesnt do anything because its cooked and its not a muscle anymore its chickens turn weve got the chicken wired up the same as before this ones raw meat and its definitely meat so lets see what happens lets go ah ah amazing I cant believe it works so it took a few seconds for the muscle to contract again but lets just try it again and see if it still works or whats happened there it goes amazing but before we find out more about that its time for a quick ad from the video sponsor which is brilliant understanding the Big Ideas behind data and computer science is essential to building a successful career in Tech brilliant has thousands of lessons from foundational and advanced maths to AI data science neural networks and more with new lessons added monthly the math courses are great if you want to learn math fundamentals or more advanced topics they start with the graphical example so you can get to grips with the basics before more advanced learning brilliant is an essential tool for professionals Brilliance visual HandsOn approach is such an effective and engaging way to master the key Concepts behind todays technology which is critical to staying ahead its built for busy people the lessons are bite size so they break down important Concepts into understandable Parts brilliant is the most effective way to learn to really learn anything youve got to go through it and do it just like the way I learned to do science its like having a personal learning coach helpful explanations along the way never leave you guessing whats being discussed in the lessons and Brilliant helps you build analytical skills with brilliant you learn how to think to try everything brilliant has to offer for free for a full days visit brilliant.org James Brewton or click the link in the description the first of you will get of brilliants annual premium subscription right lets find out what we can do with this piece of chicken have you ever really looked inside the body of a chicken probably not well we have these diagrams here today to show you something you probably dont know which is the chickens are an incredibly powerful bird their muscles are the strongest of any modern bird and that is because they are the closest living relative descendant from the TRex now how can we use this power to help us in our modern lives so Ive made this amazing test rig which has got this joint on the top thats like an arm and then weve got two muscles so one of these is like a bicep so if I put Power on the red and green then we should find well that one shrinks up and the other one stretches so thats like your bicep but what about your tricep and then if I go for the black and blue wires then we should get the other muscle pulling it the other way there we go just like a human hooray it works you really can use meat instead of a motor to build a robot I knew it oh I should wash my hands now yeah you dont want to get food poisoning it doesnt press on mean fish no its chicken not fish but I dont want to get salmonella I just said its chicken not salmon what are we gonna do with this so I was thinking we could reassemble a chicken made out of chicken why would you do that its a bit weird isnt it like a Franken chicken I dont know about that do you have any other ideas yes my other idea is a pipe organ where chicken opens the valves and plays music oh but we already made a paperwork and remember Music foreign Music foreign Music this chicken onto some D prints right so I built this controller thats gone Arduino Mega in as usual for my projects its got some relays that you can see turning on here and weve got some connections out here that go to the two chicken mussels so basically this is turning on one muscle and then its waiting for five seconds for the muscle to contract and then its turning on the other one so Ive made a kitchen cleaning robot that pushes this glove balance and forwards to clean the surface using two bits of chicken as mussels thats to stop us getting food poisoning Thats pretty clever can it do anything else yeah every so often it squirts this squirty bottle of kitchen cleaner using a third chicken muscle so the chickens have a nearly sunlight helping out around the kitchen wow we did science you really can use meat as Motors laughs what was the last part but I dont want to get salmonella I dont want to get a salmonella but I dont want to get a salmonella sorry its really conceptually getting to me pipe organ remember Laughter ah sweet memories

write a youtube script

have you ever been to a sushi restaurant where the food comes along on a conveyor belt and have you ever wondered how the conveyor belt can go around corners well there are various offtheshelf solutions for conveyor belts that can go around corners that are used in packaging and Automation and things like that this is some video from Dorner conveyors which shows how to assemble the conveyor belt and as you can see there are some bearings on the end of each track link and those fit into a Groove which guides it along whether its going straight or its going around the corner the track is driven by a sprocket just like a tank and theres an idler at the other end and that seems to run smoothly the entire course of the conveyor I recently made an omnidirectional tank that can move in any direction but could we make a tank which has a single track which bends to steer just a quick ad from my D printing sponsor thanks to lulsbot for supporting my Channel with D printers foreign Music thanks to D fuel for the filament for this project all of these parts are printed in Pro pla plus Im using some steel rods to make the pivots in between each piece of track there so those are just threaded through the middle there and those can pivot and what Ive done with the pivots here is that those are kind of a V shape essentially with a taper point in the middle and that should mean as the track bends that it keeps a constant length rather than just being completely loose theyre constrained in the middle theres also quite a lot of stretch in this because those axles are actually quite bendy this is some cheap steel I get from an educational supplier it does deform very easily though so for this project Im going to upgrade to some three millimeter diameter stainless steel rods with those cut to the right length we have something that feels a lot better and the stainless steel is a lot tougher basically so theres a little bit of Bend in them but its much much stronger and its not going to deform permanently if I stretch it we need a sprocket to drive that so Ive got this sprocket that fits into the gaps in the track and those should wrap around perfectly and fit onto the teeth of the sprocket so we can drive the chain that sprocket fits into the middle of a wheel and Ive made that in three pieces so that I can change the sprocket and change the tooth spacing or size if it doesnt run well in the future without having to reprint the whole thing which is quite a good way of Designing things each side is just screwed on with three screws that are recessed into the pieces there and that means I can change the sprocket really easily if I need to in one of the side pieces Ive got a bearing there and on the other side Ive got a hole and theres also a grub screw and a captive nut I recently built another tank project which was the Omni tank that can move in any direction with omnidirectional tank treads just like Omni Wheels but a tank but the motors I put in here were a little bit underpowered feeling so for this project Im going to use quite a more substantial motor which is a big motor with a bigger gear head on it on that motor is a D printed collar which is going to slide all the way down onto the body of the motor almost all the way to the back and that collar fits perfectly inside the bearing in my wheel and that means I can put the motor right inside the wheel the actual output shaft of the gear head on the motor is fitted into the hole on the other end of the wheel and held in place with the captive nut and the grub screw so this is quite a handy hint if youve got nowhere else to put the motor on a drivetrain with a belt driver or Gear Drive because now the motor is inside the wheel and perfectly balanced across it and that motor and wheel is held between two pieces one has a bearing for the smaller end of the wheel to fit in there and the other end has basically a clamp which is going to hold the motor body instead of my hand these parts are a bit special though so Ive got lots of little bearings with a little D printed cap attached to them and all of these fit onto those parts thanks to Simply bearings for the bearings for this project check out simplybearings.co.uk so theres a bunch of those fitted at degrees to each other on each side and I have put a washer on the inside there to space them out so they dont rub on the plastic so we now have this unit which represents one end of the tank and that has the motor fitted inside the wheel there so weve got our powered sprocket fitted between two of these pieces and weve got all of these bearings in there which are going to serve as a liners for the track the other end is very similar only weve got this passive wheel which doesnt have a motor in were just going to drive the whole thing at one end and youll also notice it doesnt have the sprocket although Ive left a blanking plate there so I can put a Sprocket in if I need to but for now thats just going to fit in between two bearings so it can idle as the other end of the track is driven by the motor so that just pops straight in there and I think this idler will run in the track just fine without the sprocket it should run in the track held in with the two edges which I thought to print into the track parts so what weve got is two ends to this thing one is active with the motor in the other one is passive so it just spins round and then weve got these sets of bearings here and the idea is that these two pieces are going to bend in the middle and the bearings hopefully are going to go and align the track so by the time he gets to the sprocket and the other idler here is straight and it doesnt come off The Idler so theres a little bit of straight section and then basically the bend is in the middle so now we need a way of bending these together that gives us a nice smooth curve in the track so Im going for two partial gears which rotate against each other and youll notice theres bearings in those which are top and bottom and that means weve actually got two pivot Points spaced away from each other instead of just having one hinge in the middle I fitted These Bars and Ive got eight mil bolts with lock nuts on the bottom there so that holds the whole thing together and keeps the two gears aligned so now when I put the track on we get multiple Bend points mainly where those two bearings are which allows that track to bend smoothly instead of there just being one pivot point in the middle but before we see how well that works its time for a quick ad from the video sponsor which is fan home fan homes mission is to inspire fans with high quality build up models and collections from your favorite entertainment gaming and pop culture Brands all the products are original designs you wont find elsewhere each shipment includes fully Illustrated magazines full of inspiring content and packed with information just choose your collection or build up model and every month youll receive exclusive products along with their magazine fanhomes Marvel superheroes and villains is an Incredible scale collection of all the characters you love from the Marvel Cinematic Universe the figurines have been produced with painstaking detail created digitally by experts and then cast in metallic resin and hand painted in your first package youll receive an Ironman and Captain America just as they appear in The Avengers all subscribers will receive a display stand Rocket Raccoon figure a marble metal box and three posters each figurine comes with a magazine full of information about the character including their comic book Origins and how they were brought to life in the movies to start your fan home Marvel superheroes and villains collection now use the link in the description and promo code Bruton to get an early bird gift right lets see if this tank track works I put all my track segments together so Im just going to put some power on that motor and see how it runs seems all right on the table lets try it in a bigger area so pretty glad that my sprocket has stayed in the track so theres a liners look while theyre working with all of the bearings but lets see what happens if we drive it and then try and drive around a corner so Im just going to put a little Bend in it at the moment this is totally passive so lets just power that up so its okay if we dont bend the track too much basically although it tends to drive straighter by itself as we go so lets bend it round again obviously that track will only Bend to a certain extent so yeah not too unhappy with that that looks like its not working too bad well theres a bit of a jam so lets just straighten that out the track isnt very tight all the way around and also in the middle of the bend its kind of loose and theres no way to tension it so thats something we need to look at but it does fundamentally work as long as were on flat ground and basically we fluke it by trying to keep that belt tight so my track isnt overly tight its quite loose in the middle it could do with tensioning but worse than that is if I bend it then it gets super loose because theres nothing basically pushing it into that corner also if I pick the whole thing up the track falls off the bottom and if you then bend it it pretty much comes off the aligners completely because theres really nothing tensioning it and holding it on there as they should be so Ive upgraded the bars that fit on top and bottom of those bearings in the middle of that gear and these have got bearings attached all over them so the Sun that the track can roll on to stop it falling down until on the sides there that should hopefully hold that track into the bend Ive also fitted some extra guide bearings on the end of that straight section and thats bonded permanently to it just to help support that track as well so now my track is held up by the guy bearings and its also held side to side by the other guide bearings so now as I bend the whole thing that track should stay forced into the bends which should keep the track tight and thats the same on the top and bottom it can still just fall off though either lift off the top or it can of course fall off the bottom if I pick the whole thing up so to fix that Ive got these extra little pieces which are going to screw onto the sides and these are going to hold some extra bars with some extra bearings on so those fit just there its a bit of a nasty solution but basically Ive got some bearings on the outside and Im pulling that track right down to keep it tensioned over the middle section so now the track doesnt fall off the bottom and it should stay on the top the tracks a bit of a funny shape but it does allow it to bend still so hopefully this is a good solution I dont really like this solution of keeping the track down because it means the track isnt flat anymore the driving service isnt totally flat but it does hold the track down very well ideally what we need is a double layer track so we have some little bearings that go into essentially a Groove in the flexible track to hold it down and the outside driving surface is smooth that would probably be a complete rebuild though so for now were just going to stick with this and see how well it works I didnt really think about how I was actually going to drive it so Ive got this linear actuator which is stuck on the outside which has got more than enough Force to push it so I think thats gonna have to be the solution for now I also didnt think about where I was going to put batteries or any motor drivers so Ive made this controller which is an Arduino and two motor drivers on the end of a long lead so this will do fine just for testing whether this thing has got traction and can climb over things and steers and drives okay it seems to drive on carpet okay and that actuator is more than strong enough to overcome the friction so I can dry steer but of course my tracks are smooth at the moment and they dont have any grippy rubber on or anything like that so this seems a perfectly adequate actuator and a perfectly adequate Drive motor to drive the whole thing along my track doesnt Bend quite as far as I wanted it to and also the actuator doesnt move quite as far as it could if I actually bent the track as far as it would bend so I can turn around but it takes quite a few multipoint turns to turn the whole thing around and do a and head off in the other direction should I want to of course I can drive in both directions so I only ever actually need to turn degrees but can it climb over obstacles lets try this plastic box lid so the track itself grips which is good and if it were a smooth driving surface throughout the whole length of the tank then it will probably work fine but actually weve got really stuck on where those bearings are and where the dips are in the track at the bottom lets try some aluminum Extrusion which of course has exactly the same problem and Im basically stuck on there and the motor is jammed so yeah its just down there by that bearing and the bit of plastic that holds it and I can hardly move the motor if we try three of those at different spacings um yeah the same thing happens again so Im perfectly grounded there on the two little wheels you can see on the bottom those bearings lets put them further apart surely thatll work so it looks okay till of course it pulls one of those along the ground and then everything gets stuck again yep still one of those jams in exactly the same place so other driving surfaces are fine provided there are no big lumps and bumps which is the whole purpose of a tank really it would be quite good if I actually had some rubber grips on the outside and I thought about sticking rubber feet on all along the way on all of the tracks but actually we can have real problems because of those bars on the outside that are holding the bearings theres only about three millimeters clearance there so Im pretty sure that rubber pads stuck on are just going to get jammed under the bar and then the whole thing will jam up and it wont run at all so that works fairly well for a concept theres a couple of improvements Id like to do in a version two one of those is getting rid of these bearings on the outside and having a smooth driving surface as I mentioned earlier and the other one is having more Bend and less straight sections most of this is actually straight and weve only got this little bendy piece in the middle if we had two of these joints with these part gears then we could have far more Bend maybe its a bit longer and well just keep those aligners at the end which seem to work perfectly well but for now Im going to publish all the Canon code on my GitHub as I usually do for all my projects so if youd like to support me through patreon or YouTube channel membership then those links are in the description below and patrons and YouTube channel members can get access to all the videos up to a week early including sneak peeks and pictures of whats coming up alright thats all for now

write a youtube script

weve all seen Dr Otto Octavius or Doc Ock in the SpiderMan comics and movies doc Oak has four tentacles and four of his own limbs so that makes nine in total and thats the same as an octopus lots of cosplayers have made Doc Ock cosplays including Adam Savage although mostly these tentacles are in fixed positions or supported with strings Jake laser built one that was articulated and it was also apparently brain controlled Jakes system used stepper Motors and foam tentacles to keep the Maslow although you can still see some weird kings in the tentacles and they look tricky to control precisely as Jake pointed out in the video most commercial tentacle robots use a huge Bank of high power actuators and thats mainly because these actuators have to be able to support the mass of the whole tentacle a scientific demonstrated theres a lot of friction in cable control robots but you should also check out his tentacle robot which used a totally different mechanism most cabledriven tentacle robots have their segments all Bend in the same direction as the actuator on that side of the robot Titans unless there are multiple actuators for each side and some cables fed through Boden tubes further down the tentacle Ive also previously built a robot with tentacle legs which used angle swiveling sections and this was inspired by a Lego build from akiyuki brick Channel although this method could use relatively small motors and still hold itself up we were stuck with a rotating Motion in each section so this time Im going to try a different approach using six actuators in each segment of the tentacle a Stuart platform is a parallel manipulator platform typically used in flight simulators it uses six linear actuators to move the top platform in six degrees of freedom which is three in Translation and three in rotation this is typically the same as my robot dog builds where we can slide the body back to front sideways up and down and also rotate in Pitch rolling your the Stuart platform can do the same and the rumor is that calculating the inverse kinematics to make it do so is fairly simple so what would happen if we make several of these and stack them on top of each other we should be able to get some fairly independent movement of each tentacle section foreign just a quick ad from my D printing sponsor thanks to lolsbot for supporting my Channel with D printers Music Applause Music Applause and thanks to D fuel for the filament for this project all these parts are printed in Pro pla plus Im just using some standard RC style servos these are kilogram torque servos and these ones have metal brackets for building robotics assemblies Ive mounted six of those on the base and Ive mounted them all on their ends with the output levers facing outwards so I can get them nice and close together Ive also extended that lever with the D print and that has another hinge piece on which is going to face upwards mounted onto those are the main levers which are going to drive the next platform up and these move in all directions now weve got those two joints and the reason those orange parts are so long is so the gray part can lean right over the servo when we do some quite extreme moves with both rotation and translation and the whole mechanism hopefully wont Jam those are of course attached to the next platform with another intermediate stage with two axes so heres the structure together thats looking pretty good its not too unstable despite all the joints what Ive noticed is if I take opposite pairs of servos spaced out by two servos on each side then that gives me the uh translation axis there so that will always move sideways and if I take another pair of those servos then obviously we get exactly the same thing and then to tilt or roll this all we need to do is basically put the servos lower on one side and higher on the other side and then we can also control the translation still with the other pair of servos and that just means we can tilt it and that works of course in any direction so now all we need to do is work out how were going to control this with some joysticks so my joysticks of course move up and down and left and right so each one only has two axes but we need to translate that motion for rotation and translation to the six servos that we have in the mechanism Ive translated joysticks to similar mechanisms before though Ive built quite a few omnidirectional robots that use omnidirectional Wheels spaced degrees apart this was a simple case of using trigonometry to work out the vectors in the case of a threewheeled Omni robot we use the cosine of which is the difference in angle from where the wheels would be if they were all facing in the same direction that gives us a value of . and although it seems like we should run the angle Wheels . times the speed of the straight wheel in fact we need to run the straight wheel faster to untwist the robot as it drives so the angle Wheels run at . times the speed of the straight wheel and that makes it go in a straight line sideways I think this robot is going to be quite similar but this time weve got six Motors so Im going to try moving some of them . six times and some of them . times the distance we want to move and hack some numbers together around that but before we see if that plan works its time for a quick add from the video sponsor which is brilliant understanding the main ideas behind computer and data science is really important not only for building robots but also for advancing your career in Tech brilliant has thousands of lessons from foundational courses to Advanced maths to AI data science neural networks and lots lots more and they add new courses every month if you want to get better at math for instance then those courses are great and you can start with the fundamentals before moving on to more advanced topics most courses start with graphical examples so you can get to grips with the basics before trying some more advanced learning lots of professionals use brilliant its visual HandsOn approaches a much more effective way to get to grips with the key Concepts behind todays technology so if you want to stay ahead in your career then you should definitely check it out and Brilliant is built for people who are busy the lessons break down important Concepts into understandable parts that you can do when you have time the best way to learn things of course is to actually do them and thats the way Ive learned to do most of what I do in my videos and how I produce so much content just basically trying things out and seeing what happens mostly brilliant has great explanations along the way so you can actually do the exercise to learn the skill and its like having someone teach you as you go to try everything brilliant has to offer for free for a full days visit brilliant.org jamesbroughton or click the link in the video description and the first of you will get off brilliant annual premium subscription right lets check out the control on this first stage of the tentacle my servos arent spaced at exactly degrees either but basically that seems to work and it gives us the rotation axis in both directions and of course they all mix together as well so that we can move the whole stick around and we can tilt that top in any direction the translation also works using a similar combination of numbers and I can move that in all directions although actually its not quite right front to back because the platform gets higher and lower so I probably need to think about the maths a little bit more we also have the heave control which is going up and down basically and that mixes with the others too so these joysticks have a third axis where you rotate the top and all that does is add something or subtract it from each servo these all mixed together as well of course so I can use both the translation and rotation joysticks at once and the heave control and that looks like it works and we get some quite good motions so I think this is going to work out pretty well once weve stacked two or three of these sections together and we can independently control them so it looks like it works but actually my math is quite hacky I pretty much just hacked numbers together from the control sticks to make this move as it should one of the things I havent done is actually worked out the actual length from the top of the servo to the top of here as we move each one up and down basically we need to work out another triangle thats from here to the point of the pivot on the lever down to the servo I havent done that Ive just assumed that the rotary motion here is linear if it was a linear actuator at work but of course this isnt actually a linear relationship because this is a rotation and it makes another triangle we need to work back which I havent done the other thing I cant do with this one is your which is rotating the whole table round which you can in other Stuart platforms and the reason for that is that it doesnt physically move like that and to do that I would need to have an actual vertical rotation axis on each of these joints here so basically another linkage that would actually allow that rotation then all we need to do is make every other Servo high and every other one low and that should push it round in each Direction so for now Im going to leave it like this because its already wobbly enough with all the joints but maybe well come back and do a version two so all we need to do now is stack some of these together to make a tentacle so I left places to attach the next set of servos on that platform so thats pretty good and if I move them all in the same direction we can see those moving as well I have fitted some switches which are going to reverse some of the sections so we can control them independently so if I switch that second segment over and reverse all the positions then we can see it tries to keep it level as we do the Tilt motion basically those servos are just moving in the opposite direction so the base servos and that means we can get the second section to move in the opposite direction as well so Im pretty happy with how thats worked out so Ive now installed the second section and of course if we operate that heave control to move up and down both sections go up and down and the tentacle gets longer the Tilt control tilts them both in the same direction at the moment although we dont get quite as much of a tentacle action as I wanted translation helps though because that will actually Slide the whole thing over and if we start using the translation and rotation together we get something more like a tentacle with quite a lot of reach and of course with both sections moving in the same direction we can operate that with both of our joystick axes and move the tentacle all around Ive now switched over so that the two sections move in opposite directions so that means the heave control will move both sections in opposite directions and the top platform should stay nearly still it also means if I do translation or rotation moves then they should bend in opposite directions that means we can get some quite interesting moves it looks quite organic a bit like a dancing Groot were going to put another stage on top of here with another six servos and Ive got another reverse switch so we can independently reverse the second and third stages from the base I am using some motion filtering on my sticks to control the tentacles so we get nice smooth motions I made a specific video about this a while ago using some cheap servos and an bit Arduino to show how you can make cheap servos move nice and smoothly for animatronics and basically what were doing here is taking a large percentage of the previous value for this Servo or the stick by bookmarking it in the loop and then using a small percentage of the new value and doing that each time and that means the value cant change too quick so we get this nice deceleration as the robot moves so thats three stages put together its actually quite big and awesome its about halfway to a dock Arc tentacle and I can get some quite good moves out of it so lets give it a spin and see what we can do with it heres all of the heave axis together so everythings moving in the same direction which of course makes that tentacle longer and shorter if I reverse just the middle section then we get some other sort of effects and Im wondering if we could make a worm that moves along through peristaltic motion like this I can also lean if I use all of those motions together so this is the translation and rotation and every section moving in the same direction with three sections we get quite a lot of motion so you can imagine perhaps if it were five or six sections long how far it will be able to reach how far it can bend obviously the strain on the bottom servos will be high as they carry the weight of the whole tentacle but its not doing too bad so far with three sections so I think some ball screw or lead screw driven sections would be fine here Im reversing just the middle section and that gives us that weird dancing Groot look again but with three sections and thats actually quite organic its quite a weird feeling watching it doing its motion and Im moving the translation and rotation sticks out of sync here this sort of motion isnt that useful for actually reaching and grasping things Id imagine although it does make the tentacle slightly shorter when I bend them like that but basically it does show that we do have independent control over each section if I move only the bottom one in the opposite way to the other two then that gives us something thats actually quite useful so we can bend on the spot there with the bottom section bending one way and the other two bending back above it in the other direction so this is actually quite a good tentacle feeling which should be quite useful to Dr Octopus foreign get away with just the Tilt motion or the rotation motion on each stage here to make a tentacle that bends but the translation really helps with that reach and makes you do weird other things I really want to put LED strips all over this and tell everyone there is a piece of art yeah so pretty happy with that Im pretty happy with the moves and all the things that we can do with this its quite flexible like an actual doc OG tentacle obviously we probably need about six stages and then the servos of course at the bottom stage are going to basically have a lot of strain on them where that whole tentacle is leaning on it but if we change these for linear actuators with a screw drive thatd probably be much better and then of course we wouldnt have these pieces sticking out as well because we just have that linear actuator straight up so it actually probably get down to something like this which would be quite docked octopus tentacle shaped so Im probably going to do another version of this Im not going to wear this one right now but Im going to publish all the cad and code as usual and that link is on GitHub and the links in the description below so if youd like to support me through patreon or YouTube channel membership those links are below as well and patrons and YouTube channel members get access to all the videos up to a week early and sneak peeks and pictures of whats coming up all right thats all for now

write a youtube script

do you remember when I made this robot bin that could autonomously navigate around and pick up the trash well the next step for that was to make a robot that could go from room to room and deliver me a cold drink from the fridge but then lots of people said in the comments there is no use unless he can go upstairs I found various robots on YouTube that can go upstairs of course including Boston Dynamics spot mini my open dog project isnt really ready for that though even though it can walk over lumps of wood I also have these bends in my stairs which means theres limited turning space in two places so its gonna need to be something much smaller I came across a few tracks robots and robots with multiple wheels on pivot so they can just drive upstairs these look good on test steps but theyd need upscaling significantly to get up actual stairs which tend to be larger and quite steep then I found this stair cleaning robot which splits itself into three vertical sliding sections to lift itself up on each step the middle section has four wheels to hold it stable and it also looks like it has Omni wheels so it can slide sideways to clean each step as it goes my stairs are a lot shallower than that which means the robot would need to be much thinner and the middle section would be much less stable when it lifts its back or front up so Im going to adapt this design with an extending section thatll allow it to span three steps at once to keep it stable so now it must be time to D print some parts the quick ad from my D printing sponsor thanks to lolsbot for supporting my Channel with D printers Music and thanks to the D fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com so weve got some Motors that are going to drive the main axes and Ive used these motors quite a few times before I got them from gimson robotics in the UK theyve got motors with gear heads and Hefty mil output shafts those are fitted onto a D print between two bits of Extrusion which is Extrusion with vslots and you can see weve got them fitted there one back and one front the two main axes are driven with ball screws and those are fitted to a CNC coupler to the motor those things just pull apart though so we need some way of restricting it so Ive got a plate with a bearing thats going to fit above that and hold the coupler together essentially with the ball screw poking through it so theres a ball screw for each motor for each of the front and back axes we with bearings top and bottom and of course if I run that motor and hold the ball nut still it will go up and down the ball screws have got quite a good linear axis there which is what we need and of course the one on the back is exactly the same so thats the core of the robot in the middle and its fairly Hefty because the aluminum and also those Motors are fairly weighty and of course the steel mil ball screws so all we need now is a piece that fits on that on the back and front to slide up and down Ive done that with a piece with some vwheels on so that slots into the vslot and theres four of those wheels and thats a pretty good fit and it runs pretty well so of course what we need to do is couple that to the ball nut and Im doing that with another D print which just sits on top and will bolt onto that nut putting that together gives us a lovely linear sliding axis its a bit slower than I wanted but with that bubble screw and that fairly Hefty motor thats going to be pretty powerful and of course the robot needs to lift itself up the drive wheels of the robot are going to be attached to these fairly Hefty worm gear driven gearboxes that Ive used to quite a few times before I use those in the robot bin and of course weve got one of those either side so we can do differential drive this robot has four Drive Wheels though because we need to drive the front wheel section as well so theres another pair of those Motors that are going to be fitted onto the sliding section we just built and that looks like this with four wheels which are just gonna skid steer when they drive with differential Drive Ive got some TPU ties to fit so we can get grip on the carpet and those of course are just a push fit around there but theyre pretty tight so theyre nice and grippy so with that assembled weve got one fairly substantial robot theres quite a lot of mass in this with the ball screws and the six Motors what Ive actually done here is made sure most of the mass is in this middle section so weve got the ball screws and four of the motors ahead of it weve got two Motors and weve got the battery in a pocket there thats going to power it and that means most of the mass is going to be ahead of it going up the stairs as it goes up and then weve got that trailing section to fit on the back next which is going to be much lighter so hopefully having hardly any Mass on the back means it wont get dragged back down the stairs backwards so the front section as youd imagine uses the other ball screw to make another sliding axis and for now weve got two forks on there made of more V side Extrusion that looks a bit like a forklift truck that has a sliding section on with some vwheels so we can extend over more than two steps so at the moment its looking like thats going to be a bit back heavy and we still need to add some Wheels to that so Im using Omni Wheels here which are going to work a bit like casters so it can still skid steer and we wont have any problems with drag so these are wheels with little the wheels all around the outside Ive used these in several projects before that whole assembly of course extends out the back so we can reach over two or three steps at once and thats going to be driven with a rack and pinion type assembly so Ive got a rack in there and Ive got a motor with the pinion on which fits on top so as you can see as we move that you can see the pinion turning as the rack moves so of course if I activate the motor and move that with some electricity then we can move that axis but before we carry on with that its time for a quick answer on the videos sponsor which is on shape on shape is a cloud native Canon PDM platform built for business created by the founders of SolidWorks because they saw that modern product developers still experience many challenges related to their Canon PDM systems so they started from scratch and created on shape on shape have just released the new and extremely popular capability on shape simulation on shape simulation is a unique Cloud native approach to FEA and is the easiest and fastest way to examine meaningful structural analysis results such as stress displacement and factors of safety as you create and iterate in the onshaped design environment with simulation built into onshaped Cloud native architecture you can build better quality products faster with mechanical feedback that stays in sync with your assembly throughout the product development process onshape is built in the cloud so it works just like Google Docs so an unshaped document is a single source of True For Your Design data onshape is always growing new releases are pushed to the product every three weeks to add new features and functionality this happens over the air automatically which means your company will never have to manually deploy a cad update ever again I highly recommend the engineers and product developers watching to consider using onshape for their business and you can try it for free at onshape.pro jamesbroughton so now thats actually made the robot very back heavy which isnt what I really intended and if I roll the axis out it gets much much worse of course and theres quite a lot of force needed to be applied to the front there to compensate more than the mass of the batteries so what Ive realized of course is sometimes you want it to be front heavy so particularly when we pick this piece up and particularly when its at full extension so it doesnt pull itself down the stairs and sometimes actually though we want it to be back heavy when it picks the front up so it doesnt tip over forwards before its got its next foot on the step so what Im going to do is have a mask that slides up and down on some more rails up here so it can slide back with some forwards and we can either make it back heavy or front heavy as we need to that is a belt driven axis which moves two lumps of lead which are a kilogram each and these are diving weights are using the bipedal robot projects to shift mass and thats just one pulley which takes the belt through some bearings so its nice and tight and the belt goes from one end to the other so now of course we can be front heavy if we want and if we roll that weight the other side then were completely back heavy and even if I roll it to the front and move that back axis right out to full extension thats still enough to keep it front heavy as usual the electronics for an Arduino Mega with an nrfl radio chip so I can remote control it and Ive got three motor drivers on this side and three on the other side so we can drive all those Motors and powering it from a volt lipo and a USB boost pack which just Powers the Arduino I made a custom remote for this robot so weve got one joystick or a thumbstick which can drive it all around for driving the wheels and various buttons we can press to drive those linear axes inside is another Arduino Mega lots of wires and the other radio chip so I can use that thumbstick to drive the whole thing around and that works okay and youll notice the back Omni Wheels obviously they can slip sideways so its got no problem doing differential drive with the front wheels and the other ones are Slip sideways so that works all right I can also Drive the other axle these around so weve got those linear axes that I can drive there as well as that back axis there and my linear axis as well to lift those two pieces up and down so I was thinking about what sensors we could have on this to automatically detect the stairs and work out when its on flat ground at the bottom and its got to the top and when it needs to lift up until it clears The Next Step but actually I think its going to be hard enough to get it up the stairs as it is its got a quite over complicated so for now Im just going to drive it manually with this remote and well see if we can actually go up the stairs so first of all weve got a position with the bottom step Im just driving this manually thats me in the mirror Ill speed up the footage in a minute because it takes quite a while to get up the stairs but for now Im going to do it in real time so you can see how it works at least till we get the first three feet up so were gonna lift up the front and drive forward and position it there move the mass so its in the middle and pick up the middle section by sliding both of those front and back linear axes down then we can drive on with the next two sections so those are on the first step and then we can move the mass to the back and pick up the front foot again and that gets us so weve got all three feet on three different steps after that we can move that Mass to the front and then we can pick up the back section that lifts up fine now weve got our two front feet down and weve got the mass at the front and then of course we can suck that in and then repeat the process picking up that middle section driving forward and so on and so forth Music thank you Music thank you he made it back upstairs it took about seven minutes though in real time which is a little bit too long really Im quite happy with the mechanics though because it made it round those Corners which a longer robot couldnt do and does actually go up the stairs although its quite tricky to get it aligned with the steps properly and the spacing just right so it can lift its wheels so I think putting sensors on to make it do that autonomously is actually going to be quite a difficult thing to do Ive had a much better idea though called the Rumba Stair Lift which Im going to build in a future video for now though Im going to publish The Cutting code and all thats on GitHub and the links in the description to this video so if youd like to have a look or build your own you can and also if youd like to support me through patreon or YouTube channel membership then you can and those links are in the description as well and patrons and YouTube channel members get access to all the videos up to a week early and sneak peeks and pictures of whats coming up so you can be part of that discussion and Discord benefits

write a youtube script

this is the next video about coming up with an idea for something Ill eventually make big enough to ride on I really want to make a giant rideable land worm and last year I made an eightwheeled snake bike which bends to steer that had Wheels but I really like to build something that can move in an unconventional way or walk on legs Ive been experimenting with wave drives the first one used a spiralshaped former inside a flexible track and I tried a camshaft driven design a few weeks ago at the end of that video I mentioned it might be better to turn all the cams around degrees and use them to drive legs in a wave formation this leg mechanism is half of the walking mechanism I used in each unit in my omnidirectional walking machine so I decided that perhaps I should just use a whole mechanism and line up a bunch of those to make a centipede I got this leg mechanism idea from the DIY Walkers YouTube channel which you should check out I want to make this machine four or five segments long and drive the whole thing from a single motor so Ive designed universal joints to link all of the sections and a gearing arrangement to drive everything from one overhead drive shaft there are quite a few parts that have to mesh properly mainly these bevel gears which are going to be translating the motion from the overhead drive shaft down to the legs and those have their own builtin drive shaft with bearings at each end theres also several legs if theres going to be five sections with four legs then theres legs and those have some weird overhangs so they mesh properly and can fit onto the pivot Points thanks to LOL spot my D printing sponsor for supporting my Channel with D printers thanks to D fuel for the filament for this project and lots of other projects all these parts are printed in Pro pla plus and thanks to Simply bearings for the bearings for this project of which there are quite a few each segment is basically a box with four sides and Ive got those Slimline bearings embedded in recesses my bevel gear fits in there between the two long sides and of course the other side fits on there to support it at the other end those bevel gears have got recesses in each end and that allows a keyway to fit which is for the cams which are actually going to drive the leg mechanism on each side and those screw on as well of course on the top of that we have another overhead drive shaft and youll notice that has a spur gear on but that doesnt actually mesh with the bevel gear in order to mesh the two together weve got this bevel gear which is basically attached to a spur gear and thats got bearings in each end so that it runs nice and smoothly and that fits onto a bolt and it should fit right in there so that it meshes the overhead drive shaft with the big bevel gear to support the other end of that bevel gear and the bolt Ive got this little bracket and that drops into a recess in the corner so now we should find that the whole mechanism runs together so as I run the leg mechanism you can see that bevel gear turning the drive shaft and the reason weve done it this way with this overhead shaft that turns the spur gear turning the bevel gear and finally turning the motion for the legs is so that we can keep the motion going in the same direction through all of the segments if we went in at one side of the bevel gear and out at the other side then each other section will get reversed and that means that some of the legs will walk backwards so weve got a pivot point installed on each cam on each side and Ive installed the first two levers of the leg mechanism each leg fits between the lever and the cam and weve got two legs on each side and thats where those funny recesses came in which were printed with support material heres one module and of course we do have those cams facing in opposite directions to keep the legs in sync correctly I did use washers on all of those pivot Points but basically theyre just studying and bolts bolted straight through holes in the plastic and yes I made the whole thing five times and Ive decided we should have five modules in our mechanical all centipede here are the universal joints that I mentioned earlier and again those have a keyway that fits into the recess on that overhead drive shaft and that should allow us to drive the overhead drive shaft between the sections as the sections of the centipede Bend in all directions Ive synced up the legs so they dont hit each other although it doesnt leave as much turning Circle as the sections Bend side to side but there is a gap so well see how that goes and Ive also added these red feet which are basically pla theyre not grippy but that will take off the sharp Corners that might get stuck on the legs so turning that manually it looks like it walks along okay obviously its quite difficult not to pull and push it as Im doing it you can see those feet slipping slightly on a smooth surface but altogether the mechanism looks like it should work okay it doesnt work quite so well on carpet though so basically those feet are getting stuck on the grippier surface and thats causing the two sections to Kink up and its causing it to bend and I think thats happening because its rocking side to side or one axis basically is allowing those feet to get jammed and push against each other so Im going to try constraining all of the axes apart from one with this plate which is going to screw on top this has got a simple hinge on it so yeah that seems immediately much better as I push it you can still see theyre bending between the two of them but because the other axes are constrained basically it means those sections cant tip around independently and the feet arent getting jammed in the carpet so much so this thing of course means it can only move in one axis and it cant move up and down anymore but it also cant twist so I think this twisting is whats causing the problems however I do want it to be able to move up and down and bear in mind theres going to be five of these together so I quite like it to be able to walk up and over things and of course if its one rigid stick then its not going to work very well so were going to use another piece between the body sections which is another universal joint and thats got a hole in the middle so that the existing universe will join that does the drive can fit right through the middle there but this one is attached to the bodies and that means it can only move left and right and up and down and theres no twist so thats fitted there between two of the sections and of course that means that both universal joints can move in all directions but not twist and my drive shaft still works in the middle because those pivot Points are perfectly aligned with the outer universal joints but before we carry on with that its time for a quick ad from the video sponsor which is Rage Shadow Legends its completely free to play and has millions of players already theyve already been million downloads raid Shadow Legends has amazing Graphics just like a console or PC but on mobile theres over completely unique Champions and there are billions of ways to customize them there are new Champions and updates every month so theres endless content and tons of awesome bosses to take on new players can enjoy a special Valentines themed adventure with the raid Love quest all you need to do is download raid from the links below copy your ingame player ID and then head to raidlovequest.playerium.com enter your player ID and set out on this heartfelt Quest running from February the th to March the th play one of the Valentines minigames for a chance to win some fantastic ingame and real life prizes raid Champions and even Amazon gift cards worth up to a thousand dollars old Ray players use special promo codes and Valentines that everyone can use to get a small Valentines gift oh and for Amazon Prime members who just got genbo keep an eye out for the next drop with some powerful Savage gear is available from March the second until March the th with all this exciting stuff and more coming to raid if you havent started playing yet then what are you waiting for use my link in the description or scan my QR code to get insane bonuses were talking an epic Champion Kellen the shrike and other useful things like energy refills magic potions and XP Brews so with three segments together thats looking much happier so I believe taking out that twist has meant that the legs are hitting the ground in the right order theyre not straining against each other so much and causing it to King up its hard to tell driving it manually though so its time to add a motor and that just sits on top there and drives one of the overhead drive shafts which of course will drive the whole thing all the way through so pretty happy of how thats looking its moving at quite a good velocity but lets try it on carpet so apart from having no idea where its going so if one section starts to bend then basically it leads it and it starts going in that direction which is pretty random but all together Im pretty happy of how the mechanism is working thats looking pretty convincing for a mechanical centipede you can see that the feet are actually picking up and clearing the ground quite well there is a bit of foot slippage but well see how that works out for us later as we add more segments so it works okay with three modules but weve actually got five in total so what Ive done is put a motor at both ends theyre still mechanically linked all the way through with those universal joints but now weve got two Motors driving it instead of one because weve got all that extra friction of the extra two modules Ive also got three extra things on the top of the middle modules so theres our head motor weve got two servos which are for steering well look at later and a motor driver to drive the motors an Arduino Mega there on the middle section weve also got a battery there for now just a s lipo to power the motors and heres our tail motor so that makes up all five segments Ive wired that in now I in fact added another battery for the servos which is a s lipo because they wont run on a s lipo lots of wiring there is a DRX unit there which is the radio control receiver so that we can radio control this and drive it around like an RC car Ive got the motor driver wide in and the servos there are linked to basically two bits of bungee cord so Im going to be using the universe personal remote thats a DSM remote that I made in a video a few weeks ago which Im now using for all my robots so basically for steering were just going to pull these reins between the first and second segments which are elastic so the legs dont get jammed together and it doesnt Jam the mechanism with that hid unit held in roughly a straight line with the bungee cord it seems to walk in a straight line now rather than going in random directions but I can steer if I want to so lets just steer to the left there you can see me pulling the left Servo and now we should go to the left and Im pulling the right Servo so I should head around to the right there and well see if we can walk in a complete Loop well I got jammed on a table leg so Im just going to go and move that around a little bit more and well continue round the Turning Circle isnt great but its not too bad for something so long and of course eventually the legs will hit on each other and thats why Ive got those elastic bungee cords so that the bungee just stretches and it doesnt Jam the mechanism but Im pretty happy with that so far I found a bigger area for it to walk on and since it seems happy with carpet and a smooth floor were going to try it on a smooth floor so yeah you can see that the legs do slip a bit but basically its pretty good at steering as long as that first section steers then the whole thing just follows it around so Im able to make quite a convincing Bend there to come around the kitchen cabinets and head back out of the door and all those leg sections are still lifting up quite convincingly each section of the centipede seems to support the next one so that holds all of the sections level of course and it looks like a happy centipede walking around not sure if its a hungry centipede or a Hungry Caterpillar though its time to try some obstacles but its feet dont lift up amazingly well so if you dont get it in sync then it gets jammed but if Im careful where I start it and we can get that first foot up on the ledge then pretty much all of the rest follow because the next section gets picked up by the last one and its quite good at climbing over things I can only steer the head section though so if the body starts to sort of fall off the edge there I cant do anything about steering it back so perhaps I could have servos on every section so the whole thing would bend but Im not going to bother for now if I hit the ramp properly then it all just climbs over fine but its a bit hit and miss my driveway Im testing on here is slightly sloped so Im going to have a quick walk up and downhill and see how well it deals with the incline you can see the steerings working okay there but we still have some foot slippage so Im not sure how much traction its actually going to have walking uphill but anyway downhill its pretty happy its a pretty happy centipede so lets try and make that turn now and try and climb back up the hill its struggling quite a bit there and you can see at least on that front section those feet are practically moonwalking theyre hardly getting any traction at all so despite legs trying to push it in the same direction its not really working very well at all and its not coping with that very slight incline well its struggling quite a bit there it is making very slow progress but yeah not working really very well at all not as good as it was on a flat surface or downhill so lets turn around and go downhill again which should give it a significantly easier job so there is a tiny bit of foot slippage but now were making much better progress and I think this is because the legs dont move with linear velocity for the linear rotational velocity so theyre actually acting against each other and thats really whats causing the legs to slip which it does on a smooth surface lets try it on a grippier Surface before we change all of those feet for rubber feet so here we are on grass but you can see thats really struggling and the motors really laboring and only really works if I grab it and help pull it along a better mechanism for this might be the Strand Beast mechanism which has slightly more complicated linkages but the feet move back in a much more consistent linear fashion so they dont Jam against each other several people have made rideon versions of this and versions with legs pushing bicycles and things and they dont seem to be too many problems with the legs moving out of sync so this would probably be a much better mechanism for the centipede with lots of strand beasts or walking in a line as usual Ill publish the Canon code for this which is on GitHub and the links in the description to this video if you want to download it and build your own so if youd like to support me through patreon or YouTube channel membership then those links are in the description as well and patrons and YouTube channel members get access to all the videos up to a week early and sneak peeks and pictures of whats coming up so you can be part of that discussion all right thats all for now

write a youtube script

Music dont go away Music Ive been working with World of Tanks and theyve sponsored me to recreate an allterrain vehicle concept from the s called the Rhino and the unique feature about this vehicle was its huge hemispherical Wheels I got myself a little wonton dumper stripped of all it Parts tilted the axles at degrees and then built these bulbous Builders now then is he going to do what the original inventor wanted it to and the more it sinks into the mud the more traction it gets lets go and test it lets go and find some mud this is optimistic Colin optimistic column is looking forward to testing this wonderful looking machine only optimistic Colin is a full optimistic Colin is gonna have a bad day I didnt think about how youre getting it did I so squeezed into the machine I drove it down to this water filled ditch to give it its first offroad test well that was I was full of water a week ago but not the turd I carried on as it was still kind of money Music things werent looking too bad you know we was making progress the wheels were turning but then we got to what is this machines arch nemesis the slight gradient back we go Applause Music what is that noise I have no idea what that noise I think Im still Laughter basically it wont crawl up this very slight Hill which is not great but its not the problem of the whale concept the original inventors idea is still sound Im calling him the original inventor cant pronounce his name now if all four wheels turn this thing would plow out of here no problem but my donor vehicle little hydraulic dumper had slots smaller wheels on we put huge wheels on it basically doesnt have the torque to spin them Hydraulics are just pumping oil it cant do it saying no Colin no so Im gonna have to try and get it out of here Im going to attack something slightly more on the level thank you with the tally Handler at full reach lifting the kilo machine was causing the back wheels to flick up in the air but we got it at so where next Colin water waters flat and the original inventor intended these massive bulbous Wheels to actually help the vehicle float you know the amount of displacement they give now my vehicle is quite heavy so I dont think itll ever float but its certainly going to help when were in the river oh yes we will move it the only thing that could stop me now it felt like was bushes damn you Hawthorne I feel hes doing a lot better in the river oh my God its getting deep here oh hes all coming weve got water let me just uh first back a minute okay this things definitely a home in the river but is it going to get out the river Music no no is the answer three two one very nice jump Colin but this machine still hasnt ended a test without the use of a JCB such a commercial World of Tanks is a free to play online action game and if you love tanks theres over of them at your order disposable to control in a range of environments now you can go and stealth and snooky or you can go in Fast and smash everything up in your path its up to you and theres a huge Community play in it as well now if you use my link in the description of my download code oh youre gonna get yourself free tanks credits and seven days premium access boom whats better than that and you wont get stuck like me right how am I going to get out of this situation okay this is still very nice it falls into ditches it goes splashing through the river but weve not really tested it yet because lets face it everything weve done we can do quite easily if not better on the quad so Im thinking when I did the screw tank video there was a bog up at Toms which I went in in the screw tank and that was proper sloppy it was flat so hopefully you havent got any Hills to worry about I think we should take it in there armed with me well hes on see how we do with this one Music I dont think this thing works does it Music pull the welding on the roll bar is obviously good well Im in new territory here people for the years Ive been uploading projects to YouTube ever since the Wall of Death made out of pallets back in this is the first thing I think Im actually gonna have to class as a complete failure yes it looks absolutely brilliant I could see this driving around on Mars I could see Dr robotnikbot bombing around and it looks like something that the leader of the Daleks would drive around in and its a little bit like a Dalek itself you know they were kind of menacing but they couldnt go upstairs and this cant this cant do a lot at all now I did try a few things drain the oil out of it check that put some new oil in didnt make a difference and I thought right lets put the old wheels back on and if it still doesnt climb up banks with the old wheels and I know Ive broken something so Ill swap them over and when you saw the old Wheels next to the new big ones you realize how big these new ones are and obviously that is the main problem with this Ive put two bigger wheels on it basically so swap them over sure enough it worked just as it did before nothings broken I thought Ill take it for a thing I still got it stuck now Ive asked way more of this dumper than it was ever designed to do and its done well to do what its done really but it was a design of the wheels which inspired me to do this project now I can see the logic and the fatness when youre driving into the mud you do feel like yes this is the way forward however you have to tilt them degrees and this is one of its downfalls if you do a sharp turn the inner wheels will pick up and you feel like its just going to roll over on them also the reason why the Army never took on this design of course if they get shot and pierced they can fill up with water and I didnt get a shot but even for a tiny little weld hole with me sat in the river I filled one of them half up with water so I understand their concerns now I wouldnt normally uploaded this because it doesnt work but you know it is good show your mistakes and we can all learn from it but also you guys can have a bit of fun with it I saw I want you to roast me in the comments about this now keep it clean because Im going to pick the best ones and put them on my Instagram page for everyone to look through go follow them in there of course and there we are so you know have a go but what about the pros I mean Ive mentioned it looks good which it does it can go through prickles frico yep precos now then of course thank you for World of Tanks for sponsoring this theyve made this wonderful waste of time possible use your link get you free stuff boom we love it now normally Id ask you a subscription but Im not sure if Ive earned it with this video Ill let you decide right see in the next one which I promise will work oh I cant even walk past it with that bloated Applause

write a youtube script

this video contains an ad for ag by athletic greens more on that later its time for the Makers secret santa every year makers get together for the Days of Christmas and we all make something for each other so Im going to make something and send it to someone and Im going to receive something from someone else now the whole thing goes around in one massive Loop so you can follow the whole thing all the way around and back to the beginning and discover all those channels you havent heard of before this year though its a leap year so theres only nine of us and the makers this year are Colin Furs xyla foxlin Becky Stern kids invent stuff Jimmy diresta this old Tony Emily the engineer look Mom no computer and me and this year its a random draw and Ive pulled Emily the engineer to make for so Emily the engineer has a quite a small YouTube channel but also three and a half million followers on Tick Tock so check that out and Emily does Iron Man cosplays and lots of other D printing and Engineering content so what am I gonna for her well I looked on Emilys Tick Tock for inspiration and came across this rather interesting video about an automatic boxing arm that she made apparently in two days which looks like its D printed and this is a reciprocating boxing glove that she tested with various other YouTubers that you might recognize well I dont think its a bad invention William Osmond so Im going to turn it into the game of automatic boxing arm BattleBots so Im going to make the reciprocating arm and put it on a base with wheels so it can drive around and make two so they can fight each other Music just a quick ad from my D printing sponsor thanks to lolsbot for supporting my Channel with D printers foreign and as usual thanks to D fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com the first part is two Runners which are screwed onto another D print to make a slot so that I can put the sliding Motion in theres no bearings or anything like that but its pretty free moving so that should be fine for our purposes and that screws onto the top of the whole robot so that weve got a kind of base for it to fit on Ive got these small motors which are six volt Gear Head Motors and Ive used these loads in lots of other projects so theres two of those one to drive each wheel and thatll give me my differential Drive setup so those sides screw onto the top that weve already got now we need to make the reciprocating motion so weve got this D printed gear which has an axle and also has an offset bolt bolted through it that fits onto some actual bearings so it can spin around smoothly on the slider weve got a bolt and not some bearings to put on one end of the rod and so that means as the rod goes around on that gear its going to make that slider go in and out in a reciprocating motion so now we just need to mount that block with the bearing and the big gear on the end of the top of the robot and weve got that motion we can see which is pretty much the same as Emilys original boxing arm so now we just need to add a motor which is one of the motors the same as the wheels and that fits onto a plate which fits on the bottom with a small gear to mesh with the big gear and that should drive everything along yep that seemed to have worked out quite well so its time for some Wheels which as usual have some TPU tires printed which fit right on there and that means theyll grip the ground nicely on the bottom Ive got two Caster wheels and that means weve got the differential Drive in the middle and the casters there just touch the ground so we dont lose traction so all of that seems functional and because weve got three separate Motors we can drive forwards and backwards or turn and operate the reciprocating arms separately but we need two so we can fight and also we need some Electronics but before we carry on with that its time for a quick hand from the video sponsor which is ag by athletic greens ag by athletic greens is a daily supplement that has me really excited about simplifying my health routine it has different ingredients including vitamins minerals superfoods probiotics and adaptogens I work flat out making content for YouTube and Ive been trying to cut down on caffeine so I love the energy supporting ingredients in ag which include a natural form of vitamin B Biotin and bioavailable folate its your daily multivitamin and multimineral pre and probiotic immunity support and it provides Your Daily Greens blend too its really easy to get into the habit of drinking ag all you do is one scoop of ag foreign ounces of water which is just under milliliters give it a shake and thats pretty much it and do it every day and it tastes pretty good too go to athleticgreens.com jamesbroughton to get started on your order and athletic greens is going to give my community a free oneyear supply of immune supporting vitamin D and K and free travel packs with your first purchase so weve got three ibt motor drivers which are more than sufficient for those Motors an Arduino Nano and weve also got an nrfl radio chip on the end there and everythings kind of stuck together and that fits neatly in the front so that we can control all three motors with those motor drivers over radio control but each robot is going to need its own transmitter so Ive got another Arduino Nano and another NRF l radio chip and that fits into a nice little handset which has a switch on the front which is going to activate the reciprocating punching arm and it also has a joystick on the top there so that we can drive it around and Ive got a little cap and a little cover for the side made as well these are just going to be powered by USB boost Banks and the serving suggestion is to just drop that into your pocket I have to ship these to Emily in America so this has got a ship internationally and I didnt want to ship Lithium Polymer batteries so Ive used nickel metal hydride for the robots themselves so with the controller all configured and the code written I can drive that around and it seems to be working pretty well well I can turn on the spot or drive backwards and forwards and the trigger there operates the reciprocating arm to punch with so that looks like its going to be super fun but now we need something to put on the front I left holes for magnets and Ive made some fists which have some metal screws in and that means they can attach or be knocked off by the other robots Music Music foreign Applause Music so well get that packed up and sent off to Emily and you can check out what happens when she receives it in her Channel and what she makes for someone else follow the loop all the way around and back to me for now it must be time for me to receive something is that the sound of Santas sleigh is that the sound of the doorbell all right all right better get that foreign first though because at him last year Music right Ive got my Christmas tree up because its Christmas so now we can see who sent this Secret Santa Music its a rocking chair but not just any rocking chair this ones got wheels hang on a minute theres a card all right if youre sitting comfortably then I shall begin right its from kids invent staff and the card says happy secret santa UNL the proud owner of the party chair designed by sixyearold Samaya from Pakistan it has speakers lights and confetti cannons and it drives have fun love roof and Shawn and it looks like theres a picture of the party chair originally drawn by sixyearold Samaya on the front Music thank you it also came with this wonderful pair of headphones but theyre not actually headphones theyre earmuffs and theyre covered in crystals all right Music oh doesnt do anything if you want to see how kids invent staff build the party chair then dont forget to check the build out in their Channel and dont forget to check out Emily the engineers channel to see what happens when she gets my gift and see what she makes for someone else Ill put the link to all of the makers in the description to this video and Im also going to publish the Canon code for the automatic boxing arm BattleBots project if you want to build your own that Links in the description as well alright thats all for now

write a youtube script

do you think static camera shots are boring with the camera sat still on a tripod well some YouTubers have camera operators which makes the footage look super fun and they can interact with the YouTuber too William Osman has a camera operator called cameraman John but who exactly is cameraman John the longawaited reveal of cameraman John give me that okay now the only problem is it turns out this whole time you have been cameraman John well Id like to make my footage look super fun too so today Im going to build a robot camera operator that can do lots of fun things just like cameraman John including looking in a mirror robot cameraman John needs to be mobile so Im using this omnidirectional bass which is originally part of a promo for Lego it was a tall robot that could drive around in all directions and on top were nine robot arms were displaying the parts of a Lego art set its a pretty robust base because of all the steel axles and heavy motors which Drive ring gears to drive the wheels the wheels have little wheels around the outside so it can move in any direction we need it to I D printed some extra parts with a . millimeter nozzle that makes them really tough and those fit over the existing motor mounts and that allows this tripod to stand on top I got this from a car boot side and it looks like its from a telescope eventually it will get zip tied down so I can remove the wheels easily if I need to on top we can have a pan and tilt mechanism and some other special features right its time to D print all the parts that didnt work now its time to D print all the parts Music just a quick ad from my D printing sponsor thanks to lolsbot for supporting my Channel with D printers and thanks to D fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com so I need to build the mechanism that can pan and tilt the camera and thats how the camera can automatically track me for this project Im going to be using some Dynamix or servos which are fairly Hefty and fairly quiet and these were given to me by robotis some time ago for another project and that project was another pan and tilt mechanism which is actually from a promo for EA Games to put an autonomous Nerf blaster on my open dog project which a lot of people didnt like because they thought it was scary that could automatically Target things and shoot at them but in this video were building exactly the same project but were going to shoot footage instead of shooting guns it works pretty well as a pan and tilt camera Gimbal and I had a GoPro on top of the dog anyway so the dynamixels attached to two blocks and those are D printed of course which fit onto the bottom of the other D printed Parts I can screw it on and that means that its aligned perfectly with another big bearing thats going to take the load of the pan mechanism theres a platform that fits on that and thats got a kind of knob on which fits of course into that bearing and then and that gets screwed down onto the top of the Dynamix also it can turn it and that means weve got the bearings inside the dynamixel and the other bearing the Tilt axis is pretty much the same with another dynamixel and another bearing attached to a piece and that screws onto the side of the pan mechanism so our actual tilt mechanism is of course a kind of Sling that fits in there between two round pieces on the end so one is fixed to the dynamixel and one is seated in another bearing which supports the other end of it so that of course gives us a very smooth moving pan and tilt mechanism using those high quality Dynamix or servos and multiple other bearings to support the axis so the load isnt just on the servo bearings and that seems to move pretty well on the Tilt mechanism is a quick release camera mechanism which I use on all my cameras to switch them between tripods so we can just clamp the camera in there which is a Panasonic Lumix gh I can still flip out the screen so I can see it on the other side and I can also access the SD card in the side so thats pretty handy I can get to those things without taking it off we need to fit the whole thing to the tripod Ive got so Ive got this base which has got an M bolt and a stopper on it and that fits through an existing hole on that telescope tripod does up nice and tight so it doesnt fall off and then the rest of the mechanism will just screw on from the bottom I do need to adjust the zoom of the camera dynamically as well so Ive got another little Dynamix or Servo and I D printed this gearing which goes around the zoom mechanism on the camera so it can turn it so thats all mounted on the tripods but now we need to put some Electronics in so Ive got this handy box here with some switches and knobs and inside is an Arduino Omega with a Dynamix or shield and the Dynamix or serial bus connected and also a Jetson Nano which were not going to use just now for now were going to do a basic test so its unplugged Ive got some switches for modes and also six knobs which means I can adjust the pan tilt on Zoom so for now we can just do a basic test driving those dynamixels Ive got some motion filtering on here which basically means the values cant change too quickly so we get a nice deceleration and we can operate all three axes there so thats the zoom and obviously weve got the pan and tilt the other three knobs currently allow me to drive the base around so using the Omni Wheels I can drive directly forward and backwards by using two of the wheels I can drive sideways by using a vector and driving all three wheels at the right velocity and of course I can rotate by turning them all in the same direction down the bottom is the unmodified base from the Lego project so theres a pocket for three batteries and theres three connectors Ive just got one in there for now each motor has a BTS motor driver and Ive got a volt regulator so weve got cables that go up to the top which allow pwms to come down for the motors and volts to come up for the electronics Ive used the Jets and Nanos in quite a lot of projects before and its really easy to retrain one of the deep learning models using transfer learning with the scripts and the GUI tools provided by Nvidia all you really need is a lot of pictures of the things you want it to recognize and theres some stuff on the Nvidia GitHub about how to categorize it and make it work in the past I trained the model for hand gesture recognition so I could recognize four different hand gestures and that allowed me to drive open dog around by pointing at it or showing it the gesture including jumping in the air I also train the model for facial expression by pulling a bunch of faces into the camera and I use that so I could activate Wolverines claws when you pull Wolverines face and in the last video I trained the model to recognize Vision markers so my robot bin could drive around the room aligning itself with the vision markers and collecting the rubbish but the Deep learning model comes pretrained so if we dont retrain it to recognize something else then it already recognizes lots of objects including people Im using the HDMI straight out of the camera that Im actually using to film with that goes into an HDMI to USB video in device and thats plugged into the Jets and Nano Ive also got a USB to serial converter so I can get the data out of the Nano and send that to the Arduino controlling the servos I mounted a screen on the front of the tripod so I can see the detection and I can check everythings working well and check that its actually keeping me in shot its pretty good at detecting people having been trained on probably millions of images already so you can see thats working quite well as I move around but before we carry on with that its time for a quick ad from the video sponsor which is on shape on shape is a cloud native cadmpdm platform built for business created by the founders of SolidWorks because they saw that modern product developers still experience many challenges related to their Canon PDM systems so this started from scratch and created on shape on shape have just released a new and extremely popular capability called onshape simulation onshape simulation is a unique Cloud native approach to FEA and is the easiest and fastest way to examine meaningful structural analysis results such as stress displacement and factors of safety as you create and iterate in the onshaped design environment with simulation built into onshaped Cloud native architecture you can build better quality products faster with mechanical feedback that stays in sync with your assembly throughout the product development process onshape is built in the cloud so it works just like Google Docs so an unshaped document is a single source of Truth For Your Design data onshape is always growing new releases are pushed to the product every three weeks so add new features and functionality this happens over the air automatically which means your company will never have to manually deploy a cad update ever again so Id highly recommend the engineers and product developers watching to consider using onshape for their business and you can try it for free at onshape.pro jamesbroughton right lets see what happens with this Auto cameraman project so this is my initial testing at getting it to move the camera and remember weve got three axes here so in the bottom left you can see the detection window and in the bottom right you can see the camera moving around so you can see that its zooming as I move further away and if I get wider the detection gets wider so it zooms out again if I come closer to the camera it zooms out as well so that seems to be working pretty well I have got some heavy motion filtering on here and you can see the cameras moving quite smoothly so Im actually filtering for motion on the input of those variables to the Arduino theres a PID controller proportionally moving the camera a different distance depending on how far it has to move and theres also motion smoothing then driving the servos so theres quite a bit of lag we probably need to tune it up a bit sharper but it looks good for now Ive also got this translation mode so instead of using the pan it actually drives the robot base around using the Omni Wheels theres probably not space in the room to use this extensively um just there where it nearly drove off the edge of the floor is in fact some stairs so I dont want to drive down there and I dont want to crash into things but nonetheless we can actually drive the wheelbase as well by flicking one of those switches and putting it in a different mode and the Tilt and zoom work just the same in that mode so Ive tuned up so its a bit more responsive so I dont want to make everyone seasick but if I move then its a lot quicker to catch me so that seems to work okay the zoom is just the same pretty much its a little bit more responsive but it doesnt sort of oscillate or anything like that which is gonna make everyone feel weird so lets go over here to the computer I cant help looking at that screen thats on the bottom there which is why Im not looking at the camera lets type some things on the computer if I make myself a wider person itll zoom out so you can see whats this drill lets do some drilling on something maybe Ill do drilling on the ceiling yeah just about in shot whats down here oh yes its open dog two here we go thats the emergency stop now Im wider so it seems out if I now come and walk to the camera it doesnt cut my head off for so long thats as far out as we can zoom so I was going to have a mode that made it drive further away when the zoom runs out but theres not really space in this room so I dont think I want it to do that Im pretty happy with the translation but actually Im just using the pen here and that works okay so weve now got something that can track me and it looks a bit like a camera filming me but it cant interact with a camera so we need some more features so Ive got a bank of three foot switches which Ive wired into a nineway drange and these are normal foot switches which Ive put on a D printed base and those plug into the front of the thing with the drange connector and that just goes to the Arduino so I can trigger some other features right so Ive got my foot switches plugged in so if I press the first one we can do something like whats that oh do you like hat jokes should I do more hat jokes great more hat jokes it is then yep thats literally a robot arm on the side of the tripod using a Dynamix or Servo to pop it up when I press the foot switch and youll notice the camera also zooms out so its in shot before going back to a bookmark value so you can pick up image tracking again and that leaves neatly out of the way underneath the electronics so it doesnt get in the way of the switch and the knobs and this is one of those nail training hands so practicing doing nails and I found that I can manipulate the hand into quite a few different Expressions that we could use in the future so one thing we can do without looking away shot is I can show you what Im doing down here by pressing the pedal and going down to that static shot set with the knobs on the front of the electronics on the tripod show you what Im doing down here for a bit keep talking to you press the pedal again and it comes back to the wide shot and it does that by bookmarking the tracking values that weve got now when I press that foot pedal it goes to the preset values and when I press it again it goes back for about two seconds to those bookmark values so we dont lose the tracking basically because when were down here it cant see me or detect me so we have to give it a couple of seconds going back to being able to see me before we can pick up the tracking again so the one thing I did have to do though was Implement a little filter to take out objects that are too small because theres quite a few things in the background which basically the camera interpreters being humans and so therefore when I was panning down and coming back again then sometimes it was shooting off to my left there to go and identify Tickle Me Elmo as a human or one of these mannequin heads or something like that and you can see the camera getting really confused there trying to focus on all of them at once so basically Ive said if its less than pixels wide then ignore it and as long as Im in the foreground then it tracks me fine Im pretty happy with how thats turned out I think its worked out pretty well and I feel like its quite organic actually the way it follows me and allows me to do stuff to camera and the cameras not just static so hes a bit like a person filming me I can of course not do that if I want so I can just switch this switch and then I can use the knobs to um assign any kind of sorts of static shot I want there wherever I want it and we can fix it in position so we dont have to have a moving camera all the time and of course I can just take the camera off with that quick release and go and put that on a normal tripod so Im going to publish a Canon code for this as I usually do which is on GitHub and the links in the description so if youd like to support me through YouTube channel membership or patreon then those links are in the description as well and YouTube channel members and patrons get access to all the videos up to a week early and sneak peeks and pictures of whats coming up all right thats all for now foreign Music

write a youtube script

write a youtube script

do you ever find that you have trouble getting rubbish in the bin because the Bins too far away or if you cant be bothered to get up the lids in the wrong position its a real problem affecting millions of people were ready to believe you so today Im going to show you how I made binbot the robot bin home assistant device that comes to you when you call complete with a lid thats open or shut when it should be I started with a bin its a pedal bin where youve got a pedal you push which opens the lid but sometimes the lid gets jammed and it doesnt go back till you bash it so we need to do something about that and we need to put the whole thing on Wheels so it can drive around so Ive got a bit of a frame to build and Im going to do that mostly with D printed parts Music just a quick ad from my D printing sponsor thanks to lolsbot for supporting my Channel with D printers Music and thanks to D fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com Im using structural aluminum here which is with a t slot so that I can screw on some D printed brackets which have got tnuts on so those drop in and turn to do everything up tight and Ive also got some metal Corners to hold them Square heres my frame and the bins going to sit on top like that but we still need to attach it and put some wheels on so Im gonna have a single Caster wheel at the back so it can run in any direction and two wheels on the side with these worm gear gearboxes which are pretty powerful so that means that I can have a two wheel differential drive and with that Caster fitted we can move the robot in any direction pretty much Ive got the wheels fitted which fit onto the mil shaft on that motor and theres a grub screw and a bolt to hold them on and of course theyre going to have a flexible TPU tire so that they grip the ground nicely so two grippy Drive wheels and one Caster wheel and the bins fitted on with more tslot bolts through the bottom and that means I can pick the whole thing up by the frame and move it around which is quite convenient we need to make a new lid opener mechanism and I dont like the original one so Ive made this D printed part thats going to go on the top and allow us to pull the lid open with a big Servo on the back of the frame Im just using a string to pull it which means theres no load on the servo when the lids shut and the servo moves degrees for the whole range of motion so we need some Electronics to control the wheels and control the servo so Ive got an Arduino Mega with a breakout and four switches so I can test the functions and that lives on the back of the frame along with a pocket to put some batteries in Ive also got two motor drivers which are the ibt or BTS which are rated at volts and these are volt Motors and the motors have wheel encoders which Im going to use so I can work out how far Ive driven so we can read those with interrupts and I got the code from the Arduino playground website so we can see the numbers updating as the wheel turns and that means I can drive a set distance and then program it to do something else foreign so all were doing here is driving forward until weve exceeded a number of encode accounts then turning until weve exceeded a number of encode accounts driving forward again until weve exceeded a number of encode accounts then opening the lid shutting it again and thats all it does the problem with this is though that its just dead reckoning which means were just using those wheel encoders to drive but if theres any wheel slip then eventually well lose our position which means its not repeatable to do the same thing over and over again and find its way back home so we need another solution I built some robots in the past which use lidars which is a scanning Laser Rangefinder so it spins around projecting a laser line and can work out where the distance is to all the objects so combining the wheel odometry which is the encoder positions it could drive rounds compare the two and basically draw a map of the environment and then use that map to navigate so it could also as well as just navigating from A to B update the map as it went so if objects got put in the way it could mark them on the map and when they got removed it could clear them on the map and it could do range finding and it could do path planning and do all of those things it needed to do to navigate round complex environments now this was using roles which is the Robot Operating System but I found that was quite complicated to configure theres a couple of videos of building this robot in my channel which goes into those details it is open source but the documentation is very widespread over multiple versions thats really hard to find out what configuration you need to use so for this project Im going to do something a lot more simple I am going to be using an Nvidia Jetson Nano though which is a proper computer running Linux with USB ports ethernet and HDMI and everything else youd expect now Nvidia provides some really good tools for retraining a deep learning model which can do Vision recognition so basically theres a little GUI and then you can highlight the object that you want to recognize and categorize it and then you can keep doing that for multiple objects and basically you can retrain one of the offtheshelf vision recognition deep learning models and this is called transfer learning so already knows how to recognize objects but you can insert some new objects essentially into its database using lots and lots of images and then at the end it will recognize them fairly reliably obviously its down to the data youve got but its pretty easy to train using the scripts that are provided and the tools which you can find on the Nvidia GitHub so Ive mounted the Jetson Nano onto my robot bin and Ive got a USB hub there so I can plug some extra stuff in and a little USB WiFi adapter because the Nano doesnt have WiFi on weve got a Raspberry Pi camera on the front of the bin there so you can see where its going another volt lipo which is dedicated for powering the Jets on Nano and Im using a to volt regulator because we need volts for the Jetson plugged into my USB hub is a USB to TTL converter which gives me serial data that I can take back down to the mega and Ive used the mega because its got lots of Serial ports in this case Ive trained for one target which Im going to use in this project and you can see that Im sending serial data out of the Jetson Nano and into the Arduino on the console on the right there is the Arduino serial console using debug to send me that data so as I move around we can see those coordinates changing which is how left and right and up and down I am and the width of it its better if I use the serial plotter though so we can draw pretty lines so you can see the red line changing as I move left and right so thats that value and if I move up and down then we can also see one of the other lines changing the green line so that seems to be working pretty well and as I get closer the blue line moves up and down as the width of the detection gets wider or narrower so I can draw some pretty lines there if I just wave the thing around and that seems to be working pretty well are you looking for a job in robotics if you are then you should listen to the ads from the video sponsor which is pipe dream pipe dream is building the city infrastructure needed to make super fast delivery near free and emissionless leveraging autonomous robots running through pipes underground these autonomous robots run on small underground tracks so theyre essentially building a tiny train system under a city theres lots of information on the pipe dream website about the concept including the robots themselves and the portal where the robots will be loaded and unloaded everything is based around inch pipes which is the same as current utilities pipe dream needs the best makers in the world and is looking to hire two or three more mechanical engineers even if you dont have a mechanical engineering degree theyre really looking for the best tinkerers roboticists inventors and makers of the world to apply go to pipedreamlabs.co apply with James and describe a project youve built recently no resume is required right lets see if we can get Vision assisted driving working on this bin so Ive stuck my target onto a tripod so I can move it around and do some experiments so the first one is were just going to put the target some distance away and then were going to set off the robot and using that serial data were going to turn until the target is in the middle drive towards it until its a certain width and then pop the lid and you can see the monitor in the background there is actually what the robots seeing thats plugged in with a long HDMI lead into that Jetson Nano so well just repeat that and that seems to be working pretty reliably even though theres other items in the background which are round like those round lights and various other things that Targets fairly distinctive though because its a circle in a circle and its got a big white border so hopefully thats pretty reliable we can now think about putting the bin to use in the kitchen so Im going to have two Targets one on the stand and one on the opposite side on the cabinets so Ive now programmed the robot so it will drive forward until it sees the first Target at a certain width then turn left now its just using dead reckoning using the wheel encoders to drive a set distance open the lid and when I press the home button it shuts the lid and it still uses dead reckoning to turn around degrees and drive forward until it thinks its roughly in the middle of those two Targets Music he then turns until the home Target is in the middle of its field of view and drives forward until that Target is a certain width in its field of view and then it turns around degrees but then slows down until it sees the Target on the opposite side until thats in the middle of the field of view and then it knows its a home so that seems to work okay it can also do the same thing and drive to the table by again driving forward until that Target is a certain width turning roughly degrees and driving for the set distance again this is using wheel encoders so its just dead reckoning pops the lid I can throw my rubbish in and then I can press the home button which of course makes it spin around degrees to head back to where it came from and drive forward a set distance with the wheel encoders again until its roughly in the middle of those targets and again its going to turn till it sees the home Target and home itself and now because basically its using those two Targets to always get home it means its quite repeatable so even if theres Wheels slip when its using dead reckoning as long as there isnt so much wheel slip and it doesnt get stuck on an object so it completely loses its position it can pretty much find those two Targets on home itself so this is what the robot sees driving forward until that targets a certain width and its not even trying to home in on that Target its just driving forward with default motor speeds which means it doesnt even go in a straight line but it doesnt really matter yeah the lid blocks the camera but there we go so turning back round were using dead reckoning here so its ignoring that Target driving for the set distance and then turning right until we see the home Target thats roughly aligned with the middle of the camera and it dries forward till its a certain width again not even homing in on that Target just driving forward just using default motor speeds but as long as it spins round until it sees the second Target in its middle of its field of view then it knows its roughly in the right position and the Arduino code for this is basically a state machine with a flag that gets incremented each time that we progress from one stage to the next and thats basically checking different conditions either the variables coming from the Jetson Nano to see how wide the image is some time expiring where theres a little delay or the encoder position being at a certain value the whole point in using Vision recognition as well as the encoders is to make sure this is repeatable so I thought Id run this several times and see that it does actually home and it doesnt creep and it doesnt get lost so this is five times speed but it seems to be working pretty well and you can see it pretty much always aligns itself with that track of tiles in the middle of the kitchen where that Target is this Music Music Applause Music thank you Music but its not much good if I have to keep getting up to press the button to make the bin come so we need some way of activating it remotely faster thank you left left more do you remember when I made that voice controlled car well that video was sponsored by Deep Graham this video isnt sponsored by Deep ground but Im still going to be using deep RAM for the speech recognition and deep Graham is a service that runs in the cloud which is trained using deep learning so its basically AI speech recognition and its pretty reliable and Ive still got some free credit in my free deep gram account you get when you sign up right left that is a separate script running on the Jetson Nano but scripts cant share serial Port so I have another USB to TTL adapter plugged into my USB hub the other end of those wires of course go to another serial port on the Arduino Mega and Ive also got a microphone which is one of these omnidirectional mics which is a USB audio device location one Music go home Music location two foreign Music Music thank you this project is ultimately a test for navigating just with wheel encoders and a little bit of vision recognition to help us on the way its much simpler than configuring the navigation stack and everything for roles what Id like to build next is the ultimate robot that can navigate through multiple rooms in the house and bring you a cold drink from the fridge perhaps we can use some image targets Ive also thought about putting those up on the ceiling or two of them on the ceiling so it can align itself so we dont have to pin these everywhere or actually using the machine learning to try and identify different aspects of the house Im not sure how tricky that will be so watch this space if you want to see an update Im going to publish all the Canon code for this as it is though and all of thats on my GitHub and the links in the description to this video so if youd like to support me through patreon or YouTube channel membership then those links are in the description to this video as well and patrons and YouTube channel members can get access to all the videos up to a week early including sneak peeks and pictures of whats coming up alright thats all for now Music foreign Music

write a youtube script

Alan pan recently made a video called giving snakes their legs back but I didnt think his legs were very snakelike so this is giving legs to their snakes back I found a video from akiyuki brick Channel with a very interesting looking Lego tentacle assembly it looks like it has quite a lot of motion and it only has four Motors each motor controls a stage of the machine driving what is essentially a turntable but with each motorized section set at degrees based on the combination of each rotary section the tentacle is able to do some quite interesting motions as it reconfigures itself there are lots of other really great builds in this channel so you should check it out a similar mechanism has been used in the RollsRoyce lift system to thrust Vector jet engine and you can check that article out on Wikipedia Ive built quite a few walking robots in the past but I really like the way the tentacle moved so this made me think about whether I can make robot legs that works in a similar way and how easy it would be to control them I started with a simple CAD design and some simulated joints ideally I want to put all the motors inside each segment and I want to use RC Servos so my sections are slightly more elongated than the original Lego build Im using these servos Ive used in a few other projects theyre pretty hefty and theyve got these metal brackets which makes it really easy to attach things to them Ive set up an Arduino Mega so I can control those servos Ive got some pots wired in so I can just position them for now by reading the analogues and driving the servos and Im using a s lipo because the servos need at least six volts so Ive got the first Servo attached to the base of the first stage of the tentacle so thats attached by its horn and that leaves the other piece attached to the servo to attach to the casing theyll attach to the next segment and thats just screwed on from the inside before I put the horn on the servo the housing for that segment fixes onto that and theres a piece built on the inside so the flat of the servo fits onto the flat of the housing now the plan was to screw that metal bracket on first and then I lift recesses in the bottom and the top so I could get to the screws to attach it to the servo the only issue is once Ive screwed the next section on then I wont be able to get to those anymore because the orange piece is going to block it and I need to put it together in a specific order so to help me with that Ive made a little shim that fits into this handy Gap here on the bracket and that means I can screw the housing on from the outside so that now rotates quite nicely and that leaves us with a nice piece at degrees to stick the base on for the next section so there it is in place I put a base on this to hold the whole thing stable as well for testing and you can see that second Servo moves so if we put the next segment on we can now see that weve got some quite interesting motions that we can make even with just those two segments so so far Im pretty happy with how this is working out if I keep building this up in the same configuration then we end up with a situation where the foot sticking out has quite a big overhang from the base of the robot and that means that any Force driven from the ground is going to skew all those Motors and put extra load on them so Id much rather turn this stage around so the foot is directly above the shoulder and the force is driven directly into the joint so for that reason Ive made an alternative piece which has the slant going the other way to the screw holes and that allows me to put the foot on the other way around so now the foot is almost directly above the base of the robot and now I can do all of these motions by moving all three segments now Ive stuck at three segments because thats generally how robot dogs and things work so that should allow us to have quite a good degree of freedom and so far Im not too unhappy with it so it looks like I can move all around and that should be okay to make a walking robot with any luck however if we try and sequence a walking motion then its a different story its actually very difficult indeed to pick the foot up and move it and then basically move it back again with it touching the ground so weve got three joints the same as a robot dog which can move in any direction and XYZ Cartesian coordinates but with these three joints weve got a lot of problems obviously my servos only move degrees but thats much more than the robot dog joints would move to achieve a better motion with the foot so it looks like weve got some problems here but before we carry on with that its time for a quick ad from the video sponsor which is jlc PCB jlc PCB are at the Forefront of the PCB manufacturing industry jlc PCB provide a OneStop service for PCB design PCB prototyping PCB assembly and also D printing its easy to order from jlc PCB just click on instant quotes and upload a gerber file the jlc PCP website will show a preview of the board and then you can select various options for manufacturing if you want the board assembled then you can add a bomb and CPL file jlc PCB currently have an offer where you can get five one to four layer pcbs for Just Two Dollars with free smt assembly and with the new gel C PCB parts manager you can check available stocks order parts or even request specific parts for your project jlc PCB ships worldwide and they have fast build times so you can get your pcbs in as little as three to five business days jlc PCB also provides D printing services from little as cents per gram so you can get enclosures and other parts manufactured in a variety of materials D printing processes include SLA mjf slm fdm and SLS in both Plastics and metals the ordering process is very easy at gel cpcb so use the link in the description to this video to check it out now so your typical robot dog has three joints it moves this way on its hip its got a shoulder and its got a knee and with those three joints I can pretty much move that leg into any position I want to and those joints are doing nothing like degrees theyre probably not even moving degrees and then we use a system of inverse kinematics which means that I can position the foot in any position I want to in XYZ Cartesian coordinates essentially and I can put that anywhere and then we use maths using inverse trigonometry to work back all of the joint angles so that we can position those joints in the right position to achieve that position on the foot Ive talked quite a lot about this in the build Series this is open dog and theres as the name would suggest various other versions this is a bit more complicated though theres probably some positions I cant even achieve for instance I cant move this down directly very easily I can try sort of twisting it but I dont really achieve getting that end it goes off in another rotary position position so working out inverse kinematics is going to be very hard and theres going to be no solution for quite a lot of positions so with this we need to be a bit cleverer but I also realized theyve been holding this wrong by trying to make this be completely upright having this foot directly below the base all these joints are at degrees so of course what we need to do is turn this base joint where the whole thing rotates to degrees as well so that the leg is more like this and then this faces the ground and this is degrees and all of these are degrees that does mean weve got quite a big overhang now which is what I wanted to avoid so Im going to take this section out turn it back around the other way and turn the whole thing upright this way so turning around that last segment so all the segments are identical weve now got something that looks a lot better and if I run a sequence of motions on those which Ive just hackerly put together we can almost get a walking motion so the foot will scrape along the ground to push the robot back and then it will lift up and move again and pretty much clear the ground now we need to adjust some positions here but thats looking pretty convincing to me so Im pretty happy with that but obviously to make a robot which is going to walk along it needs more than one leg so I better get printing to make the robot really stable for this experiment Im going to be making six legs so that means making housings and of those gray bases and all the other parts but fortunately Ive got more than one D printer so just a quick ad for my D printing sponsor thanks to lulsbot for supporting my Channel with D printers Music foreign Applause fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com yes I made six of them which took quite some time so thats servos in total and Ive put these brackets on the top so I can attach each one onto a piece of Extrusion to build up a structure to mount all six legs Ive got three legs on each side and some of these Bridge brackets between them and that means Ive got something that fairly convincingly looks like a hexapod with all six legs and its pretty stable so if I twiddle the knobs we can see the first axis moves at degrees the second one moves parallel with the ground and the last one moves at degrees again so trying to do this manually to try and make a walking motions actually pretty tricky I can just about pick up the foot and move it and try and move it from one side to the other without scraping the ground but really I need to spend some more time trying to sequence this to make a convincing motion so first go at a bit of a sequence to try and lift up the foot and move it which it does quite convincingly although it makes quite big motions so with the other feet taking steps as well theyre probably going to hit each other so Ive just sped that up a bit and also reduced the motion so now weve got something where the foot tips back as it moves forward and then it hits the ground again as the foot moves backwards to push the robot forwards so pretty happy of how that looks so far what we really need to do is wire in all of these and sequence them up and see if itll hobble along youll notice one thing is that the foot wont stay parallel with the ground as it moves and thats purely because it cant reach with this mechanism so with some of the previous robots Ive made generally that Servo and the mechanism is directly on the joint so if I push this back I can back Drive the servo and these are geared down two to one as well so all take the load however with this system I can push down pretty hard on here and it doesnt back Drive the servos which is really good so the robot is pretty big the legs are quite long for the strength of those servos and thats because weve got this rotary motion instead of the legs directly back driving the load of the servos on the load of the robot so I fitted the battery onto the robot now and the Arduino and also a whole lot of wires on a breakout board and I made that on stripboard to wire in all servos and Cascade all the power out yep theres loads of wires obviously one for each Servo and one for each joint but if we sequence up a walking motion its not looking too bad it took a while to get all of these legs coordinated some of them do the same thing and some of them move in the opposite direction and then in the opposite direction again once theyre on the other side of the robot but thats looking pretty good to me so it does hobble along it reminds me a bit of a sloth the way it moves its legs I guess that might be the pointy feet but we can see hopefully that the legs are lifting off the ground as it locomotes and its not doing a bad job there obviously those legs dont move in a straight line along the ground or when they pick up and thats just because of these rotary motions and my inability to pretty much do any inverse kinematics or move the feet where I want Im pretty constrained by the structure of the robot here and the way these joints rotate I did reduce the stride length here slightly so the legs dont hit each other and also as they move in an arc it means the robot Bobs up and down if the legs move too far so thats not looking too bad its not looking too unstable although it could be much better of course if it had inverse kinematics and those feet would move in a straight line along the ground for that reason its also not the most efficient it could be at walking you can actually see the legs look like theyre slipping slightly as they move back and thats because the front leg picks up and sort of catches the ground on the forward Strokes is quite a lot of friction with the carpet you can see the legs are actually picking up and clearing the ground mostly even though theyre slipping back at the end of the backstroke so Im not too unhappy with how this has actually turned out at least it functions on a smooth floor its actually a much better story I guess theres less friction so when legs are definitely pushing it Forward by pushing the legs back then it goes forward and otherwise any slip on the legs pushing forward any friction on that forward stroke slides along the smooth ground a lot better so that seems to work actually much better on a smooth surface than one with more friction even though the feet are smooth now Ive only coded it up to walk in a straight line if I wanted it to turn on the spot properly then we really would need those inverse kinematics and the way we would do it is to basically plot a circle from the center of the robot move the foot around the circle in whatever position we need to and then work out the X and Y coordinates from the old position and the new position and use inverse kinematics the ability to move those legs in XYZ coordinates to move the foot to the right place and then move the legs off the ground in the opposite position and Alternate them like walking and make it turn so for now Im not going to program it to try and turn on the spot or anything it does walk in a slight Arc but that due to there being friction on one side more than the other and thats probably due to the fact the servo positions arent the same between all of the servos so they would all need calibrating if I were actually doing it properly so one drags more than the other well I took it for a walk in the garden and it doesnt seem too unhappy about that cant really see if its picking its feet up or not but it seems to work okay Music Applause Music Im not too unhappy with that Im mainly impressed by how tiny these servos are relative to that mechanism and how big the machines turned out and how it still works well and thats all owing to this rotary mechanism and not putting the joints directly on the servo to back drive it or cause too much force of course now we cant move the legs where we want so we cant warm properly and its probably useless for a manipulator as well to move a robot arm where you want of course it would work better if we had lots and lots of these Im not sure how many youd need to be able to solve every inverse kinematic position probably more than five Id estimate though or if you had a really long arm then it might be quite good it might get quite wobbly so I think theres definitely some potential here for the future which needs investigating for now Im going to publish all the cad and code as I usually do on my GitHub and the links in the description to this video so so he loads to support me through patreon or YouTube channel membership then those links are in the description as well and patrons and YouTube channel members can get access to all the videos up to a week early including sneak peeks and pictures of whats coming up so you can be part of that discussion all right thats all for now Music

write a youtube script

in this video Im going to speculate on how the animatronic mechanisms work in B emo from the new Star Wars Andor and also try to build my own version there are no spoilers for the Andor Series in this video but obviously Ill be looking at the B Droid in depth B appears to be a practical effect rather than CGI a real B was featured at the recent premiers of Andor and several influencers have posted footage on social media as far as I can tell this was built by the same team that built bb which includes Josh Lee and Matt Denton and you can see the bb internal animatronics reveal from Star Wars celebration in the Star Wars YouTube channel at first B doesnt look like anything special its just a beaten up looking Salvage Droid during the Andor Premiere however the first thing I noticed was that it had mechanum Wheels which makes it omnidirectional Ive done a few projects with standard Omni Wheels in my channel but never mechanism wheels these wheels have little wheels around their circumference which are angled at degrees mirrored pairs are usually installed on the robot and depending on the combination of wheel rotation on each wheel the robot can move in any direction B is even more complicated than that though each foot can lift up and down independently and also the whole wheelbase can extend to make it longer and presumably more stable there are also a number of other features in the body which can boost itself up to make the Droid taller there are rotating parts inside the body and the head can all also move independently in multiple axes and all of this is a real piece of engineering so I made a simple CAD model to explore the drive mechanism of the Droid each foot has a wheel which must have a motor for the mechanum wheels to work properly then each foot can slide up and down independently so the Droid can lean in any direction or move its whole body up and down so thats another four Motors lastly the wheelbase can extend so thats probably another powered axis although this could be achieved by just driving the back and front wheels in opposite directions and keeping them there by using wheel encoders to regulate the wheel velocities and positions so they dont creep as the Droid drives around looking at the internals of bb that were revealed and the size of the motors for the animatronic motions in that such as the head Id guess that some expensive robotic servos are in use in B like the dynamixels Ive used in some previous projects courtesy of Robotics the sliding motions are probably servos like these in continuous position mode driving lead screws and linear rails to move the axes then there would need to be some inverse kinematics going on to take into account the wheelbase Lane length on the height of each foot to drive the droids body to a specific angle in pitch and roll this is really similar to my robot dog projects which can keep all of the feet planted on the ground while the body can rotate in both pitch and roll I dont really want to build this with so many expensive servos though because you can build almost anything with enough money Ive built quite a few Star Wars droids in the past including my own bb which use scrap windscreen wiper motors from cars and those only cost two pounds each and my walking Gunk dry which had a number of small cheap Motors in it so Im going to look at an alternative approach to building B and achieving similar motions with budget parts mechan and wheels are quite readily available although they tend to cost quite a bit even for moderately sized ones let alone the big ones so I decided I design and make my own with D printing each of my mechanon wheels is going to have rollers around it so weve got rigid cores and TPU tires so we get some grip on the wheel weve also got the two sides of the wheel to make and a couple of other bits and pieces so just a quick add from my D printing sponsor thanks to lulsbot for supporting my Channel with D printers The Hub or the two sides for this wheel are made in two sections which screw together and that gives us the complete wheel with axle holes so we can fit those rollers at degrees and the rollers have a rigid core a TPU tire and those just push fit together and youll notice theres a slight raised part in the middle of the core so they grip that of course spins on a four millimeter piece of Steel so it runs nice and smoothly and sticking of those into one of the wheels gives us our mechanum wheel with rollers all at degrees so Im pretty happy with that thanks to Simply bearings for the bearings for this project theres one of those on each side with an eight millimeter internal diameter and I put a drive pulley on here as well and yes I remember to make mirrored pairs of these for each corner of the robot so that we can move just like the example earlier in the video if I turn one of these on its own then it moves a bit like a screw so we get that sideways motion and if youre in any doubt heres some footage of B emo moving sideways at the premiere and also rotating so its a truly omnidirectional robot moving on to the rest of the drive system Id really like to have the feet slide up and down just like the original however this would involve some quite compact actuators which will ultimately be costly but then I remembered a previous build the PED rail wheel vehicle this had wheels of feet which helps the machine climb over obstacles I mounted four of these wheels on a chassis with fourwheel drive and suspension which worked pretty well as one wheel finds an obstacle it moves upwards and the whole thing averages itself out theres no active control in that build but what about if we built a hybrid passive active suspension system Ive still got the chassis from that build if I bounce it on some rolls of tape we can see that its pretty responsive to being pushed around which feels like quite a natural movement if I take one piece of bungee cord off and lift that corner then the other suspension arms comply with wherever I put it only takes a little bit of force to move it around so what about if we balance the Droid on some suspension and then instead of needing powerful actuators that can lift the whole Droid we use small the cheaper actuators to coax it around all of the wheels would comply with the ground so we dont need to calculate any complex inverse kinematics either there is one issue here which is as the feet move up and down theyll get closer and further away from the body however this distance is less than three millimeters so it would be easy to ignore and once each wheel is on suspension the back and front of the Droid can slide on linear rails to extend the wheelbase so its time to get printing again to print each of the four feet for the Droid the suspension arms and also those Central sections with that sliding core for now Im just going to build the drive mechanism and well come back to the rest of the body in a later video so weve got two sides for those boxes for fitting the feet those are the central Parts which are eventually going to have linear rails sliding on vwheels and also the suspension arms thanks to D fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com heres one side of my foot and the wheel slots in just like this onto some late more studding thats bolted onto each side to help hold the whole thing rigid if I slide that foot around we can see that the mechan and wheel turns when I move it sideways and backwards and forwards and that explains how a combination of these motions makes the Droid omnidirectional for now Im using some pretty small six volt motors with gear heads Ive used these in quite a few projects and theres quite a lot of choice out there for similar Motors so with a belt drive onto the Mecca and wheel that seems to run okay Im just running that on a s lipo and yes of course we need four of these with opposite pairs of mechanon wheels for the whole Droid but before we carry on with that its time for a quick ad from the video sponsor which is on shape onshape is a cloud native Canon PDM platform built for business created by the founders of SolidWorks because they saw that modern product developers still experience many challenges related to their Cad and PDM systems since onshape was built from scratch in the cloud there are many unique advantages like builtin PDM on the GitHub inspired branch and merge model to test new ideas onshape is accessible across all operating systems and works like Google Docs so an onshaped document is a single source of Truth For Your Design data one shape is always growing new releases are pushed to the product every three weeks to add new features and functionality recently onshape added PCB Studio to connect ecad and mcad designs and they also just acquired a company called Cloud Milling which means cam is coming to onshape in early . all of these updates happen over the air automatically which means your company will never have to manually deploy a cat update ever again onshape also just released the connection feature with Arena PLM which synchronizes Engineers manufacturers and suppliers enabling the instantaneous sharing of product design information at the click of a button I highly recommend the engineers and product developers watching to consider using onshape for their business and you can try it for free at onshape.pro jamesbruzen right lets get the rest of this together the next Parts were going to look at are the central sections which slide on linear rails so those have got v slot rollers fitted in there and some v slot Extrusion that fits nicely in the middle and slides backwards and forwards and then my suspension arms link all of those sections together at the moment Im just using some bungee cordless suspension Springs but Ill need to replace that with proper Springs in the future and as the mass of the Droid builds up thatll need to be correctly rated the four suspension arms allow the Droid to lean backwards or forwards or left or right and or in any direction really basically by moving any combination of those suspension arms and that gives me quite a lot of motion so Im pretty happy with how thats turned out so far and my sliding section allows the wheelbase to extend so that we can make it longer or shorter just like the real Droid the only problem with this is that the central piece doesnt stay in the middle it slides up and down by itself which is going to cause a bit of a problem its impossible to match the friction between those two rollers perfectly on the two sliding pieces so generally that middle piece will stick on one side so to make sure it always stays in the middle Ive got a new piece fitted to another piece of Extrusion and that consists of a spur and two racks and that means that itll always pull out and theres the spur rotates itll keep that middle piece perfectly in the middle at all times no matter what I do so this seems like a good solution there is one challenge that those racks are quite bendy and they might pop out of the gear so to stop that happening weve got another bracket that fits on which is actually holding the servo thats going to move the suspension and that constrains the racks in place so they cant pop out from that spur gear in the middle and the held nice and rigidly so that seems to be working pretty well but now to get it working we need to put some Electronics in for now Im using an Arduino Mega to control this which has got enough I O for the servos and Motors although as we build the Droid up were probably going to need something with some more pins on as usual Im using the nrfl radio chip so I can use the open dog remote to control this but Ill probably build a custom controller in the future and Im using a USB boost Bank to power the Arduino and a s lipo to power all of the motors and servos each motor has a BTS motor driver which is complete Overkill but itll be fine if I ever have to upgrade the motors so the right hand stick on the remote makes the wheels go battles and forwards as youd expect as I push it backwards and forwards and also goes sideways using those mechanum wheels and well put it on the ground and demonstrate that in a moment if I turn the stick then it will turn on the spot by running the left and right Wheels in opposite directions and if I switch this switch on the remote and turn the other stick then it drives the wheels in opposite directions backwards and forwards to extend the wheelbase the servos Im using here are hightech hs BB Plus servos which cost me about pounds each so theyre not too expensive obviously theres four of those to operate each of the four suspension arms and each one of those is attached with a piece of string which goes slack of course if I push it manually and that means if the Droid goes over a bump then it doesnt smash the servo to pieces and that Servo isnt directly supporting the load of the Droid so now if I switch that switch to the other position and turn that left hand stick instead of driving the back and front apart it now pulls all of those servos down I can also push the stick backwards and forwards and left and right to operate those servos in pairs and those numbers all mixed together so I can do any combination of motions with rotation or moving the stick diagonally for instance whatever I do the chassis complies with the ground due to that bungee cord and the loose strings attached to the servos if we stick axes has motion filtering on and I did a video about how to make robots move smoothly these are just cheap servos and basically an Arduino Uno controlling this head so check that video out basically it just means that the value doesnt change too quick and that gives a lovely smooth motion so we get deceleration as we get towards the targets so its time for some tests on the floor driving backwards and forwards and driving left and right with those mechan Wheels which is working pretty well and rotating and of course I can mix all of those motions together at the moment Im controlling all of this with the right hand stick so it takes a bit of getting used to normally I put the rotation on the left hand stick but we need that for the other two functions and I really want to keep that separate yes we can spread that chassis apart by driving the front and back wheels in opposite directions with that left hand stick and that seems to work pretty well now theres no motor actually driving that independently and theres nothing to lock it in place but it doesnt seem to creep too badly the motors are just being driven with more or less current to vary the velocity I could use the encoders which are actually built into them to drive the speeds precisely that would mean that it would perfectly stay with the chassis spread apart altogether but for now that seems to work okay in current mode so Im pretty happy with that so lets have a look at those servos Lets Twist that lefthand stick first of all so we can move the whole body up and down and we get about millimeters of travel if I push the stick backwards and forwards then those servos operate in pairs so that we can bend backwards and forwards or we can lean sideways and that seems to be working okay obviously I can use both sticks at once and I might put some Auto mixing on this so it always leans in the direction of travel but that seems to be quite organic especially with that motion filter and of course that works sideways as well having those sticks independent is quite useful because I can also Bend away from the direction of travel just by doing that manually if I want to so Im really thinking about a custom controller here with some extra joysticks or some extra kind of thumb sticks or finger sticks as well as the two main joysticks so that I can control all the functions once the bodys on and it extends up and down and the heads on then we can have all sorts of things that I cant control by myself so were probably going to need some intelligent mixing to make some of those functions automatic in terms of scale my body width is going to end up being about millimete

write a youtube script

Tesla recently announced their humanoid robot product called Optimus it can walk around and do stuff and Elon Musk says its going to sell for twenty thousand dollars which is cheap I mean the parts in there cost a lot more than that not to mention the r d costs and we still havent seen the launch of the Cyber truck but how hard is it to build a robot that walks on two legs Ive built quite a few walking robots in the past if you head right back to the start of my YouTube channel there are quite a few large robots that could just about hobble along made out of wood and driven by windscreen wiper motors and other parts that I could get hold of I built these earlier this Century in a room of a house here I used to live in which is why theres a bed in the corner of the shot Ive moved on quite a bit since then to some robot dogs that Ive built recently but a few years ago I built some bipedal robots like robot X and my walking gonk Droid these actually had a sense of balance using an inertial measurement unit to read how far they were leaning over and make the motors respond accordingly to keep them upright that can be quite tricky though and and most of whats on YouTube is the good cuts of footage where they work just about okay Ive taken some of these projects to events like make Affairs and youll always find that one person oh hello I was watching your YouTube channel with my dog and I did a stroke of Genius green Millions think alike James why dont you make a simple robot that just all it does right is lean that way right moves that leg okay then it leans the other way right moves that leg and then leans the other way right then moves that leg and then moves leans that way and and moves the other leg so in this video were just going to make that simple sounding robot that shifts all its mass One Way moves the other leg and shifts it back the other way and moves the other leg and then it will walk along Im gonna make this as simple as possible using a lot of linear sliders each foot can move up and down in a straight line each leg can also move backwards and forwards in a straight line to take steps and theres a carriage on top that moves side to side with a lot of Mass on it theres quite a bit of Hardware like Extrusion and nuts and bolts and so on in this project but the majority of it is D printed so just a quick ad for my D printing sponsor thanks to lulsbot for supporting my Channel with D printers and thanks to D fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com the majority of my sliding rails are basically V Wheels which are attached to some D prints with some bolts those have got bearings in and that allows v slot Extrusion to slot in and that makes a sliding piece so Ive got a foot with some rubber feet on and weve got some linear axes that are attached to those to allow that foot to lift up and down the various parts to cross brace it including some plates on the back and the front so theres our foot with our linear slider on and of course that v slot still fits in and thats the main mechanism thats going to make those feet lift up and down so it can take steps those fees are going to be driven up and down with a bit of lead screw and a lead nut which goes on there and obviously it goes up and down as it rotates this bit of lead screw came out of an old project which is why its quite dirty looking but it seems to move okay and of course theres a coupler on there and Im just going to drive that with a DC motor with an encoder so that we can drive the screw and position it that fits into the extra little block on the foot so as I drive it up and down then obviously that will drive the foot up and down so my motor fits on the vslot itself and goes into there and that means as the motor turns as youd expect the vslot moves up and down which we can attach to the next piece of the leg one thing I didnt mention earlier was that we actually need to make the legs rotate so that we can turn on the spot and walk around in circles essentially rather than just going in a straight line so that piece of Extrusion fits into a bearing with some D printed bearing blocks its double braced with two bearings to hold it nice and securely and that allows the leg to rotate and youll notice theres some more V wheels for the next linear slider which will come on too shortly so Ive made two of those legs which both rotate and somehow we need to actually make them rotate those so they dont just go wobbly and to do that Ive attached a little lever and a Servo onto that v slot and the servo onto the mounting there so with a little push rod in between that means as we move the servo we can rotate that and we can lock it in place when we want to so it doesnt just Freewheel and have the legs twist as its trying to walk so Ive put those on the middle there and made sure they dont knock each other as the legs slide past each other by putting the levers at an angle so with some more Extrusion this time so we can hold it nice and rigid weve got the next sliding axes which is going to slide those legs backwards and forwards those s fit into some more D printed blocks and Ive left lots of holes there for some tnuts and bolts to attach them and those blocks are fitted to some more which makes up the frame of the robot and itll also allow a place for that Carriage to move side to side so you can see my legs have quite a lot of movement there to move backwards and forwards and thats how its going to take steps so we just need that Carriage now which is going to move side to side to shift the mass and again that has some V wheels at degrees to each other so it slots into the vs on two sides of the Extrusion on either side of that frame and because of the V Wheels which are horizontal that locks in nicely so I can pick the whole thing up by it and it doesnt fall off I fitted some more D printed Parts on the end of those rails which serve two purposes one is as an end stop to stop the carriage falling off and the other one is Ive got somewhere to tie a piece of belt on so I can actually drive this side to side to do that were going to be using another one of these motors with a D printed T pulley on and that pulley has a captive nut and a little grub screw to hold it onto the flat of the shaft which is six millimeter on these motors Im using open length T belt which is about six millimeters wide and thats going to go around some idlers and onto some little clamps on those D printed end parts so now as that slides up and down the motor will turn and as the motor turns it will cause the carriage to slide up and down to shift Mass side to side and those idler bearings are just a pair of skate bearings roughly with an eight millimeter internal diameter those legs of course need to be driven backwards and forwards so this is the sideways shot of the robot and to do that Ive got another motor and another idler so its a bit like the carriage of a D printer Ive made this clever little belt clamp that we can put two bits of belt in or at least the two ends that wrap around theres a little block that locks in there and stops that pulling out so with both open ends of that belt installed we now have a belt drive with another motor and thats the same on both legs The Carriage has some platforms on which Im going to install electronics on and probably also put some Mass on we can put some batteries in there although lipos tend to be not very heavy because theyre designed for drones essentially but I can actually roll up one of the feet now and move that Carriage over and test whether its actually going to balance so yep theres definitely a gap under the foot and it will just about balance on one foot even with no additional Mass on that Carriage however yeah its pretty sketchy it will probably topple over especially if we try to move fast so Ill probably install some Mass a bit later on and see how that goes but for now its looking pretty hopeful I can slide the legs back with some forwards as well and it doesnt tip backwards or forwards so thats looking pretty good but before we carry on with that its time for a quick ad from the video sponsor which is established titles now what Ive got here is a proclamation that says my name is Lord James Bruton and Im going to tell you how you can get one too established titles is a fun and novel way to preserve the natural Woodlands of Scotland while helping Global reforestation efforts the project is based on a historic Scottish custom where landowners are referred to as lards or in English Lords and Ladies the title pack gives you at least one square foot of dedicated land on a private estate in edleston in Scotland and an official certificate with a Crest your certificate features a unique plot Number with which you can see the exact location of your land established titles plant a tree with every order and work with global Charities one tree planted and trees for the future to support Global reforestation efforts you can officially change your name to Lord or lady and get it put on your credit card or anything else pretty much they even have couple packs that come with adjoining plots of land the first people purchasing a title pack using my link will effectively be next to my plot and depending on how many of you want to become a lord or a lady we can build our little James Bruton Kingdom it makes an amazing last minute gift and established titles is actually running a Black Friday sale plus if you use the code James Bruton you can get an additional off go to establishtitles.com jamesbroughton to get your gifts now and help support the channel Im using the encoders that are built into these motors and theres going to be five Motors and therefore encoder wires I need to read those who have interrupts in the code to make sure that thats very accurate which interrupts everything as the name would suggest so Im using a Teensy . which is a very fast micro controller so we can deal with reading all those encoders and not missing any steps there are a lot of wires to wire in all of the encoders the motor drivers which are BTS s to actually drive the motors with their pwm wires and the power that needs to run to the encoders and motor drivers and the servos Im using some code from the Arduino playground website which uses two interrupts per encoder because theres two wires and that gives you the most accurate answers and you dont miss any encoder stamps so you can see all the values updating there as Im turning the encoders so thats pretty good and it looks like theyre all working together so now I can position that Carriage using the encoder value and driving it to a demand position for now Im just using the remote control for open dog that I showed at the beginning of the video and for now its just trying to get to the position as quickly as it can I move on the stick but thats rather jerky so Ive moved on and written some Arduino code to make a step sequencer and Ive also got some proportional control so as the distance between where the encoder is and where it needs to be get smaller it decelerates and that means I get a nice smooth motion and no overshoot so you can see that motor is turning away nicely there and you can see its decelerating as it gets towards its targets Ive also got some motion filtering on here which I talked about in another video and that video was just using cheap servos and an bit Arduino Uno to make an animatronic head and you can check this out on my channel Ive actually got some additional motion smoothing here which basically takes a large percentage of the old value in each Loop and a small percentage of the new value thats coming in from the remote that means the value cant change too quick and so we get a nice smooth motion you may have noticed Ive also added some extra mass in the form of two one kilogram lead diving weights I got off Amazon so now theres definitely enough Mass to over balance the robot onto one foot while we pick up the opposite foot Im using the encoder control again on those feet Motors to turn those lead screws and Ive used lead screws here rather than belts so theyre not easily back drivable they can be back driven a bit but with the motor holding power in place that robot can definitely hold the mass of itself plus that additional lead that we now have which is another couple of kilograms spinning backwards and forwards on top so that was an early design decision to make sure this thing would definitely work and it looks like its going to be a okay so now its time to probably use those sliding axes to slide the legs backwards and forwards and move them in opposite directions when one foots off the ground and hopefully we can take steps so lets give that a go we can move the mass to the right there and move the foot up and slide those legs in opposite directions and then we can take a step and now our robot walks along just like we wanted it to its pretty slow though and well talk about that in a moment but essentially it cant move too fast because then it will over balance because of that huge inertia from that mass moving which would basically be enough to tip it over so it is statically stable that means it hasnt got any sense of balance and that basically means that we have to go quite slow to make sure the mass is over one foot before we do anything Ive also Incorporated the servos here so we can rotate the legs although you can see that theres quite a bit of wobble and thats mainly down to the fact that theyre holding that whole rotation with those D printed levers and the servo itself which only has a certain amount of holding power we have all that Mass on top that we need spinning battles and forwards and you know essentially being an offcenter load from where the rotation is so we definitely get some wobble there and again we have to go quite slowly otherwise weve got no chance of it working whatsoever foreign Music difference Music well it looks great in five times speed but why cant we make it go faster in real life the problem with accelerating Mass fast is that mass has inertia which means it wants to stay still if I move a large Mass from a stationary point or move it repeatedly side to side then it causes a reaction force that pushes me in the opposite direction so as well as the mass of the robot being enough to offset the mass of the opposite leg while it takes steps if we accelerate it enough it would also cause enough of a reaction that it would tip the robot over to manage that were back to measuring the angle of the robot and dynamically controlling the velocity of the mass to keep the robot from over balancing as well as all the other axes which are essentially Mass moving around my robot axol Gunk droids look like they were moving more fluidly although they were often temperamental and thats when you realize how much work has gone into robots from Boston Dynamics and other large companies although they still have their issues at some point Id like to build a bigger bipedal robot again with some of the things in it that Ive learned from doing the robot dog projects like the cycloidal drives those back drivable axes and a bit of dynamic stability and the inverse kinematics mainly but for now Im pretty happy with that for a twoweek project Im going to publish all the cad and code and thats on my GitHub and the links in the description to this video so if youd like to support me through patreon or YouTube channel membership then those links are in the description as well and patrons and YouTube channel members can get access to all the videos up to a week early and sneak peeks and pictures of whats coming up alright thats all for now Music

write a youtube script

youve probably all seen AI generated images by now there are quite a few videos on how this works but simply put an AI is trained on billions of images from the internet and you can make a draw what you want from a text input so if you want an image of Donald Trump fighting Darth Vader or a polar bear riding on a paddleboard then you just type it in and how you press though or make some nice images its particularly good at drawing portraits of people who are really famous and Ill have a pretty good go at some midlevel celebrities too even some YouTubers its kind of funny how all the pictures of Colin Furs have a crazy invention or a funny looking car in the background though there is this type of image generation of any practical use to us besides generating art or funny pictures to laugh at Im using stable diffusion for this project which is open source and you can try it for free and Ill put the links in the description for the actual images in this video I use dreamstudio.ai and thats because its the same algorithm but much faster you get some free credit but after that it was about ten dollars for a thousand image Generations Ive been thinking about building something in real life that AI draws for a little while and I posted some of my previous results on Twitter and Instagram it started with a human making machine and a dog wash machine but the first results were too weird and I dont own a dog I also tried inputting words around practical things that could be useful like machine that helps humans make food and domestic robots most of the machines it drew just looked like something from a factory and most of the robots look pretty generic I guess thats what all the Souls images were like and thats what it was trained on so thats what we can expect to get out after some more experimentation I found that the phrase experimental robotics equipment came up with some absolutely fantastic looking stuff although it was hard to determine what the machines would actually do if I tried to build one so using that phrase combined with some other things I got some more interesting looking project ideas youll notice that whenever theres a person in the image theyre typically a white male but thats the result of the images the AI was trained on so I guess its a reflection on the tech industry and lab environment where youll find experimental robotics equipment so its societys crime rather than the AI itself it only knows what we tell it after all I settled on experimental robotics equipment for playing music which made some crazy looking stuff there are plenty that are obviously based around robots playing keyboards or other things but my favorites were the ones that didnt look like traditional instruments and I settled on AI telling me to build this one by the look of it its got five things you can grab which would match a pentatonic scale quite well so this seemed like a good start there are also some caster wheels as well as some other junk around the base which Ill probably ignore its a pretty basic shape to draw in CAD but what will the AI tell me about its hidden features Music foreign just a quick ad from my D printing sponsor thanks to lolsbot for supporting my Channel with D printers Music thanks to D fuel for the filament for this project and lots of other projects so check out my channel for more D printing projects and check out dfuel.com before I try to assemble what AI had wanted me to build I wanted to think about what other features there were specifically the handle looking things that hold the ram grabby bits so I asked AI to show me equipment handle operating instruction and it looks like most of the images describe some sort of rotary action so I designed some parts to measure rotary position using magnets and hall effect sensors these consist of two parts that rotate against each other one with a hall effect sensor in which is a little Electronic Component which can sense how far a magnet is the other part has a magnet in and then we can get an analog signal that we can read with an Arduino that tells us the distance from the magnet so if we get a magnet and put that near there and read an analog in we can see that we get quite a nice smooth curve and its very accurate depending on the distance if its really far away it doesnt work at all so we need to be quite close but this is a sort of thing thats used in twist grips on ebikes So the plan is we can move these levers of which there are three in total on the device and the whole effect sense will measure the position of each one so if we plug that into the Arduino we can see that we get quite a nice curve there and it seems pretty accurate to a point right at the top it peeks out but thats good enough for our purposes the squashy piece is attached to each handle look a bit like oldfashioned car horns and I was going to use oldfashioned car horns but I decided to ask AI instead it came up with some pretty cool stuff that probably could have done with some further investigation but in the end I decided just to make the simplest thing I could to squash a magnet near a hall effect sensor so I just split the ball in half in one half Ive got the whole effect sensor and two Springs and on the other half of course is the magnet so the two holes fit together and you can compress the Springs to push them closer together and they spring apart naturally but I really wanted it to look like the ones that AI designed so I decided that we need a rubber skin to go over there so I got some of these stress balls that have these little water balls inside cut them open and got the balls out so that were just left with the skin and those things are pretty stretchy so its really easy to put them all over my D prints and that fits pretty snugly and that gives us something which we can squash and that will hold the whole thing together as well now that rubbers pretty grippy and that means some bits of blue plastic have got stuck in there but if we go back to the original AI drawing its got a little Flex in so its almost like hey I knew that I wanted to do this the main enclosure for this fits together in two halves Ive got three handles and five squashy things which are there to play the five notes of the pentatonic scale or at least thats what I think AI wanted me to build and the levers lift up as well all three of them two of them will actually stay in position when you put them there and the other one wont just due to the angle that theyre fitted on there yep I decided to put the legs on with casters on and these are casters off office chairs which arent terribly great in all honesty but that seems to be what AI wanted and I think those are there so we can spin the whole thing round to access those squashy things and access all the handles so thats probably quite a good piece of design that the AIS invented but before we carry on with that its time for a quick ad from the video sponsor which is on shape onshape is a cloud native cadmpdm platform built for business created by the founders of SolidWorks because they saw that modern product developers that experience many challenges related to their Canon PDM systems since onshape was built from scratch in the cloud there are many unique advantages like builtin PDM and the GitHub inspired branch and merge model to test new ideas onshape is accessible across all operating systems and works like Google Docs so an onshaped document is a single source of Truth For Your Design data onshape is always growing new releases are pushed to the product every three weeks to add new features and functionality recently onshape added PCB Studio to connect ecad and mcad designs and they also just acquired a company called Cloud Milling which means cam is coming to onshape in early all of these updates happen over the air automatically which means your company will never have to manually deploy a cad update ever again onshape also just released a connection feature with Arena PLM which synchronizes Engineers manufacturers and suppliers enabling the instantaneous sharing of product design information at the click of a button so Id highly recommend the engineers and product developers watching to consider using onshape for their business and you can try it out for free at onshape.pro jamesbroughton so theres a lid that fits on there but before that we need to put some Electronics in so yes weve got the same Arduino Mega and thats fitted into a box with a midi shield and a midi Shield of these ports on the top here which are five pin dins and these are the sort of things that you get on musical keyboards that allows you to convey musical data so we can control musical instruments from this so if its just in there which is quite neat and Ive got a breakout there to take all my analog ins and get my midi outs so with that all assembled weve got our ports neatly presented at the back now midi is only data about the notes we want to generate so to make actual sounds we need to plug it into some sort of synthesizer and Ive got a little synthesizer here thats got midi in and audio out and it can make the sounds of various instruments so we can select from a little dial on the front thank you foreign Music I Put A Ring of LEDs on the front just like the original and as I lift the lever the LEDs change color but as well as change color this is actually an octave shift function for the music Music foreign as well as the pitch shift lever weve got two other levers that we can use as separate midi controllers for Midi effects but before we demo the whole thing out I think we need to find something better to plug it into than this tiny synth Music yes I went to visit Sam from look Mom no computer who owns this museum is not obsolete down in Ramsgate in Kent where hes got the Furby all gun and loads of other projects from his YouTube channel like the Game Boy Mega machine and an organ made of owls right youve seen what AI Drew I have heres the real thing so this is experimental robotics equipment for playing music designed by AI my gosh what the heck is that sheet off wow that looks friendly that looks really friendly oh is that hello how are you doing buddy is it it looks like a vacuum cleaner thats so amazing what the heck oh theyre all oh they feel very strange oh ais got thats what AI told me to build yeah and I tell you what it has its got a definite personality it sort of looks like youre James have you been its been looking at you look at the image oh thats beautiful I cant wait to plug this in are you what are we plugging it into I think the church organ is going to go great with this yeah you did Buddy Music Music foreign Music hahaha Music foreign Music Music hahaha Music Applause Music foreign Music Music well thanks Sam so AI really does know how to make experimental robotics equipment for playing music especially church organs that was pretty good good job Music hello hello is anyone there oh is it this one hello is anyone there no

write a youtube script

in this video im going to see how far i can ride on a cardboard bike yes with cardboard wheels in the last few videos ive built quite a few vehicle projects most of these were made from steel plywood and d printing as well as the electronics i generally recycle most of the parts from my old projects the motors and electronics can be reused steel cut up and welded back together again to make something else and d printed parts get mashed up by handy dans d prints and turned into flower pots but what if i could make a project that could be even more easily recycled i previously made a cardboard skateboard which was pretty tough it was made of five cardboard tubes glued together and attached to some d printed ends which i had normal skateboard wheels attached to but could i make the wheels of a vehicle out of cardboard too so i bought a bunch of cardboard and that comes in a big bale there are pieces here of single corrugated cardboard and of course none of them come fixed together or anything like that theyre just wrapped up in plastic if i hold them together though then it looks like it can support my weight okay if i dont hold them together they skew and i fall over but it looks strong enough that all of those bits of cardboard together will be able to make a wheel that will support my weight so the plan is to cut out circles of cardboard and mount them on a cardboard tube that should be strong enough for the axle and glue them all together to make a wheel on the bottom of the wheel of course there is a curve and a smaller piece that touches the ground where all the force is going to be so well have to see how that works out when the edge isnt straight like it was in the bale so ive made a template to cut out all of the circles which im going to do by hand itll be great if i had a laser cutter that would accurately and precisely and quickly cut them but i dont so im using a knife to make all of the cardboard discs and of course we need two wheels so we actually need a hundred of these i designed a jig so that we can get the wheel built straight and also so that we can get the cardboard discs in the middle of the axle which is a cardboard tube with a bit of cardboard tube sticking out either side so we can mount it so i see and seed out some mdf and basically what ive made is a platform and that allows part of the cardboard tube to stick into the platform so we start building the wheel halfway up we stack up cardboard discs and then that allows us to have a bit of cardboard sticking out the other side as well so we can mount the wheel so first of all i stuck down some plastic sheets so that i wouldnt glue the cardboard straight onto the mdf put down the first cardboard disk and then i sealed that on with hot glue and thatll stop the other glue that im using basically running down the tube and gluing the whole thing to the base so now its time to get gluing and im just using pva here which is just white glue to glue all of those discs basically in a massive stack to make that wheel i cut out another disc and thats got a slightly bigger hole in the middle and thatll allow me to push down all over fairly uniformly to glue the discs together and eventually ill put weight on the top of that when all of the discs are glued down so it can hold the whole thing in place so yeah we just keep going adding more discs adding more glue and pushing them down and sticking them all together im just using whatever pva glue i could find in a large container so we keep going you can see the thickness built up here we need to glue down of these for each wheel so heres my first wheel and thats discs glued together that gives us a thickness of roughly millimeters for the entire wheel with the axle sticking out either side so now i can put my pusher piece of wood on and put on some heavy weight so we can let that glue up and hopefully itll all come out solid so the glues all dry and yes ive made two of them because we need two wheels for the bike so yeah it seems pretty tough pretty sure its going to be okay obviously weve got this square edge so it might be a bit weird to ride i think whats going to probably happen is the edges are going to get compacted down and ill end up slightly curved i am considering sealing the surface with something well probably wear it in a bit first and then think about that later before i carry on with the project i thought i should just do a quick test and check itll actually take my weight without the cardboard wheels immediately folding and buckling under me but so far thats looking pretty good i am on soft carpet but those wheels are looking absolutely intact and the surface is pretty hard actually with all of that glue on there the rest of the construction for the vehicle is going to be the same tubes as the axles which are pretty strong tubes and these are roughly millimeters in diameter with about a five or six millimeter wall thickness so this is roughly what its going to look like i am going to use d printed junctions to join the tubes together because its already got cardboard wheels and most of its made of cardboard which is sketchy enough and i cant think of an easy way to actually join those tubes together these tubes are quite a big diameter so the parts are quite a big diameter to fit over them as well so ive printed them all with the . millimeter nozzle so theyre all pretty tough theres only two perimeters and about percent infill so just a quick add from my d printer sponsor thanks to lolzbot for supporting my channel with d printers there is support material in these which breaks out just the same and that should leave us with some really good looking parts and thanks to d fuel for the filament for this project and lots of other projects so check out my channel for more d printing projects and check out dfuel.com the first d printed part were going to look at is this plastic shim which a bearing fits on very neatly thanks to simply bearings for the bearings for this project check out simplybearings.co.uk these are some pretty substantial bearings and these are for the wheels because i didnt want to run cardboard on cardboard and of course that shim is just to make it the right space for that cardboard tube because i couldnt find any bearings exactly the right size im sticking all the plastic parts to the cardboard with some contact adhesive i found on the back wheel weve got two of those and one of them actually has a pulley on so we can drive the whole thing along the other half for that bearing is also another substantial print printed with the big nozzle but i made a part with a finer nozzle to actually fit the bearing in accurately so i dont have to reprint the big part if it doesnt fit properly so of course that will fit onto the other shim that i already made thats fitted to the wheel so now we get a really substantial wheel bearing and in fact these joints and the wheel bearings are probably overkill theyre much more substantial of course than the cardboard wheel my contact adhesive appears to have gone off a bit so its quite stringy it should be more runny than that but it probably works perfectly well and it means i can get loads on the spreader to get it in the hole there so basically ive already primed the cardboard tubes with pva which is why they look a bit shiny now and that helped me with the cardboard skateboard so obviously just sticking that in spreading the glue around as much as i can and then letting it go off and well see how strong it is but hopefully that should be pretty good some of these parts ive made in two halves mainly because they didnt fit on the print bed but also it means that i can build the bike in two halves and then fit them together and ive left bolt holes as well as gluing them onto a central piece in the middle when i put the two halves of the bite together so yeah the bikes pretty much designed to be split in two and that means i can easily assemble the forks for each wheel in two halves and then put the two halves together and youll notice theres at least two tubes for all of this structure the two pieces in the middle and the two for each of the forks so that should make it nice and strong the closest thing ive got to steering bearings is this top hat piece that fits in so im running plastic on plastic instead of plastic or cardboard on cardboard i was going to try metal wheel bearings but i thought this would probably do i thought id try some hot glue on some of these to see what happens so ive got a very short piece here with two ends on which have notches in and that piece is basically a pillar between my two crossbars that holds the whole thing together most of the load is from the top onto the bottom and thats because ill be sitting on the top and ive got my two steering pieces which basically make up the steering column with those two high hats so basically i need to push the two down and then both of those steering linkages will apply force to the front forks so i did half at a time and then went and hot glued it well see how well this sticks but basically thats the best way to kind of get this aligned without having to clamp it in place so now weve got the steering column and i put a cap on the top as well to stop it falling out the bottom so that looks pretty good yeah the hot glue just broke off the edges it didnt stick to the d print very well so i put some zip ties on but ive got a cosmetic plan for covering that later most of my other joints are really good but the cardboard is basically very tight in that hole so the glues work really well the next part its really much more critical the tube doesnt slide because theres no end cap on these so im using that expanding gorilla glue glue which expands out and fills all the gaps to make sure its stuck properly so my front forks are all glued up which are of course a symmetrical pair which fit together and glue and bolt around that rotary piece ive already made for the steering column ive also made the back forks as well with those funny angle pieces and you can see that gorilla glue has foamed out and filled all of the gaps but before we carry on with that its time for a quick ad from the video sponsor which is zometry i make most of the parts of my own projects but sometimes there might be a part that i dont have the tools or equipment to make zometry solves that issue and its accessible for anyone from large corporations to the hobbyist anyone with a cad file can upload their design to zometrys instant quoting engine and within seconds get instant pricing lead times on production and dfm feedback zometrys quoting engine is quick and simple to use zometry offers an exceptional range of manufacturing options with over a dozen manufacturing processes including various types of d printing and cnc machining sheet cutting injection molding die casting and many more zometry also offers thousands of materials and finishing options you can select from a number of economic domestic and global shipping options to impact lead times thanks to zometrys immense network of over global manufacturing partners ordering with zombatry is more affordable than most other options pluszometry offers a number of smart and easy financing options so you can get the parts you need now and pay later so no matter how big or small your custom manufacturing project may be you can start your manufacturing project with zometry visit xometry.com brewton to get your design quoted produced and in your hands in no time so ive glued and bolted those front forks onto that rotary steering column that id already made so that seems to be working all right and ive also fitted some handlebars so that i can turn the wheel and thats looking okay at the moment i think these d prints are pretty excessive i probably could have got away making them slimmer i think these cardboard tubes are probably excessive as well the wheel still feels a bit soft in comparison to everything else being corrugated cardboard but at least we know the rest of the structure probably isnt going to fail right the backs on and it looks like a bike so the key thing is here that these four joints are all glued up and theyre rigid and that holds these two tubes at the right spacing so the force from the top is pushed through here and pushed down to the bottom and everythings loaded properly the front is slightly different though which is why weve got this piece in here this upright so when i push down on here it takes the load onto the lower tube as well and thats because these pieces arent attached rigidly to the front so the steering works and that means when i sit on this crossbar were actually putting the load onto both of these which are resting on both of the pieces below there so the force is conveyed equally onto that steering column yep it seems to be definitely strong enough to take my weight as well so thats pretty good and its pretty easy to move the steering wheel so were all good there time to sort out the drive mechanism so ive attached a piece which has the motor mounted and thats d printed and glued onto the tubes and that has a slider in with a piece of studding that i can tension to pull that belt tight so we can always adjust the belt tension yep so thats looking pretty good and obviously weve got a d printed pulley on there and thats because my metal one didnt arrive in time so this is a brushless drone motor which is a turning g aero drive kv and ive used these in loads of projects to drive the motor ive got a vesc or vesc controller which is for skateboards and ive got a s lipo which is a high drain one and its amp hour both of those fit in the tubes at the back conveniently and now i can run that wheel which seems to not be running too bad its pretty much on center for something i made out of a bunch of cardboard so obviously that jig helped quite a lot and to control it ive got an ebike kind of thumb stick thing thats supposed to go around the handlebars but it cant because the handlebars are too fat so we seem to have enough torque at least to push me along and were using magnetic braking or regen braking but thats looking pretty good so im pretty happy of how thats turned out so far but we need to make sure people watch the video so its not just a waste of time youtube algorithm really likes it when people click on the video if that click through percentage is really high then it shows it to more people who hopefully click on it so im saying in the youtube title that its a cardboard bike but the d prints have turned out a bit more overwhelming than i expected them to be so basically im going to cover some of the plastic with brown paper tape so it looks more like a cardboard bike and hopefully people watch it weve got cardboard wheels though but yeah i know its cheating but thats societys crime not mine well it looks a bit like something fred flintstone would ride on but it also looks like a cardboard bike so hopefully that makes it really popular i put a foam pad on here so im not sitting on that d print so its nice and comfortable well i thought the hot dry weather would be a good time to do this project so its dry outside but as soon as i finished i need to test it its raining and theres going to be thunderstorms forever pretty much so i need to do something so i can still test it ill take my chances with the weather but i probably need to seal the wheels so they dont go soggy if i run for a puddle so im using some liquid latex which is the nearest thing i could get to cardboard basically because it comes out of a tree there are obviously plenty of other types of sealant but theyre generally solvent based and this is basically water and rubber so when the water evaporates it just leaves the rubber which is about percent and it dries a nice cardboardy brown as well i left that overnight and the latex has darkened up a bit as you can see the sides of this are still a bit squashy but the latex has still got a fair way to go it will darken up and go a bit more solid so hopefully that will be okay Applause Music Applause Music well heres the first casualty i managed to snap the handleb

write a youtube script

i recently built an eightwheel snake bike and that was based on a smaller version that i built some time ago before that i built a worm robot which moves by expanding and contracting this is known as a system of peristalsis which from wikipedia is defined as a radically symmetrical contraction and relaxation of muscles that propagates in a wave down a tube while i was researching these projects i came across another experimental robot which uses a wave motion with a wave running down its body to propel it forwards this is a bit like how snakes actually move well at least the ones without wheels or legs of course my original plan for this project was to make a surfboard you can ride on the land but using a mechanical wave like this to propel me along and use a large set of mountain board wheels for steering however i thought i tried to make a smaller version for this mechanism first to see what the potential problems could be it looks a bit like the mechanical wave is formed by a thing that looks a bit like a giant corkscrew which simply rotates inside a flexible track so i d printed a former which i used to try to bend some wire around i wanted this to be quite stiff because in my version it has to support the whole weight of the machine so i used four millimeter diameter stainless steel rod this was pretty hard to bend into shape by hand though and my resulting screw shape doesnt quite match the former or the original cad design but well see what problems that causes later i then d printed lots of parts to make the flexible linkages and the other parts of the machine just a quick ad from my d printing sponsor thanks to lolzbot for supporting my channel with d printers Music Applause Music and thanks to d fuel for the filament for this project and lots of other projects so check out my channel for more d printing projects and check out dfuel.com so weve got lots of parts to make the linkages for all of the track pieces so weve got shorter ones and longer ones so those can go together and then it can be articulated quite a few of those go together to make a nice bendy track and that has a cavity down the middle that we can put our screw piece in so thats nice and flexible all of the linkages are just m bolts screwed into the plastic so now we can insert our giant cork screw and rotate it and see if that looks like its doing anything that it should be so with a bit of help it looks like it runs in there quite freely without too many problems at least and thats looking like weve got that sort of wave motion so im pretty happy of how thats gone so far the rest of the chassis for this was too big for the print bed so its made in two pieces with a bridge section screwed on to hold the two together those screws need to run on something so ive got this special pulley which has got some eight millimeter escape bearings in the middle and thats got a groove in that allows that piece of steel to rest in there and it can be zip tied or glued in so hopefully that can be rotated by the motor and idle at the other end each of those fits onto some m studding glued into some end sections so with that all together weve got quite a satisfying screw motion so at this point weve got a screwdriver that seems to work pretty well and of course you can see the motion moving from one end to the other as i turn it so we could actually just drive it on this and you remember that colin first did build a screw tank and it had two giant screws that drive it along due to this motion the only downside of that is that we do have some sideways friction with those screws turning which would rub on the ground doing it this way with this bendy track over it means we can manage the friction within the track using bearings or sliders or something like that and then we only get the motion pushing front and back as the track wiggles so now ive fitted the track onto the cord screw and you can see we almost get that wave motion but its a bit sticky and it gets stuck in certain places which means its not going to rotate very easily with a motor so it appears that putting the call screw on center has caused some issues with the tolerance of the screw itself so ive printed some slightly different tops and bottoms for the track segment so theres slightly more space for that screw to rotate and now it seems to move much more freely and we get that nice wave motion that you can see there so we can see that weve got that wave going backwards and forwards as the corkscrew turns and moves the flexible linkages so hopefully that should be sufficient to drive the whole thing along i want to try and make a tank out of this so ive made two of those assemblies and ive also printed a top plate that holds the two parallel with each other so now hopefully we can drive backwards and forwards by attaching motors to it and using differential drive so ive put pulleys on my motors as well and with those fitted and some electronics fitted we should be able to drive the whole thing along ive got an arduino mega in there with two motor drivers that should be more than sufficient and an nrflo on radio chip so i can use the open dog remote to remote control it so it looks like the waves move backwards and forwards so we can drive them in either direction as youd expect for a differential tank drive so they can move in the same direction and also in opposite directions so hopefully we can turn on the spot but it looks like we might have a couple of issues as well as not having much grip on the surface its also rotating and going sideways when it shouldnt be so we need to investigate why that is yep that really isnt very good at all so i think some of the problem is that those motors are massively heavy and thats probably causing quite a lot of issues pushing down the mechanism and flexing that wire that i have in there so we can see that the pulleys are actually running on the ground at the bottom which is causing it to rotate sideways i think those motors are probably overkill so ive got some much smaller lighter ones here and ive used these in quite a few projects before these are volt motors and theyve also got a gear head on them and an off center shaft which is quite useful because we can rotate them into different positions to get the belts tight with the new motors fitted we can power that up and see how it performs and initially it seems much better those tracks are moving much more freely im actually rotating deliberately here and that seems to be working not too badly theres not much grip with the surface though so ive got lots of little rubber feet and im sticking those to the tracks all the way along hopefully that will help with ground clearance as well as traction with the ground but before we carry on with that its time for a quick add from the video sponsor which is jlc pcb jlc pcb are at the forefront of the pcb manufacturing industry and they provide high quality low price pcbs its easy to order from jlc pcb just click on instant quote and upload a gerber file the jlcpcp website will show a preview of the board and then you can select various options for manufacturing if you want the board assembled then you can add a bomb and cpl file jlc pcb currently have an offer where you can get five one to four layer pcbs for just two dollars with free smt assembly jlc pcb is a onestop shop with a team of trained engineers to manufacture and assemble your board under one roof and with the new jlc pcb parts manager you can check available stocks order parts or even request specific parts for your project jlc pcb ships worldwide and they have fast build times so you can get your pcbs in as little as three to five business days jlc pcb also provides d printing services from as little as cents per gram so you can get enclosures and other parts manufactured in a variety of materials d printing processes include sla mjf slm fdm and sls in both plastics and metals the ordering process is very easy at jlcpcb so use the link in the description to this video to check it out now so we still have that wave motion and it looks like those rubber feet are pushing in one direction or the other so hopefully that should help it locomote along so this is deliberate rotation which seems to be working much better than it was before it seems to have much better traction with the surface and driving seems to work quite well too and we can see now weve got much better ground clearance that pulleys well off the ground so thats not dragging it sideways when it shouldnt be so it seems to drive okay although it looks like theres more friction on one track than the other because one of the tracks is running much slower so here you can see the inside track sometimes gets stuck and it kind of pauses whereas the outside track thats going much faster is running much freer and thats running constantly so its time to have a look and see whats going on there this track is running fine and this one is not low so it looks like somethings getting stuck or theres too much friction so if we run it we can see that sometimes that pulley pops off so it looks like theres some extra forces in the tension of that screw and its not quite the right shape so the more i kept testing it and i tried running it slower as well so theres less power on the motor but then it completely jammed so i think that that shape isnt quite right i spent around an hour trying to reshape that screw by hand to try and get it to run freely so this is about the best i can do so it seems to run okay if i run the motor slower so theres not as much power it runs in both direction both my pulleys and idlers stay on and it doesnt jam it runs much better now although one track still does have more friction because it still turns slightly in one direction but i have differential drives so i can alter the motor speeds manually and steer it back again so its not perfect and its definitely very critical how those screws are shaped so theres not too much friction on one side so im pretty happy of how that drives along though i think it probably goes about the same velocity it would if it were a traditional screw tank like the one colin built but the added bonus is that those screws dont have any friction when they rotate sideways onto surfaces like the carpet for instance so hopefully managing that friction with the metal running on smooth plastic makes it more efficient than it would be if it just ran on the driving surface and i really like the way it looks when it drives it looks pretty peculiar so i think theres quite a lot of potential here for using this mechanism in some other projects whether were driving a vehicle whether were driving a surface with this mechanism upside down but the question you really want answering is can it drive over obstacles so here are my pieces of test wood that i use in lots of other robots and you can check out some of the other weird wheels and things in my channel and yeah its not too bad im having to steer a bit here so only one track was running for a bit but that was intentional to try and steer into the wood but yeah it will climb over stuff and doesnt do too much of a bad job of it we dont have too much ground clearance due to those pulleys and those idlers at each end of the screw but on the whole it does climb over things and im not too unhappy with the way it works Music Applause Music im not too unhappy with that there are several improvements we can make one would be to have shorter links on these flexible tracks and thatll make the track more flexible so it would fit around the screw better i think thats probably something we need to do for a tractor this short in the original one it looked like it was much more flexible and those links are much shorter the other thing would be to form the screw better of course which was quite a challenge and probably the reason this didnt work as well as it should a better approach might be just to form the wire around something round the way springs are made and then stretch it out so it looks like a giant compression spring getting the diameter right would be quite hard of course because it would get smaller the more you stretched it out but that might be a better approach although itd be quite hard to get it to run on center perfectly which is probably where most of the problem is i still like to make one that i can ride on using these mountain board trucks in that case most of my mass and the motor driving it will be supported by the wheels so it wouldnt be quite the same as this but then wed still need quite a strong screw in case i lent back onto it so you probably need something bent out of tube now i have bought a pipe bender recently but its going to be quite critical still how that spiral is formed its quite going to be quite hard to form out of tube so it might be something for a specialist company in any case i am going to try and make one i can ride on like an artificial wave with a surfboard on it because i think that would be quite fun to see how it propels me along so im going to publish the canon code for this if you want to have a go at building one or have a look at any of my differential drive code or any of that stuff thats on github so if youd like to support me through patreon or youtube channel membership then those links are in the description below and patrons and youtube channel members can get access to all the videos up to a week early including sneak peeks and pictures of whats coming up so you can be part of that discussion alright thats all for now Music Applause Music Applause you

write a youtube script

this is the future of pizza delivery Music the first thing you do when youre in a hurry for a nice fresh pizza is pick a pickle phone and make an order pizza Demario are gonna take your order then of course they prepare it they mix the dough they need it cover it in all the beautiful they slide into the oven it Cooks perfect they slide it out stick it in a box put it on a delivery bike and on its way to you it goes but this pizza is not as fresh as it could be depending on how far you live away from the pizza place and if you live really far they might not even deliver to you at all have a solution okay heres the plan rather than cooking the pizza and delivering it were going to cook the pizza while were delivering it now this video is sponsored by insta more about them later every time I use their cameras to film Ill put this little logo in the top corner and if you like things that turn up in boxes you might be interested in this conifers jigsaw puzzle Im doing a final run of these once theyre gone theyre gone and then this hightech World a puzzle is actually a refreshing change its brilliant its got all the inventions on its got the bunker on absolutely fantastic Link in the description get yourself a piece of first history and thatll be nice right pizza bike so we need to take our bike shut the back off add a pizza oven add a pizza chef and stick the back of the bike back on again now before we check this wonderful Suzuki apart we need a pizza oven now Im gonna mount it just behind me keep me nice and warm but if you buy one then they tend to be like this flat type with a semicircular top and the chimney very nice but I like like domed brick ones obviously Im not going to make it out of brick so Im going to make it out of metal and Im quite good at making metal spheres and domes so I think we need to do a bit of hydroforming Music now to sum up hydroforming quickly its essentially making a D metal shape filling it with water pressurizing it and using that pressure to force it into a new shape and it is absolutely perfect for making metal spheres now there are more videos on my channel I will link them in the description lets see how we get on today oh its big Music moving on it Music Applause Ive made PacMan its not quite what I wanted but this is a very good example of why youve got to get all the air out of these things before you hydroform it when I fill it up I fill it up to the top and I make sure there is no bubble of air at the top here which can be compressed because obviously had that have had air in it when it goes bang it would have been a much bigger and more powerful explosion because its just full of water which is noncompressible you just get a little bit of a pop it splits open thats it makes you jump thats about it were all still standing Ive still got a garage right Im going to clamp me down weld you up and have another go but while Im doing that lets talk about these cameras the insta is an action cam but with so much more its ability to take degree images added with its inbuilt software to not only seamlessly Stitch the two lenses together it can cut out the mounts that hold it giving you dronelike footage in the palm of your hand image stabilization is impressive even at the end of a three meter pole the insta has a range of mounts and attachments to suit your needs from bike mounts to get unreal perspectives in a high intensity environment Music to backpack braces to capture experiences and moments like never before the camera itself is waterproof but for the more extreme depths there is a dive case available editing the footage can either be done on desktop or through the insta app editing options are huge with the ability to track faces pan through a shot keeping what you want in frame you can also use your phones accelerometer to track through in real time and best of all you can choose the aspect ratio to suit long form or for social content for any content creator this is a musthave piece of Kit so check out the instax link in the description and get yourself a free invisible selfie stick Music right Colin how are we doing with this Hydra for me Music okay welded all four joints again Music thats pretty good I dont know if I want to go much better than that Music laughs Music okay were about there its a little bit taller than what it is wide but its okay because Im gonna make this oven double skin this is going to be the inner skin so you wont actually see this so next thing we need to do is to get the templates for a slightly bigger ball weld them up with this in the middle and then hydroform it again and hoping well have a ball in a ball and then I can cut my little pizza entrance at hold it up like have a chimney sort of go into it I know I know youre thinking Colin you are making a rod for your own back here I was shot gone for a blooming semicircular one right lets have a crack at that oh my God you guys are geniuses so Ive been struggling with cones havent I like putting them in pulling them around sliding it around pulling it around mauling it around and I remember a comment on a past video where somebody said oh just drop a little bit of angline in between the rolls and it will assist you much better for the life of me I could not find this method anywhere on the internet but Ive just seen it I know what you mean when you shove this in here measure the distance between the inner part of the cone and the outside of your slip roll cut a bit of angle iron drop it in here just in that little bit there shove it in here up against that bit of angle iron starter turning and it just turns it round look for you the person that put that comment you are genius you have transformed my life a little bit that is why I love the internet oh its just one big learning circle isnt it just imagine the amount of things we have learned from other people what else can we pass on to each other never mind AI just you know whatever that is right Conan it up we do coning it up we do now I got so excited about this Ive made a whole video on special Workshop technique Ive Linked In the description that makes me so happy look at it foreign Music Applause okay Im still smiling over that slip roll technique right weve got two thirds of our secondary sphere this one now is gonna go in it Ill put you in here it fits thats a good sign what I was gonna do second coat on the top like that very nice lets make a little chimney piece which goes on the top there I was going to weld that around then and I was hoping that this would become a ball and then that would tuck in and blend in hes not going to do that is it because obviously for that to go in the pressures gonna be on the outside here its not going to work so were gonna have to do that bit separately okay weve got a sphere and a sphere but the next bit is the tricky bit got a round pizza stone to get in the center ball a chimney to weld to the center ball and a doorway to go through them both how Im gonna weld to the center one I dont know I dont know where Im gonna do the next bit thank you once plasma cut through the outer layer I could see my inner ball sitting in there all nice and then I knew what I had to do I just pulled the inner ball towards the outer ball at the entrance weld the two up there and therefore the biggest air gap will be at the back where my back is this is if I planned it like that okay very nice got a pizza stone in there got two blow torches piped up with some gas awesome tried to make a nice little separate swoopy bit for repair but essentially made a butthole now Ive had it running above the Flames gets hot but the bit at the back doesnt get hot so were all good for the double skin thing okay now this door is quite big were going to close that off a little bit lets give it a test see if it works with the front door fitted I then rather foolishly turn my mic off and try to make a pizza Music here we go there we go Music thats good I need to let the stone warm up a bit more because hence why the okay I need to let the stuff but will it work while its moving along so with it all safely strapped to the quad got the camera set up fired it up slapped a pizza offering off I went found the only hole in the field Ive broke the stone the pizzas a mess I drove down a bloody old Dodge this house actually really deep I think I got away with that quite Lightly for the burnt cheese scraped off the stone and the pizza oven set back on the quad it was time to give it another test with no holes all was going well the pizza was cooking but unknown to me as it cooked the harder it got the more slippery it got it was going on a journey of his own I nearly lost the Blooming Thing but it looked good and it tastes good okay Pizza Oven works now it just needs a bike shove it into it Music now this part of the build has been made a little bit more complicated because while Ive been messing around hydroforming a pizza oven Ive been riding in Suzuki I like the Suzuki so I do not want to cut the Suzuki up so all USB lovers out there fear not gonna treat it with respect okay Ive mastered it we are not going to butcher this at all Ive got a chassis extension Ive got some extra little bit of seats Im stretching it out Im bolting it all onto stuff this thing wont know its even happened to it Music getting youre not afraid that theres this rope I hide I never use it but it makes me feel like safe Music and though I know its pretty stupid oh I need some pride no no get in there oh no ah theyre the ones laughs yeah oh its so long Colin yes it is but dont worry because in long time subscribers will know I broke the world record twice for the worlds longest motorbike foot one and a foot lock so this is nothing anyway never mind all that come have a look at this okay this is a very long chain that is only going to cause problems with tensioning and its flopping all around all over the place now unlike any other chain motorbike chain is a funny size and you cant just go and buy a little blank sprockets like you can in other chain sizes so these tiny little ones here which Ive got here for the machine tension Ive made myself yes with the blank up on the lathe so basically machined a bit of aluminum down and then I fixed it down screwed it down to the bed of the CNC router and then routed the teeth out into it oh yes very impressed myself there obviously machines and bearings out put it on Boom weve got chain tension in now its not all smiles because such a big chain pitch its such a small tooth you end up with a bit of bouncing and a bit of noise gonna be a lot of noise when we ride this and also this chain tensioner should probably be facing the other way but never mind that side lets go and give it a test ride because I kind of want to see what its like foreign on it a transform Pacer delivery forever Music yes yes it looks very good Colin but come on lets deliver some pizza got my first order right with our first customer we lit the oven and headed off through town Music ovens up to ten Rick gives the five minute warning now preparing a pizza on a motorbike was never going to be easy and as we know from before Im not very good at getting it round and flat I even tried bashing it with my helmet I tried extra flour I even tried a small rolling then I remembered that technique where they Chuck it in the air spinning it around that clearly doesnt work on a motorbike so what a pizzas not gonna be round Im still enjoying this so with something that resembles a pizza isnt that hard to get it in the oven so close Colin so after losing our second pizza we slowed the pace down whipped the third one up got it in the oven just in time to arrive at our customer whos in the middle of a field Music Music hes brilliant right I admit its not a circle Ive just realized we dont have any plates or any pizza boxes although not a perfect shaving actually our third Pizza the end result was a perfect representation of Italian cheese on toast this is actually a pretty good idea so with our lessons learned we headed off to our second customer this was much smoother Music we even told them to bring their own plate because this is lets get the professional pizza making try out and see what he thinks after watching the footage back its immediately obvious that I didnt know what I was doing and he does and while it was cooking ironically we rode past dominoes its looking good at the destination the customer come out with a plate fit for one slice so while he got more Mario checked out the cooking he was very much surprised he had them in the base is good very good with everyone served and happy we just kept cooking can love me ready there you go Peter the Mario on the go were calling here well I think well call that a success and there we have it now thanks to Pizza Demario and Stanford of course for Lending me their kitchen the use of their ingredients if you go and dine out in there just say Colin said yeah you wont get anything free or any discount but you will make him smile now thanks for insta of course for sponsoring this video check out the link in the description to get your free Magic selfie stick fantastic cameras offer shots and options that you dont normally get its all good now thanks to my channel members for actually naming this the name pizza demoto came from my members because I put a poster saying hey guys can you help me out with this and they always do this is what we do the old Channel membership thing bit of interaction if you want to join you can down the bottom of course it was a combination of comments from these two guys Ryan and Gordon I kind of put them together like domoto pizza Demario it all makes sense and there we are right Ill see you in the next video fingers crossed should be a tunnel video subscribe people subscribe are you gonna see this out on the streets because it does work only two downsize is it only makes one pizza you could possibly get two in there and of course if its raining hmm not gonna be too keen on doing that are they no anyway thanks for watching Applause everything

write a youtube script

veritasium made a video about how bicycles balance he had a special bicycle made so the steering can be locked so any turns in one direction or locked so it doesnt turn at all the conclusion of this video is that steering is the most important thing for balancing on a bicycle essentially you steer to catch yourself as the bike tips in either direction and bicycles can also steer to stay upright by themselves due to the geometry of the front wheel and the steering position which leads the front of the pivot point with the steering locked its impossible to balance for any length of time even if you want to ride in a straight line but what about if we make a bike with two wheels that steer it seems like it would steer and balance okay if the wheels rotated in opposite directions so they steer in the same direction but what about if they rotated in the same direction so you could drift sideways would you still be able to balance so in this video im going to make a bike which has both modes of steering i need to make sure both wheels stare at the same angle so the wheels dont try to drive apart from each other so im going to apply the leading wheel geometry to both wheels so they both point forwards and im also going to make this a twowheel drive ebike so the wheels can maintain the same velocity and one doesnt lag the other also so i can switch modes and have the back wheel either copy or mirror the front wheels motions the back wheel is going to be driven by a powered servo and that means that i can switch modes while riding it with some switches on the handlebars so i thought about modifying an existing bike or a scooter or something like that but due to the back wheel steering in the end it was easier just to build the whole thing from scratch so yeah its made out of box section steel i dont have a tube notcher or a tube bender at the moment so basically its going to be far too heavy duty but hopefully then it wont break and nothing will go wrong when im doing the testing and its just a test rig to test out the principle of how bicycles balance so ive got some box section steel for most of it which is box just going to deburr it there with a power file and an angle grinder these parts are actually going to insert into d prints so i want to make sure theyre nice and clean and clean before i weld them im going to be using inch hoverboard wheels to drive this for the front and back powered wheels so im doing the normal thing here of drilling a big hole which the hoverboard stub can insert through and then attaching a nut basically in the other way so im just going to tap that there with a bit of cutting compound that means we can put an eight millimeter nut in and we can weld the nut on essentially so that we can screw a bolt in and thats nicely backed up and strong and thatll grip the stub on the hoverboard motor and ive done this in quite a few projects recently if you check out the recent projects in my channel weve got some funny angles to make here so ive got a bit of a crank in the upright so that i can get the hoverboard wheel directly below the pivot point so ive got some degree angles which are of course and a half degrees cut in that box section and that welds up to make the degree angle so basically what we end up with after weve welded all of those funny joints is a kind of crank shaped piece and that means that the hoverboard motor can fit directly below the main upright piece and im using inch hoverboard motors which have a pneumatic tire so that of course fits into the big hole i drilled and the bolt goes in the other way backed up by that nut so that it grips the flat on the shaft and that means that hoverboard motor is directly in line with that upright shaft and everything will rotate on the spot thanks to simply bearings for the bearings for this project weve got several bearings here and theyre pretty heavyduty ones so normal bearings and weve also got thrust bearings and thatll allow the whole thing to take my weight these are made in three layers a bottom runner a row of ball bearings and a top runner and that allows force to be pressed down on the top and for it to take that load i need something for the thrust bearings to rest on though so im just going to cut a slot in the box section steel and essentially you put a tab in each side which then gets welded on and that makes a little platform to support the thrust bearing and the other side of the thrust bearing will have the frame resting on it there are some other parts with some pretty complex geometries and the easiest way for me to make those is with d printing these are pla parts but im using a . millimeter nozzle and you can see how fat the extrusions are there and that makes the parts really tough so just a quick add from my d printing sponsor thanks to lolzbot for supporting my channel with d printers these are mainly the pieces which bearings are going to go in but we also have to make some pulleys and thats so we can get feedback to a pot for the joints and also feedback so we can actually drive the servo to the right position thanks to d fuel for the filament for this project and lots of other projects so check out dfuel.com so yeah ive also welded up a ridiculously heavy duty bike frame to go between the back and front wheels its nothing like a bike frame really in that its made of square bits instead of triangle bits but its going to be really strong so i dont think theres going to be any problems at all these big pieces have bearings inserted into them and as usual what ive done is made a smaller shim piece with a finer nozzle so i can get the tolerance accurate and if it isnt then i dont have to reprint that whole part which took about hours each so bit of a hint there for d printing large parts so ive got bearings fitted into those and the down facing parts also have a place for that thrust bearing to go in so that its all held centrally and it wont slip out and of course the bottom of that goes onto the metal parts in there weve got some inserts for my steel so that i can put a square axle in essentially and weve also got the square axle shape here for the rest of the frame ive left slots in these parts so theres some flexibility and its much easier to get them in and if it wasnt for that slot id never be able to put them on at all and get the tolerances right so there are bolt holes you can see there was gonna bolt through the metal as well but i dont think im gonna have to because theyre a really really tight fit so theres that frame with its four pivot points and my two steering columns and ive also left a little stub there for the saddle so weve now got one of the steering columns in with that wheel on its crank shaped shaft and you can see the metal resting on the thrust bearing that of course goes the other way up so the wheel is facing down and the thrust bearing is on the bottom and weve got those other bearings in their other bearing holders for both the front and back steering columns so yep thats that thrust bearing sitting on the bottom and of course its the same on both ends of the bike the handlebars are fitted on with another d printed clamp section and ill come back and talk about the switches and controls later in the video so now weve got something that pretty much resembles a bike and both the wheels basically point forwards like they do on the front wheel of a normal bike so pretty sure im going to be to ride that like a normal bike but now we need to sort out the control so that we can turn the back wheel as we turn the front wheel in various directions but before we carry on with that its time for a quick ad from the video sponsor which is potentis i love building prototypes and showing how they work on youtube but have you ever thought that your ideas could have potential to be a real product if you have the first step you should take is doing some research to check if someone else had the same idea already and to look at any relevant patterns to see if the intellectual property was protected by someone else if youre the first one with the idea and you want to develop it further what comes next is filing a patent but patterns are very costly and complicated however there are other easier and much cheaper ways to protect your intellectual property and maybe it isnt as difficult as you think you can learn about this topic with the potential game app potenticus was developed to help investors get a better understanding about the topic of intellectual property rights the app offers a fun and easy to use game that will teach you how to protect your idea and help you get your ideas to the next level in the game section youll progress from a law school graduate to a highly respected patent attorney by learning about the various topics of intellectual property rights and helping your customers with their ip issues the faq section provides you quick answers and dedicated information to your specific questions the app provides easy to understand information and gives you lots of examples so far the apps content has been specifically adapted to uk and germany law the app is available from both google play and the apple app store and its free of charge and no personal data is collected so potential is a great way to get insights and information about protecting your ideas check out the links to the app store provided in this video to download the app and learn how to make the most out of your ideas note that the use of potentials is not legal of ice and should not and cannot replace it right lets get back to this bike to do that we need some way of measuring the position of the steering wheels and also the back wheel steering column so we can position it accurately so ive got a potentiometer with a big pulley attached and that runs one to one on another pulley which is on each of those steering columns so theres the one on the handlebars and weve also got the same thing on the back wheel but as well as feedback were gonna need to power it so ive got a windscreen wiper motor fitted on another d printed bracket and that has a sprocket on so there is a chain drive and yes its held on with zip ties and thats a bit like a hardware fuse because it means if theres too much force or something goes wrong then the zip tie snaps and it doesnt burn out the motor strip the gears or burn out the motor driver controlling this is an arduino uno with a lot of wires attached to its io weve got a bts motor driver a vesc on each of the wheels and yes weve got two batteries because it was just easier that way to put two lipos on the wiper motor is volts and im running that off the same battery but i only ever run an analog value on the arduino of out of so on average it never gets more than volts ive got a brake kind of lever there and that pulls a string which pulls a micro switch and that reverses the main drive wheels which is the closest thing ive got to a break ive also got a twist grip on the right hand side there which is the main throttle the two switches are for the two steering modes so the back wheel either copies or mirrors the front wheel so now with the switches off if i move the front wheel then it should ride like a normal bike with both wheels electric and normal steering switching the first switch makes the back wheel copy the front wheel so they rotate in the same direction and steer in different directions and if we switch on the other mode then basically they mirror each other so theyre rotating opposite directions but they steer in the same direction so the bike goes round in an arc turning both switches off always makes the back wheel go straight and it operates like a normal bike again but the real challenge is going to be in copying mode to see if i can actually ride on the bike and steer and balance this is a bit of an experimental thing and this is just a test so first of all weve got our control for the experiment which is riding the bike like a normal bike with the back wheel fixed in position and the front wheel steering as youd expect and both wheels are still powered so as youd expect i can ride that like a normal bike and i can turn and all the things youd expect could have done with somewhere to put my feet up really but its actually not bad having them there on the ground so that i can put them out really easily if i happen to fall over but so far that rides like a normal bike its not particularly fast thats about the top speed but im only running on a s lipo and the motors are rated for s but it seems to be okay and feels quite safe and my braking strategy also seems to work quite well right its time for the first funny driving mode and that is both wheels rotating in opposite directions so hopefully they steer in the same direction and i think they should be perfectly fine to ride because it still steers as youd expect it to well it took a bit of getting used to because you tend to over steer what you really need to do of course is steer half the amount you want to because both wheels are steering which steers you twice as much as youd expect also its a bit like a rear wheel steering vehicle where the back swings out so its quite easy to get into a sort of snake motion although you can make some quite tight turns now so just doing some practicing going around in circles but obviously the bike still balances and you can ride a bike with two steering wheels that both steer in the same direction and rotate in the opposite direction so that works perfectly well now its time for mode three which is of course the wheels turning in the same direction and we dont really know whats going to happen here but there we go lets give that a ride and see how it goes so initially it seemed to work okay and i could balance on it my feet are off the ground and im going along but then something happened to it so lets see whats happened there and it looks like the back wheels gone all loose but yeah it was that safety zip tie thats broken so im gonna fix that up and have another go and yeah i can still balance on it although it seems to always turn in one direction and that probably means that the wheels arent tracking perfectly in a straight line so one is steering more than the other and thats causing it to go around in circles but in any case you can see that i am riding the bike my feet are off the ground and im balancing on it and it seems to be going along but then there was a bit of a disaster i was going to put a gopro on this to um see that we could see the wheels turning but unfortunately some of the plastic has broken so yeah weve lost the motor mount for the steering motor and weve also lost the clamp on the handlebars so it looks like weve proved we can balance with the wheels rotating in the same direction but i wanted to do some more testing so i went away and reprinted those parts and then i tried to ride the bike by riding it normally and then switching on the twowheel steering mode and initially it was far worse than it was before however it looks like the wheels arent tracking perfectly in the same direction so after a few adjustments and a bit more practice i managed to ride the bike quite successfully so i found this not too hard to ride if i ride in normal mode and then i switch on twowheel steering mode when im already going in a straight line and then its possible to balance for quite some distance and this is probably because if i try and steer too much with both wheels turning in the same direction it makes the back swing out so im almost traveling sideways more and more like a hoverboard almost but without the active balancing which makes it really hard to balance so doing it this way means that actually its not too hard to ride and i can do some long runs well i wasnt expecting the second mode to be of much practical use but it does show you can balance on a bicycle with two wheels steering even if those wheels point in the same direction and of course you cant turn because they need to point in opposite directions or just one steer for that but it is an interesting experiment twowheel steering where they both stare in the same direction that second mode is

write a youtube script

its part two of the giant rydon snake robot this is the small one that i built a while ago and you can check that out in my channel and now im making one big enough to ride on this is half of it but this piece goes and links to another two segments which i built last time that you can check out in last weeks video so each of these segments has got a driven wheel in and its got two wiper motors that are going to act as servos so it can bend in both directions its definitely strong enough to ride on and this time we need to put the electronics in and some sort of seat to sit on and some sort of control system to drive it so the aim is that it can bend up and down and left and right im not sure if its actually going to carry my weight and be able to bend up and down to go over obstacles very well because the servos arent that strong but its definitely an interesting way of steering and well see how it works at the end of the video first of all we need to turn those wiper motors into servos so theyre going to need some feedback so ive got some heavier duty pots here than the normal audio ones that i would use on small projects and weve got a d printed t pulley with a captive nut and a grub screw that fits onto the pot and the pot fits onto a bracket and that mounts onto the mount which i made last time for the servo so on the servo itself of its steel servo horn ive got another d printed part im going to glue and zip tie down that fits neatly over the nut on the wiper motor and then the feedback pot is going to fit with a belt so that we can adjust this exact position and that fits onto the bracket that i made last time so when the wiper motor turns the servo turns its pot and we can get the feedback theres lots of d printed parts in this project including all of the pulleys all of the wiper mounts and all of the flexible linkages between the parts so just a quick add from my d printing sponsor thanks to lolzbot for supporting my channel with d printers and thanks to d fuel for the filament for this project and lots of other projects so check out dfuel.com so weve got our pulleys and our pots all fitted and thats on both sides on each of the segments so theres six in total each unit of the snake is going to be quite modular so in each one theres an electronics tray with an arduino mega a battery mount and also somewhere we need to mount two motor drivers for each of the two wiper motors and theres also an esc for those hoverboard motors so weve got three of those together and those are linked together with a serial bus we have one which has got a radio receiver to receive data from my remote control and two which are just going to receive data from that one and control the other modules there are actually four sections but the front section doesnt have any wiper motors to push another section it just has a hoverboard motor so we dont really need dedicated electronics there ive been thinking about where to put controls for this so i can actually control it of course youve got various axes as well as the throttle to control and i thought about sitting on one of the back segments and putting some sort of steering column on one of the front segments its quite bendy which means the steering wheel is going to actually move as i steer which is going to be quite tricky so instead im just going to make it wireless and use the good old open dog remote which has got various axes of control and that means i dont have to bend my body while im steering and also i can just drive it with radio control as well and thats what the radio chip is for on that arduino to send serial data between those arduinos the main one is just doing a lot of serial prints and im using megas because theyve got lots of serial ports so that makes it really easy first of all its sending which is a value that none of the other variables will be and thats an identifier so that we know where the start of the data is and after that its just sending the other values on the receivers im just checking the serial data and then doing a loop basically going round checking that we get that value and it knows that thats the first value after that we can read the other values in order and we know the data goes into the right place so for now im just printing those out to the serial terminal so if we do a plotter and i move the sticks we can see that weve got all that data and this is on one of those arduinos receiving it over serial so weve got various controls here to bend up and down and weve also got a differential control to bend the two servos in opposite directions when i steer so that it moves side to side like a snake but when i move up and down they both move together i can also move that control to the right hand stick so that i can use the throttle and the steering with one hand and thats purely so i can drive it and film with a selfie stick in my other hand in order to actually drive it though im gonna need a seat to sit on so were back to the workshop to cut up some steel and make some additional pieces to mount on top yep im going to ride it like a motorbike because motorbikes are cool so i got this bike seat off ebay i cant remember what bike its from but its got these handy hooks in the bottom for mounting so im going to make up some steel to fit on there and so it can actually hook in on the mountings that it was intended to so ive just cut up some steel there to fit all the parts approximately then im going to tack weld all of that together and try and see if i can fit it to the seat the rest of the robots already made from steel so once ive made this and fit it to the seat we can just fit that onto the existing structure so im going to cut out some slots here which are for those hooks on the bottom of the bike seat so hopefully we can hook them in and if we get it right we should be able to clip it on just like it would clip onto a motorbike presumably so weve got two slots in our piece of ushaped metal that we now have and those should fit pretty accurately into the two hooks and the rest of the steel is supporting the seat so that seems like its going to work so with that bracket attached we can now try and place it onto the rest of the robot and i can see how im going to sit on it its pretty low to the ground so im going to be leaning forward quite a bit and i need definitely to support the front of my body somehow but i really dont want to attach it to the front of the snake because then ill be twisting left and right as i steer but were going to fix the seat on first so ive added another bracket to the bottom here and now im just going to tack that onto one section of the snake of course because it still needs to flex so its going to be fitted onto that section thats the second from the back and ive put that at an angle so the seats actually level which is presumably how it should be and that means that when i sit on it its a bit more like riding a bike so i could ride it feet forward like a big old harley with monkey hangers or probably not im gonna ride it like a sports bike so yeah i definitely need to um put my feet somewhere that isnt on the wheels and i definitely need to support the front of my body because i just cant hold that position otherwise its impossible so ive made another bracket which is going to fit also onto that section and this is kind of a set of handlebars so that fits into two holes that ive drilled on the same section that the seats mounted on and it gives me something to rest on it does need to be quite high up though so the snake can still flex upward and theres still clearance and ive put some foot pegs in which are just some tubes banged into the square section now they are on the back section which is going to flex but its not too hard to move my feet backwards and forwards as it does so so i dont think thats going to pose too much of a problem so back to the electronics weve got the arduino mega in a box with a breakout board for all of the i o and power made from some strip board ive also got the two motor drivers wired in a servo connector for the esc for the hub motor and two pot wires and theres three of those all together but before we carry on with that its time for a quick ad from the videos sponsor which is on shape onshape is a cloud native cad and pdm platform built for business created by the founders of solidworks because they saw that modern product developers still experience many challenges related to their cad and pdm systems one shape is accessible across all operating systems and works like google docs so an onshape document is a single source of truth for your design data onshape is great for working with teams using hybrid or remote working you can collaborate with team members at the same time on the same document across the world data management is built into onshape theres no need for file management on your local hard drive one shape uses a github inspired version and branch emerge model for fearless design experimentation onshape has industry leading manufacturing specific features for sheet metal and frame based design as well as surfacing configurations and detailed drawings an onshape is always improving new releases are pushed to the product every three weeks to add new features and functionality i highly recommend the engineers and product developers watching to consider using onshape for their business and you can try it for free at onshape.pro jamesbruton right lets get on with the snake i fitted the electronics in one section this is upside down with the wheels in the air at the moment so the motor drivers will be on the plate of plywood which is where the motors are and everything else is on that plastic mount which goes up the other way the esc for the hub motor is on the bottom which im not particularly convinced about is fairly recessed so hopefully i wont scrape it on anything so ive got my controls for bending up and down which bends both servos in the same direction and that seems to work pretty well and i can also bend left and right by twisting that stick or the right hand stick if i switch the other switch to move the control over so so far thats looking pretty good that just leaves us with three more sections to fit and one more hub motor to wire in for that full section at the front but so far im pretty happy with the outcome i did smooth out the values so that the data cant change too quick and it doesnt jerk around too much because the whole thing has quite a lot of mass and i did a demo in my channel about this with an animatronic set of eyes that i designed a while ago so you can check that out but basically what it does is take something like percent of the previous value and only percent of the new value when the values change it does that on a loop that goes around and around and that means the value cant change too quick and so we dont get really sudden motions and you can see that servo moving and giving feedback to its feedback pot which is controlled with the arduino with a pid controller to position it so these servos look like they should be pretty good for some fairly heavyduty animatronics ive now fitted the second section which is pretty much the same as the first one with the electronics linked on that serial bus so now weve got both hub motors which move together so thats working okay and weve got all of those servos moving together which you can just see the second one in the bottom of the shot so as we can see all of those servos move in the same direction so this second segment is going to link to the third segment and that means the whole thing is going to flex in unison with all of those sections moving together the front sections are also wired in now so that third section has the electronics again and the front section just has the esc and a battery so with all of the sections moving together i havent linked them in the middle yet because ive still got to carry the whole thing downstairs and its enough trouble in two halves but we can see that all of those move in unison so that should be good so lets put it all together we can see that it can bend up and down and it can also steer left and right and it can dry steer even though those wheels are actually really grippy and its really heavy there is a problem though which is one of those servos isnt moving properly in the middle section you can see just there so it bends left and right okay but bending up and down is a real problem so im just going to take the rod off there and see if the servos working or what happens but it looks like thats working fine so it seems like theres just too much load on it with certain moves when its right in the middle of the robot so ive got no problem moving to the full extent there but when i put the whole thing back together then that one doesnt push forward anymore now those motors on that section are actually getting really hot which is a bit concerning because it means the motors are over driven the back ones are cold though so i guess its just a problem which the middle segment has got the load of the rest of the robot on when its trying to arch its back and the servos just arent up to it so its probably not going to lift all of my weight so its time to give it a test outside this is the first time ive driven it so im not going to ride on it to start with im just going to drive it around remotely and try and get a hang of the steering and the throttle i do have reverse as well so thats pretty good so we can do threepoint turns and things like that youll see the wheels are lifting off the ground slightly as it flexes but obviously with the person on thatll apply some extra mass to push them down so its time to put the seat on and ive also attached some headlamps right so with that sort of handlebar thing on which isnt actually for steering i can ride it okay the only problem is im pushing my mast down quite a lot on that lever which is tipping that second section from the back where the seat is attached and lifting the back wheels off the ground also as i steer is moving my body in the opposite direction of course which is causing me quite a lot of problems and makes it really hard to ride so yeah you can see as i push down on there its lifting the back off which youd kind of expect im pushing those servos back and that steering rest is actually moving the opposite direction to the way im steering which is really disconcerting because its like adverse camber so were going to take that off and im going to try and find some other way of riding it we need to find another way to ride it and i definitely need to rest the front of me on something i dont want to be too far forward because i dont want to bend too much but i think im probably going to have to be because otherwise its just really uncomfortable so this is the most comfortable position but im gonna have to bend my whole body left and right as i steer but anyway lets give that a go so this is actually much more comfortable its not as bad as i thought it would be and after going quite slow and just getting a hang of the steering and remembering to flex my whole body as i do so its actually not too bad i could probably do with some foam arm rests because the top of those headlamps are a bit sharp but actually its not too uncomfortable to ride and it feels like quite a natural riding position Music so Music it Music it is of course possible to sit up on it and ride it like a gocart but its a lot harder to take those corners because youre much higher up and so youve got a higher center of gravity so you really have to lean into it so much prefer riding it in bike mode Music now ive had a bit of a go on it and got used to steering and driving its time to test out my skills with some cones Music me Music Laughter Music so steering works okay but weve obviously still got this problem with flexing it in the other direction it just about works with no load on it but of course if i sit on it then it doesnt really work at all i also think that if we can have something that climbs over obstacles we probably need more than one drive wheel on each segment maybe tank tracks and

write a youtube script

this is part two of the ultimate drift cart normally drift carts have powered back wheels with less grip so that they can slide sideways but in my case im using an omni wheel on the back which is like a normal wheel but has lots of longboard wheels around its circumference that means it can be powered to drift the cart sideways but drive forwards and backwards under power from the front wheels the back wheel was originally made for the balancing omniwheel bicycle project which you can check out in my channel the front wheels are inch hoverboard wheels which have motors built into the hubs so this makes building a front wheel drive with steering a lot easier last time i got most of the chassis together including the steering assembly and i got the omni wheel mounted on the back along with a motor and belt reduction gearbox to drive it all together it looked pretty good and i think its going to drift around just fine and under power which should make it super controllable this time we need to get the steering column built but before that i need to put the seat on so i can get the handlebars in the right place and work out what angle i need the steering column to be at Music hey Music Music Music Music so with two rails attached and the seat attached for me to sit on i can put my feet up and grab a stick which is going to make up the steering column then i can work out what angle it needs to be at to mount the steering wheel i need something to support that steering column which i can mount some bearing blocks on so ive got another piece of steel which is by im just predrilling the hole so its easier when ive welded it on i need a cross brace across the chassis here as well to mount a vpiece on so that we can make basically a little pyramid with three sides to hold the steering column and hold the steering wheel and that piece looks like this which is basically that piece of that i drilled holes in mounted at an angle and then its supported with two pieces on each side which i need to cut and ive cut those out of strips of steel although its quite bendy when theyre long theyre actually quite rigid when you cut short sections so that should be more than strong enough and then we can put one of those each side and that completes our threesided pyramid to hold the steering column and hold the steering wheel the bearing blocks themselves are fairly complex geometry because we need to seat a bearing in there and clamp them onto that piece of so im d printing those with a . millimeter nozzle so just a quick add for my d printing sponsor thanks to lolzbot for supporting my channel with d printers there are quite a few d printed parts in this project including these knobs for the steering wheel were going to mount the throttle on and well look at that in a minute the steering wheel or handlebars as they really are are also printed with the . millimeter nozzle and these are just pla parts which should be tough enough thanks to d fuel for the filament for this project and lots of other projects so check out my channel for more d printed projects and check out dfuel.com each of those bearing blocks fits onto that steering column mount and each one can be bolted through with a six millimeter bolt i actually used a finer nozzle to print the tolerance part for the bearing to fit and so weve got a bearing with a bushing fitted onto each of those and that allows the actual steering column to fit in there which is a millimeter steel tube theres a clamp on each one top and bottom to stop that falling out which is another d print with an m bolt through i made another metal piece which has got a millimeter collar clamp on one end that can be tightened up on that tube and a millimeter hole in the other end and that fits onto the bottom of the steering column with a millimeter bolt through so that we can put the joints in with the levers that actually push the wheels and these are millimeter rose joints and that means as i turn the steering column our steering moves and the wheels point in the right direction the handlebars are fixed on in a similar way with another collar clamp and a piece of metal with two bolts through the d print so that that will fix on and it wont come off with any luck so now if i sit on it i can put my feet up and i can turn the wheels and that seems to work perfectly well with the load of neon and i get quite a good range of flexibility there so that i can steer left and right and of course it seems to run fine when i push it along as well im not forgetting weve got that drift wheel and of course the plan is that well be able to drive and steer with our front wheels which are those powered hoverboard motors and well also be able to drift the back wheel under power and under control as we wish to so so far it seems to move around pretty flexibly and with all of that weve got quite a tight turning circle but before we carry on with that its time for a quick ad from the video sponsor which is kiwiko kiwiko is defining the future of play by making it engaging enriching and seriously fun kiwico have super cool handson projects and activities designed to expose kids to concepts in steam thats science technology engineering art and maths and theyre a great resource for learning at home each monthly crate is designed by experts and tested by kids more than a thousand hours go into developing every single crate kiwi co offers eight subscription lines each catering for different age groups and topics from toddlers to teens and even adults i built the crawling robot kit it was convenient because everything is included that you need with detailed kidfriendly instructions and an educational magazine kiwico crates are a great resource for learning a home theres hours of entertainment for curious minds kids can learn something new every month kiwico crates offer a wide variety of topics from month to month to support my channel click on the link in the description below to get off your first month of any kiwi co subscription theres quite a few more parts to mount including the batteries and electronics so ive added two plywood plates which are screwed to the chassis with selftapping screws and that makes the chassis a bit longer as well so ive got somewhere to put my feet and my legs on so bent i found this pedal on amazon which is basically for ebikes and atvs apparently which has a hall effect sensor in and thats just a normal accelerator pedal and thats going to drive the vehicle forward on the handlebars ive got two ebike twist grips and thats going to be the controls for drifting left and right so that should work pretty well each front wheels got its own brushless motor driver which as i mentioned last time are flip sky v esc clones we need to mix the three accelerators together so ive got an arduino uno in a box and on top of it is a switch for reversing and the whole thing is powered from lots of lipo batteries weve got one for each motor at the front and one for the motor at the back as well as an emergency stop so it drives like a normal car with the front wheel drive and its got a pretty tight turning circle as it is so i can just about drive around in the workshop although its too fast to be safe really and weve also got the back wheel which we can run sideways by using those paddles on the handlebar to give us an even tighter turning circle so it feels fairly terrifying that back wheels quite powerful with the to because weve got quite a lot of torque from that watt brushless motor and what we want to achieve ideally by drifting is steering with opposite lock which means for instance steering right but drifting the back of the car right faster so you actually turn left even though the front wheels are pointed right but were gonna need more space for that Music Applause Music Applause Music Applause Music Applause Music Applause Applause Applause Music Applause Music Music Applause Music Applause Music Applause Music Applause Music Applause so pretty happy i managed to get my opposite lock on and drift by steering right and drifting right and therefore turning left which is the main thing that i wanted to do with this its a bit unstable though with that single back wheel perhaps a better design would be to have two of those drift wheels at the back so that its much more stable so we have four wheels in total but i can drift backwards which you cant do on a normal drift cart as far as im aware so thats pretty good but again going backwards and stopping with that single back wheel makes it a little bit unstable anyway pretty happy with that im going to publish all the cad and code so if you want to build something like it you can and you can see how i mix those throttles together all of thats on my github page so if youd like to support me through patreon or youtube channel membership then those links are in the description below as well and youtube channel members and patrons can get access to all the videos up to a week early and sneak peeks and pictures of whats coming up so you can be involved in that discussion thats all for now Music you

write a youtube script

what about instead of steering by turning the wheels we bend the whole vehicle i previously made a snake robot which has multiple segments that could bend sideways and up and down each segment has powered wheels and that means it could bend in all directions to get traction no matter what the terrain was i really enjoyed doing that project so now im going to make a much bigger one that i can ride on so this time were going to have to build a much heavier duty steel chassis that will carry my weight and were going to need pivot points so the whole robot can bend to steer in order to bend up and down were also going to need some extra pivot points that we had last time and were going to need some much heavier duty servos to bend the whole robot but well talk about what those are in a minute so youve probably noticed in the last couple of months ive been doing a lot of projects with steel in them and prior to that most of my stuff was d printed there is still d printing in this project but i wanted to talk a bit more about making stuff out of steel so of course we can make bigger structures that are much stronger and i can kind of make them quicker than i can with d printing but obviously theres quite a lot of manual work so most of the stuff i dont show on the video is things like cleaning all the steel so the welds are nice and clean so all of that has to be cut debird and im cleaning every piece with acetone here before we can go and center punch and drill it so obviously with d printing i can pretty much leave it unattended to make all the parts although it might take slightly longer but with making things out of steel or wood or manual processes i can probably make a bigger structure thats stronger quicker but it does involve lots of manual work so effectively its harder work this project has multiple segments of course so that the snake can bend in all directions and that means ive got a lot of parts to make of the same instead of just printing them multiple times with a d printer obviously ive actually got to manually make the parts drilling all of the identical parts and marking them out and making sure theyre accurate by hand so weve drilled lots of holes in steel thats been cut to length im now tapping a hole here and this is basically going to be the mount for the motor and as in a couple of the previous projects im using hoverboard motors and those are going to be held in place with a bolt which goes onto the flat of the shaft on that hoverboard now ive learnt my lesson in the previous projects and im actually welding a nut on to back that up so were not just relying on the thread into the steel which is only about two millimeters thick so weve got four of those because weve got four segments and were going to have four driven wheels obviously any bigger holes have to be drilled out this is a mil hole cutting saw and thats for the main shaft of the mount of the hoverboard motor to go in the basic plan is that were going to mount two bits of metal on one end and thats going to hold a passive wheel which is actually a wheelchair wheel and thats the closest wheel i could buy which is a similar diameter to the hoverboard wheel and that hoverboard wheel is going to mount into that hole which ive just put that nut on and drilled the bigger hole in so lets weld that frame up im tig welding everything because i really like it so weve got some clamps to hold all that metal in place im just going to go around and weld all the straight edges to make the complete structure and of course this has to be done four times because weve got four segments in our robot which are all pretty much identical apart from the front segment that well look at in a while and my y brush is one of my favorite welding tools so here are the four segments and youll notice that theyre actually turned the opposite way round so weve got four passive wheels and four hoverboard wheels only theyre going to be more like this with a slight offset so those wheels run in line weve still got to make the upper deck though so those servos can bend the robot up and down but ill save you lots of footage of welding and cutting and all of those repeated processes again so heres the finished thing weve got a tshape at the front and a kind of rail at the back and yes ive had to do that at least three times apart from the front segment which has only got the rail at the back quite a lot of the parts for this project are d printed though and thats due to the fact that theyve got complex geometry and its the easiest way for me to make them so just a quick ad from my d printing sponsor thanks to lolzbot for supporting my channel with d printers and im using the lulzbot hs . nozzle here which is . millimeters so i can make those parts really tough and make them about four times quicker than a normal half millimeter nozzle and thanks to d fuel for the filament for this project and lots of other projects so check out my channel for more d printing projects and check out dfuel.com and im just using standard pla to make all of these parts of which theres quite a few some of these parts have bearings in and what ive done is printed a bushing in a finer nozzle so i dont have to reprint the whole part if that hole isnt the right tolerance in the fat nozzle but before we carry on with that its time for a quick add from the video sponsor which is jlcpcb jlc pcb are at the forefront of the pcb manufacturing industry and they provide high quality low price pcbs its easy to order from jlc pcb just click on instant quote and upload a gerber file the jlcpcp website will show a preview of the board and then you can select various options for manufacturing if you want the board assembled then you can add a bom and cpl file jlc pcb currently have an offer where you can get five one to four layer pcbs for just two dollars with free smt assembly jlc pcb is a onestop shop with a team of trained engineers to manufacture and assemble your board under one roof and with the new jlc pcb parts manager you can check available stocks order parts or even request specific parts for your project jlc pcb ships worldwide and they have fast build times so you can get your pcbs in as little as three to five business days the jlc store also sells pcb coupons and offers free pcb designs in the d designs the ordering process is very easy at jlc pcb so use the link in the description to this video to check it out now the small version of this project used standard size rc servos to move all of those segments to make it bend but the bigger version is going to need something a bit more substantial yes its windscreen wiper motors because theyre super powerful motors and theyre pretty cheap so each of those is going to be mounted in a super d print printed with that . millimeter nozzle thats got two recessed m bolts and those bolt into the casing that normally holds that motor together rather than mounted on the front on the actual mounting points and ive made six of those because weve got two for each of the segments each of those motors has a spline shaft on and theres an m nut which fits on top of that and that would normally bolt down onto the cam that operates the windscreen wipers rc servos normally come with servo horns which is the lever fitted onto the top but ive got to make mine because i need something much more substantial made of steel so ive got some millimeter wide steel that im drilling holes in ive got some super hss drills here which seem to go straight through it and i decided to countersink that hole so we can taper it so that it fits properly on that tapered spline shaft using my big vise and a spanner we can tighten that steel down and hopefully those splines cut into that countersunk taper so that seems to have seated right down on the spline shaft that should give us loads of grip and i can just about back drive these motors but they are a worm gear so its pretty tricky and its pretty stiff we need to cut some cnc plywood parts out because we need to make platforms to put those wiper motors on and its pretty tricky to mount them straight onto the steel frame because they push and pull in all sorts of different directions so im cutting some parts out of millimeter ply and weve got three of those for each of the sections that those will be mounted on i painted those silver so they dont look like plywood and of course each one mounts two wiper motors so the output shafts are directly opposite each other those get bolted down and thats looking like its going to be pretty substantial im going to use these rose joints with millimeter bolts as basically levers pushing off that servo horn and on the other end of those we need that to push the next frame so ive made these bent pieces of metal with millimeter holes in each of those mounts to the frame and theres a handy cut out to go around that tpiece and some bolt holes that are predrilled in all of the parts so we now have eight bolts holding the wiper motors and eight bolts holding it to the frame so that should be more than strong enough to hinge off the frame weve got some more d prints with those bearings in and weve got bolt holes cut in both of the parts theres an intermediate section though because we need to effectively make a kind of universal joint and that has ml studying through which i was going to bolt in hence the recesses but i just glued it in and that seems to be more than strong enough before i can link to the next section i need to weld the tabs on and yes those get welded onto three of the sections which are the front three sections because those are getting pushed from the previous sections so now we can put the next piece of the d print on and that goes on to that c shape and of course the pattern then repeats from the front of that onto the next segment so now we have four segments all together three with servos on and one which is the front that currently has nothing mounted on it now we can work out the length of those rods with the rose joints on that are ml studding locked onto the rose joints with some nuts done up and well just put some nuts onto those millimeter bolts on the tabs and the servo horns its time to put the wheels on so those are just actually mounted on studding those passive wheels have their own internal bearings and im using two nuts done up against each other to hold it in place so those fit onto every opposite side theres four of those in total and the rest of them are hoverboard motors of which theres also four held in place with that nut i welded on earlier the wheels arent exactly the same diameter though but i already worked out where to place the axles and thats possible because theres not a solid axle between them so i just predrilled the holes so that one of them is three and a half millimeters higher than the other so now it goes active passive active passive active passive active passive and thats all of our eight wheels theres going to be quite a bit of electronics to be installed before we can actually move those wiper motors as servos which will be coming up next time but for now it seems to roll along pretty well and it moves pretty smoothly it also looks pretty substantial so im pretty sure this is going to carry my weight and nothings going to break and everythings going to be fine so well just give that a test though before we carry on and seems to have absolutely no problems carrying my weight those servos arent really back drivable so the whole thing stays in place i set it up for a slight curve and everything seems to run just fine im not quite sure how im going to ride it yet i thought that maybe it could be ridden like a skateboard although steering is gonna get really tricky or the other option was to just lie down on it a bit like superman but for now i think perhaps ill think about riding it like a motorbike with some sort of seat on or really its a bit like a little train isnt it but that looks pretty good to me so next time we need to turn all of those wiper motors into servos by putting feedback on them and a microcontroller so we can control the position and then we need to work out how im actually going to ride on this and how im going to control it because we can bend in two directions and of course we need some sort of accelerator and we can also do a differential drive with the wheels on the inside and the outside when we go round corners so it looks like its going to carry my weight though so not too worried about that but if you want to see me actually riding on it then dont forget to come back next time and thats happening in the next video im going to publish all the cad and code for this once its done and the canon code for the previous version is already out there so if youd like to support me through patreon or youtube channel membership then those links are in the description below and youtube channel members and patrons can get access to all the videos up to a week early as well as sneak peeks and pictures of whats coming up so you can be part of that discussion alright thats all for now Music you

write a youtube script

its pretty hard to walk on a tightrope but i thought id have a go i set up a bar which is a steel scaffold pole and tried my best to walk along it sometimes i could almost make it to the end but most of the time i did far worse than that its a bit easier if you crouch right down and use your arms to help you balance but standing up tall makes it even harder standing still on the spot for any length of time is practically impossible but could i make a device that would help me balance ive built a couple of projects that used reaction wheels to make a device balance on the spot this is a wheel which rotates and dynamically accelerates in either direction based on the angle of the device force equals mass times acceleration so we only get a reaction force when the wheel is accelerating or changing direction i used an inertial measurement unit to read the angle and a pid controller to tune the response and you can check out my previous videos for more indepth details using the pid controller allows us to tune how much the wheel accelerates and how fast it goes and also how fast it stops when the device is upright i also built a onewheel balancing robot which uses the reaction wheel to balance sideways and then balances back to front like a traditional two wheel balancing robot or a segway note that this is different to a control moment gyroscope which uses a constantly spinning mass moved in a control axis to exert a perpendicular force in another axis ive also built various balancing robots using gyroscopes including another one wheel device and you can check all these out in my balancing robots playlist on youtube but how big would a reaction wheel need to be to help me balance so i can walk all the way down the beam Applause Music Music Applause Music Applause Music Applause Applause so that leaves us with a funny looking bracket with lots of bells and whistles and knobs attached but we also need some cnc parts and some d printed parts Music Music so Music just a quick ad from my d printing sponsor thanks to lolzbot for supporting my channel with d printers Music and thanks to d fuel for the filament for this project and lots of other projects so check out my channel for more d printing projects and check out dfuel.com so that leaves us with a steel frame with some d printed brackets that attaches neatly onto a frame rucksack which im going to wear on my back so weve also cut some wheels which are going to make the main reaction wheel and thats mounted on bearings on some bright steel bar welded to the frame and ive got a d printed htt profile pulley on there so that we can move that reaction wheel and hopefully itll help me balance so far its pretty comfortable to wear and its attached to me with both the waist strap and the shoulder straps and the link in between the shoulder straps to control the whole thing ive got a tnc lc which is more than powerful enough and an mpu inertial measurement unit im going to read the angle with ive made this trigger attached to a slide pot so i can actually ramp up and down the output and that means when i turn it on i dont get a sudden jolt that you would have seen on my balancing bike project so thats just a slide pot and a spring underneath so now when i turn it up and down and we move the inertial measurement unit we can see that motor is spinning faster or slower depending on how much i pull the trigger and that means i can ramp up and down smoothly and if i let go then it should just stop the gearbox to drive this is pretty similar to the last two projects i did which are both vehicles so weve got a brushless motor with a t pulley on mounted on an aluminium plate and again its one of the aero drive kv motors theyre about watts this time though im using a two to one reduction because we still need quite a lot of velocity and i think that motor will be more than powerful enough so as before thats mounted between the aluminium chassis where the motor is and a d print which holds the other end of its axle and ive got some d printed spacer blocks there so we can hold that nice and tight and those are bolted up and that holds the pulley nice and tight so weve got our two to one reduction running between the belt on the motor small to big and then small to big back onto the main wheel so now we can spin our reaction wheel and that seems to run okay so you can see theres the intermediate stage there with both of those pulleys and both of those belts turning and everything runs pretty smoothly so now we can control it with a hand controller by ramping up and down or turning it on and off and ive already written some code with a rudimentary pid config so that that reaction wheel catches up as we tilt either side but before we carry on with that its time for a quick ad from the video sponsor which is fan home fan home is a new brand dedicated to developing unique collections and build up models from the best loved brands like marvel and star wars as well as providing fully illustrated magazines with inspiring content in every product all products are original inhouse designs paired with inspiring content including behind the scenes material little known information and rare images every month receive exclusive products and magazines with inspiring content in your first package youll receive either two figures and two magazines or two assembly stages with parts and two magazines from the second package onward you will receive three to four figures or assembly stages with their corresponding parts and magazines enjoy the amazing journey into the universe you love ive got some ships from the complete galactic flea collection from star wars each model is diecast and has been handpainted to provide unprecedented realism and accuracy in every detail and it comes with a fully illustrated magazine the collection has been reviewed and approved by disney and lucasfilm limited collect a different vehicle and starship with every shipment the millennium falcon attack the death star and many many more and to celebrate the fan home launch early bird subscribers will also receive an exclusive starkiller base model just go to the link in the description to this video and use the promo code promo jay bruton ive got my belt nicely tensioned and ive done that by means of two bolts which are bolted into that t piece you can see just above the gearbox and those are pushing the gearbox down so our main belt is tensioned against the main wheel we need to put some mass on our wheel though and what ive got in the workshop is a bunch of steel so im going to cut up some box section steel and attach that to the wheel all the way around onto the predrilled holes that ive already left and yes we had to make this piece lots of times so that weve got lots of them to attach all around the wheel i drilled and tapped those with an m tap and that means that we can basically go and bolt them onto the wheel and i already left holes that are predrilled all the way around so im just using some m socket caps and then we can go and screw those on and that will hopefully add some mass im not sure if its going to be enough mass yet though because box section steals hollow and its actually pretty light but it looks quite compact so far so its time for the first test to see if it looks like its going to tip me upright so you can see the wheel spinning back in the other direction as i lean over and i can feel a little jolt but it doesnt feel like theres anything like enough to tip me back over as im gonna fall over so i added some steel bar all the way around to fill in the gaps and give me a bit more mass so thats just solid bar instead of box section and now i can still feel a bit more of a jolt and it feels like its nudging me back again but thats still not really enough so im gonna add some more steel bar to that by welding some steel bar to those bits of steel bar and that means that ive got slightly longer pieces which fit all the way around now so we have a bigger diameter which should give us more reaction plus weve got more mass which should give us a bigger reaction so lets see how that goes so im balancing on one foot here to see if it feels like its going to work when i balance on the bar and it feels much much better its like a person either side sort of pushing me back so i can really feel a jolt there although my bodys quite squishy and you tend to absorb the load of something pushing you rather than being completely rigid like the rigid reaction wheel that was just to stick to the ground so its going to be quite an interesting technique i think to try and balance with this without absorbing the load but yeah i decided to add some more steel to that i didnt want to make the wheel any bigger and there isnt any more clearance ive just added another piece which makes up that length so weve got two pieces right on the circumference of it so those are all added now so weve got pretty much the most mass without changing all the box section that im comfortable with adding its getting pretty heavy now though so to actually get this on i need to put it on the bench and then take the load but lets give that a spin and see what happens so it really is like someone either side nudging me back now you can see that wheel spinning up far less because its pushing me back more and that means that the motor doesnt reach its maximum velocity so im pretty happy of how this feels i dont really want to carry any more on my back and without increasing the diameter of the whole thing i think thats pretty much as good as its going to get but yeah it weighs probably more than kilograms now so im not really adding any more mass just because i dont want to carry it but can i balance on the bar well thats better than i could do without it lets try and stand on the spot im just going to get myself at the zero point on the ceiling there and then try and stand still so it feels like its working and i could stand still for a lot longer than i could before but ultimately i cant stand still for a very long time and yeah walkings not too bad its definitely helping me but its still not that easy to walk all the way along without falling off about halfway through so just retuning the pid controller to give it more proportional gain so that its more aggressive and pushes me back more well im able to stand still for a lot longer completely upright which i just couldnt do before now there isnt quite enough reaction because there isnt enough mass and the diameter isnt big enough so eventually i managed to push through the reaction force essentially and it doesnt hold me anymore but its much better than it was without it walking along the bar though isnt too bad it took me a couple of goes but basically ive learnt a technique which is pushing against the reaction on each side in sync with my steps as i said there isnt quite enough reaction to hold me totally upright but if i imagine theres people pushing me either side and i push off one and then on the other and then on the other on each side in turn then basically i can walk across and i feel like its not too unnatural it feels a bit like im walking like a robot like hondas asimo or something like that but basically after a few goes practicing with this pushing either side against the imaginary virtual people that are pushing me back i can actually walk pretty well im pretty sure i could walk much further if the bar was longer and it probably worked for quite some distance that is of course as long as i actually keep my feet on the bar im actually not too unhappy with that it does fundamentally work and i can walk on the beam better than i could without it its actually quite hard carrying kilograms on your back and stepping up onto the beam as it is so probably what i need is a much bigger reaction wheel with less mass on it to give a bigger reaction although actually theres a practical limit of course to how big that can get im going to publish all the canon code though if you want to have a look although i dont recommend you try building this so if you want to support me through patreon or youtube channel membership then you can and those links are in the description below to this video and patrons and youtube channel members can get access to all the videos up to a week early alright thats all for now Music you

write a youtube script

ive recently been watching animatronic videos on youtube there are lots of great mechanical assemblies for heads hands and other types of creatures built with d printing and other techniques if youve ever built a project using rc style servos in arduino you know that its really very easy to move servos around using the arduino servo library but youll find that the servos start suddenly move to their new positions as quick as the motor will go and then stop suddenly again this can be quite noticeable and it can make the project look less organic than it could so today im going to show you how to make servos or other actuators move much more smoothly using only two simple extra lines of arduino code Music Music thanks to d fuel for the filament for this project and lots of other projects so check out my channel for more d printing projects and check out dfuel.com i decided it would be more interesting to show you how this works if we had an animatronic device to actually demonstrate it on so im putting together a relatively simple thing with two eyeballs and some other axes that move so weve got a pivot for each eyeball and youll see ive left slots in the top there so we can get the screwdriver in for various things theres another lever thats going to drive each one of those pushed by a servo it only needs to be pretty simple so the eyeballs only look left to right theres no up and down motion although well look at that later in the neck you really dont need expensive hardware to make the motion of your animatronic project smooth so im using some really cheap nine gram servos which are branded with some random name and i got these off amazon these all fit into d printed brackets and we can have one of those to move the eye on each side all of the pivot points are just screws straight into the plastic ive used some of the smaller screws that came for mounting the servos as well so here we can see one of the servos mounted on its d printed bracket and thats going to move the eyeball side to side quite convincingly i thought it would be a nice feature to have a pair of eyelids on each eye that just gives it a bit more character and ive designed the top and bottom lids to fit together and hinge around the same pivot point and again these are just attached with a single selftapping screw straight into the plastic and into that single pivot point on both eyelids again ive added another cheapo servo thats going to push both of those lids closed and therell be one of those on each eye so i made some d printed levers with a slight crank in and that just means it can reach over the bracket thats holding the pivot point of the eyelids so that the servo can push and pull both of them at the same time without any obstructions and that seems to work pretty well i have of course be making two of these and these are mirrored pairs both of them are attached to a single bracket with a pivot point in and that means theyre both fixed together just like they would be in a creatures skull there are several pieces that make up the neck mechanism theres this intermediate piece that has two servos on and it also has a pivot all of this is d printed and all of its running just on the thread this particular piece is on an m bolt this is attached to the first section that has the eyeballs mounted on so that we can move around in two axes and that will of course be controlled by the two servos with two levers pushing the upper stage and the combination of the position of those two servos will move the two axes that whole piece has a piece of m studding screwed into the bottom and that fits into some bearings mounted in the actual base theres a washer here acting like a little spacer just to shift that top part up so it can run freely on the bearings and that gives us the side to side motion and thats controlled by a third servo with a lever and these are just pretty cheap servos as well that i got off amazon i built a breakout board so i can plug all of the servos in and distribute power and ground to them i just built this on call components perma proto board so that i can break out all of the wires and all of those are of course attached to digital pins on an arduino uno which is all you really need im powering the servos from a amp adjustable regulator turn down to volts and any old battery i had lying around and that should be plenty of current for this type of project to start with im going to use a switch to control this which just switches on and off ive made some really simple arduino code that reads that switch into a variable it reads it in and then multiplies it by a hundred so we get quite a big number and then im typing it out to the serial terminal so that we can see it and im running this every milliseconds so that it runs times a second so if we open a serial plotter and have a look at the switch we should be to see that we get a square wave as i switch it between and and we can see thats quite sharp so if we to use that to move a servo then it would move as fast as it can to its position and then stop abruptly but what we really want to do is not let that number change too quickly so we get a smoother motion so ive created two new variables one called switch smoothed and one called switch previous and all of the code is in here which is just two lines so what were doing here is taking only five percent of the new value that were reading from the switch and multiplying by and of the previous value we then create the previous value by bookmarking it here so its the same as the new smooth value weve created im then typing those out to the serial terminal so we can see both the original value and the new smoothed value so now in our serial plotter we can see the same original blue square wave and the new data which is the red line and we can see that this slowly decelerates as it gets towards its target and thats because the number cant change that quickly because its largely using the last value and only a small fraction of the new value on each cycle of the loop but before we see how that goes its time for a quick ad from the video sponsor which is pcb jlc pcb are at the forefront of the pcb manufacturing industry and they provide high quality low price pcbs its easy to order from jlcpcb just select your shipping destination and click on quote now and upload a gerber file the jlc pcb website will show a preview of the board and then you can select various options for manufacturing save it to your cart and enter your shipping information jlcpcb currently have an offer where you can get five one to four layer pcbs for just two dollars with free smt assembly jlc pcb ship worldwide and they have fast build times so you can get your pcbs in as little as three to five business days the ordering process is very easy at jlc pcb so use the link in the description to this video to check it out now so still just using the switch to get my data input ive now mapped the data output to one of the servos ive actually turned up the ratio so were now using of the previous value and three percent of the new value and that makes it even smoother and you can adjust the ratio as much as you like but you can clearly see that that servo is decelerating as it gets to its target and that means we get a really smooth organic motion i thought i should just mix in the eyeballs as well so im still using the switch and you can see my two waveforms so that my eyes also decelerate as they move in each direction so im pretty happy with that so far but i thought we should breathe some life into it so i got a pair of joysticks mounted in a box from an old project and these each have three axes as the tops turn as well so thats wired into six analog ins on the arduino uno although im actually only using five of them ive applied the smoothing to those values and now you can see no matter how fast i move the stick we always get that deceleration in the animatronic and that makes a really organic motion that im really happy with my left hand stick is controlling the three neck axes and my right hand stick is controlling the eyeballs and the eyelids obviously ive got one axis spare if i actually could move the eyeballs up and down but for now i think that demonstrates the concept pretty well of course we still get a sharp motion when we start the movement and thats pretty normal because most creatures and people will actually move fast to start with and decelerate as they get to the target so that actually makes it quite realistic and quite lifelike and if i suddenly let go of the sticks you can really see the motion smoothing on all those servos if we want to run our animatronic or prop on a fixed sequence of motions we can still do that but we must keep the loop running all the time so that we keep that smoothing over each iteration of the loop and that means we cant really use delays which would block the code ive based my code on the adafruit learning center multitasking arduino tutorial so yep lets ditch that delay and the idea here is to use the system clock so we can check the time see how much time is elapsed and trigger events without using delay which would block the code for us that means seeing if a certain amount of time is elapsed incrementing a flag to the next value and resetting the clock to the current time then the next time round the loop we can see if the flag is at the right value and another set of time is elapsed and then we can increment the flag again and reset the clock again to the current time and in each step of the sequence im setting my pot values to whatever positions i want instead of moving the analog pots on the controller and after this ive got the same filtering and smoothing code applied so the result of this is that the loop keeps running round and round at or hertz or whatever we set it to but because the loops running the smoothing still works on each iteration of the loop decelerating those values each time the loop runs it checks for that step flag to be at the right value and for the time to have elapsed and doesnt do anything else otherwise and multitasking means of course we could run multiple step sequences at the same time because our loops always running really fast and its only kicking off the steps if the time is elapsed and the flag is at the right value so we could have several sets of things running on one arduino and that would all be fine its particularly important though because the motion smoothing relies on that loop to keep running so that it can keep iterating on each loop with the previous value and the new value and chopping them up to give that deceleration even though theres no new inputs so i think thats been quite interesting if theres any more little tutorials youd like to see about how to do something with an arduino or something else then let me know in the comments im going to publish all the cad and code for this if youd like to have a look at it so check that out in the link in the description below its on github and if youd like to support me through patreon or youtube channel membership those links are below as well and patrons and youtube channel members can get access to all the videos up to a week early as well as sneak peeks and pictures of whats coming up to be part of that discussion alright thats all for now Music you

write a youtube script

once upon a time i built electric rockets with ivan miranda and the idea was that they would fly up and then land on their bottoms just like spacex well mine went up in the air but then it came down again not exactly as intended uh right well i think weve got glue but i didnt feel too bad about it because ivans crashed as well so now im going to attempt another complicated flying project im going to try and build an ornithopter which is something that flies by flapping its wings not quite as big as this one though so i bought a brushless motor and an esc off amazon this is a kv motor which means it does rpm for every battery volt thats quite fast but the smaller ones tend to be faster it came with a amp esc which should be enough power and the two were supplied together this plugs into a battery and an rc receiver its also got an rc set which is a pretty standard thing and again i just bought the first one that i came across we need some sort of gear reduction so im going to use a belt drive so ive d printed this t pulley and youll notice ive also got a bearing mounted on it i designed the pulley to fit on the collet that comes with the motor and that would normally be to attach a drone rotor that just screws down and it clamps itself tight onto the motor shaft youll notice this whole arrangement is quite long and since the body of the motor turns and its mounted only on the bottom thats what the bearing is for so that we can double brace it so my wing arrangement is of course two levers that we can attach to something that turns rotary into linear motion so weve got two levers one for each wing and then weve got a big pulley that turns around and thats going to have two levers to operate the wings so ive just put some levers on and youll notice ive got multiple holes here so we can adjust the leverage angle but well start with that ideally we need a rotary motion for each wing and you can see as i turn this theres a bit of a funny wobble as the pivot point on the big pulley moves side to side but this will do to assess whether weve got enough motor power and velocity to start with ive mounted the motor and youll see theres a d print on top that holds that bearing so that we can double brace it and we can put a sideways load on from the belt tension so there it is all in one piece lets power that up and that seems to run pretty well it probably runs far too fast but for now that looks okay so its time for some wings and ive got some carbon fiber rods and these are eight millimeter diameter and half a meter long those fit into the two holes and ive just packed them out with tape to make them a nice tight fit so lets see how well that runs yet that motor really isnt happy with the load and the mass of those wings so were gonna put a double stage belt reduction in so we move up to around a to one reduction Music Music thanks to d fuel for the filament for this project and lots of other projects so check out my channel for more d printing projects and check out dfuel.com so this time my motor is held either side of a plate which is the chassis for the whole belt drive arrangement on top of the plate is a bearing which double braces the pulley again so we can put a sideways load on it this time weve got our two stage belt reduction ive used some t belts which are nine millimeters wide the whole things a bit chunky but thats what i had the final pulley is supported by two similar bearings and that gives me two rotary outputs one for each wing which will make it perfectly symmetrical we should have a lot more torque this time and it should run more than fast enough in fact that looks like its too fast so i changed the motor for another one i found thats kv which will give us the same amount of torque but with dissipating less heat and itll be happier running a lot slower ive attached my two levers from the rotary output to each wing and these like all the other parts are just printed in pla that seems to run pretty well and i can vary the speed okay so it looks like its going to go fast enough hopefully with enough torque and i can vary the speed over quite a good range of motion and im not really sure how fast the wings need to move at this stage so carbon fiber rods back in and lets cautiously test it to see whats going to happen this time so that looks pretty good my wings move far enough and i can run it full speed with no problems well i just need to fix the rods in better my wing fabric is just some rip stop kite fabric which claims to be airtight that i got off amazon its just attached with packing tape but if we now power up the mechanism this is a true load test because weve got the air resistance of the wings as well as the mass of the carbon fiber rods but before we see how well that works its time for a quick ad from the video sponsor which is bittrex matrix helps companies grow by providing affordable and quality instruments for customer relationship management or crm including a contact center task and project management website building internal collaboration and communications project planning can be tricky but bittrex can help by providing strong project management functions including gantt stretch and deadline bars you can move stickers on kanban boards and track progress visually you can also organize dates and events with calendars and scheduling and set sequences dependencies and subtasks bittrex provides a fully functioning crm for free its unlimited and suitable for goods and service based businesses in different industries both big organizations or small teams and entrepreneurs there are free tools to generate new leads from the contact centre and your website process deals on a sales funnel quote and invoice your customers and receive online payments run targeted marketing campaigns and more bittrex allows you to manage your customers and your employees remotely with one platform it provides tools for work and to supervise full transparency and accountability are available access rights adjustment can provide privacy and security for several teams working simultaneously on the accounts so check out bittrex now using the link in the video description below and get gigabytes of extra online storage for a year right lets check out this thing and see if it flies so i thought i might as well test it outside but the main problem seems to be that its front heavy or at least it pulls itself into the ground all the time so my wings dont quite reach the end of the pole so we can make those bigger to get more lift but i actually put the battery right at the back here so i was expecting it to at least be balanced and if i do try and balance it with the battery in the same place it seems pretty well balanced in the middle or at least its not front heavy so time for a slightly different iteration of the design ive now moved the holes where the carbon fiber wing poles mount further forward we should move the rest of the mass back ive also added this tail section which will allow us to put a rudder on for steering the rudder will have a bigger flap attached but for now ive just put the servo and everything on so we can move some of the mass back and heres my new wing mounts which seem to work just as well as they did last time the prints are probably a bit chunky though those wing mounts also give me around millimeter extensions on the wings so now my wings are even bigger but the whole thing still seems perfectly happy well it definitely doesnt dive into the ground anymore so im going to try moving the battery around and see if i can make it a bit more front heavy well it still has a tendency to lift at the front so at least there is lift although it seems pretty unstable in the air ive added a tail fin which looks like some of the others i saw on youtube with a former that causes it to curve and im just not going to bother with the rudder for now i started with just a gliding test to see if its well balanced in the air and it looks not too bad i think that tales definitely helped and some of these dont seem too bad but its still pretty unstable so its time for another tail fin to hopefully keep it going in a straight line and ive just made these parts out of some cards stuck on with tape so still quite a few challenges staying stable im not sure if the wind is catching it but it does look like its lifting at the front so there are some obvious problems but first of all whats good about it well the structure held up and despite all those crashes into the ground and the mechanism held up thats operating the wings despite only being those pla levers and so on held on with screws at each end as pivot points and nothings actually broken also theres enough power with that reduction that weve got and that motor to drive the wings they move far enough with enough power so thats pretty good and that part of its been pretty solid for the next iteration though we need to remove some mass so im still going for a belt reduction because i think thats pretty forgiving gears can get mashed whereas belts just slip a bit and then they carry on usually so im going to make the belts a bit thinner and use five millimeter belts and also remove lots of mass from the pulleys and make those much much thinner im also going to mount the whole structure on three carbon fiber tubes and make really minimal d printed parts to hold the mechanics so as well as having to remove mass i also think my wings are a bit too floppy so ive looked at lots of these on youtube and lots of people have made them theres lots of kits available as well and the wings seem to be much tighter so ive got some thinner carbon fiber rods which are probably going to be quite flexible to try and brace those wings out a bit like a bats wings and hopefully tension that fabric so hopefully thatll make a more rigid wing that should flap properly im also not sure if these rigid carbon fiber tubes ive actually got holding the wings should be rigid or whether they should flap like a bird so ive also got some carbon fiber rods the same size it claimed to be flexible so well see how that works out so there is going to be a part two to this and hopefully we can at least get it flying quite convincingly at least in the straight line so dont forget to subscribe for more updates on this project and lots of other projects and like the video if you wish me luck you can support me through patreon and youtube channel membership and those links are in the description below and patrons and youtube channel members get access to all the videos up to a week early as well as sneak peeks and pictures of whats coming up so you can be part of that discussion alright thats all for now Music Applause Music you

write a youtube script

this is my tabletop bowling lanee its kind of like beer pong except literally not at all and is instead bowling okay so this week we worked on the automatic ball return last week we built a working conveyor belt and now were motorizing that conveyor belt to drive the pins and ball against the divider that separates the pins from the balls as you can see the ball is getting stuck also the gutters are directly facing each other so the ball just slides out of one and goes into the other all we have to do is change the ball exit to give it a little angle also the divider is slanted to force the ball into the return slide after after printing out my new parts the automatic ball return was working really well also our little gutter fix worked perfectly however the ball is getting stuck in the opposite corner I may just have to increase the angle of the wall divider next were going to work on returning the pins follow to see the rest of my build

write a youtube script

this is my tabletop bowling lanee previously I added these corner pieces to the gutters because I didnt have a printer big enough to make a conveyor belt as wide as the lane plus the gutters but I recently got the prusa XL and Im updating the conveyor belt I also added these side walls so the pins will be forced to go onto the conveyor belt on impact and thatll make it easier for the sweep to clean up the pins that fall into the gutters last week the ball was getting stuck in this corner but thanks to all your great ideas including making the divider jshaped and angling the conveyor the ball has a much easier time getting to the turn also the pins fall over too easily I dont know how to describe it its not as satisfying as heading real pins and I think they have to be heavier right now the pins are solid but I tried only having the bottom solid and the top at infill to lower the CG it didnt really work next Im going to try putting these tungsten weights in the pins follow to see the rest of my build

write a youtube script

this is my tabletop bowling lane today were going to work on the pins because right now its a little too easy to get a shrike they fall over a bit too easily from vibration so an easy fixes just to widen the Basse a little bit Im also going to print the Bas in TPU to increase the friction between the lane and the pins Ela doesnt bond to TPU but I can print them on top of each other with interlocking geometry I also added tungsten weights to the pins to more closely control the weight and CG of the pin I kept lowering the tungsten weight in the part until it sturdy enough to resist small vibrations however I turn this knob a bit too far and messed with the CG of the the pin notice how these pins sort of just tip over and are tightly fixed around the base of the pin but in reality the pin should be able to rotate midair to be able to make strikes easier the pin should be about a quarter the weight of the ball and the CG should be about a third the total height from the bottom of the pin spacing is also really important as the ball should be able to almost fit between two pins so I printed this guy to show me the spacing of the pins this final version is much more accurate to reality but the pins are still a little too resistant to wobble follow to see the rest of my build

write a youtube script

this is my tabletop bowling lane so in the last part we added some major functionality but Id like to minimize the amount of electronics use instead of two Motors well be using one with the gear train to control the conveyor belt and the pin wheel at the same time a little Lube never heard anybody since the conveyor belt rotational direction is perpendicular to the pin wheel Direction I Incorporated some bevel gears to change the direction of motion this was my first time printing bevel gears and I dont know its interesting to learn about these things in school but its actually so cool to use them in practice I also added these little pincer hands to the divider to funnel the pins into the pin wheel so they dont get stuck in the corners the pin wheel has perfect geometry to hold the pins in place but would otherwise fall out naturally due to gravity so I added a kidney just like in re bowling to hug the pins perfectly until they reach the apex of the wheel follow to see the rest of my build

write a youtube script

now Im quite keen on model aircraft as you can probably tell with the plane and the drone on the wall behind me and a lot of my early videos on this channel were to do with male craft however recently Ive been so tied up with trying to get my electric bike working I just havent had time to design build and fly anymore lair craft however I might know someone who can help me hey man hows it going yeah yeah good good so Im thinking of a new project I havent built a plane in a while and thought Id come to you for help because you obviously build planes more ofter than me what Im thinking is something like some crazy thrust vectoring idea right one thing you could do is have instead of having canards at the front of some sort of delta wing theyre not delta wing but some kind of yeah some kind of jet light thing you could have two motors that tilt and then maybe that would be a bit of an alternative to like just having a normal thrust vectoring okay yeah so replace canards with two motors yeah I like that something something different I havent seen that before okay I tell you what Ill design the two mechanism d printer or something like that then do you wanna build the plane yes yeah I got to work sounds good so for those of you that dont know that was a guy called James and he runs a YouTube channel called project air Ill be linking his channel down in the description below and he builds all sorts of experimental aircraft and he also works for another YouTube channel called flight test which you may have heard of if youre interested in will air craft and watch youtube videos theyve got quite a large following and he actually writes the articles for their website now he just sent me the drawings some sketches of the design of the plane that were going to be building Im not sure how well they show up in the camera so Im moved to the side and put an image here and what it looks like is a delta wing style fighter jet were usually canal odds would go either side at the front of the fuselage instead of having clouds there is two spars with motors mounted on the ends of the spars now as Im going to be designing this tilt mechanism part Ive had a quick think about how Im going to do it now it would be quite nice to have the motors tilt individually for roll control during usual forward flight however I just think the complexity of this would be quite difficult to get around and also the added weight because Ill need to renter the strong electronic servos to tilt these motors so one thinking is that the motors will tilt together and so there wont be any roll control from the thrust of the motors but also these kind of fighter jets they have quite high roller thority in comparison to the pitch so yeah what Im thinking is that the motors will tilt together for exaggerated pitch control so that will be the thrust matching at the front of the plane rather than at the rear of the plane which could bring some interesting flight characteristics to the plane now in terms of materials I believe James is going to be building the plane out of foam board to keep it as lightweight as possible but this section that Im building has to be a bit more chemically strong so whats almost like you do is use aluminium spar section like square box section forward the spa at the sides where the motors will mount to and then Ill most likely d print the hinge section of it and the servo mount for rotating this bar so what Im going to do is Im going to jump into the CAD program and see if I can make a design lets go heres a quick design which Ive made in fusion as you can see the motors are mounted on the ends of this aluminium square spar section and Ive designed lots of holes to be cut in this aluminium section which Ill most likely be doing on the CNC machine the wires for the motors will be routed through the center of this aluminium spar section and come out through the d printed part where Ill mount these speed controllers moving onto the d printed part Ill most likely be printing this out of d prints UK PLA plastic and as you can see here Ive modelled the servo which will be used to tilt the tilt mechanism I havent finished modeling this part but basically there will be two push rods going from the servo to this small servo horn here which will allow these servo to move this tilt mechanism now what I need to do is d print this blue part and see and see the holes in this aluminium spar section Music right so just before I send this up to James I need to test that everythings working I think Ill add it all up correctly I only have one receiver and I havent even program my transmitter so Im going to be testing everything individually so lets plug this in so now still down Music and that sort of spin the right way everything seems good so what Im going to do is Im good to neaten up these wires and stick in the post and Center to you James boom its fun so basically what we have there James can probably run us through yeah so the the fuselage its all made of foam board so from flight test and flight test foam boards that they use over in this in the US but yeah so this is just it starts off with cutting the wings out and then went on and made like a sort of flat bit for the fuselage stuck that on the top its all very like not all like computeraided design like on those but its more a proper way well I would say that with pencils and stuff and then I did edit the performers to make this sort of round roundness of the fuselage and then yeah stuck a nice coating on it and put a fin on the top and calls it an aeroplane thats base lets talk about it a bit about sort of the weight in the center of gravity the two mechanisms quite heavy isnt it yeah so thats thats gonna move the center of gravity forward so we need to put the battery further back right in the back here cuz I think when I weighed it before I sent it to you it was grams yeah Im not sure how much the plane is because we see that okay thats good idea I did it in such a rush yeah so well say about grams and yes grams and about quarter to a third is so what were gonna do now is get the plane ready to fly we need to hook up or like program all the electronics and we yeah program the transmitter and stuff this will be on Jamess channel the project air tunnel so yeah just head over there and well be doing all the programming and then getting sending this thing into the field and then put it up in the air a little bit yeah and please go subscribe to change the channel thank you he makes all sorts of planes on this yeah he works with flight tests as well so its an interesting content so there will be a either a box here with a video yeah or therell be a link in the description yeah so go check out James channel leave a thumbs up and subscribe make sure you match see in James video see that Music

write a youtube script

so last week I partly finished the motor mount for the electric bike version and was waiting for the police to arrive so since then I made a start on making all the rest of the parts for the motor mount now because the plates of the motor mount have to be separated perfectly and also perfectly parallel due to their being bearings of seats in it and also a shaft that goes through I had to make some spaces that held the two plates exactly millimeters apart now fortunately my granddad owns a lathe so I was lucky enough to use his lathe to make these faces out of some aluminium tubes the lathe also allowed me to shorten the length of the motor as I mentioned in last weeks video the stub on the end of the motor got in the way of the pedal cranks so being able to cut this off and make it look relatively nice was quite handy he also has a grinder so I could grind some flats on the end of the secondary drive shaft so that when I mount the pulleys on to the drive shaft they have a flat bit for the grub screws to grip onto the pulleys arrived shortly afterwards and it was time to assemble the rest of the bike but here comes the first problem my camera ran out of battery now Im not sure whats the best way to kick myself but I definitely deserve a kick right now I basically decided to take the bike for a quick test ride and yep the pulley failed and I have no footage of its destruction so what happened well basically I rushed a design to try get done in a week if youve watched the pulley video I actually went through two verses in one week and I overlooked a very simple design mistake this is what one of the pulley spokes looks like before its broken and this is what all these six pulley spokes look like now it seems there is just too little material around the bolt holes at the bottom of the spoke which when torque is applied from the motor puts an upwards force on the righthand bolt and a downwards force on the lefthand bolt ripping the lefthand bolt out of its hole now the reason for there being so little material on this point is that the pulley has to fit over the disc brake hub on the wheel which is millimeters in diameter so just before I finished designing the spokes I had to cut out this millimeter diameter hole in the bottom of all of spokes so that they fit over this hub now I do in fact do some structural analysis or simulation in fusion threesixty on this spoke before cutting it out however I got some unreliable results because I have no idea how much talk this electric motor actually produces I had to look back on one of my version three electric blanket which ran the same motor I will also put a link to the motor in the description below because a few people asked in my last video and what I found is that the bike accelerates from zero to thirty miles per hour or thirteen point four meters per second in about seconds this gives an acceleration of about . meters per second squared and as the mass of the bike and me is about kilograms it gives a forward force of Newtons with the tire radius of not . to meters this gives an average talk of . Newton meters now because this is an average talk Im not sure exactly what the peak torque is it could be quite a bit more than this and plus the electric by bedroom fall is geared for extra torque but less top speed so what Im going to do is Im going to double this torque value as a safety factor so lets use eighty five point eight Newton meters so because the radius of the pulley is naught point one four meters and this talk has to be shared between six different spokes basically one pulley spoke has to withstand Newtons of force at its tip which when these values are entered into the fusion simulation gives a safety factor of two point one six which basically means the pulley spokes should be able to withstand double the amount of force without braking now for those of you engineers out there I did in fact use the smallest mesh size possible and this took quite a while for my computer to compute it but I had to do it just to get the most accurate result possible but it still wasnt correct so obviously after the real pudding destroy itself I decide to go back to the fusion simulations and I realized I was carrying out the experiments completely wrong Id originally fix the pulley at the bolt holes at the bottom here and then applied the horizontal force at the hole here but what this didnt take into consideration is that these bolt holes can be ripped out so what I had to do is I had to change the simulation and basically use a bolt constraint in fusion to these holes a fake hub and really the simulation this gave her a result that looked a bit more accurate with a safety factor of not . basically meaning it was going to break I took this testing a step further and I actually took the old spokes mounted them to the hub and then clamp the hub in the vise I put them in the opposite direction so it would rip the other bolt hole out and it would give a pretty accurate test even though one side of the spoke was already broken I then hung weights from the end of the spoke and tested how much force it could injure okay this is the first test with kilos deforming I dont know how well you can see but theres definitely some deformation around this bolt hole here okay so Ive raised the weight to . kilograms and slowly lowering the weight and guess what they failed at . kilograms which is about Newtons of force which is also only slightly higher than the fusion prediction so I got on with designing some new pulley spokes which not only had more material around where the bolt holes were the bolt holes were actually spread further apart to reduce the amount of tension and compression or forces when its put under torsion I also designed and cut an aluminium ring which would reinforce the d printed section of the pulley but before cutting out all the bits and assembling the pulley I decided to recreate the same structural test by clamping new pulley spoke in the vise and hanging some weights from it and it managed to hold kilograms without budging even with a few hits from a hammer the next day or so it was spent cutting tapping and bending all of the new pulley parts as well as printing the new d printed tooth profile section on my new d printer which was sent to me by any cubic its their new Chiron model printer I think I pronounced that right and its pretty nice so far its got a huge build platform and its got some really nice features that my old gramma TCR didnt have Ive only printed a few things with so far so I cant give a concrete verdict however Ill be posting a link to it down in the description below so go check it out if youre interested in a new printer the Polly was an assembled on the wheel and then mounted to the bike and it was ready for another test ride but of course this time I was going to record it yes its past the first test of exceeding walking speed then a bit more talk cause about to slip time to add some more attention I then slowly increase the throttle inputs each time to see how the pulley handled and it seemed to hold up pretty well Music however the issues dont end there the pulley seems to withstand all the forces now so thats all solved but I tried to remove some of the smaller police to threadlock League grub screws and they werent budging even with both the grub screws removed from each pulley and I also heated up the pulley to try and expand it on the shaft it wasnt coming off basically what happened is that the stainless steel shaft that was quite a soft stainless steel and the grub screws that clamp it to the shaft had dug into the stainless steel and from the torque of the motor the grub screw had basically chipped some of the stainless steel and this chip had lodged itself in the aluminium of the pulley basically locking the pulley to the shaft after quite a lot of hammering and using plies on the pulleys was try not to wreck the teeth of the police I managed to get them off eventually however I then needed a sauce hardened stainless steel or hardened steel shaft for a replacement I then realized that I still have the old motor with the old millimeter shaft which is a Highland steel shaft so I decided to put that in the bike and all was good now the bike is ready for me to mount all the electronics in properly and finalise it and then go for a proper test ride well not really I also noticed was riding the bike at a certain speed at about miles an hour the bike started to vibrate quite violently and I think this was some kind of resonant frequency caused by the small belt or the secondary driveshaft I just need to pause the video for a few seconds I came down to the shed to film the vibration of the initial pulley stepdown and fortunately its not actually something wrong with my design but unfortunately there is a design fault or something wrong with the new motor that Ive got I apologize if that audio is a bit bowing just filming this on my phone but let me try and show you Ive disconnected all the belts and theres still the vibration occurring so it seems like the motor is unbalanced so as you can see here this is the motor output shaft and theres no belt connected to this step down and then the rear wheel is also disconnected from the belt for safety reasons and if I throttle up certain rpm youll hear a really violent vibration Music Applause Music Applause now Im not sure how Im going to edit this quick video that Im filming on my phone into the main video because I think I explained that Im going to go back to the old single pulley setup but now I found out that is the motor then I need to try and solve that with the motor and that my design with the motor mount is actually okay so we might still be good with the high talk dual pulley stepdown setup anyway and obvious talking lets go back to the main video Im coming tempted to run the main drive belt straight from the motor so get rid of the dual pulley stepdown but then that removes the whole idea of having lots of torque on this electric bike and my last weeks video would be completely pointless because I have to remake the whole motor mount so thats where I am with the electric bike Im not really sure what to do next I hope the whole analysis of the pulley the structural destruction of that was of interest to you I need to take a step back from the whole project and do some more thinking about what Im good to do next I would rather do that then you know do another step forward and then two steps back again so yeah I want to spend a bit more time trying to engineer it properly I am I didnt say at the very beginning of the project that I wasnt going to rust this but the excitement of having an electric bike just got the better of me and I ended up rushing it I also need to mention that there wont be a video next week I want to start going back to doing fulllength project videos every you know couple of weeks or so mainly because Ive ended up working weekends where I get these projects done because it normally takes a day or two to and the rate and then edit the videos really for upload so only leaves me about three week days to work on a project which just isnt enough to make a decent length video especially with the various projects that I want to do so Im quite looking forward to spending a good week or two on a project and then editing it you know finalize project make sure everything works and then upload it to youtube so that you guys dont have to watch all these fails as much i will still mention the bits obviously when something fails but instead of making a whole video on it it will be you know its a little segment of the video because I think its important now obviously I show the mistakes I make yeah Im just looking forward to spending a bit more time on each project also Ive got a few collaborations coming up in the next few weeks so it might not be that long before the next video probably two weeks or so so Im not completely stopping the weekly videos Im just saying that Im not going to rush a project to get out for a week if its not finished and its not working so if you dont if you dont see a video from me on Friday then dont worry Im just spending a bit longer on a project so stay tuned for the collaborations that I have coming up also for the later ebike its not dead just yet just might need to spend a bit more time on it and yeah Id like to thank you very much for watching and if you enjoyed this video please leave a thumbs up if you need to my channel please click subscribe a huge shoes thanks that all of my patrons for supporting me you guys make these not so weekly videos possible I will be uploading behind the scenes stuff you know what Im up to photo wise on patreon so not for videos every week but I will be at least posting photos up on patreon if you want to know what Im up to any current time thanks once again for watching and Ill see you in a few weeks good bye

write a youtube script

Music I built an autoentry turret that is capable of tracking humans and shooting them on sight killjoy is valerent genius German engineer in the yellow puffer she uses her Sentry to play lock down defense and pick off enemies and I made a fully functional prototype of her turret since you guys blew up my last project Boomie I was able to partner with Red Bull for this project in months I brought the turret from game to reality just in time to travel to La film with Red Bull and Kio and attend the valerant world champion let us begin were going to start with the shooting mechanism I wanted to have enough power to shoot someone across the room I ordered this gel Blaster and I think its going to be perfect Lets test it out this thing shoots super fast but it definitely jams so were going to have to Tinker with that to fix it my plan was to put the gearbox inside the main compartment of the turret there also needs to be a large Hopper to put ammo in this cylinder now the question is how do we transfer rounds from the back in the storage container to the front to be shot I found this motorized mag aine and I thought I could recreate my own that loads Orbeez through a feeding tube also the Orbeez Jam super easily and chicken skunk from my Discord gave me some amazing advice Im new to the Blaster community and he recommended checking out Nerf rival out of Dart sells a proton pack which feeds into a Jupiter Blaster which looks perfect for a KJ turret the way it works is that a feeding motor pushes rounds one at a time through the chamber where they meet the fly wheels that spin super fast and Propel the round forward it also uses a vacuum hose to transport rounds Jam free to the Blaster I purchased my own proton pack and Jupiter and I am so impressed with the quality the people at out of darts are freaking awesome the motor spin at RPM or almost times a second wow the Jupiter is normally actuated by a trigger in the handle but I replace that with the relay to control how often to shoot with software Lets test it out now lets work on the loading mechanism I wired up this brushless motor and wrote some code to control its speed basically it spins super fast and has an impeller fan on the end this creates a uction force that feeds the rounds through a vacuum tube now lets take the thing that shoots and put it inside turret these are printed parts that I dont have the files for so I thought I might invest in a D scanner except theyre really expensive so basically I decided to do it my own ratty way and scan the parts on a d scanner it gave me highquality resolution pictures of the profiles that I could then bring into CAD to trace and nail down the position of the mounting holes then I started modeling the turret head the overall turret height will be a little over ft which feels pretty accurate to the game also the turret head just fits on my print bed lets go I hooked up the proton pack in Jupiter to Timmy mark one and let me tell you Timmy Packs a Applause punch for the mark one weve used this green area and for the mark were going to use this red area to create a recoil mechanism I ordered a linear actuator that wasnt as Snappy as I would have liked so I tried a solenoid instead and it seemed promising the turret head will be attached to this linear bearing to allow for smooth smooth railing I printed out Timmy Mark I and it works kind of it seems promising in a quick test but when slanted downwards the solenoid spring is not strong enough to return the turret head Against Gravity Im thinking Ill add a stronger spring or add two solenoids opposite one another so it could deploy and retract John from Instagram recommended having an outer shell that slides along an inner shell instead of sliding the entire system to save weight honestly the solenoids kind of suck for this because they just dont perform well under any form of weight this is a slip gear mechanism the pinion gear mesh is with the rack and its driven forward as it rotates only the pinion doesnt have all of its gear teeth and the back has a rubber band attached to it so once the gear runs out of teeth that mesh the rack slings back and resets its position this is perfect for recoil I printed out the new parts and assembled the mechanism I played with different variations of gear teeth until I reached the desired recoil effect I was going for also before you guys say anything yes The Recoil is going forward usually it goes backwards the turret in game actually recoils forwards and I was just matching that once I got the slip gear Mech working it was all about nailing the timing and matching the game next lets work on Timmys ability to look around I hopped into solw works and created arms that hold the turret head up the turret head should be able to swivel up and down and also left to right I hid Motors in each arm and designed a one to two gear train basically all that means is that the turret head will move twice as fast as the motor driver gear I did this so that it can quickly snap the targets however this cuts the strength of the motors in half lets see if it works and it doesnt instead of a to gear is IO for increased speed I tried a to gear ratio for increased torque the turret head is can levered which means its only supported at one end I decided to slide over the arms as much as possible without it looking too strange in the model I also upgraded the servos to these big boy Motors and installed a counterweight and the turret is much more stable now I then assembled the parts for the Left Right swivel this one is way less complicated because it has a shorter lever arm now were cooking now lets work on the tracking this is a husky lens its an AI Vision sensor that uses machine learning to recognize its environment my initial plan was to use facial recognition to track enemies but what if the enemies turned around or really far away it also has object recognition which includes people for some reason this mode doesnt only include people it also detects dogs sorry lucky the plan is to detect a person but when a person is seen switch to color recognition mode to check to see if theyre wearing yellow if theyre wearing yellow dont shoot theyre friendly if they arent light them up to me started teaching the sensor what the yellow KJ jacket looks like so could remember it later its best to get varying angles and distances the rounds are the same yellow so that may be an issue later Timmy will steer until the target is centered but wont move it all if the color check is positive for yellow however I want the human box and yellow box to overlap to not pick up random yellow in the shot I first thought was to create a matrix of all the points for each box and then check if any of those points are the same then I realized you could only check the corners because they represent the minimum and maximum of ranges for each box I actually cheated and used chat GPT to write me an arduin script to see if two boxes are overlapping all I had to do is modify it for my code and it actually worked it sees a person and does a quick color check boom right there is he friendly yes dont shoot him now move the balls to the side of Ross it looks for human sees yellow on the side not overlapping is he friendly no shoot him next I added back in the firing Music me I know it doesnt look like much but theres actually a lot of software going on thats telling it not to shoot because of the Yellow Jacket check out Timmy Mark III since last time I added the barrel and ammo canister I ditched the counter weights and now my ammo canister is the counterweight next I wanted to work on the Tracer or blast of purple light I made a compartment for two single cell neopixels at the top and bottom of the barrel theres a hole for the light to exit out of however the plastic isnt opaque so it glows with the light making it look like a cheap toy the ammo canister has a threaded lid so it can easily replace also switch from yellow to purple and pink rounds to avoid the color recognition mode on the Husky lens version two of the Tracer design moves the neopixels to the front of the barrel and triples the amount of Lights used in terms of brightness this is so much better combining this with the colored ammo this just looks so sick when Timmy sees someone he pops up and a third leg comes down Im not going to even worry about the support leg right now we need two legs that can move up and down for when Timmy sees an enemy the common way robot dogs provide lift is with motor in the hip and in the knee I did some quick math to figure out that even with these big boy Motors we wont have enough torque to lift the approximate lb load of the upper body these motors are also pretty expensive and since all the legs have to do is provide like a couple inches of vertical lift I wanted to try to use only one motor in each leg this design hides a motor in the hip and uses a to gear train to provide increased torque for lift the hip is attached to a four bar linkage mechanism and as the hip rotates it drives the whole mechanism up I think this first design is super clunky and it didnt work exactly as expected since the turd is topheavy Im grounding the feet down with these lb weights they look super wide and goofy as hell right now I think Im going to make a concrete mold to get a heavy piece in a custom shape to weigh down the feet one of you suggested using stick on tire balance weights as ballast weight these were great because theyre Domino sized and I could arrange them better to the shape of the foot I upgraded from these regular Springs to gas Springs theyre used in furniture doors all the time theyre filled with pressurized nitrogen and theres a p and attached to a rod inside since the bottom has a rod its area is slightly less than the top which produces a net extending Force these guys are newtons or about lb and Im going to put one in each leg the turret head weighs about lb so these will be perfect for stabilizing the weight and keeping the system pretty close to equilibrium a hid Motors in the lower legs and the gas Springs will make it so that the motors have to do less work to move the entire assembly up and down we only have to move up a few inches when an enemy is seen so Im hoping well have enough lift without throwing off the B balance of the entire system it looks so good tip over wait its actually so good wait can it walk is it going to be no its not going to walk lets power the motors up and see what were working with so we definitely have enough torque for the popup I started using the ramp library to smoothly reach my Servo end position instead of moving there in one swift motion a design and printed little rubber foot pads to prevent Timmy from slipping they friction fits slide into the plastic and give Timmys feed way more grip strength I also added all the details in the feet like these little feet horns and little circles now lets add the third leg I modeled and simulated my design and solid works and then printed the parts out from there it was just about sinking up the servos to the right angles as the turret pops up when an enemy is seen I swear this thing looked way less sus in the CAD software before I printed it out what do you say how about we fix that little buddy if we could angle the foot upwards and sort of Tuck the lower leg in as much as possible I think that would help a lot I added a torsion spring in the ankle so its always angled upwards but it can still flatten out when it lands on the ground on top of that in tweaking the servo motor angles I think I was able to hide the third leg pretty well inside the base so I noticed some smoke coming out of the front barrel of Timmy and I really dont know what to do like I think I burned out one of the motors um and the place that I ordered them from are coming from California so its going to take like a whole another week just to ship them like Ive already put so much time and so much effort into this I dont know what to do like this is so frustrating every time I try and solve something and like I get it to work it feels like another five problems arise I feel like I dont want to do this anymore sometimes like I dont know its its hard to feel inspired to like finish a project that takes so long and so much effort but after taking Timmy apart I found this one screw loosened and started searing into the plastic flywheel causing the smoke sometimes its hard to be as detailed as Id like to be especially while working on a deadline it sucks that the motors broke but there are so many other issues with the turret its kind of okay because I have to fix those in the meantime The Recoil motor also broke so i redesigned the turret head to house a larger one the three symmetrical tooth design was now getting jammed though I was able to fix it by moving to only two teeth I also replaced the rubber bands with springs because some of you said the rubber bands might rot over time for the ammo canister lid I flipped the threads and added this plastic cap that squeezes down and acts as a gasket preventing air from the vacuum from escaping I also perfectly aligned the handle like in game the new flywheel Motors are finally here and its time to hook them up next we need to fix Timmys scanning cuz its really really slow it uses while Loops to slowly move the swivel Motors until the camera is centered on the target what we need is a way for the turret to see someone and quickly snap to them like in game Matt from my Discord had a great idea where we develop an equation to predict snapping distance I needed to make my own shooting range so I got a lifesize cardboard of Dwight the Husky lens wasnt picking him up though so instead I just printed out Ross and taped them together that seemed to work next we need to gather data points where the x value is the distance from the camera being centered and the Y is how far the motor has to be turned to B centered I built this little test apparatus for calibration and after getting like data points we can plot the data in Desmos and draw a curve of best fit with a quadratic equation now Timmy has a way to predict how far to turn all in one swift snapping motion based on previous data sets right now the equations currently calibrated for a Target at ft but the Husky lens can actually recognize someone at up to ft finally after months of work I started to run out of bugs and problems to fix heres a look at everything working together so its going to scan for a Target its looking around once it SE someone pops up swings the leg out and its going to look for me and then Music shoot also Ive made a package of files free for download in the description if you want to make your own turret thank you guys and back to the video her wristband ingame is pretty Compact and she uses it to spawn her turret I found reference images of The Wristband and hopped into solid works to mock up my design the wireless transceiver had a larger range and was much more stable with an antenna but I actually kind of like the look of it I printed out my design and hooked up Electronics I also added some lights that look pretty awesome in my opinion lets give this thing a Music try Okay so by default Timmy will be in an automatic scanning mode and well be looking for humans to shoot but if I flip this toggle switch in the back it will then look for my RC wristband and will be ready to fire on command at this point Ive added all the functionality I want and its time for paint this is the most amount of Parts Ive ever had to sand Prime and paint in a small time frame I almost went insane I always start with the low grid on my orbital sander then I apply a coat of wood filler to fill in any large gaps after this super tedious process I start to apply layers of filler primer I repeat the filling process until Im happy with the surface finish there were so many parts it took like a whole can of filler primer to apply a layer to every part then I hit up the promised land of paint Home Depot I got two different yellows because I wasnt sure which one matched the canister lid more closely Applause luckily most of the parts are one solid color but a fe

write a youtube script

for Music Music sh Music Music hey guys Toby is here it has been a while that I didnt publish any kind of video because I am busying with my master feces however recently I got a little bit time so I built this this is a c useless robot this robot only do one task which is when you try to push the switch to the on position it turned itself off by pushing the switch back to the off position Ive always been wanted to build one of these robot for a long period of time but building one of these robot requires so many skills including Electronics PCB design and firmware programming that it takes me a lot of time to learn and practice this skill before I can really build one in this video I will tell you where the ideas come from how this robot works and walk you through the process on how I design and manufacture it lets start with the origin of this robot design around years ago there was a Japanese make named Kyoshi he created the world first Kawaii useless robot that moves and display facial expression using a dot matrix display the isue with his design is that his design is not reproducible in his design he tear down an existing LED Matrix display module and reroute the PCB using UW moreover the stepper Motors and Battery system he used is also special made that means if Im going to build something similar Im going to replace these part with more modern off the shelf Parts this design help reduce the complexity when people trying to reproduce my design by my open source schematics as well as reduce the cost of the overall build during my research on how to build this robot I found out that this robot indeed have a mass manufacturer version however it is using some simple mechanism that doesnt provide LED Matrix animation or movement so I decided to build it from scratch and this is how I made the same robot years Music later when designing this robot I decided to start from the front side to the back side so the first module you will see on the front side of the robot is the LED Matrix display there are total of LEDs in this robot and obviously I cannot control it one by one using one gpio per LED from the MCU that is why a dedicated I are required for each of the LED Matrix module on this Matrix board after considering the required circuitry for the display module and it is really complex so I decided to make a PCB for it the PCB consists of two side include the LED Matrix module on one side and a current setting resistor located in the gap between the LED Matrix module and PCB and on the other side of the PCB it consists of eight Max LED Matrix driver IC this IC has builtin shift register so you can chain them together and require as minimum as three gpio to control the signal behind the LED Matrix module is the step promoter driver circuitries and Servo connector the driver PCB is designed to sit just mm above the display module and on the side here you will see five cables connecting the display module to the driver board this helped simplify the connections designed between the driver board and display module to the MCU s two board can share the same FFC cable on the bottom of the robot and connect directly to the MCU board from the final PCB it is the MCU board it contain a charging circuit which is is based on the IP power management IC and a programming circuit based on s and on the edge of the PCB you will find an esp this is the esp E variant which contain two cores and onboard WiFi function all fre PCB combines provide the required electronic controls drivers in programming space enough to power up and operate this kwaii robot to do his kwaii Music things talking about the structure of this robot this robot is definitely one of the most complex project Ive ever built with D Music printing as you can see the internal component are tightly packed with nearly no space in between in the center compartment you will see two stepper motors used for driving the wheels and two Servo for pushing the top cover and the switch on this angle you will see better on how the pushing arm is drived by the Sero to push the switch and the wheel of the robot sit in this indented space on the edge of the robot structure when the case is installed the wheels will get hidden inside the robot body and for such a complex structures obviously I cannot design it with just pen and paper I use cat software to design the structure and test fit everything before I start my production process and for some of the parts that requires light tilting like the ster Motor installation right here I printed out the cat files and compare it with a physical step promot us to make sure everything aligns with my cat designs and the process of of making mockups and aligning all parts and make them fit inside these robot bodies takes me around months which is one of the most time consuming procedures when building this robot by the way the wheels are not actually Wheels but a D printed plate with a rubber ceiling ring that usually used in PVC pipe interestingly though it do acts like a wheel and and works perfectly as a wheel Music the software is written in C with adrenal IDE and utilizing the free Ros features so that I could use both of the course and do different things on different cores in the primary call the primary logic is handled there including wireless connections steper Motors control and more and on the secondary core the animation rendering and display driver are handled so the second core will handle loading animation frames from the SD card and rendering the frame buffers to the LED Matrix display the communication between cores are done with M Tex and semore where a few variables are storing the animation code and allow the animation core to handle the rendering by itself one of the difficulties in writing firmware for this robot is the number of shift register required to operate shift register are really helpful when youre running low on GP pins however they also require lot of cycle conss in order to shift data into it and it usually block other process running when the data is being shift into the shift register especially when you are programming in multi core and multi situations as the motors and the display module are both powered by an independent set of shift registers two dedicated cores are required to operate them in order to prevent blocking over each others rendering process Music lastly lets talk about the animation the animation is done by a really naive approach of writing the bites into a file stored on the SD card when the logic want to render an animation frame the target frame is R from the SD card and loaded into the frame buffer and the animation rendering routine will render the frame buffer onto the LED Matrix module by shifting each of the btes into the LED Matrix driver I each of the frame contain bits of data which is not possible for you to manually plot them into a file so instead I develop a web UI to generate this file using JavaScript on the web UI you can simply select and click the LED you want to light up and draw an Pho expression on it afterward you can export the emotion file as a binary file and put it on SD card this way anyone can create their own facal expression for this robot one more interesting feature is that if you push the switch before you power on the robot the robot will enter wireless control mode and you can connect to it using your phones browser this feature is really helpful when youre trying to debug something Music mechanical every year I will build some cool and interesting things to challenge my programming and electronic design skills and this is my project of the year a learned quite a lot of skills regarding programming in rtos how to design and develop prototypes with high density to layer boards I hope this video interest you and if you want to build something similar you can check out my kit repository see you in the next video

write a youtube script

imagine a world free from nasty emissions where energy comes straight from the Sun today were not solving any environmental issues in fact the solar charger Im about to make might have a negative impact on the environment considering all the electronics and D printed Parts but maybe this video could contribute to a growing interest in clean power and possibly Inspire few people to explore ecofriendly projects on their own so buckle up because were about to turn sunlight into Pure Unlimited Power Ive always had a preference for working with mechanical systems such as gears and linkages but as the world increasingly moves towards electrical systems I as well have to adapt so I wanted to find a project that would blend the two principles of engineering and I came up with the IDE of a solar powered USB charging station which will look something like this with four solar panels raised up from the ground arranged around the central Hub kind of like the pedals on a flower the panels will track the sun in order to maximize efficiency throughout the day but this concept is not new Ive seen at least a few other videos showcasing similar designs what will set this one apart is its ability to automatically pack itself into a small volume and this is where the challenge lies to achieve this I plan to stack the panels on top of each other and have them lowered and raised into the base and beyond all of that I still have to figure out the horizontal and vertical alignment for sun tracking that was pretty much all I had in vision for the project at this point but I took these ideas and worked them into concept and eventually I created a circuit diagram with this the systems operation became clearer in essence an Arduino is used to regulate motor and Servo actions by responding to input from sensors and switches additionally sensors monitor the current and voltage produced by the solar panels which is then displayed on the screen as it turns out the system needs a battery to find the sun before becoming selfpowering but there are few ways to implement this battery either we go for a large battery essentially acting as a power bank to run both the electrical system and charge the external device or a smaller battery dedicated for the electrical system and then directly charge the external device with power from the solar panels I went with a second option to keep things compact but theres a catch the solar panels can charge everything at once nor can they ensure that there is always a minimum charge in the battery to address this I monitor the battery voltage when it drops too low a relay switches to charge the battery once the voltage rises again the relay switches back to charging the external device with that the groundwork is laid out lets get started on the build I began the design process working from the top down recognizing that the dimensions of the panels would dictate the size of the base that they were going to fit into my IDE was to use a small motor to drive worm gear rotating the panels around the central Hub the worm gear would connect to the panels via an axle that rotates freely in the hub center once the outer panel rotates it latches onto the panel below it initiating a cascading effect when closing the effect reverses micro switches will be positioned to signal to the aren when the panels reach their fully extended or retracted positions to handle the vertical alignment of the solar panels I will use a simple Servo I then printed out the parts for prototyping redesign reprint printed redesigned and reprinted and each time the part shrank the first iteration features symmetrically placed micro switches but I abandoned this approach for two reasons first off the D printed part had to accommodate both switches which made it quite wide secondly stacking the switches motor and Servo on top of each other resulted in an excessive height and in the final version I replaced the bulky switches with smaller push buttons and repositioned the motor completely off to the side with these changes made I was finally satisfied with the overall dimensions of the assembly I then conducted a test by hooking it up to an Arduino excitingly it seems to work in order to track the Sun the system must be able to determine the direction of the incoming light to achieve this I will use four light dependent resistors positioned on each face of a pyramid shape at the center of the Hub these components change their resistance based on the intensity of the incoming light which the Arduino then can interpret by positioning the them at an offset to each other the sensors will see different light intensity the system then adjust the rotation of the panels until all sensors record similar light levels but these components are not manufactured perfectly leading to slight variations in resistance to manage this I bought several of them measured their resistance in similar light condition and noted the values I then selected the four with the closest readings and installed them onto the central Hub it was around this point the bulkiness of the cables and their constant interference with moving Parts became increasingly apparent no matter what I did a cable always seemed to get caught or pinched to overcome this I had to do several design changes first off I created better defined Pathways for cables through designated channels in the rotating Hub and linkage I also redesigned the Hub so that the cables from the light dependent resistors ran through a fixed tube in order to keep them stationary as the panels unfold and all of this was just to make the cables behave in a more reli and predictable manner it took some experimentation but eventually I had it figured out I could now shift attention to the lifting mechanism my vision was to use a four bar linkage to raise the solar above the base using CAD I determined the lengths of the individual links to achieve the desired movement I aimed for the panels to be centered over the base when raised and I completely flat when lowered my initial plan was to use use a simple server to drive the mechanism however this setup proved to be far too weak I had to apply a significant amount of manual Force before it would start moving on its own back to the drawing board next approach involved using a threaded Rod to move a nut back and forth the servo would be attached to the base plate and the nut would be embedded in this joint of the linkage when the servo turns the nut would pull on the joint and thereby raise the solar my thinking was that this would be a lot stronger since its both geared on from the threads and from the gears Music Music themselves after refining the design and providing a means to attach it to the base plate I tested it out given my fullness for mechanical systems I couldnt help but think this look cooler than having just a simple Music Servo this design showed an improvement in lifting power yet was still not strong enough to lift the mechanism on its own the issue laid in the design of the linkage which in the bottom position is not in an optimal position for pulling on to address this I raised the servo up to achieve a better angle for pulling on the mechanism additionally I redesigned the gear assembly to reduce friction despite this efforts success remained elusive however progress was evident since only a slight touch was required for the linkage to start moving on its own to provide the last bit of pulling force needed I added a spring that would be pretensioned in the mechanisms Lower State this seemed to do the trick and the server was now able to lift the mechanism with a working solution in hand I could finally move on to the electronics relieved to have overcome this hurdle I decided to solder all the electrical components onto two breadboards this would sit nicely on either side of the linkage mechanism under the solar panels using cat I figured out how to arrange the components on each breadboard then I started soldering Music once the bread Works were assembled I measured out the length of the cable harness that would connect them together and double checked everything against the circuit diagram turns out it was time well spent because I found a few mistakes that I had to fix with wiring mostly done lets discuss horizontal movements for sunt tracking I want the solar panels to have complete freedom of rotation this way I wont have to adjust the rotation of the solar charger manually or consider the suns position when setting it up yet this creates a problem with cables leading to the panels as they can twist indefinitely and components like the USB port would be really inconvenient to have rotating as it would drag the charging device around to prevent this issue Ill be using a slip ring a slip ring facilitates transfer of signals and power between rotating components without the need for physical cable connections the best Slip ring I could find that was still relatively small only supported six cables but there are way more cables coming from the top assembly than that so I had to find a workaround I decided to move most of the electronics above the slip ring so that they rotate with the panels keeping only essential stationary components in the base I reviewed the electrical diagram to determine where I could split the electronics and put the slip ring eventually I decided to keep the USB port display and the button for evolving and devolving the mechanism in the stationary base this setup would result in exactly six cables running through the slippering if I solder all the ground wires together so I jumped into Cad and started designing my plan is to use a large internal gear in the stationary base paired with the smaller gear in the rotating base at the center Ill install a large bearing and a slip ring to serve as a central rotating Hub and lastly in the stationary base I made a small compartment for housing the screen USB port and push button this marked the completion of the design face lets make the last few parts and put all pieces together Music oh Music m Applause Music Applause Music Music and there you have it the solar charger is complete and the track in function works like a Music charm a slide switch activates the system allowing us to see the current and voltage generated by the panels in real time interestingly Taking It Outdoors unlocks a significant power boost this is because sunlight offers a broader spectrum of light frequencies compared to artificial indoor light however under clear Sunset Sky we only see around Watts not quite enough to juice of this l little battery but the answer is hidden in the formula for calculating the power generated by a solar panel efficiency is typically about and unless I plan on moving to a sunnier Planet I cant do much about the intensity which means that the easiest way to get more power is to make the panels bigger meaning this is more a proof of concept than a practical solution as the saying goes with great power comes great responsibility or in this case great surface area this project has been a fantastic Journey into the world of miniature solar power we learned that the surface area is a key factor in power generation and while this prototype might not be a Powerhouse all the electronics are in place ready for future scalability thanks a lot for watching see you next time

write a youtube script

in my last electric bike video the ebike . prototyping video I designed and prototyped the motor mount which would mount the motor just behind the seat post here and in front of the rear wheel as well as the arm for the idler pulley to keep the belt up and away from the rear frame now in this weeks video Im going to be designing and making the large pulley for the rear wheel now in my version version and version Model electric bikes the large pulley on the rear wheel was a hundred percent d printed and it was in fact mounted to the rear wheel by cable tying it to the spokes via different mounting points which spread the load quite well but it still put unwanted lateral force on all of the spokes now as this bike has a disc brake I want to utilize the mounting point of the disc brake mainly because its obviously a really strong point on the wheel and the hub because its you know it has to stop the bike really quickly so if its designed to stop the bike really quickly surely it should be able to work well in Reverse and be able to accelerate the bike really quickly the only issue of having the disc brake on the rear wheel is that it doesnt leave much space for the pulley to fit into its not so much the disc brake more the caliper because the caliper runs very close to the spokes so Im gonna have to squeeze a pulley in there somehow without you know using the spokes as a mounting point so I think what Im going to do is Im going to show you the constraints the space constraints between the caliper and like spokes and then show you how I will get around the design and then finally manufacture it lets go this is a bike I used for the previous electric bike versions and as you can see there was plenty of space between the frame and the wheel making it easy to mount a pulley however on the new bike there is a disc brake on this side of the wheel which restricts the usable space by quite a bit its not necessary the disc itself this issue but instead the brake caliper leaving around millimeters of space between these spokes and the brake as I dont want to use the spokes to structurally reinforce the pulley its all need to be made out of something stronger than just d printed plastic and I also realize that the spokes on the previous bike are quite a bit thicker than so thats maybe how I got away with it for so long I have in fact changed the design of this pulley three times now due to various reasons but mainly due to the ease of mounting it to the disc hub its very tight to fit a pulley on to the end of the same bolts that mount the disc so I designed a plate which mounts on opposite side of the disc and allows bolts go through the disc this doesnt put any unwanted load on the disc other than a clamping force which should only aid in reinforcing it the next stage was to cut the parts out of aluminium on my CNC router which I havent done before as aluminium is more difficult to cut than the wood Ive been previously cutting I raised the cutting bed to reduce any unwanted lateral forces in the motor and the gantry I then attached the aluminium sheet to the bed using four screws and proceeded to set up the machine as usual before this I had done a whole lot of research into various cutting bits and feeds and speeds I think I tried four to five different feeds and speeds calculators before finding a result of felt right I would add the feeds and speeds information in the description on this video so here goes the first cut Music Music and its working for those of you who dont know what feeds and speeds are when cutting material such as aluminium you want to be able to cut clean chips off the metal so not too fine dust but also not huge chunks cutting a fine metal dust can cause overheating of the cutting bit and material and cutting big chunks can put unwanted stress on the bit both cases controls on the life of the cutting bit its quite a bit more complicated than this and there are lots of information online if you want to learn more Im going to do a bit of skipping through this project because I made a few mistakes with the first pulley essentially I wanted to cut the outer ring of the pulley and its spokes in one piece then bend the police spoke so it would avoid the brake caliper but it proved to be quite difficult to evenly bend the spokes of the pulley to the correct same angle the bouncing of the pulley works well with the clamping solution through the disc with the two six millimeter spacers to shift it as far from the disc as possible however when mounted on the bike it didnt run very true do CD pulley bent spokes the final solution I came up with was to cut the spokes to the pulley individually which would allow me to prototype just one and bend it until it fit the world properly then once I knew the clearance of the brake caliper was correct I could repeat this process another five more times and have a relatively straight pulley now Ive received one question quite a few times on my previous electric bike videos how do you design the large pulley well when I initially designed it it was a lot of trial and error by printing pulleys seeing whether the teeth meshed properly and trying it again so to save you the hassle of this trial and error here is how I designed the pulley for a hasty v belt obviously if you use a different style of belt these dimensions wont work but youll get a rough idea of how I designed it in fusion CAD software draw two construction circles with a diameter of roughly how large your pulley will be anything between and millimeters will work just dont make it millimeters or two millimeters then draw a horizontal construction line from the center to the outer circle and another construction line at a slight angle constrain the diameter between the two circles as . millimeters then draw a line upwards from the end of the horizontal line this line needs to be very so I tend to draw it at a slight angle and constrain it to be vertical then draw two arcs the first being tangent with a vertical line and the second being tangent with the first arc the end of the second arc must be constrained to the slightly angled line and the inner circle they mentioned the radius of the arcs as shown and then dimension the distance between the ends of both construction lines as two point five millimeters this is now half of a tooth and needs to be mirrored in the horizontal line now you need to determine the diameter of your pulley depending on the number of teeth required this is done by dimensioning the angle between the two construction lines define this angle divide by the number of teeth you want on your pulley which in this case is teeth so divided by is one therefore this angle is degrees and the circle should automatically adjust their diameter in this case the diameter is roughly two hundred and eighty six point five millimeters then youve used to make the other teeth appear by using the circular pattern tool and selecting the individual components of the tooth as well as the center point of the circle then enter the number of teeth you require and click OK they have it a sketch for a tooth pulley now you can extrude and do whatever you want with it in terms of printing there are two main settings that are required to achieve a good tooth to bail fit number one is to print with only two outer walls and number two is to print these two walls with the outer wall first oh and you need to have a pretty well tuned printer for a pulley of this size because if the belt doesnt fit properly it can cause all sorts of issues once I was happy with the design I proceeded to print the pulley ring using ptg plastic from my filament sponsor d prints UK this was printed with in field to make sure it was strong enough to hold its shape once on the bike it was in just a case of assembling all the parts and mounting it to the bike so the large rear pulley is now complete it was quite a bit more difficult than I initially thought trying to squeeze it between the caliper and the spokes but its turned out really nicely its it has a bit of flex in the lateral direction you know towards the camera and away from it but it would need to flex by quite a bit probably at the tip of the pulley probably a good millimetres before Hitler caliper at the root of the spoke in terms of the talk direction it virtually has no flex at all theres no way the talk from the motor will produce enough force to flex it in that direction so Im really pleased with the way thats turned out there are a few other things I want to address which you guys are probably wondering if you havent written a comment yet is that Ive changed tires on this bike the original tires in the last electric bike video were offroad nobley style tires and they were the ones that came with the bike they in fact really thin tires and ran really thin inner tubes because these are inch wheels are really big you want to keep them as lightweight as possible so that you can pedal them quickly and accelerate fast however simply wheeling this bike down my garden to my shed I managed to get a puncture so that was a bit of pain since the moment Im purchasing this bike I had wanted to put Road style tires on just because I prefer the way that these handle round corners you can lean them over a bit better in comparison to nobley tires and also less rolling resistance so hopefully more efficiency and more range now these tires are called schwalbe marathan plus theyre quite nice ties I think they retail for pounds per tire however I managed to get a deal with both tires for pounds because these are really odd sized tires theyre forty seven millimeter wide c tires so forty seven by six to two tires so theyre essentially really wide road tires and obviously no one was buying them so I bought a set for this bike these tires also have a thick rubber layer in them I think its rubber to prevent punctures and Ive seen some reviews claiming people have ridden miles on these tires with no puncture or wearing issues as well as that they also claim theyre ebike ready up to kilometres an hour but this bike might go slightly quicker so might have to be careful of that the other thing I want to address is that some of you may have realized from my last electric buy video that I mentioned I was going to do a pulley ratio of I think it was to or . to so that would be a to through tooth pulley on the wheel however in this video you probably would have noticed that I said Im putting a tooth pull it on the wheel and thats because Ive completely changed the design of the selector bike I am in fact going to be doing a twostage step down on the belt so instead of doing a to or tooth pulley Im going to be doing a point to a pulley then a foot to tooth pulley Ill explain the reason to this in my next electric bike video but its basically to reduce the amount of bend in the belt when the belt flexes round a pulley it creates quite a bit of resistance and going to such a small pulley on the motor is probably going to cause resistance issues as the belt bends around it not only that but with my last electric bike version I had to have the belt really tight to prevent it skipping on the motor pulley so having a large motor pulley will engage as many teeth as possible which means I can running a lower tension on the belt which hopefully increases the other decreases the resistance caused by the belt and also increases the lifespan of the belt because having a belt tension to high cause it to wear out quicker so I better stop talking about all this because this is probably to be covered better in my next weeks video really happy with the way this pulley is turned out its really really strong it runs really nice and true and it misses a brake caliper all good Id like to thank you very much for watching if you enjoyed this video please give a thumbs up if youre new to my channel please click subscribe a huge huge thanks to all of my patrons for supporting me you guys make these weekly videos possible and I couldnt do it about you so huge huge thanks for that and I think thats it see you next week goodbye

write a youtube script

Applause hey guys welcome to the first test ride of the ebike . I thought I would film this video whilst riding the bike rather than at the back of my chair just to make it a little bit more interesting now I left off to the last ebike video with the pulley being fixed and then I ran into an issue with the motor vibrating now Ive managed to balance the motor since and I run into a number of other issues with the motor the secondary aluminium plate on the left here was actually causing some weird man magnetic issues with it because its running too close to the motor so Ive had to cut a new motor mount and I couldnt get the whole full amount of power out of the motor it was a it would only output about amps whereas now this can go up to amps thats the peak obviously it doesnt do that constantly butter than the ebike working really well I need to show you something really cool you ready how cool is that Applause so now you guys can see how much power Im using and also how fast Im going Im sure theres gonna be a lot of comments asking how Im doing this on the screen display so I just talked about that cooking first basically the V ESC this is a speed controller controls the RPM of the motor you can buy Bluetooth module for it and it locks the data on to your phone Ill be posting links in the description of how you can do this I know I always praise the risk six it is just an amazing piece of equipment which you can purchase from trampled boards dont credit UK or is it calm Ill link it down in the description below now Im still getting used to this bike a bit because it has so much torque that wasnt even a full throttle and I cant see what speed Im doing right now but you guys are good to see on the screen and gotta be a bit careful around some of these parts yeah its probably its probably outputting just over kilowatts Ive had this motor peak at three point nine five ish almost full but thats literally just a peak it doesnt average that amount of power Im still not going full throttle yet therell be a more open path in a minute where I can put some power down now the bike still isnt finished but its there Im trying to show you some images on the screen whilst Im talking about stuff Ive had to do some other modification to the motor the shaft was slipping and the motor shaft was slipping inside the motor and because the whole housing of the motor spins and it has to grub screws and grip it to the main shaft but that was about slipping so I had to CNC machine a plate with a semicircle like cut out of it which would sit onto the flattened shaft and how sort of solved now I also had to have also d printed a case for the motor as you can see here I need to neaten up these wires a bit but theres d printed cases preventing leg going in the motor so yeah its looking pretty nice now what Im going to do is Im going to ride to the next biggest hill and well see how well it climbed hills lets go now Im coming up to whats probably the largest hill in my local area whenever I fly my drones you can see it from miles away but for some of you guys this is going to be just like a little bump I believe the total elevation of the hill is about meters feet I think thats about right and the path length is about meters I think theres not mega steep but its quite a long hill if I believe Im tracking this data the GPS wise as well so I might be able to load the elevation data onto the screen but if not then youre just gonna have to take my word its quite steep so meters from here to the top just check everything its nice and cool Motors cool batteries are cool belts tight right lets see how quickly I can get up this hill holding GoPros recording its low on battery so were gonna make it up before the battery runs out lets go Applause Applause all right thank you Applause so yeah as you can see Hills no issue that was fun alright Motors a bit warm it doesnt get much cooling inside this housing what I want to do eventually is put a temperature sensor in here link to the speed controller so yeah good a ride back now lets take it easy down the hill probably the only time Ill ever ride down a hill slower and I came up so Im back from the bike ride now and I really wanted to talk a bit more about bike on the descent from that hill however the GoPro ran out of batteries and I also got stung by a wasp or a bee on my neck so I had quite a sore neck for the last few miles on my ride but other than that the electric bike version four is working really well what Im going to do is Im going to run through specification of everything on the bike like Ive been doing with my previous electric bikes starting up at the handlebars theres the twist grip throttle which is I think its designed for electric mopeds if you just search electric bike twist grip throttle and eBay theyre about pounds pretty cheap moving a bit further down is the cell amp hour lithium ion battery that I made in total its cell so its about volts amp hours when I discharge it from . volts down to like a comfortable voltage I think like . volts its about watt hours I think about watt hours gives me about to miles and range depending on you know how I ride it obviously then moving the feather down from there is the ves see the vests six which I mentioned earlier on this video to be the genius bit of equipment that allows me to do that you know data logging and also show the OSD the onscreen display for you guys to watch now I did mention also Im going to be d printing cover for this I will be using a rubber material just really print rubber case for it and this will be most likely TPU for my filament sponsor d prints UK they supplied me with this black filament which is ptg plastic and its its super solid I think I printed it with infill and its theres no way that thats going to hit the motor or your knees cant hit the motor inside there so Im really happy with that I know I always mentioned d prints UK they are my filament sponsor and they do make great quality filament I dont if you can see the quality of that print but its rather nice so Im looking forward to d print the TPU cover for this to keep the wires protected and hopefully also keep any kind of water off of it you know from splashed up from the front wheel hopefully it wont causing the issues now in terms of the motor mount obviously its covered up now by this plastic piece but this is the newer motor mount where its spaced further away from the motor I briefly mentioned it whilst riding I was concentrating a bit too much on the riding but essentially when the motor rotates close to a metal it doesnt have to be a ferrous metal copper would do the same copper and aluminium it has an effect called Lenzs law I believe where the magnets induce a current in the plate and then that produces a force in the opposite direction most people learn at school during physics where you drop a magnet down a copper tube and it you know slows a descent rate and its the same with this and it was just causing issues when the magnets are trying to be spun above I think was above amps current row of the motor so with the new mount where its spaced about Tim mill away from the actual magnets it seems to be fine so I thought Id mention that just in case any of you guys are building an electric bike make sure that you have the motor away from any kind of metal plates now on the other side I can get in here there is the other this is where the motor actually mounts to down here and theres a on the motor and a tooth pulley on the secondary drive shaft so its a two to one ratio here this this belt I bought a short about the original belt was millimeters long but I had to put tension in here or is this is a millimeter belt and doesnt need a tensioner so thats quite handy I also machined the guards off the edges pulley because its a nightmare to get this belt on and off with guards on both of them seen as I couldnt move the police closer together so Ive just machine the end of that pulley and its relatively easy to pull it off now it never comes off when riding because the amount of guard on this pulley keeps it going straight with that pulley now going to the main drive output pulley to the rear wheel this is a point and it drives a . meter belt millimeters and here you can see the I guess you call it an idler bearing is mounted in a slot with a t slot nut on them on the other side of this plate so all I have to do is I have to tighten up this sorry loosen the bolt I can slide it up and down and tension the belt and tighten it and it locks in position so thats quite handy for tensioning the belt and to get the belt on and off I just loosen that bolt slide it down and the belt just slides off the side of this pulley so this pulley is has teeth and its spinning at half the speed of the motor because a twotoone ratio on the other side and then its teeth down to sorry up to teeth so overall from the motor to the rear axle is a gear ratio of to I think so motor spins times per revolution of the wheel now moving to the rear if youve watched all my ebike v videos youll be very familiar with this pulley its simply bolted to the disc brake mount here via a lots of different bolts and then six separate aluminium pieces which go out to this d printed ring now it seems to be holding up quite nicely I havent had any issues with it yet it does need a bit of adjustment because it doesnt run quite true but I havent had the belt slip off or anything and no cracks or anything anyway so yeah Im pretty happy with that its overall apart from obviously my GoPro battery running out and being stung by the wasp or bee Im pretty happy with the way the electric bikes going I need to do a few more test rides on it I also plan to d print a cover for the pulleys on the other side and the pulleys are not much of an issue of get your leg stuck in it because your leg will just hit the side of this large pulley and its its not really got any sharp edges on it so it shouldnt be an issue if your leg just touches it and youll just bring your leg away however if I let other people ride it just just so they dont like you know Jam their ankle in it I will put a cover on it so I want to thank very much fortune I think Ive gone through enough details on this bike not quite sure of the power output exactly in the video I think it peaked at about three point three kilowatts and Im not sure the range either because I need to do some more riding obviously the range will vary depending on the riding style and also whether Im riding on tarmac or offroad so theyll be interesting to test out also when I was riding I mentioned I want to put a temperature sensor inside the motor ive ordered one and i believe it does still datalog through the speed control it to my phone and I can also a program in temperature cutoff so if the most it gets too hot it all starts a decreasing acceleration which I probably dont want because I like to have my acceleration nearly fell over but it could be quite handy just to save the motor from burning out a tool so thats pretty much it thank you very much for watching if you enjoyed this video please give a thumbs up if youre new to my channel please click Subscribe and a huge thanks to all of our patrons for making these weekly videos possible I couldnt thank you any more so I thank you once what time thank you very much and Ill see you hopefully next week or maybe the week after Im not quite sure goodbye

write a youtube script

this trebuchet is still one of my favorite YouTube projects to date however despite having wheels its not very portable now this is a smallscale model of my larger trebuchet that Ive built when prototyping the larger trebuchet and this is a bit more portable as you can see its quite a bit lighter quite a bit smaller and launch is a very small projectile however its still a bit fiddly as it doesnt have a proper trigger mechanism and also trying to hold the counterweight in the right position is a bit awkward as well as that the projector doesnt go very far so what I want to try and do is make an even smaller trebuchet than this that will shoot a bit further range by using some sort of slingshot rubber to store the potential energy rather than large counterweight I designed the trebuchet to be cut on my CNC machine in two halves from millimeter thick plywood so when the parts are attached together theyll make a trebuchet frame of about millimeters wide and roughly millimeter tall axial height mounted on top of a pistol shaped grid after a bit of sanding I installed the axles for the trigger mechanism and the two halves slotted together quite nicely the trigger mechanism was d printed due to the required small dimensions however these could have been cut on my CNC machine with the right cutting bits when connected by the steel push rod you can see how the mechanism will work where the rear laps will hook onto the main arm and simply pull away to launch the projectile Ive been installed the main axle main arm and a length of slingshot rubber which will be used to store the potential energy and by holding the slingshot rubber with my thumb I was able to give the release mechanism a quick test with the slingshot rubber properly tied down to the frame I made a test projectile using two nuts and a piece of string this added up to the same weight as the millimeter steel ball bearings I eventually want to fire so its time for a test launch now trebuchet is occasionally misfire which causes the projectile to be launched backwards at high speed so with the trebuchet loaded and roughly aimed I tilted my head to the sides to prevent any unwanted head shots fortunately it released perfectly the next step was to make the hand grip a little more rounded and then it was time for some optimization right so Ive been doing some tinkering over the past couple of days and Ive made quite a few major modifications so let me bring you up to speed with what changes Ive made the original main arm design had this curved section to allow the tension from the rubber to act at a large distance from the main axle when the trebuchet was loaded to create a larger talk but this only allowed for a very short piece of rubber which meant the tension throughout the release reduced significantly as the rubber got shorter I then designed this newer arm were two pulleys either side this allows the rubber to be far longer as well as always acting at a constant radius from the main axle this applies a more constant torque on the arm during launch and results in a similar release velocity as the previous rubber whilst being stretched far less this means we can always increase attention to get higher release velocities later now for probably the most difficult part of this whole build the sling the sling is often the most frustrating part of a trebuchet to build especially if youve not made one before or have lack of sewing skills like myself on my larger trebuchet I build the slings using nylon straps so together with nylon thread into a threepiece cup which perfectly fits the required projectile but to replicate this with a millimeter diameter projectile is unfeasible eventually I settled on a solid sling that can be d printed its simply two halves of a d printed Bowl with the string hinge in the middle its made perfectly to hold the millimeter ball bearing and only requires two attachment points two final modifications include a magnet to keep the projectile in position before launch so the trebuchet can be tilted without issues and finally a magnet used to store up to balls inside of the grip for later use now youre probably thinking other than having a miniature trebuchet that you can hold in your hand is this useful for anything well lets put it to the test starting with an accuracy test yep the accuracy is reasonably terrible the main issue is the height and angle of projectile release but even then the projectile is released at at least centimeters or one foot above the original grip height which at this short distance makes a huge difference to where you need to aim in terms of projectile velocity it releases the . gram ball bearing at sixteen point eight meters per second giving the projectile a kinetic energy of about . joules if I were to use a counterweight in place of the slingshot rubber it would be raised about . meters so assuming the trebuchet had the same percent efficiency as my larger trebuchet for the same performance wed need a counter weight of about . kilograms which considering the whole trebuchet weighs under grams and a . kilogram counterweight being larger than the whole frame itself the slingshot rubber is a far better solution so what about the range the range depends a lot on what angle in the project how is released but at best it seemed to fire about to meters which might not seem that far by considering its scale of my larger trebuchet the scale range isnt that bad though it will probably struggle to win against a nerf gun especially with the next test the reload time do I even need to comment on that Im not really sure how to conclude this project I mean where else can you say about a mini trebuchet I suppose I mainly wanted to build this thing just to sort of see what it would like in slow motion you know like a mini pistol gun with a weird trebuchet just whipping out the top it sounds really powerful at short range I think its its actually putting dents in my shed from where Im only about a metre and a half away but yeah I hope you enjoyed this short little project it was good fun to build and its definitely something to hang on my wall so if you enjoyed this video be great if you can leave a thumbs up if youre new to my channel and youre interested to see more projects similar to this then please click the subscribe down below a huge huge thanks to all of my patrons for supporting me over on patreon comm and thanks once again for watching see the next video you Music

write a youtube script

this is my D printed wedding flower so Im getting married next month at a Vineyard and we like the blue hydrangea a lot so I thought itd be cool to print a little something people could take home as a souvenir I wasnt sure how the petals would attach to the stem of the flower at first and mess around with different versions I planned on printing the flowers only mm thick and realized since theyll be flexible I could bend them around the tip of the stem I created this neck feature on the petals that snap into place my first version was uh questionable I realized instead of random branches they actually need to form the shape of a dome I also de to create multiple configurations some with three pedals and some with four I also Siz down the pedals about and Im printing on a textured surface to give the petals some surface finish it took a bunch of different versions but I actually kind of like this one I took a heat gun to the petals to give it some natural curl were almost there I felt I needed a few more petals to cover the sides and this final version in my opinion is awesome theyre a little Annoying to make so maybe a few flowers on the welcome table and then petals with their names and the date on it at the tables I dont know F to stay updated and comment what you think

write a youtube script

right so check this out i have a fan mounted inside of a duct here and a fog machine mounted behind it and what i can do is i can blow air through the fan and we should get a stream of smoke coming out of this nozzle now ill give you a quick demo i should be able to deflect the air a bit with this piece of foam so lets spin it up okay so hopefully you could see the smoke getting deflected by that piece of foam but watch what happens when i used a curved surface and put it near to the stream of air Applause you should be able to visibly see the air being pulled downwards around the outside of the curve and this is due to an effect called the coanda effect this video is sponsored by wix go to wix.com forward slash go forward tom stanton to build your free website today so what is the coanda effect when air or any other fluid exits a nozzle at high velocity it will pull surrounding molecules with it as it disperses into the stationary surrounding air but when a surface is placed near to the flow the air between the surface and the fast flow is dragged along with it this causes the low pressure zone between the fast moving air and the surface and the surrounding atmospheric air above it will push down on the flow attaching it to the curved surface so you might be wondering how am i going to use this coanda effect to produce thrust and lift a drone off the ground well if you search coanda effect drone into google youll see many results that look like flying sources and initially it seems as though theres a center propeller that blows air directly downwards to produce lift but in fact there is no hole through the center of these flying saucer type drones instead the airflow from the propeller flows on the outside of the dome and follows the outside of the curve due to the commander effect now these things are pretty cool but i dont really see much point in putting a bowl shaped dome underneath a propeller you may as well just have the air from the propeller flow directly downwards so instead of using a propeller i want to use an impeller which will basically blow air directly outwards horizontally in a disk and use the coriander effect to then follow the outside of the curve of the dome so lets d print some impellers since i have no idea what kind of impeller would be most suited for a propeller replacement i printed many different types to test some couldnt handle the rpm of the motor and some were too much for the motor to handle i finally settled on this design as it was relatively close to the drone propeller in terms of current draw the next step is to make the curved dome that will be mounted under the impeller to deflect the air downwards for lift now im going to d print the dome because its the quickest and cheapest solution to be at this point in time in order to make the d print as lightweight as possible itll need to be hollow but its almost impossible to print this on a regular fdm d printer due to the huge overhang in the center and even with support material itll be very difficult to print with such a thin layer so ive printed this test dome in two pieces and weighing in at grams plus the layer lines being perpendicular to the airflow it wasnt ideal so my next plan was to cut the dome into slices like a cake that way i could d print the dome in eight pieces and attach that to a center hub this saved quite a bit of weight and im not sure it could be much lighter using a d printing technique the next step was to mount the motor and impeller to the dome and see if the air really does get directed downwards due to the coanda effect and it looks surprisingly hopeful then i set up a thrust test stand with a conventional drone motor and propeller to get some control data this setup produced . newtons of thrust while drawing watts i then replaced the propeller with the impeller and attached it to the dome this produced . newtons of thrust whilst drawing watts which is extremely disappointing but at least it looks awesome in slow motion so lets get on with the build repeating the dome construction another three more times wasnt very enjoyable with all the ca glue fumes but at least i didnt choose to use lightweight carbon fiber tubes for the arms which would require cutting and making carbon fiber dust on top of that i recycled the motors from a working drone to build this thing once the arms were attached to the domes and laid out like the quadcopter shape it was designed to be it looked like a formation of chinese lanterns or a few salad bowls with the electronics installed and the motors soldered up it was time to fit the impellers and see if this thing will fly okay here we go first test of the coanda effect drone Music that kind of worked it needs full throttle to get off the ground so im a bit worried about the motors overheating Music Music one of the impellers is loose on the motor so im going to go and tighten that up as well as let the motors cool down a little bit that was a very short flight but i have a feeling theyre going to get quite warm because i was at full throttle just then just to get it off the ground Music so the drone does fly but its not great the all up weight of the drone is grams and its only capable of producing about grams of thrust so its only able to hover in ground effect at full throttle which heats up the motors very quickly i then decided to try different impeller designs with no luck including this weird design which kept the motors cooler but didnt increase the thrust at all Music i then tried reducing the weight by flying with the battery on an extension wire Music but again it still struggled to get out of ground effect Music so i think those few test flights uh went pretty well uh it proves that you can use the commander effect to deflect a horizontal uh airflow down vertically to produce lift to lift up a drone however it is incredibly inefficient um no matter what i did to this thing uh it just wouldnt get out of ground effect even without the weight of the battery and all sorts of impeller designs ive gone through i dont even know how many impellers so many different impellers theres still a part over there different angled blades different height different diameters different number of blades no matter what i do it just doesnt have enough thrust so i think as a last ditch attempt to get this thing to fly still using the comander effect but just getting rid of these impellers is im going to use conventional drone propellers ill need to raise the motors up a little bit and uh put some kind of scoop to scoop the air from the propeller outwards and then hopefully it will still use the coriander effect to pull the air downwards by just using a slightly more efficient form of producing thrust all the motors were desoldered and removed from their mounts and bolted to a d printed extension once reattached to the original motor mounts they were raised slightly to allow the propeller airflow to flow out sideways they were all then resoldered back up and ready for another test flight right here we go last attempt with conventional propellers oh it sounds so smooth oh thats so much easier to fly Music i would say thats about ground effect and im only at about half throttle so its going much better this so its still using the coanda effect but it flies miles better it takes off at about half throttle oh im so pleased and police were also annoyed that my impellers didnt work very well Music it may seem as though this is now just a regular drone with a few large tomato things mounted below it but it is in fact producing lift due to the coaanda effect the diameter of the domes are far larger than the propellers and therefore the thrust can only flow around the curved surfaces one thing i noticed after a couple of flights is the strong ground effect on the craft a conventional drone experiences ground effect but ive never flown a drone that floats around on such a cushion of air its able to fly out of ground effect but requires quite a bit more power consumption as well as getting buffeted around in the wind Music Music whilst i was doing research for this project i read somewhere that apparently using the commander effect to direct a drones airflow downwards to produce lift actually increases the efficiency of the drone which after carrying out this project i struggle to believe however the theory is that the fast moving air over the top of the dome is at lower pressure than the air flow or the stationary air underneath the dome and this pushes up on the drone in a similar way that the difference in pressure above and below an aircraft wing works now this may be possible if you had a really efficient propulsion system like a really efficiently designed impeller and also possibly different size domes uh if you increase the diameter of the dome you will increase the surface area which will make the effect of that pressure difference greater so it will push up a bit more also i assume you would have to compare it to a drone of a similar weight um so that the weight of the domes are negligible because uh if i were to take these domes off the drone would be so much lighter even if i manufactured these using vacuum forming technique or carbon fiber or something like that the added mass would still decrease the efficiency i assume of the of the drone also the flight characteristics in high winds or or at high speed would probably be a nightmare with this thing its not the most aerodynamic thing i was really impressed with the ground effect that this drone experienced uh theres a really strong cushion of air that it floats on and im what im wondering about that is uh whether the airflow that comes around the top of the dome with the conventional drone as the airflow hits the ground it spreads outwards but with this i wonder whether some of it goes underneath the dough and almost like pressurizes underneath im not sure exactly how uh whether that does work but the cushion of air that it floats around on ground effect is really strong itll be interesting to see if this could be modified into almost like a ground effect hovercraft or a coanda effect ground effect hovercraft and also see whether that still requires gyros for stabilization as hovercrafts dont require active stabilization they just sit on a cushion of air above the ground so there we have it a drone that uses the coanda effect for propulsion its a bit of a disappointment that these impellers didnt work uh better than i thought they would um i really wanted to get this thing to fly properly and then compare the efficiency to a conventional drone however halfway through the project i realized that there is just no way this is even comparable to the efficiency of a conventional drone um the efficiency is terrible um but i did learn a lot and i definitely want to explore the ground effect thing a bit more if you enjoyed this video itd be great if you could leave a thumbs up down below if youre new to my channel and want to see other crazy drone projects then click subscribe down below id like to thank my patreons for supporting me over on patreon.com you guys make these projects possible so thank you very much for your support and ill see you in the next project goodbye

write a youtube script

in the past Ive made a few air power projects using d printed compressed air engines and Ive also made a few electric bikes but I think its about time I combine the two yep Im going to be making an airpowered bike the plan is to use a pneumatic cylinder which will be mounted inside of the bike frame this will convert compressed air into a linear motion this linear motion can then be converted to rotational motion which will drive the rear wheel this video is sponsored by Skillshare but more on that later lets go make a cab model to make the cab model I have to take a couple of photos of the bike frame from different angles these photos can then be imported into fusion CAD software and adjusted to scale I can then use these images to trace the angles and lengths of the frame tubes fortunately the bike that Im using has a pretty basic frame designed so all the tubes have constant diameter along their lengths and now we have a pretty basic model of the bike frame and then modeled the pneumatic cylinder and after positioning it within the bike frame I was glad I havent ordered anything larger it then took quite a bit of work to configure the design to the point where everything would fit the pneumatic cylinder had to be mounted at the perfect distance from the crank axle so that the piston wouldnt hit the end of the cylinder at either end of the stroke and the position of the main drive sprocket had to be perfect so the chain wouldnt hit the rear frame tubes when driving the rear wheel when I was happy with the design it was time to CNC cut all of the parts Music and d print the frame clamps which I used PLA plastic filament supplied to me from my filament sponsor d prints UK once all the parts are ready and the necessary hardware had arrived it was time to check if my design would fit to the bike then the bearings for the crankshaft were added which will be used to support and shaft designed to connect the pneumatic cylinder to the drive sprocket the crank offset was unattached which will be used to convert the linear motion of the pneumatic cylinder into rotational motion for the drive sprocket and now for the important part the pneumatic cylinder which seemed to fit within the frame perfectly as per my design the next most important part or possibly the most important part the chain also I cant resist mentioning how perfect this chain length is I was expecting to need a chain tensioner but fortunately the drive sprocket position was so perfect that both the chain from the pedals and the chain from the crank were almost perfectly tensioned then came the ridiculous task of mounting the huge air tank to control the airflow from the tank to the pneumatic cylinder there must be an arrangement of valves Ive chosen to use electronically controlled solenoid valves for several reasons the main reasons being simplicity and cost if I were to design and machine mechanically operated valves the time required to build this bike would be far greater and still give me the same end result and because I have no idea what that end result will be Id rather keep the input effort to a minimum not only that but electronically controlled valves will give me the opportunity to optimize the bike at a later date as I would like to program the valves to open for different lengths of time depending on the supply pressure hopefully resulting in the exhaust pressure being as close to atmospheric pressure as possible for maximum efficiency I then mounted the valves as close to the cylinder unless as possible using flexible rubber hose to allow for the movement of the cylinder during its cycle a quick pressure test of the plumbing was successful and connecting a battery up to one of the valves definitely got me excited to get this bike running so finished installing all of the electronics and also the plumbing for the valve system theres our four valves up here which are controlled electronically from the Arduino which is in the back here theres a relay system which will apply voltage to these and that opens the valves to let in all exhaust air out now the way that the odds we know knows when to time these valves I have an infrared beam sensor down here which basically counts how many teeth of the sprocket pass through it this way it can work out the angle of the crank and therefore know when to open or close them theres also switch down here which is like a zero ring switch so when the switch is pressed the Arduino knows that is zero angle on the crank and then that can therefore count round to reset to zero count rounder reset to zero so that never loses its and the valve timing goes out of sync so what Im going to do is a well first of all plug in the air supply Ive made a very simple code for this at the moment which will just have the air in it open for one side and then the air in it open for the other side theres no special variable valve timing yet so well just give this quick test to see if we can get the wheel spinning so I plug this actually I need to turn the electronics on first so plug in the battery and then plug in the air supply alright so weve got about only about psi right now not very much pressure but lets see if this well be able to get it going so what I need to do to start the bike is I need to just rotate the wheel around slightly so that the homing switch is pressed and then I hold the throttle which is a like a thumb throttle here Ill put that to max and then give it a tiny push oh good its working right so now if I let go of the throttle it should stop I think the wheel was rubbing down there Im so happy right I think we no need to go and do some proper tests right so I gave the bike a quick test on the grass last night it was a bit dark and it was raining so I didnt film it but I can tell you that nothing happened the sprocket that was on here was a through akut and I just dont think it was creating and Ive talked to the rear wheel so it couldnt push me along Im not sure whether this is still going to be enough this is a so theoretically will double the amount of talk but antistick on grass currently if it can push me on grass that it can definitely push me on the tarmac but I want to test it just quickly here before I go out onto the tarmac so Im going to regulate the pressure to psi currently its psi the tank psi through the regulator and try and get my leg over this probably not a good thing to kick the valves right so throttle on and give it a small push to click that switch yeah its not really got much power in the grass I can feel it pushing me there we go okay its struggling a bit on the grass I think I think its time to go test it on some tarmac right so Im here at the flattest quietest Road I could find Ive got psi the tank Im probably going to regulate that down to about and then well give it a spin all right so give it a small push just to get it started and then there we go come on its pushing me not very quick looking out for traffic Oh its actually quite hard to balance with this massive tank on the back regulate the pressure up to lets try all right here we go next time around Ill click the throttle on and one push get up to speed racing now Im into the wind as well right now if I get up to speed a little bit like that and I can just take my finger and the pedal just to get up its a bit more speed just to see how it works your speed it works its just not very powerful Im not sure if this is exactly roadlegal I would assume it is because otherwise police would be pulling over cyclists for releasing some other kind of gas out there the back side give it one more run Ive only got psi so I doubt this is going to do much but might as well use all the air I have I bet you thought I was a right widow sometimes the timing goes out I think thats due to the infrared the infrared beam sensor so I should probably use some kind of proper encoder at some point well that works pretty well Ive run out of air pressure so Im going to have to go back to my shed and fire up the compressor Im happy with that though Im very happy with that I cant say Im not yeah psi and it was pretty much dead now that was a good first test but theres definitely a lot of improvements to be made on this bike but first I would like to thank the sponsors that made this project possible most of my projects such as this air pad bike require knowledge of CAD design d printing Arduino coding electronics and more but most of the time when Im working on these projects I dont have the time to put together tutorials Skillshare is an online learning community with thousands of classes in design I do we are coding and more a premium membership gives you unlimited access to highquality classes from experts in their fields so that you can use these classes to expand your skills and open more opportunities instead of browsing the internet for hours looking for tutorials school should have a whole wide range tutorials all on one site and its also one of the most affordable learning platforms out there with an annual subscription working out at ten dollars per month if youre interested in Scotia then check out the link in the description below the first people that sign up will get a two month free trial and will also help support my channel so thanks very much to Skillshare for sponsoring this video and if you want to support my channel check out the skill share link in the description below and lets get back to the bike so this bike has taken about four weeks to design and build and therefore this isnt the end of this project think of this video is more of like a part project video I would like to do a followup video where I go into depth of like how to optimize this bike because that first test was although it worked it was far from the performance Id hoped for theres a few modifications I need to make terms of mechanically to this section down here the current crank shaft through here cant survive more than psi I did a pressure test before I completed the build of this bike and I managed to snap this axle at about psi which is why I didnt run that test above psi I also want to make a lot of changes to the valve timing if you watch back the GoPro footage which was a camera which was capturing this rotating round it was quite clear that the valve timing was quite a bit out the air inlets would fire late or open late and then the exhaust valves would open early so in terms of the actual angle that the cylinder was actually putting force onto the crank it was it was probably only about degrees of the full potential degrees it could pull or push so there was a lot of energy lost in that cycle briefly during the build of the bike I mentioned that I want to adjust the valve timing depending on the supply pressure from the tank and what I meant by this is that I can basically calculate how much the volume of the cylinder will expand so depending on the pressure will let more air in or less air in so that once its expanded the end pressure of the air will be equal to the outside atmospheric air and I think this is the best way or most efficient way at least to transfer the energy from the air because any pressure flowing out the exhaust is wasted pressure from that first test that I ran you probably would have heard that the air pressure coming out the exhaust was quite high it sounded like a little egg going off every time that the engine would cycle round so if I can optimize this so the exhaust pressure is as low as possible or as close to atmospheric pressure as possible then I think that would be the most efficient way to transfer the energy from the tank to the rear wheel so yeah theres a lot of improvements to do but Ill be doing all of those improvements in a followup video so if you enjoyed this video itd be great if you can leave a thumbs up if youre new to my channel and once you follow this project then please click Subscribe down below a huge huge thanks to all of my patreon supporters for supporting my projects and if you want to find out more like behindthescenes updates of these projects I upload videos occasionally on my patreon page of what Im up to so itd be great if you could check out the link in the description below to my patreon page thanks once again for watching and I will see in the next video goodbye you

write a youtube script

A conventional drone is controlled by varying thrust on each one of its motors and propellers and by increasing the thrust on one side of the drone it will cause the drone to tilt now this works nice and well down here on earth where the atmosphere is nice and thick but up in space things like rockets and large satellites require a different form of control now probably one of the most common forms of control for the larger rockets and satellites is gas thrusters these work by expelling highpressure gas out of very small nozzles mounted at various points on the spacecraft now what I want to do is try and apply this gas thruster reaction control system to a drone but first lets talk to someone who might know a bit more about this subject than I do Joe hello Tom Hello Joe Hey Tom oh wow that was really loud. Hows it going man? Tom Good thanks and you? This is Joe Barnard he has a YouTube channel called BPS.Space where he builds model rockets. but the rockets that he builds arent your average model rockets yep this is a replica of the SpaceX Falcon Heavy and it uses thrust vectoring for control Joe has been developing this thrust vectoring system to eventually land a model rocket propulsively thats right this video isnt being played in reverse more recently hes been experimenting with a gas thruster reactor control system so I had to ask him a few questions if you could split up the hole of an RCS project into three components what would they be? you wanna start with the plumbing umm that actually is pretty accurate you you probably want to start with the plumbing so you can figure out which valves you need to actually well okay the problem is its all interdependent if you start with the valves you also want to consider the actual moment that you want to impart on the vehicle if you want to do the math about it youll have to consider the flow rate in your valves but mostly youll start with the plumbing then I would focus on how are you gonna control this so like what is your flight controller going to be your power source your wiring and maybe a little bit of layout stuff and then the last part of the project which in my mind or at least in my project has taken the longest is the control side of things what does the control logic look like when does the computer decide to fire a valve and how do you keep your vehicle from spinning out of control or not controlling itself at all and crashing and crashing right yeah like how do you minimize your your thump factor so as Joe mentioned its best to get the hardware sorted first this will include things like the pressure vessel the valves for controlling the airflow and also the nozzles for exhausting the highpressure gas now with Joes system he has to fit his into a model rocket which means it has to be quite compact whereas on a drone I have a bit more space to play with so for my pressure vessel Im going to be using plastic bottles this is a litre plastic bottle and it can handle upwards of psi its really lightweight and theyre really cheap also my compressor only goes up to about psi so these bottles are pretty ideal in order to turn this plastic bottle from a drinks container into a pressure tank I need to connect a sealed tube to the cap this is done by d printing a heavyduty threaded cap then an oring is inserted to keep a tight seal with the throat of the bottle and after a quick pressure test it seems to hold fine because the control of the drone is somewhat limited by how much air can be stored in the bottles its important to optimize the nozzles for the maximum thrust possible without wasting too much air so Ive built this thrust test stand its comprised of a load cell to measure the thrust the thruster nozzle a set length of tubing and the solenoid valve thatll be used on the drone the data from the load cell is then logged to an SD card using this Arduino board so that I can plot a graph of the thrust later the air is supplied from a single two liter plastic bottle so I can get a rough idea in terms of thrust duration its then filled using this larger tank via a pressure regulator so that all the tests are as consistent as possible I chose to run the tests at psi pressure as it would allow multiple tests to be run without filling the tank every time plus Im more interested in how it will perform at lower pressures as thats when Ill lose control to find the best nozzle design I chose several options none of which had any theoretical calculations behind their design but thats the advantage of having a test rig the first nozzle was a straightthrough nozzle which would act as a control the second was a straight cone nozzle the third was what Id call a conventional rocket nozzle with its distinct bell shape and the fourth is a reverse or a horn style nozzle each nozzle has the same throat diameter of four millimeters as well as the same exit diameter of millimeters and a length of millimeters take a guess which nozzle you think will perform the best in the comments down below I ran each thrust test six times per nozzle to get an average and imported the data onto the following graph. Now before I tell you which nozzle performed the best lets just admire how little thrust this system actually produces yep thats right the best nozzle produced a peak thrust of . Newtons or the equivalent of lifting about g so heres the lineup did you guess that the straight nozzle would perform the best? also the bell shaped nozzle performed the worst out of all of them now because these solenoid valves have a small orifice of about mm I thought I should probably try a smaller diameter nozzle so I made this two millimetre diameter straight through nozzle and it outperformed all the previous contenders in both peak thrust and duration so Im gonna go with this lets talk about the drone design so the plan is to use two counterrotating propellers mounted inside of a single duct this will be mounted in the center of the drone and wont provide any pitch or roll control it will be able to provide yaw control by adjusting the torque from each of the motors then for pitch and roll control Im going to have three nozzles mounted on three long arms extending out from the center of mass the reason for choosing three nozzles is because its the minimum amount of point of thrust you need to have stable control similar to how you cant have a stall with less than three legs now the reason why a regular drone will have four arms or more is because they need equal amounts of propellers to counteract the torque but because these nozzles dont produce a torque in the yaw direction I should be able to get away with three the frame parts consisted of a few d printed parts using PLA filament supplied from my filament sponsor d printz UK as well as some aluminium spars when assembled it created a triangular frame around the center duct then I mounted the electronics starting with the relays that would apply power to the valves then the valves themselves which one of them weirdly had a hole at the top which would leak air from but the leak was very slow so it shouldnt be an issue then came the flight controller Arduino board and radio controlled receiver which allowed me to actuate the valves from my controller I then mounted the arms which the nozzles would attach to and temporarily mounted the nozzles using tape in case they need to be removed later for modification so now most of the hardware and electronics are complete apart from mounting the bottles but before I give it a test we need to consider how the flight controller will control the valves with the regular drone the throttle can be increased or decreased at different rates depending on the required control output but these valves can only be fully open or fully closed therefore we need to configure the system to know exactly when to trigger the valves which is where once again we need to seek Joes advice so this is like uhh... this has been the constant theme of like all of the rocketry stuff Ive done its its actually comparatively easy to develop all the hardware and build everything but the devil is in the software umm so right now I have like if you were an actual controlled engineer you wouldnt want to hear this but I its just a gated and saturated saturation limited PID its fairly primitive if you do enough simulations so I have a simulation built up for how the vehicle should actually be able to control itself in simulink that helps you dial it in but its there are much better ways to do it if youre willing to get into like full state space control or something that actually considers the torque that each thruster can put out but thats also kind of a nightmare so its easier to just set it up on a test stand I dont know if weve covered this yet but thats kind of how Ive been doing it you have basically uhh you can see actually two holes in the vehicle right around the centre of mass here and so its actually slightly below the center of mass so the vehicle wants to fall over and what Ive been doing is just setting it up in this mount where it can rotate on this axis and this axis and just you just start loading PID values and giving it a shot Tom just trial and error yeah thats not how it really works in the aerospace industry but these are model rockets so Tom gets the job done yeah yeah haha all right well I think thats enough information for now hopefully Ill keep you posted on how this thing goes good luck man control systems are hard but it seems promising well yeah youre far better at control system than me well thanks a bunch Tom good luck Tom yeah you to yeah all right Ill take you soon bye now Im going to take a slightly easier route to Joe as I want to use a conventional drone fly controller this will save a lot of time as the code in these things is far beyond my knowledge level so to convert this fly controller into the similar gated PID system that Joe mentioned Im going to need the Arduino to run a small bit of code essentially the fly controller will output a throttle value as if it were communicating with a drone motor the Arduino will read this signal and convert it to an on and off signal by gating it with a set threshold when the output signal from the fly controller increases above this threshold the Arduino will open the valve now this threshold will depend on many different factors so lets program it into the Arduino and give it a test so Ive hung the drone from the ceiling so that I can sort of simulate it in midair so it can now swing forward and backwards and as you can hear the valves do actuate now I cant test the roll in this configuration Ive just hung it roughly about the balance point for the pitch axis and you can hear the valves actuating when I move the drone so what Im going to do is Im going to pressurize the whole system using my large tank down here on the ground and see if this see if this works it should work all the valves are actuating I just havent tested it with the air yet so I should be able to push down on here and this nozzle will fire thats actually really well tuned already oh I dont what happened there its working! Lets try it the other way so if I push up the two jets at the front should fire So it will try and hold a specific angle okay it slowly gave up then what I can do is I can also flick this switch and it will always try and self level its trying to self level so because its hanging slightly nose forwards as you can probably see apologies my remote keeps talking to me yeah when it hangs nose forward it will try and lift the nose to bring it level if you watch .. .. that is working way better than I expected Wow err right that is so cool it sounds absolutely amazing aaaaand level right cool now I think we need to install the bottles for its own air supply and add some motors the motors were mounted bottom to bottom and soldered in a way that they would rotate in opposite directions so they provide a net zero torque the bottles were then mounted to the frame using large zip ties and then it was a case of plumbing all the bottles and valves together this was done using flexible PVC hose that would push fit into various connectors which makes the plumbing super easy and also reversible if something needs changing I then hooked up to the air tank for a pressure test and to check for leaks which was successful okay so the drone is all ready to test I havent pressurized it yet but Ive got all the electronics turned on Im actually going to run it on a tether that way I can use the full volume of the large tank and I dont just run out of air really quickly because for his first test I wanted to sort of gauge how it feels on the ground so lets hook up the pressure tank and give it its first test Im slightly nervous about this but you know its got to be tested so lets armed the motors lets check the valves okay what the valves are working increase the thrust lets try that again the motor and propeller I chose should be capable of . kg of thrust per motor and this should be plenty to lift this drone but for some reason it didnt want to lift off the ground so I tried raising the duct off the ground a bit to see if it improved the airflow not far off that worked it definitely needs the bit off the ground just to get the airflow yeah I had control though not far off yeah its controllable these initial tests were run at about psi the reason for this is when I fill a litre tank to psi and then fill the litres of bottles from that the pressure drops to around psi then after messing around on the ground for a bit the pressure drops some more okay that seemed an issue with the tether it seemed like it pulled on drone so it was time for the first untethered hover for most of the following test flights I was adjusting the values inside of the flight controller to make the drone react quicker and hopefully achieve a more stable hover much better that was a lot better much better that was fantastic yeah the battery dies like instantly I was full throttle there you can charge the battery yeah I think I need to get a better battery really with the new battery and the bottles filled to around psi it started to look a little more promising so I think I need to tune the gyro values up a little bit more because at the lower pressures it does start to lose control lets just try that again with the tanks relatively empty yeah it just has no control and the lower pressures pressures bottles filled to psi was to connect it straight to the compressor okay this is the first test at psi bottle pressure YEEEEEESSS!! Psshhh Psstt Pshhh Psssshh Psstt Pshhhh Pssssssh yes Im typing this by hand Psshhh Psssst repeat a few more times now I know its a bit disappointing that the noise from the main lift propellers are louder than the gas thrusters but theres not much I can do about that so just imagine it sounds like this Psshhhhh Psst Pop Psh Pop Pssst ah please no more subtitles I dont know about you but Id call that pretty good success the motors are a bit toasty and so they dont melt through these d printed mount I should probably end the testing there for now Im honestly so happy with that if youd like to learn more about how highpressure gas flowing through a nozzle nozzle causes this drone to move through the air or how a rocket accelerates into space the most effective way to learn is by doing which is the reasoning behind carrying out this project but a far simpler method of learning is to challenge yourself with interactive content brilliant challenges you to solve problems to broaden your understanding of various topics they have a wide range of interactive courses from basic forces and acceleration to realworld challenges such as finding out the correct sized battery for your drone each course contains clear diagrams and animations to help you learn your way to becoming an engineer to support these projects and learn more go to brilliant.orgTom Stanton and sign up for free and also the first people that go to that link will get off the annual premium subscription so there we have it a drone controlled by gas thrusters Im real

write a youtube script

Im sure many of you watching are familiar with how hovercrafts work they force air down underneath the fuselage creating a pocket of highpressure air that they can float around on and this allows them to slide over various different surfaces but Im also sure that many of you are wondering one earth is a condor effective of craft the Coanda effect is the tendency for a fluid to flow around the outside of a curved surface and I explored this Condor effect in a previous video hes a short explanation of how the Coanda effect works when air or any other fluid exits a nozzle at high velocity it will pull surrounding molecules with it as it disperses into the stationary surrounding air but when a surface is placed near to the flow the air between the surface and the fast flow is dragged along with it this causes the lowpressure zone between the fastmoving air and the surface and the surrounding atmospheric air above it will push down on the flow attaching it to the curved surface if you havent seen my previous Condor effect video I used the Condor effect to redirect airflow around a dome to produce lift for a drone in the end I found it to be terribly inefficient however I did notice a very strong ground effect acting on the drone and said the following it would be interesting to see if this could be modified into almost like a ground effect hovercraft her a condor effect ground effect hovercraft and also see whether that still requires gyros for stabilization as hovercrafts dont require active stabilization they just sit on a cushion of air above the ground so here is a very basic cross section of a hovercraft there is a fan up top which blows air down underneath alpha craft and creates a pocket of highpressure air lifting the hovercraft off the ground now my plan is to redirect this downwards airflow to do a degree turn and flow out horizontally above the hovercraft then using the Condor effect the air will flow around the curved surface and produce lift not only will the downwards moving air produce lift but also the fastmoving air above the hovercraft curved surface will be at a lower pressure than the stationary air below the hovercraft therefore producing lifts similar to an aircraft wing also my second theory which may or may not be correct is to curve the surfaces through a full degrees in the hope that a small portion of the air flow will flow under the hovercraft and pressurize the underside of the fuselage therefore creating a greater pressure difference between the top and the bottom and producing more lift so here is their design to mainly fans and mounted just d printed brackets which attach to an aluminium frame under the fuselage this will keep them aligned and prevent any propeller collisions as I plan to build the red curved sections from foam board this aluminium frame also acts as a strong mount for attaching the rear thrusters high up and out of the Condor effect zone I havent worked with foam board for a while and this is probably one of the most complex shaped Ill produce a date so in order to get the angled sections correct I use the sheet metal tool in fusion to make a template which can be printed and used to trace parts onto the foam board each section was then covered in packing tape to make the surface smooth and also add some strength to the foam board as well as adding some color now because producing this curved shaped fuselage from sheet foam became repetitive very quickly I kept entertained by listening to an audio book called how to build a car by Adrian Newey via audible who are the sponsor for this project it touches on many engineering fields but I found the aerodynamic development of a Formula One car really interesting an example was a decision whether to move the regulation nose cameras due to a trend of other teams doing so but data from the computational fluid dynamics software conflicted with the data from the Windtunnel tests so they chose to run the test at the next race one of the drivers didnt like the new mounting position and the other was noncommittal but the data showed that the new assembly was slightly up in overall Aero load this is really relatable to me as when I build such projects for YouTube I always like to test things before coming to a conclusion audible content includes an unmatched selection of audio books an original audio shows you can listen to audible on the go whilst doing CAD work or tinkering on your latest project and if you want to sign up go to audible.com forward slash Tom Stanton or if youre in North America text Tom Stanton and start listening with an exclusive day free trial one free audio book of your choice and audible originals absolutely free so thanks to audible for sponsoring this project and its time to fire up the CNC Music Applause the aluminium parts will cut from two millimeter sheet aluminium and in total weight about grams once the top plate of foam was attached to the curve sections the aluminium frame can be clamped to the foam board using two d printed rings I then initially attached the motors directly to the aluminium frame as I wanted to test whether impellers would be enough to lift his craft but the motors just werent strong enough to drive them and became very hot at just throttle as each impeller took about nine and a half hours to print plus weighing in at just under a hundred grams each I moved on quickly to using conventional propellers which actually seemed very promising Music Music Music Music once the rear motors were attached and soldered up I programmed my transmitter to control all four motors and it was ready for a test okay heres the first test of the Coanda effect hovercraft just going to spin up the main lift motors to see if lift off the ground a bit Music definitely has enough thrust thats about throttle it starts to lift off Music lets see if we can get it moving forwards just got to get the right throttle control so it doesnt take off but enough to reduce the friction so lets try that that works Music Music Im surprised its working so well on grass hes a carefully tuned how much for us the main motors are producing citizen want to take off too much just about there seems pretty good yeah its very topheavy rice it was now the following day and Im going to give this commander affect hovercraft another test I made a small modification to the motor mounts at the rear they actually have a small well its a degree downwards thrust angle now I know just looking back at the footage from yesterday that when I increase the thrust at the rear it seems to want to push the nose down quite hard so Im hoping that this sort of downwards thrust angle will go slightly more through the center of mass and sort of push the rear down a bit more so Im now the suns coming out lets give it a test Music Music so it seems like the nose pitching down issue has reduced and I can pick up a bit more speed the flipping over issue is a I think unresolvable apart from my lack of piloting skills just to give you an idea of the co and effect actually working Im going to throttle up the rear motors without the front ones okay that does actually move it on this a soft grass but if I thought to them up so its not moving so its just about to move about there now if I throttle up the lift motors and keep the rear ones constant it reduces the friction another goes you see how fast we can get it going towards me this could end badly she tried to take off I wonder if it will take off if I go fast enough Music one interesting thing Ive just noticed is that it seems to be collecting all the leaves on the inside so I wonder if thats the Condor effect blowing air all the way run the surface and then filling this with highpressure air which is also picking up lots of debris on the floor I Ive covered all the electronics in packing tape so that should be waterproof because its quite wet on the grass this morning so many leaves are there Applause Applause Music Applause Music so there we have it a condor effect hovercraft is it going to be replacing conventional hovercrafts any time soon I highly doubt it does it have any advantages over a conventional hovercraft well aside from collecting leaves not really I really enjoyed this project and Im really happy that the commander effect actually worked to lift this thing off the ground sometimes a bit too much but I hope you found this project interesting if you did enjoy this video then itd be great if you can leave a thumbs up down below and if youre new to my channel and want to see other crazy projects similar to this it would be great if you could click to subscribe down below and a massive thank you to all of my supporters over on patreon comm for supporting my projects and these videos and making things like this possible so thanks once again for your support thanks once again for watching and see the next video roll the outtakes Music Applause Music

write a youtube script

as a kid i remember watching a documentary about futuristic aircraft technologies and this documentary briefly featured a small shiny flying saucer looking thing that would launch up into the air using the power of a laser beam that was mounted on the ground the laser beam is made up of infrared light as the beam pulses it strips air molecules of their electrons creating an explosive plasma that propels Applause now im not quite sure how this technology worked and i would be extremely optimistic to think that i could replicate it myself but instead of using a laser beam to launch this thing into the air what about if we used vortex rings this video is sponsored by curiosity stream more on them later now i know you cant really compare a laser beam to the movement of air but a vortex ring is kind of a focused portion of air that travels over long distances a little like a laser beam whereas a regular blaster there is more like a torch or a light bulb where it disperses out a lot sooner and doesnt travel as far so lets build a vortex cannon and run some tests to produce the vortex rings we need a rapid blast of air so ive chosen to use these cheap solenoid valves that open when hooked up to a volt power supply however to make sure the valve opens and closes within a split second ive made this control board which is essentially an electronically timed switch to open and close the valve in about milliseconds the only problem with these cheap valves is it appears they have a large input and output from the outside but if you disassemble them the actual valve portion has a much smaller diameter and therefore restricts the airflow so im going to hook up four valves together to increase the burst of air hopefully resulting in a higher power vortex after mounting the valves together on a stand i d printed a manifold that will connect all the outputs into one large nozzle speaking of nozzles we need to figure out how to turn this burst of air into a vortex ring from watching videos online it seems there are two main ways to creating vortex rings the first is to use what i call a rubbish bin because it would make a rubbish bin with such a large hole this is an easy method to make vortex rings as it only requires a plastic sheet attached to the rear and a small push to create the vortex however they dont seem to move with much velocity the second type is a cone nozzle which often uses a combustible fluid to create a sudden burst of air and these seem far more powerful i think some places even use these to break up hailstorms to prevent crops from being damaged so they have some decent range too so i d printed a cone nozzle to attach to the valve manifold and its ready for a first test well that wasnt quite what i expected maybe the valves arent providing a large enough burst of air for this huge nozzle so lets try a smaller cone equally disappointing however with a few modifications to the valve manifold it was starting to show some potential i then decided to scrap the cone design and try the bin with a hole method where all four valves were plumbed into a small chamber with a hole at the end to see if this would produce a better vortex and well it didnt i think the problem here is that the valves cant open and more importantly they cant close fast enough so instead of it being like a pop of air its more like a of air which creates way too much turbulence and therefore destroys the vortex rings but i have a theory of how to get around this if we take a look at the first cone nozzle the air will flow out of the valve at very high speed and slow down as it expands along the length of the cone and because it slows down my theory is this causes the length of the pressure wave to shorten im not sure if this is how it works but i kind of view it like when watching a race on the straights they all look really spread out because theyre traveling at high speed but as soon as they slow down for a corner they all bunch up together but the issue is this results in a really slow moving vortex ring and the problem with the second design is it didnt have much of an expansion section so all the air would flow out at high speed and cause too much turbulence for the vortex to form but what if we combine the two designs and had a long expansion section that would allow the pressure wave to shorten and then at the end there would be a narrower section with a hole to accelerate the air again and create a faster vortex so i printed this exact design and gave it a test Music and there we have it an almost perfect vortex ring the timing of the valves produced the perfect burst of air to form the ring and nothing more leaving a very little trailer smoke behind the only problem is they still dont move as fast as i would like them to and they tend to stop moving after about three meters on the topic of fluid dynamics i recently came across an educational series called butterfly effect which explains all sorts of science concepts over on curiosity stream its kind of like netflix but for nerds which is way better as theres thousands of movies and tv shows that cover science history technology and many more like this episode of butterfly effect which is all about fluid dynamics from the history of how ancient romans discovered density and buoyancy all the way up to the cowanda effect and bernoullis principle that play a huge role in how aircraft fly and there are so many more documentaries and educational episodes to choose from because as the name describes curiosity stream is here for people who want to know more so if youre one of those people you can watch these on your mobile computer browser smart tv and a range of other devices by signing up for just . for a whole year using the code tom stanton or by clicking the link in the description down below so to increase the speed of the vortex rings i made a new nozzle with essentially the same design but with the smaller diameter exit which allows me to fit different size caps to test the best size and speed of the vortex which resulted in a much faster vortex ring and even a few trailing rings that would eventually break up due to turbulence and this created one of my favorite slowmo shots where the secondary ring gets sucked into the first ring and knocks outwards slightly it looks like theres some complex and interesting fluid dynamics going on here and im not sure why there are multiple rings being produced it could be due to the valves opening at different times or maybe the pressure wave is bouncing back and forth inside of the expansion chamber either way it looks really cool so now we need to build some kind of sail that will capture the vortex ring and propel itself into the air but judging by the current performance of the vortex rings its going to need to be very lightweight so ive designed this domeshaped sail with a point in the middle with the theory that the vortex will flow into this cone and be pushed around the curvature and redirected backwards for thrust and because ive designed this using the sheet metal tool in fusion i can flatten out one of these sections which allows me to d print it as a template that can be used to cut out the perfect shape from lightweight paper then after carefully taping the sections together we get a slightly flimsy dome shape of what i designed earlier i then mounted this dome sail on a string and fired a vortex at it which looked quite promising at first but from watching the slomo footage it was clear that most of the movement was caused by the secondary blast of air and not the actual vortex so i moved the sail a bit further away to make sure only the vortex would hit it and well it was far less impressive but i also wonder if the straw riding along the string causes too much friction especially if the sail is twisted by the vortex so i suspended the sail by two lengths of string to hopefully reduce any friction this resulted in the sail moving much further with just the push from the vortex but it also revealed two more issues the first being the amount of energy absorbed by the sail flexing probably isnt very efficient and the second is the accuracy of this vortex cannon its mounted maybe a meter or so away from the sail and fails to hit the center of the sail maybe percent of the time which causes the sail to be thrown way off to the side so at any decent range is going to be almost impossible to hit such a small target and guess what i decided to do next build a smaller sail this was mostly to reduce the weight and flex of the sail then the accuracy issues can be removed by running the tests multiple times until the vortex hits perfectly which was easier said than done however when the vortex did hit perfectly it revealed some awesome slow motion shots like this closeup shot filmed at frames per second shows the vortex continued to spin even after making contact with the sail so now all we need to do is mount this whole setup vertically to try and launch the sail upwards into the sky but if youre wondering why i spent so long trying to make this electronically controlled solenoid vortex cannon its because it allows me to fire multiple rings per second for more thrust okay this is the first test with about bursts of air per second Music that sounds so good it may sound impressive but from the slomo footage it was clear the rapid fire of the rings was causing a lot of turbulence in the surrounding air which ends up with the following rings being shot all over the place so lets see if this will even lift the sail it seems it lifts the cell reasonably well with the first vortex when it hits the center of the sail but as soon as one hits off center it twists the straw on the string and causes too much friction you can still see the vortex rings hitting the sail with a decent amount of force but theyre just not accurate enough to move it any higher but one way to remove any accuracy issues is to use just a regular blast of air instead of a vortex and we havent even considered the stability of the sail yet Music Music you

write a youtube script

this video is sponsored by ren more on them later in my previous video i built a launcher that used electromagnets to accelerate paper planes up to about miles per hour which is quite fast for a paper plane but i want to make it faster with a purely mechanical launcher such as my trebuchet i can simply increase the size of the counterweight or with my flywheel trebuchet i can just spin it faster but increasing the speed of this launcher isnt so straightforward an obvious suggestion would be to just increase the input power but theres one issue electrical power is calculated by multiplying the voltage by the current so increasing either of these will increase the power output however the on and off switching of the coils are precisely timed using a number of electronic components all of which have a voltage and current limit which the launcher is currently operating at it seemed there was some confusion from the comments on my last video that i had built one of these Music which works by shorting a metal projectile between two high power rails creating a very strong magnetic field and launching the projectile so as long as the projectile or the rails dont melt the power can be increased for higher launch speeds but with my launcher i need to take a different approach between each coil is a magnetic hall effect sensor that detects when the magnetic sled passes it which is then used to time the activation of the coils however we can also use this sensor data to measure the speed of the sled by timing how long it takes to travel between each coil and as theyre all equally spaced we can divide the distance by the time to get a speed measurement at intervals along the rail which also allows us to plot a graph of velocity versus distance each dot on the graph represents a coil on the launcher and you can see the speed gained by the first few coils is huge and starts to flatten out as the sled gets faster now i had a theory as to why this was happening but as of editing this video i now think i was wrong in short the rpm of a dc electric motor is limited by its input voltage and because my launcher is essentially a linear motor i thought the same thing was occurring the theory is that the magnets part in a coil will generate a current in the coil and once this generated current is equal to the input current the projectile wont accelerate any faster the only way to get it going faster is either increase the input voltage or decrease the voltage generated by the magnets passing the coils which can only be done by reducing the number of turns of wire per coil and seeing as i cant increase the input voltage i decided to replace all the coils starting with the original turns of wire in the first set of coils and reducing down to tons of wire in the last pair of coils so this is the first test of the new rewound coils as you can see got the super capacitors plugged in got the computer ready to do the data logging so im gonna press the launch button and wait five seconds four three two one that look pretty fast the data logging says . meters per second thats uh thats what over increase in speed not bad so if replacing all of the coils resulted in a speed increase why do i now think my original theory was wrong well if we open up an excel spreadsheet we can create a column for the distance which is the distance that the projectile travels down the rail we can also create a column for the acceleration and we can create a column for the velocity in the distance im going to put and then im going to do it in meters because i know each of the coil coils are spaced by . centimeters so if we drag that all the way down to there that should cover all coils for the acceleration i know the acceleration is roughly meters per second squared so i can type that in there and then for the velocity we need to use an equation which if youve studied physics or maths mechanics before youve probably heard of the suvat equations theyre essentially a set of equations that describe the motion of an object for this instance we want to calculate the velocity using the distance and acceleration so were going to use the equation v squared equals u squared plus a s so if i enter that into this box solving for the velocity we put the square root of the initial velocity squared plus two multiplied by the acceleration multiplied by the distance between the coils because we want to calculate this thing for every instance uh along the rail and i can drag that down and we get an end velocity of . meters per second now whats interesting is if we plot these values as a graph with the velocity on the yaxis and the distance on the xaxis we get a very familiar looking graph in fact it looks a little too familiar so lets copy and paste over the actual velocity data and paste it in this column actual velocity now if we import this into the graph wait for it it looks almost too good to be true we have the blue line which is the simulated velocity using that suvat equation and then we have the orange line which is the actual velocity which was measured by the magnetic sensors on the launcher and they almost perfectly line up and this makes sense for the previous coils as they all have the same number of windings so produce the same magnetic forces and therefore the acceleration is constant but to be totally honest i think i was lucky that the new coils were faster this is because reducing the number of turns of wire would in general reduce the strength of the electromagnet but fortunately reducing the number of windings also reduces the resistance of the coil allowing more current to flow which produces a stronger magnetic field not only that but after some research i found that reducing the length of the coil also results in a stronger magnetic field and if i calculate the relative magnetic field strength of each individual coil based on their length i can multiply this by the known acceleration of the previous coils resulting in another very familiar looking graph for reference here is the actual data gathered from the new coils but there is one side effect of these lower resistance coils and thats the current load it puts on the power source in this case im using super capacitors which are designed to handle high discharge rates and i actually use these same super capacitors on my four kilowatt ebike so they should handle a small paper plane launcher with these fortunately i can log the input voltage to the launcher using the main control board and during the launch we can see how much the voltage of the super capacitors drop under load here you can see they are initially charged to . volts and immediately drop to about volts when the first coils turn on and theres a very clear spike in the voltage between each coil switching on and off which i thought was quite interesting as each spike gets closer and closer together as the sled gets faster but what does this data show in terms of performance well power is equal to the voltage squared over the resistance which i know these first coils have a resistance of . ohms so if the voltage stayed at . volts we should be getting watts of power but instead at volts were only getting just over a thousand watts i then thought i should test my lithium batteries that i use for my electric bike and just from the sound i could tell the sled was faster they were charged to the exact same voltage as the super capacitors but on average provided percent higher voltage which is a huge difference and its even clear from this graph that it was faster because the final coil switched off earlier meaning the sled reached a higher velocity of . meters per second at the end of my last video i mentioned that i want to have the coils not only pull the magnets along but also push them the only issue is that the electronics required to flip the polarity of the coils will massively increase the complexity of the system so my cheat solution was to make a sled that has two sets of magnets where one is a north pole that will be attracted to the coil and the other is a south pole that will be repelled by the coil effectively doubling the forwards force without requiring any more power from the coils however by doing this we also double the mass of the sled so even if the forwards force is increased the acceleration is still the same and the end velocity isnt any faster but whilst designing this double sled i noticed the magnets on the original sled were slightly higher than the center of the coils which could affect the strength of the magnetic attraction and it could also be pulling the sled downwards and potentially increasing friction so i redesigned the sled with the magnets at the perfect height and adjusted some of the dimensions so it didnt wobble so much on the rail allowing for the sled to be one millimeter wider so the magnets can sit closer to the coils which i think should help as electric motors always have really small gaps between the magnets and the coils this resulted in a top speed of meters per second so after changing all the coils using a better power source and redesigning the sled weve increased the speed from miles per hour to now launching over which is definitely too fast for a paper plane so i designed a mini d printable glider that attaches to a sled using these mounting arms which only allow the sled to pivot off with this direction of rotation so even if the plane wants to take off before reaching the end of the launcher like the paper planes did it will be held down until the sled falls away Music thank you very much for watching this video whether its uh useful to you or not i hope you found it interesting and uh be great if you could leave a thumbs up down below if you did find it interesting thanks to all of my supporters on patreon.com for making these projects possible i couldnt do this work without your support so thank you very much i guess ill see you in the next video goodbye

write a youtube script

when I was a kid I had one of these Air Hogs planes that was powered by a small plastic engine this engine used compressed air as its Fuel and I remember pumping it up to high pressure with the small included pump before throwing it into the sky and never seeing it again and ever since then Ive wanted to build my own version but a little larger this video is sponsored by oppra so I started D printing some lightweight AER foil ribs in ABS plastic that will make up the internal structure of a wing I could have used other materials like Bola wood or carbon fiber but those are more difficult to manufacture and wouldnt save a significant amount of weight as these ribs are very thin once glued to the carbon fiber Spar I glued a carbon Rod to the Leading Edge and a carbon strip to the trailing Edge as the wing will be covered in a lightweight fil so these will help hold the a full shape now Ive never covered a wing like this before so I had a quick read of the instructions this stuff is very thin and de delicate so it creases easily but on the underside it has a heat set adhesive which can be ironed onto the wing structure which is why theyre printed from ABS due to its high melting point once signed onto the ribs a quick blast with a heat gun will shrink the film down to create a smooth Wing surface but there are some issues the first being the ribs at the wing tip Bend inwards due to the tension in the covering film and this tension also causes the air foil to shrink between the wing ribs making an inconsistent profile along the wing span plus I may have held the heat gun in one position for a little too long causing the covering film to melt I just want to talk about Wing design quickly as this is the fourth air powerered plane Ive attempted to build and fly the first plane I built was the heaviest but used a very similar Wing to what Im building now though it was constructed from foam board and despite the plane using a very inefficient engine it actually holds my personal best flight time record of and a half half seconds whereas my last air power plane which had a far more efficient engine and was much lighter only flew for seconds which is partly due to it hitting a fence but if the wing had a high lift to drag ratio it may have cleared the fence Ive been trying to make these planes lighter and lighter whil trying to make the engines more and more powerful but what I havent been looking at is the drag caed by the aircraft the last Wing I built was very light but also had a lot of drag and because the thrust has to oppose the drag the very little thrust produced by the engines struggle to get the airplane up to speed so this new Wing design uses a known eror for profile which with the right air speed and angle of attack can achieve a lift to drag ratio of to so a little extra weight wont be much of an issue and if you want to design a custom Wing like this with your own eror foils I highly recommend checking out on shape I simply import the airo profile from a tools.com then with a few added design features its basically ready to print these ribs are the same color as my tshirt to fix the issues caused by the covering film I decided to use a thin sheet of Boler wood to reinforce the Leading Edge this stuff bends really easily and if you take it outside in the British weather it will bend even easier so much so that I could bend it around a D printed mold of the wing and pin it in place until it dries once dry it holds the curved shape of the Leading Edge and I can print some new ribs with a small step where the bolster will Mount to I then gave the wood a few coats of bright orange paint to make it look nice which gave me some extra time to build a new Wing structure this new Wing is similar to the previous one but has much thinner ribs as Im hoping the bols of wood will add some strength as well as some Thin carbon rods to prevent the wing tips bending inwards once the paint was dry I glued the bolster wood onto the wing structure with some Thin CA glue and the wing was looking really nice with with the orange Leading Edge and gray ribs then the covering film can be applied to check how the airall profile turns out which looks much better also with the thinner D printed ribs this Wing is actually lighter than the first wing and with two of these Wings making up a m wingspan its over and half times larger than the air Hogs plane so we need a powerful engine to fly it fortunately Ive been developing D printed air engines for a while now and Ive tried many different Des designs such as spring operated Inlet valves camon push rod valves diaphragm actuated pistons and the latest design that uses a custom molded piston seal that expands Under Pressure to make the Piston airtight which is very similar to the air Hogs engine and this engine has been a huge step up in performance compared to my older designs producing G of thrust from an engine that only weighs G the engine is very simple with only a few moving Parts High Press air flows in through the top of the engine and is sealed by this ball bearing then as the engine turns a small pin on top of the Piston moves the ball upwards opening the valve and letting highpressure air fill the cylinder but this seal on top of the Piston is a vital component to the engines efficiency as it expands due to the pressure and seals against the cylinder walls this causes the pressure to push the Piston down closing the ball valve and turning the crankshaft then as the Piston reaches the bottom the air flows out the exhaust ports and the seal contracts removing any compression as the Piston travels back up now the airworthy version of this engine is slightly different but works exactly the same it just has this D printed thread to attach to a plastic drinks bottle and the air flows up this thin tube into the cylinder head I also decided to buy a highquality propeller to hopefully maximize the thrust output and surely the fact that this propeller comes with its own socks must mean its good I know many of you are thinking why dont I use CO cartridges instead of plastic bottles as they can store far more pressure well this CO cartridge is pressurized to about PSI or about times atmospheric pressure which is a lot but its volume is very small at just ml so if we let the CO gas flow out of the cartridge to equalize at one atmosphere of pressure wed have just over a L of gas but if we pressurize a L drinks bottle to PSI or about times atmospheric pressure we can store L of air this means wed need CO cartridges to equalize the storage capacity of the drinks bottle and considering the single cartridge weighs more than the bottle the air storage to weight ratio is over times greater for the drinks bottle so moving to CO cartridges would be a huge huge downgrade but you know what isnt a downgrade trying out the Opera browser Opa is the latest version of oppra desktop browser and it really helps me with my workflow when creating these projects for YouTube like whilst researching and designing a project my browser tabs can really stack up but oppra has these tab Islands so I can arrange the tabs into groups like heres the Island full of tabs for researching Wing error foils and heres the Island full of product tabs for purchasing the parts required to build it also so with the builtin ARA AI finding an answer to anything has been made far easier for example if you read something that you need more information about simply highlight it to bring up the ARA AI prompt then select between explain briefly explore or translate depending on what you need to know personally I love using the explain briefly option as I come across a lot of technical words whilst researching my projects and this feature saves me a lot of time and it also has a command line where you can ask it various questions which I find really useful if I need a calculation done fast with many other features such as the free builtin VPN ad blocker and messenger apps I highly recommend you try it by using my link below to download Opera today and thanks to Opera for sponsoring this video so to increase the runtime Im going to mount two L bottles back to back using a nylon bolt this requires drilling out the underside of the bottles which is the thickest part and shouldnt reduce strength by much as well as hollowing out the nylon bolt but in order to attach the bottles together I need to get the nylon bolt inside of the bottle which requires cutting up an old Arrow shaft and attaching a D printed socket which perfectly fits the bolt and just squeezes through the bottleneck this allows me to sandwich a large rubber washer between each bottle before tightening it down with a nylon nut which should hopefully create an airtight seal but before we test it for leaks we need a way to pressurize it the design Ive come up with is a D printed bottle cap that has a small ball valve inside and to fill the bottle with air I can use this hose with a D printed latch attached to it this hose fits perfectly through the Oring in the bll valve and the latch locks into the bolts with a small turn once the plane is up to pressure I simply twist the latch pull the hose out and the ball should seal the air inside so I fill the bottles with water to reduce the potential energy if something were to go wrong and pump them up to psi and both the nylon bolt and the D printed filler cap held up without a leak so its time to build the tail of the aircraft for this I chose to use bol of wood sheets but this mm stuff is a little thick so I ordered some mm sheets that were believe it or not shipped in a paper bag so lets reduce the weight of the thicker sheets by drilling some holes first with a small drill bit and then with a large counter syn bit to reduce splintering and then finally with a D printed bit wrapped in sandpaper to make the holes nice and smooth this reduce the weight by about and after a few coats of paint it can be laminated with the covering film to stop air flowing through the holes because leaving a bunch of large holes drilled in the tail of an aircraft would be a bad idea I then added control horns and attached the control surfaces with packing tape as its lightweight flexible and sticks well to the covering film now I want this tail to be controlled via cables to keep weight away from the rear of the plane so I used some spring steel wire as a torsion spring within the control surface this causes the surface to be sprung loaded in One Direction which should keep the control cables in constant tension now you may notice theres only two Tail Fins here instead of a separate elevator and Rudder and thats because Ive chosen to go with a vtail design I then printed this oddl looking part that the tail boom will Mount to and I can attach servos here to control the tail which works really well once the cables are attached this Central Hub also mounts to the bottles with a flexible printed ring and a small bracket on the tail boom to complete the planes fuselage now all it needs is the wings to be attached to these points and we have a completed plane lets take it out for a test flight right were at PSI but I can hear a leak out of the rear bottle cap its not a good sign cuz I pressure tested this already okay quickly releasee that here goes nothing check the controls and start the Applause engine thats brilliant and can I land it can I land it it worked it worked it looks like the propeller is snapped off though lets go to inspect right so the uh crankshaft snapped but uh that is definitely the longest air powerered plane flight Ive ever achieved so now that the record flight time for a D printed air powered plane has been increased to seconds I took to Instagram to ask how long you all reckon it will fly for and the majority of you chose more than a minute so I cant give up just yet to increase the flight time I fixed the leak in the bottle cap by reprinting it with a larger oring and it was ready for another Music test and one for good luck please run properly oh the batterys come off come on we can climb we can climb come on yes so good oh thats so much better oh look at it its still going like really well oh my God its Music Applause working mate its still doing well its still going just not its so I know what you mean your and Rollo yeah wait its still CL wait the winds coming from our right isnt it so its com from my back I think okay Im trying to turn it its coming its coming Applause down the battery fell out can you see it hanging out the side and I sto it oh no I sto it oh we have to do another one oh yeah for the next flight I decided to add a small onboard camera but due to my excitement of the previous flight whilst tightening a bolt I accidentally pierced a hole in the bottle you see that its just hissing just hissing out that hole fortunately I had some spare bottles so it was an easy fix though throughout this period of adding the onboard camera and bottle replacement I forgot to do one last vital check and that was whether Id push the tail boom back into its Mount properly because this means the control cables are too short causing a lot of up elevator trim oh no no no no no this may look like the end of this plane but after a quick depressurization I was surprised to see how little damage there was the only damage apart from this D printed bit here as far as I can see this Wings fine just this Wing here which a little bit of packing tape thatll fix that so I fixed the wing with a small patch of tape and reprinted a new Central Hub which I honestly reckoned saved the plane when it snapped and by the following day it was back up and flying though Ive yet to beat the minute second flight record again most mostly because it really struggles in the wind oh no no no watch out watch out watch out sorry I cant control it its way too windy I have had a few flights over a minute but I think it needs some weight shaved off before I get near the minute mark even if I can reduce the overall weight of the plane by about that will be enough to add another bottle which will help extend the flight time thanks for watching

write a youtube script

this engine doesnt run off a combustible fluid or gas but instead runs off the air that we breathe but before we feed it this air the air must be compressed which leads to several differences between this engine and a conventional combustion engine in terms of power output a combustion engine can produce its maximum power right up until the fuel tank is empty whereas the power output of this engine is determined by the pressure stored in the tank which means over time the power will decrease gradually as the engine uses more air and the pressure decreases one option to get around this is to optimize the engine for a much lower pressure and then run it off a regulator so the pressure into the engine will give a more constant power output throughout the pressure decrease in the tank when doing these engine tests I use a large compressor tank and have a regulator to adjust the engine pressure however because I eventually plan to stick this engine on an aircraft this regulator is far too heavy so Im going to make my own a pressure regulator has three main components a spring a diaphragm and a valve which all work together to control the output pressure when air is supplied at the input it flows through the valve into the diaphragm and out the other side and because of this pressure the diaphragm starts to move up which closes the valve now the pressure at the output is determined by the compression force of the spring so to reduce and regulate the output pressure we need to adjust the compression of the spring until its just a bit less than the outputs pressure on the diaphragm this way the valve wants to stay closed when the pressure is set correctly but if the output pressure drops slightly the diaphragm will move down and open the valve bringing the pressure back up its essentially a mechanical feedback loop that opens and closes the valve depending on the pressure at the output now I bought this large pack of Springs a while back and the compression springs are relatively weak so instead of using a diaphragm Im going to use a small diameter syringe as the pressure force will be far less due to the small piston area and this regulator is absolutely tiny and only weighs about seven and a half grams which would be perfect for the plane Im building assuming it works of course which it does sort of work the pressure is being regulated but it causes the engine to pulse the reason for this is the friction caused by the syringe piston requires about psi of pressure difference to start moving so I think I need to go with a diaphragm design side note using a piston in a regulator isnt actually a radical idea as I later found out my compressor regulator uses a piston design too after many iterations and some new stronger Springs I finally got a regulator that was working the input pressure is PSI and the engine is running at about PSI and I can adjust the engine speed by turning this bolt to increase or decrease the compression in the spring but does this increase the engine run time this is a graph of the thrust produced as the air in the tank decreases and as you can see it outputs the maximum thrust at the beginning due to the high pressure and then it slowly decreases until the pressure is too low for the engine to run now because the regulator is set to keep the pressure at about half of the maximum pressure we might assume the thrust output should look like this with a lower Peak thrust but a far longer run time however it actually looks like this which it does have a slightly longer run time but not by much and thats because it still runs at the same efficiency so the area under this portion of the graph is just spread out over this portion but maybe this engine isnt optimized to run at the lower regulated pressures so I decided to build a new engine with some slight modifications the first being a new cylinder using an acrylic tube to hopefully reduce some friction and I also chose to D print the new piston from pla on my prusa Mark III as its faster and cheaper than printing the Piston out of resin and because the cylinder is far smoother now the ridges on the fdm print shouldnt increase the friction so to optimize this new engine for the lower pressure Ive designed it with this large volume above the Piston the idea is that when the Piston is at the top of its stroke the volume above the Piston is onethird the volume of when it reaches the bottom this is because the input pressure is about three times atmospheric pressure so expanding it by three times should bring the pressure down to one atmosphere at the exhaust which should hopefully reduce any wasted pressure but this engine has far less thrust than the old one which I couldnt understand at first as the piston and cylinder should have less friction as well as using more air per cycle due to the expansion volume which should increase power but then I realized the other major change to this engine is the spring within the cylinder head which I recently added as it helps to keep the ball valve closed when its not being pushed open by the Piston from watching the slowmo of the two engines it doesnt look like theres much difference between the two but if we analyze frame by frame we can see the inlet valve without the spring stays open for a split second longer allowing more air into the cylinder which seems like it would be less efficient as more air is wasted out the exhaust ports however I now think my expansion design might have a hidden inefficiency here we have two engines with identical piston and stroke Dimensions but the engine on the left has the expansion volume above the Piston which allows the air to fully expand within the cylinder resulting in very little air have been wasted out the exhaust whereas the engine on the right has zero expansion volume and there is lots of high pressure air exiting the exhaust but believe it or not the engine on the right is actually more efficient and heres why when the engine on the left reaches the top of its stroke the inlet valve opens and about cubic centimeters of air enters the cylinder but for the right hand engine to use the same volume of air the Piston must move down so now both Inlet valves have closed and each engine has used the same cubic centimeters of air however the engine on the right is producing far more torque due to the angle of its crankshaft and even when the left hand engine rotates to the same crank angle the pressure in the cylinder has now decreased and therefore the torque is still less so throughout each rotation the engine on the right consumes the same volume of air but produces about twice as much torque as the left engine which seems counterintuitive at first but its all to do with the matching of the press your expansion and the offset of the crankshaft angle so I built this new engine which doesnt have a spring and the Piston almost touches the top of the cylinder at its highest point and I also added this two millimeter set screw to the Piston instead of the previous Steel Rod so I can easily adjust the valve timing but before we look at the data lets talk about the internet data that is collected about you and how the sponsor of this video in cognee can help out there are these internet data Brokers that collect information about you like your name age address your online shopping habits and more and businesses can purchase this information now you can reach out to each of these Brokers directly to erase the information however the sheer number of them makes it very difficult and time consuming to do so which is where the sponsor of this video in cogney can help out in cogney reaches out to all the data Brokers on your behalf to request removal of your personal data and because many data Brokers continue to collect your data after a removal request Incognito will monitor the market and request repeat removals if a new record pops up so thats what makes their yearly subscriptions so important because its not just a onetime wipe of all of your data but instead a constant automated service to make sure youre protected online the first people to use the code or the link in the description down below will get off of incognity so thanks for incognitive for sponsoring this video and go check them out in the link in the description down below so aside from this new engine performing really well I decided to test how the Piston pin affects the performance here is one of the first thrust tests where the thrust decreases linearly with the pressure decrease but see what happens if I shorten the pin by just . of a millimeter it reduces the peak thrust significantly but also increases the run time and what I find interesting is the flattening of this curve at the higher pressures all these thrusters are carried out with a two liter drinks bottle pressurized to PSI and thats because when I was testing the early versions of this engine I would pressurize these bottles in my bedroom so for safety reasons I didnt want any of them exploding and thought PSI was a safe control variable but now looking at this thrust curve I wonder what would happen if we tried some higher pressures so lets run the tests at PSI Music here is the long piston pin result with a peak thrust of . Newtons and a run time of about two minutes but check out the shorter pin graph at the higher pressure the thrust is actually lower the very early versions of this engine had curved channels printed into the cylinder head to allow better airflow around the ball but I found this would cause strange misfire effects but the ball in the valve would be launched far too high and levitate for a split second due to the highspeed airflows around it and at the time I didnt have any Springs to prevent this so I designed the internal diameter of the cylinder head to be just a little bit larger than the ball this way the air flowing into the engine will tend to push the ball back into the valve and this was enough Force to allow the engine to run upside down but with the latest engine the power comes from the ball lifting off of the valve when the Piston hits it so at higher pressure the downwards Force caused by the airflow is far stronger therefore the ball doesnt lift as high and less air enters the engine and as the pressure slowly decreases the ball is pushed higher therefore increasing the thrust at least to a point where the pressure is then too low to maintain the engine RPM so this is almost like a selfregulating Thrust output built right into the cylinder head without the need for an external pressure regulator and with some more tuning I reckon we can flatten this line even more which will be perfect for flying an aircraft now I know we cant really compare this engine to the older engines as weve changed one of the control variables by increasing the pressure but this new engine is producing Newton seconds of efficiency which compared to the previous engine in my last video is about twice the amount and compared to the last time I made an air powered plane about times that old engine and that thing was able to maintain level flight so Ill be very shocked if this thing doesnt fly now even though I reckon this engine will fly what happens if we make a twocylinder version the easiest way to achieve this is to Simply Mount two engines back to back so they both share the same crankshaft to do this I had to print these connecting rods that are slightly curved so they attach to the same rod in the center yet still aligned with the Pistons and they fit into the crank case just like the single cylinder version but in opposed directions the next step is to attach the cylinders and pistons which are basically the same as the single cylinder engine apart from the cylinder head as they need to be plumbed together using an air manifold which splits the incoming air into two channels and flows through these thin white tubes to each cylinder then the ball valves can be inserted into each cylinder head with a D printed screw cap to keep it airtight well at least more airtight than the air manifold which leaked a lot through the Orings due to a slightly loose design but that can be fixed later lets see if it runs Applause although it runs something isnt quite right the engine doesnt seem to have any more power than the single cylinder engine and it looks like there is only condensation collecting on one of the cylinders anyway it turns out when I glued the tube into one of the cylinder heads I accidentally blocked it with glue so one of the cylinders wasnt getting any air so with the new cylinder head and air manifold the engine was now perfectly airtight and sounding much better look with a peak thrust of . Newtons at psi or equivalent to lifting grams this engine has a power to weight ratio of nearly to . which might be an excellent option for a dragster but with an efficiency near half of the single cylinder engine I think I know what Im going to stick on a plane thank you very much for watching

write a youtube script

thank you over the past six years Ive tried to develop the most powerful and efficient engines that run off compressed air from spring activated Inlet valves to diaphragm actuated Pistons so whats the next step to increasing the engines performance make it simpler this is probably the simplest design of air engine as it only requires this ball valve at the top and a pin on the Piston to open the valve now Ive tried building this design before and couldnt get it to run but there are these engines which run off high pressure CO canisters and they use this exact ball valve design and I think its because they rely heavily on an airtight piston which Ive yet to achieve as the Piston moves towards the top of the cylinder it pushes the valve open fill in the cylinder with high pressure air then as the valve closes no more air can enter the cylinder which is why its essential that none of it leaks around the Piston as the Piston moves down it passes below the exhaust ports where any leftover pressure is released and the momentum in the flywheel pushes the Piston back up to restart the cycle but if the Piston pin is too long itll open the ball valve early which could cause the engine to backfire so it needs to open the ball valve as late in the cycle as possible also as the Piston travels back up it compresses the air above it as the volume decreases and this air cant be at higher pressure than the Air Supply or the air will flow back into the tank and the large volume above the Piston reduces this compression so although its a simple design theres a bunch of things to get right so first lets try and create an airtight piston seal to achieve this I D printed a cylinder and piston from resin on the form plus printer to make the surfaces smooth with the Oring fitted to the Piston it seemed to create a seal but it was far too tight against a cylinder even with some lubricant there was a lot of friction so I just needed to adjust the Piston size to reduce the Oring contact area however Im not confident itll be completely airtight the next step is to print the crankcase and crankshaft which I printed with pla on my prusa Mark III I could then slide the bearings onto the crankshaft and insert it into the crank case for the con Rod that connects the Piston to the crankshaft Im using just a small bent piece of wire as this has proven to work well in my older engines then the Piston can slide onto the wire and the cylinder can slide over that and finally the ball for the inlet valve and a high pressure Air Supply is attached to the top but will it run okay I did not expect that to work first try Ive just noticed something very interesting the Oring doesnt seal at all uh when it was at like PSI but now this is at psi when I push the Piston up to the top it seals it leaks to start with and then the the pressure causes the Oring to expand and actually see it against the wall though because this engine wont run at pressures lower than psi I printed another piston that is slightly larger but still has low friction this allows the Oring to seal at much lower pressures and the engine runs surprisingly well for being so basic but it could be better by filling the engine with excess lubricant we can visually see the airflow out of the exhaust and because the Oring takes a split second to seal there is a significant amount of wasted air passing by the piston and out the exhaust so whats the solution to this Im going to make my own piston seal this new piston seal has a flange on the top that expands and contracts depending on the pressure above it so when it moves up towards the valve it allows the air to flow around the Piston then once the valve opens the air pressure will cause it to expand and seal against the cylinder hopefully better than the Oring did to make the seal Ive D printed these molds and bought some twopart silicon its basically the same as using resin glue just add two parts together and mix then I can slowly drip the Silicon into the molds and let it settle from one side to prevent any air getting trapped the following day I can use a blunt knife to release the seals from the molds and they look really good but will it work it does work but also requires a lot of lubricant on the plus side it seems there is far less air leaking around the Piston when the inlet valve opens it just needs some design changes to reduce the friction I think the cause of this friction is due to the whole seal making contact with the cylinder instead of just the top of the flange so this is the new design seal which is tapered towards the bottom and Im hoping even if the whole seal expands the bottom portion shouldnt expand enough to make contact with the cylinder I also noticed that the old seal contained a lot of bubbles which were probably caused by the mixing process a common solution to this is to place a silicon in a vacuum chamber to expand and remove the bubbles but unfortunately I dont have a vacuum chamber so Ill have to use the next best thing a syringe by sucking the Silicon into the syringe I conceal off the end and pull back the syringe piston to create a vacuum and the expansion of the bubble seems to be working after a few minutes I can let the air back in and the Silicon seems to be bubble free plus the syringe makes it far easier to pull the Silicon into the mold the following day I can remove the seals and check if my cheap vacuum solution worked which judging by the lack of bubbles it definitely did the engine still runs but when taking a look at the slow motion footage the seal is still expanding far more than expected creating a large contact area against a cylinder so I think were going to need a slightly less flexible material for the seal this new rubber material is a lot higher on the shawl hardener scale so where the previous silicon was about the hardness of a gummy bear this is more like a pencil eraser but I quickly found out that this also makes it a nightmare to remove from the molds even with the addition of these removal tabs I did eventually get one out and managed to test it however it was too stiff and the engine would require nearly psi to run just like the original oring design I carried on testing a bunch of piston seals using the soft silicon and the hard rubber but wasnt making the best progress then I remembered this process is called research and development for a reason and so far Ive only been focusing on the development side so the other day I was scrolling through eBay and managed to find this the original Air Hogs Sky shark air powered plane speaking of sharks this video is sponsored by surfsharkvpn a VPN is a virtual private Network that you can download onto your computer mobile and tablet to keep your online identity safe by encrypting all of the information sent between your device and the internet and it can also swap the real location of your device with a new one so you can virtually travel to any country across the globe this also allows you to change your virtual location to access libraries and streaming services from other countries so you dont miss out on shows that dont work where you currently are surfshark doesnt monitor track or store what you do online but it can be set up to alert you if any information such as your passwords are breached through weak security websites so you should really consider signing up for surfsharkvpn and if you use the promo code Stanton or use the link in the description you can get three months extra for free when you sign up so go check out surfshark today the condition of this is basically brand new as it hasnt even been unpackaged which is surprising for a year old toy and I havent seen one of these since I was a kid well actually I had a later version as I was only three years old when this model was released and it was the whole inspiration behind my compressed air engine Journey but I quickly realized I have been wrong for the past six years my first D printed air engine that I built back in was based on this Air Hogs design or at least how I thought it worked the idea was that a small spring on top of the Piston would keep the inlet valve open whilst the Piston traveled down which was a really simple design but had some reliability issues due to the spring being quite prone to failure taking a look at the actual Air Hogs engine the spring is definitely there but so is this interesting seal design on the Piston to see exactly what was going on I removed some of the fuselage foam as it was attached with hot glue as well as removing the original bottle I can then D print a threaded adapter to run the engine off my compressor air supply Applause then it can be fixed in position with another bracket so I can get some slow motion footage and guess what I was completely wrong about the Piston spring meaning my first ever engine was actually a unique design but also meaning that my latest air engine is essentially a copy of this Air Hogs design from years ago now I didnt want to have to take this engine apart as I was worried about breaking it but somehow a piece of dirt got into the cylinder which eventually got stuck in the inlet valve preventing the engine from running so I had to take it apart anyway it seems the Piston seal is far thinner than any of my designs and uses a hard rubber similar to the clear stuff I have also the Piston isnt actually attached to the connecting rod its just a loose ball and socket joint which I might also have to copy as I found wear marks on one side of the cylinder after extended testing due to the Conrad putting a twisting load on the Piston I then built a Thrust test rig to see if this Air Hogs design is actually worth pursuing further foreign this is a Thrust data from my previous best diaphragm air engine with a Max thrust of . Newtons and a run time of about seconds and this is the Air Hogs engine understandably it produces a lot less Peak thrust as the propeller is far smaller than what I used on my old engine but the runtime is far longer and if we measure the area under the graph we get the engines impulse which is a measure of its efficiency I previously explained it like its miles per gallon but instead its thrust time per liter bottle at pounds per square inch this resulted in an Impulse of . Newton seconds which for reference my old engine had an Impulse of just . Newton seconds so lets take this information gathered from the Air Hogs engine and build our own to create such a thin wool piston seal I printed a positive mold of the seal and use silicon to create a negative mold this means the final mold will be flexible which not only helps with filling all the gaps without air getting trapped but once the rubber has dried I can easily peel the Piston seal away from the mold then the seal can be glued to the Piston to make sure only the flange will move and this new piston also has the internal socket dome for the connecting or not so connecting rod to seat within only issue is it doesnt run yet the Piston seal expands as expected but just a little too much to the point where it almost turns inside out whilst working on a slightly thicker design I decided to test the engine again the next day and well it fired up perfectly the cause for this was the rubber takes a few days to fully cure so it hadnt fully hardened when I ran the first test its still not perfect as it very nearly turns inside out but its just enough to hold on and keep an airtight seal so how will it perform on the thrust test stand before I look at the data I can already tell its performing very well as the RPM is quite high and it takes a good two minutes to run out of air now remember this graph is my best ever performing air engine well here is a Thrust from the new engine not only is it producing higher Peak thrust but the run time is nearly twice as long resulting in an Impulse of . Newton seconds which is roughly three times the efficiency this seems like it shouldnt be possible but if we take a close look at my old engine its clear why the old engine required this initial air storage section as part of the twostage inlet valve which means if we have PSI flowing into the engine by the time it flows into the cylinder the volume almost doubles meaning the Piston only experiences about PSI plus a lot of energy is wasted in stretching the diaphragm which just Springs back after each cycle and you can definitely hear the difference in performance when running the engines back to back foreign but after several thrust tests I started to notice the peak power was reducing and it seems this is due to the Piston pin wearing down from impacting the inlet valve at such high RPM so I modified the Piston to use a metal pin and added this small lip to prevent the seal over flexing this new engine produced the same Peak thrust but ran for longer which as you can probably guess by now increase the efficiency even further to . Newton seconds nearly five times my diaphragm engine and three times the Air Hogs engine of course the next step is to mount this engine on a plane but before I do so I want to know how much thrust I can squeeze out of it so what happens if I increase the travel of the piston this engine now has a piston travel of millimeters instead of the previous millimeters which should increase its torque as you can imagine producing more thrust will result in lower efficiency but just look at the peak the rest of this thing at . Newtons Peak thrust thats equivalent to lifting about grams which considering the total weight of my old air powered plane was grams thats enough to make it Fly vertically of course the runtime is a bit shorter at just one minute but its still producing an Impulse of . Newton seconds so the next logical step is to build a lightweight plane and tune the engine to balance efficiency and thrust but youll have to subscribe if you want to see that in the next video thank you very much for watching

write a youtube script

if you drop a magnet down a copper pipe it seems to experience a resistance force acting against it despite the copper being nonmagnetic this is caused by the magnets generating electrical current in the copper called Eddy currents and these Eddy currents create a magnetic field within the copper as oppose the permanent magnet almost like an invisible no contact frictional force acting between the magnet and the copper now my plan is to use a metal disc and a disc full of magnets mounted very close to one another to potentially create some kind of magnetic clutch to replace the gear system on a bicycle I started by cutting two discs on my CNC router one for aluminum and one from copper the reason for this is Im not sure which is the best material for this application my intuition would say copper as its far more conductive than aluminum but if you ask Google it claims aluminum is the best and has a research paper to back it up however the majority of any current demonstrations use a copper pipe but maybe this is because copper pipes are easier to Source from plumbing stores in a recent video uploaded by veritasium about the worlds strongest magnet he drops a sheet of copper and aluminum at the same time within a strong magnetic field and the aluminum sheet seems to fall far slower indicating a greater resistive Force so Ive D printed this axle as part of a test rig to allow the discs to freely spin like a then mount a set of magnets either side of the disc with without making any contact then I can hang a weight via a length of string to apply a constant torque on the disc which should provide a reliable way to measure the resistance produced by the magnets for reference this is The Descent rate with and without the magnets the reason the nonmagnet test is still slower than freefall is because it has to overcome the inertia of the disc but how does the aluminum compare to the Copper the copper clearly has the best braking Force as a disc spins slower than the aluminum disc which makes sense as copper is about more electrically conductive over aluminum but the reason why the copper sheet fell faster in the veritasium video is probably due to Copper weighing over three times more than aluminum so although the resistive force in the magnetic field might be a bit stronger the weight of the sheet skews a results in fact even if I run my test without the magnets at all the copper will still win due to the extra inertia of the heavier disc because of this I only measure the RPM of the disc at the end of The Descent when the disc has finished accelerating and is spinning at a constant speed and its clear that with the grades and materials I have available copper is definitely the winner at least in terms of braking performance because at nearly six times the cost and three times the weight maybe aluminum would be better for other applications the next factor to consider is the magnet orientation in this first test I have all the magnets mount hunted with their poles aligned so the North Face is pointing towards the camera but what if I flip one set of magnets so well have a north south north configuration this slows the disc even further because the Eddy currents in the copper disk are produced by a change in magnetic field so having the magnetic field change with the flip magnets generates stronger Eddy currents in the copper and therefore more resistance I next wanted to test these smaller Cube magnets as I can pack them tighter around the perimeter of the disc and these produce so much resistance that I think we need to increase the weight I also tried the cube magnets in a checkerboard configuration but this seemed to perform worse since Ive been using magnets I had laying around from older projects I need to find some new magnets for this setup the best option I could find online were these millimeter Cube magnets which should outperform the previous eight millimeter magnets so I ordered of them but once mounted on the test rig they performed very similar to the smaller magnets despite being larger in volume which may be considered the quality of these magnets are skewing the results of this test after a quick Google search it seemed there are many grades of these neodymium magnets and these seem to be on the lower end so I had a search around to find a reputable seller of n grade magnets which appear to be the strongest they were double the price but its clear why attaching the cheap magnets to my bench vise I was able to suspend a single . kilogram weight below it but with the n grade magnet it could hold about two and a half times that so I then printed amount for these new magnets and as you can see they produce so much resistance that the disc barely moves so now we know the combination of these high quality magnets and this copper disc produces such a high resistance force how are we going to fit this on a bike my initial plan was to mount the copper disc to the rear wheel spokes as Ive previously mounted pulleys for electric motor drives but one issue is the spokes arent designed to be a Precision mounting point so it will be very difficult to get the copper disc spinning perfectly straight to avoid hitting the small gaps between the magnets the only other option is to mount everything to this thread on the Hub which is designed for just a single sprocket so we need to fit a copper disc a disc full of magnets and a sprocket into this and a half millimeter wide thread which also has two different diameters and thread directions the largest of which is apparently a standard bicycle Imperial thread of . inch diameter and threads per inch which obviously makes things a little difficult as I cant exactly machine this thread but fortunately we have a solution as form Labs have sent me their form plus printer and some of their rigid K resin this D printer works by curing a resin with a UV laser and this allows for much more precise and hopefully stronger Parts than my other D printers once the print is complete it needs to be cleaned in isopropyl alcohol before going through a final UV cure to make sure the part is as strong as possible then simply break off the support and we have a mount for the copper disk with the fine Hub threads already printed in once threaded onto the Hub its clear that this makes for a much straighter Mount but the question is is it strong enough well I know how we can find out as I also printed this rear sprocket from the same resin to see if it can handle the torque through the chain without stripping the thin threads that lock it to the hub and surprisingly it worked quite well for about a mile of mixed flat riding and slide up hills until this happened big problem for the project however it gave me some useful information as we can see from the footage the crack started from the sprocket teeth and moved inwards towards the Hub suggesting that the Hub wasnt the weak point but instead the design of the Sprocket was as it cant be thicker than three millimeters and still mesh with the chain so the next step is to cut a large copper disc that will be one half of the magnetic clutch and we will directly Drive the rear wheel then we need a lot of magnets mounting to a freely spinning hub that will be driven by the chain but how many magnets do we actually need I went back to the test rig and used this container of water to vary the weight hanging from the axle which directly varies the torque on the disc and this showed that the RPM and the disc is proportional to the torque applied so if we plot a graph of torque versus RPM we get a nice straight line the next thing to test is how the RPM is affected by the number of magnets so each test had a different number of magnet sets with each set containing four magnets as I have two on the front and two on the back and this result wasnt linear so increasing the number of magnets didnt reduce the RPM proportionally but from the first test we know that the weight is directly linked to the RPM so we can actually calculate the required weights we need to reach RPM for each set of magnets this now gives us a constant RPM value with a varying amount of torque and can be used to calculate the dissipated power which if we plot a graph of the power at RPM we get a straight line so this means the number of magnets is directly linked to the power transfer through the clutch and not the RPM I then found this calculator online for working out the power required to ride a bike at a given speed so lets say we want to ride at miles per hour we need a power output of about Watts from the test I know that five sets of magnets can produce . Watts at RPM this means that Ill need sets of magnets and because theres four magnets per set thats a total of magnets but theres a catch this RPM is the speed difference between the copper and the magnets which at miles per hour the rear wheel and therefore the copper disc will be spinning at RPM so the magnets need to be spinning RPM faster than that at RPM this might sound like a lot but actually through the sprocket ratio on this bike I only need to Pedal at about RPM to maintain the speed difference which is definitely achievable I then realized it was cheaper to buy these magnets in packs of so I ordered just to be safe which seemed to wipe out their stock and having magnets means theres sets which should fit perfectly around the perimeter of the copper disk I then press fit some bearings onto an adapter that matches the smaller thread on the Hub then these bearings can be pressed into the rear sprocket that also has mounting holes for another disc that will be used to attach the magnets which I cut from an aluminum sheet once everything is on the bike its easier to understand how this will all work the chain spins the aluminum disc which will eventually hold the magnets and it isnt linked to the rotation of the rear wheel or copper disc at all due to the bearings so the pedals and rear wheels spin completely independent of one another and the only way to drive the rear wheel is through the magnets inducing current in the copper disc I then D printed a magnet bracket using clear resin on the form and these turned out really nice the magnets can then be pushed into position with the flipped polarity configuration and attached to the aluminum disc with a couple bolts by spinning the pedals its clear the magnets have an effect on the rear wheel which is very promising as this is just of the final number of magnets but there is a slight issue the chain sticks to them to stop this from happening the magnets need to be spaced further from the chain which meant redesigning and printing new Hub parts to move the clutch towards the wheel as well as a thicker aluminum disc to prevent it flexing under the load of the magnets also because the magnets stick to the spokes too I had to squeeze the spokes together with cable ties to move them away from the clutch which isnt structurally ideal but this is a prototype and now the magnets are able to spin freely without sticking to the chain spokes or steel bike frame and we can get a sense for how this clutch both accelerates the wheel when the pedals Spin and applies a braking Force when the pedals are stopped all we need to do now is Mount the remaining magnets to hopefully increase the torque transfer through the clutch and with the rear wheel locked there is a huge amount of resistance on the pedals but the magnets are still able to spin and with the pedals locked the wheel is also able to spin and experiences the same resistance through the clutch however when both are able to spin the pedals Drive the rear wheel with ease and the braking Force stops the wheel almost instantly I think its time to test this thing and if we can actually move through the power of magnets lets go okay its kind of working if it just feels like Im in a really low gear lets see what happens if I stop pedaling oh I can really feel that resistance through the magnets turn around okay hill now struggling a bit its still able to move cant get much speed at the moment drop a pedal quicker oh right I mean it works I think I might need to adjust the chain tension because it sounds like its sticking to the magnets and then starting to jump off the sprocket so lets tighten that a bit in theory the faster the more torque it should output according to my previous tests but like now Im just matching the speed of the copper and then if I stop pulling oh I can really feel that resistance its a really weird sensation because it its uh Im not directly linked to the to the rear wheels it almost feels like Im wheel spinning if I put too much force in its very hard to keep a constant pedal speed with the with the rear wheel okay here we go and lock the pedals so its not quite as strong as a regular break uh but it does slow down a bit not bad I think we need a larger sprocket on the front to increase the gear ratio the reason why it feels like its in a low gear is due to the difference in RPM between the magnets and the copper because in order to generate Eddy currents the magnets must always be moving faster than the copper therefore to reach a higher speed I need to adjust the sprocket ratio to spin the magnets faster so I removed the original sprocket and machined my own to sprocket which should increase the ratio significantly and has a diameter of millimeters which is the size limit for this bike frame fortunately when buying new chains they are slightly longer than most bikes need so I didnt need to join Two Chains together to make up the extra length this really helped when cycling as I was able to maintain a pretty decent speed when riding on flat ground its difficult to explain exactly how this magnetic clutch feels as its so unique when riding on smooth flat ground as long as the magnets are spinning faster than the wheel the toolk at the rear wheel is determined by how fast you pedal also the braking Force seems to be a lot stronger at high speed due to the difference in RPM of the magnets and the copper so its not great for bringing the bike to a complete stop but if you wanted to bleed up some speed it definitely has some advantages over a conventional break due to its no contact friction which would save on brake pads and thats the summary of The Good features of this system so would I recommend you build this for your own bike definitely not you see when climbing Hills it really struggles the best way I can describe it is its like riding on Sand every of the pedals requires a lot of force and you barely move forwards it might look like Im riding in a low gear due to the pedals spinning faster than the wheel but because the RPM difference between the magnets and copper is so large the resistance is also really high so Im putting a lot of force through the pedals and going nowhere speaking of nowhere I was exhausted after riding up this relatively easy Hill and if youre wondering where all my energy was going well when I got home I took out my thermal camera and hopefully this makes it a little more obvious yep the copper disc has heated up by about degrees Celsius which using the specific heat capacity of copper and the mass of this disc is about joules of wasted energy I had hoped it would somewhat simulate a continuously variable transmission where you could pedal at whatever speed you like and the clutch would vary the torque output depending on the inclination of ground you were riding on however the wasted heat energy makes it so much harder to ride and therefore I cant see it having any advantages over a direct chain drive but hey I had an idea and I wanted to test it which was made possible by onshape which is a professional grade computer aided design and product data management tool for businesses and in fact onshape is used by Form Labs who create the formlabs resin printer onshape was actually created by the same Founders as SolidWorks because they saw the engineering teams needed product development tools that were more modern so they created onshape from scratch and unlike other CAD programs onshape uses Cloud native architecture which allows for realtime collabo

write a youtube script

Im making a real life fallu course but with Lego my idea is designed around the roll out course where you have to run around on a spinning cylinder that is speeding up and pushing walls at you its going to be a twoplayer game that uses remote controls to slide your minifigure along a pulley track as you dodge obstacles and aim to be the last man standing I was really trying to make the pulley track a recessed feature within the rotating cylinder but I couldnt figure it out so I thought of a new idea that slides the Minifigures back and forth by their head think Monsters Inc door Transportation I used to play with Lego a lot as a kid and I forgot how expensive it was since Im getting back into the hobby I ordered some pieces to get started and it was a lot I also started recreating some of the standard Lego pieces in D space so I could D print out my own versions and change it if its needed basically Ill use all of these pieces as a foundation but if I ever need to create a custom part Ill have the freedom to do so I also really want to make my own Fall Guys Minifigures and I use mid journey to get some inspiration images for my design follow to see the rest of my build

write a youtube script

part one of this project covered the building and testing of a two rotor aircraft that used D printed helicopter rotor heads for control these rotor heads were simple by Design with each blade being freely hinged at the Hub and with some precise motor movements they were able to provide sufficient control to keep the aircraft in the air however it didnt fly first time due to some vibrations with the frame but this was fixed with a stronger wingspar and redesigned motor mounts I then bought these slightly larger blades which are a bit longer in length and theyre also completely unbranded in the hope that these will provide more thrust because once we add the fuselage to this thing its going to weigh quite a bit more but this brought back the vibrations and I think the cause of these vibrations were more serious than I first realized from this slowmo it seems the small vibrations caused by the blades pivoting causes the frame to flick slightly which is then fed back into the rotor head causing crazy unwanted blade pitch movements these extreme blade angles then cause unwanted load on the motor which then amplifies a vibration to the point where the whole roads ahead rapidly loses lift bear in mind these vibrations are bending this millimeter carbon fiber tube like its an uncooked spaghetti and these tubes are very strong I could go back to the old blades but if such a small change in the blades causes such a drastic change in performance then I hate to think what would happen once I get the fuselage on this thing for this reason I started designing an actual helicopter rotor head with a swashplate and linkages but whilst developing this I received an email from the original designer of this freely hinged rotohead system if you havent seen my older video on the swash playlist helicopter the original idea was invented by a guy called Jimmy as part of a research project at the University of Pennsylvania and his motor controller was developed by his friend Matt much has since developed this motor controller further as part of his business vertic and hes actually designed his own freely hinged roads ahead system that uses a single Central hinge instead of one hinge per Blade the clear advantage to this is obviously its simpler having less hinges in the rotor head but more importantly to my specific use case is that all of the blades are fixed together let me show you why this is important imagine this is the roads ahead and all three blades are evenly spaced apart by degrees the center of mass should be in the center of the rotor head meaning its perfectly balanced and when the motor wants to create a control output it either accelerates or decelerates to cause the blades to lead or lag in an Ideal World the rotor head should stay balanced as all the blades will swing by equal amounts but if there is a slight variation in hinge friction or one blade experiences more drag than the other they might lead or lag by different amounts which will shift the center of mass away from the axis of rotation causing an imbalance and therefore a vibration so having all these blades fixed together will hopefully remove the vibrations Im getting and Matt also told me that he successfully tested a threebladed version of this rotor head with a single Central hinge so with this information I have to build my own this rotohead has a central hinge at degree angle which means when the motor accelerates or decelerates the blade in line with the hinge will change an angle of attack but the other two blades will actually move up or down this looks a little odd when rotating as one blade maintains the same planar rotation but the other two seem to follow a different plane in rotation once on the aircraft it was clear that the previous vibrations were gone which was very promising however it doesnt seem to have much control so the aircraft could only hover for a short period before pitching over and crashing but then Matt showed me another road ahead hes been working on which has the same Central degree hinge as well as another horizontal hinge which means the rotor head can tilt in any direction almost like a universal joint but because of the degree hinge it can still be controlled by the motor speed to Vector the thrust of the blades I found this fascinating that adding a couple of hinges to the rotor head can cause this thrust vectoring effect by simply adjusting the motor speed so of course I have to build my own this is very similar to the previous design but features the extra horizontal hinge through the center and whats interesting is this pivot Direction works fine due to the horizontal hinge but because of the degree hinge to get it to Pivot left and right the blades have to twist side to side and its this exact movement in reverse that will be used to control the thrust vectoring when mounted on the motor it only requires a small amount of motor pulsing to get the rotor to Pivot which is really cool to watch also because of the reduced vibrations of this rotor head I removed the bulky motor mounts and went back to the millimeter Wings bar to save weight but how does it perform in the air not great after lots of tuning and tweaking of the gyro stabilization it wasnt getting much better but then Id have thought what if I turn the gyro stabilization off my theory was that the gyroscopic stability of the rotor would tend to stay in position as the aircraft would rotate so is this fighting the electronic gyro stabilization the fact this thing flies with no stabilization on the pitch axis is really interesting and its something I never considered would be possible the only issue is it relies completely on the position of the rotor blades to keep it stable so if a gust of wind causes the blades to adjust their angle the whole aircraft will follow ultimately I have to make a decision between are we building a strange tworotor test vehicle or are we trying to build a v Osprey scale model and from the title of this video I think you know the answer I really wanted to hold on to this freely hinged rotohead system but in reality its not feasible for this application the final rotor head was the closest to success but because its essentially a Thrust vectoring design this can be achieved with a Servo actuated motor pivot and a server actuated motor pivot has the advantage of being able to be locked in position during forward flight Im sure I could spend a few more weeks or months trying to tune this road to to perform as well as Matt from vertic has done with his but I also feel like thats following the sunk cost fallacy so I removed the hinge rotor heads and have gone with a simple Servo actuated motor pivot which because these will only tilt about degrees in either direction I can use a small Servo with a three to one gear ratio then there will eventually be a large Servo in the center of the frame to rotate the whole Wings bar so this isnt just far simpler electronically but its also lighter as it just requires these small servos to provide all the control for hover removing both the thicker Wings bar and one of the large tilt servos and when the aircraft transitions to forward flight these small servos will keep the motors in a fixed position essentially just acting as regular aircraft propellers I think its probably time we move on to the fuselage portion of this project which I gave a little teaser to in the last video the plan is to D print the fuselage using a material called lightweight PLA and some unique design methods to save weight this lightweight material Foams up when it is D printed and weighs as much as a regular D print but the issue with this material is the expansion caused by the foaming process leaves behind some residue if the printer nozzle is lifted away from the part to get around this the model needs to be designed so the printer can create the part without lifting the nozzle I briefly covered this in my previous vtop project and I created a short tutorial over on my second Channel but that was for just a simple Wing shape so how will it work for something more complicated like this fuselage I first started by creating a number of sketches in CAD to follow the profile of the real Osprey which took quite a bit of time to get right but when combined together ends up looking somewhat like the real thing I then created another fuselage that is slightly smaller which will be important in a second from this I can start cutting the reinforcement ribs that will help the fuselage keep its shape currently this would simulate a regular D print infill which fills the whole part but for this application its only needed around the outer skin so I can use a smaller fuselage to subtract all the inner ribs leaving us with the skeleton frame the final step is to remove these ribs from the solid fuselage model as this will trick the printer into printing the part without lifting the nozzle if we were to model the fuselage as a shell and add solid ribs inside the printer will have to lift the nozzle every time it creates a new rib but if the fuselage is a solid model and we cut the ribs out of it the printer will draw each layer without lifting the nozzle creating a clean lightweight part also this is a new version of lightweight pla from colorfab that is able to withstand higher temperatures as Ive previously had D printed Wings melt in the sun unfortunately this is the perfect color for the Osprey you may notice there is a combination of large and small ribs spanning the inside of the fuselage and thats because I ran a lot of test prints to figure out how I can keep these parts as light as possible whilst also maintaining a strong outer shell heres just some of the prototypes Ive printed whilst designing this thing like this one which has like this one which has a high density of very small ribs which holds the skin quite well theres also quite bendy and and flexible then theres this one which has a lower density of larger ribs which is quite a bit stronger but the ribs dont support the skin very well so it flexes between the ribs so this is the final design where its a combination of the two so theres a high density of thin ribs to maintain the strength of the skin and then a low density of larger ribs to create the general strength of the fuselage so after about four days of printing I now have all the parts ready to glue together which I used regular CA glue or sometimes referred to as super glue which was a bit of a pain as it dries quite fast the fuselage is designed so its glued together in three pieces the nose a midsection and the tail which will eventually be held together with magnets I then need to remove the wings bar and motors from the frame to install it in the fuselage which is done by sliding the carbon fiber skids into these rails printed into the design once installed I need to cut an opening in the print where the wing will be mounted because this hole couldnt be designed into the model as its required to print the part using vars mode now the wires can be plugged back in and the wingspar and Motors can be remounted to the frame with access at the front and rear to tighten the required bolts I can then slide the tail section onto the skids to check how it fits as well as attaching the nose and its really starting to look more like the full scale model all it needs is a few magnets to hold the parts in position I then printed the horizontal stabilizer and elevator using the same lightweight material and design methods to create the hinge for the elevator I chose to cut small strips of rubber sheet that can be glued into this preprinted slot I was tempted to D print an actual hinge pivot for the control surface but this rubber hinge is far easier to manufacture and is still strong yet flexible then the elevator Servo can be glued to the underside which will be hidden inside of the fuselage once mounted but first the Rudders must be glued onto the tips of the horizontal stabilizer and as you can probably guess these were also D printed and the whole tail assembly can now be be glued onto the fuselage now its time to cover up the motor mounts with some scale engine sales and these things are huge just like the real thing which is great for hiding all the messy wiring and they are fixed in position with these glued on brackets as well as some magnets to attach the two halves together leaving just the top of the motor and propeller sticking out also because the motor pivots independently to the engine to sell when hovering the thrust vectoring wont be tilting in the cells back and forth hopefully keeping the aircraft looking scale so before I build the wing its probably a good time to test if this thing will still hover with the added fuselage the propellers did vibrate a little to the start as the blades were folded slightly for transport but they soon straightened out and the aircraft lifted off the ground easily to be honest Im surprised at how well this thing flies its actually more stable with the fuselage attached which is probably because the added weight and inertia dampens the movements I havent done any tuning since adding the fuselage so it could be a little better but its definitely a good start and Im also glad I moved away from the older rotor heads as this system is far simpler and when hovering its almost impossible to tell that the rotors are tilting due to the stationary engineer cells the next step of this build is to add a tilt mechanism and Wing to the Osprey but first let me show you some interesting mechanisms you can build yourself from kiwico kiwico makes HandsOn projects for kids that are designed to be much more than just a toy as they help teach science and engineering to spark curiosity and creativity Ive always been a fan of making or doing things to learn and these kits are an excellent way to get going like this kit where you can build your own lock which requires an actual key to open everyone knows how to unlock a door but do you know how they work inside the design and quality of these kits are just amazing and theyre not only educational but also good fun and with everything you need being supplied within the box there is no need to buy anything extra which I think makes for an excellent Christmas gift and trust me completing a project that actually works is the best feeling and you can experience this too by going to the kiwico link in the description to get your first month with kiwico for free so definitely only go and check them out before building the wing we need a method of tilting the motors for transition as they are currently fixed in the vertical position so I removed the fixed Wings bar mount and replaced it with a D printed Mount that will hold the Tilt mechanism I chose to use a Servo operated gear system for this tilt as it needs a lot of torque to tilt both Motors simultaneously so these Bethel gears give the servo a three to one ratio where the servo rotates degrees to tilt the wing spark from horizontal to Vertical the server I used is a cheap one I found on Amazon thats designed for radio controlled cars and apparently can live kilograms at a one centimeter radius so about two newton meters of torque but with the gear ratio this is tripled to about newton meters for the aerofoil I chose to use a NACA because its profile is large enough to fit the wings bar and its capable of producing a high coefficient of lift before stalling as this aircraft is getting pretty heavy so we need as much lift as possible once D printed I can mount the aileron servo inside and glue all the wing sections together which I still havent improved my gluing skills but Im hoping to clean some of it off later to fix the wing to the fuselage I didnt want to use glue as I might need to access the Tilt mechanism and other electronics so I made this midsection that simply rests on top of the fuselage then when I slide the wing sections onto the spa it holds the midsection in place and the angle of the wing is supported by the top of the fuselage pressing against it I then remounted the engine the cells and checked if the Tilt mechanism was still working finally we can add the ailerons which work the same way as the elevator by using rubber stri

write a youtube script

there are many ways to extract energy from water whether it be a water wheel or a more optimized turbine but ive never seen someone try to run an air powered engine using water well ignoring steam engines this video is sponsored by nordvpn more on them later without the use of coke mentos we can increase air pressure inside of a tank using a pump to push more and more air into a fixed volume or we can use a fixed amount of air and reduce the volume of the tank an easy way to shrink the volume of air in a tank is by filling the tank with water but one small problem is the water must be at the same pressure as the air or the air will just push it out however because water is roughly times denser than air we can use this to our advantage if you swam down to a bottom of a pool youll probably have felt the water pressure on your ears and no matter how large or small the pool was youll feel the same pressure at the same depth this is because the static pressure of water is only relative to the depth and is essentially the weight of water above you pushing down and i plan to use this to create water pressure in my tank the plan is to put a tank of water in the loft of my house then run a tube downstairs to my tank full of air the vertical distance between these tanks is about . meters which means in theory when i let the water flow down the pipe the tank at the bottom should experience the pressure of being . meters under water which when we calculate the density of water multiplied by the acceleration due to gravity and the depth we should get around . bar or . psi of pressure then we can use this pressurized air to hopefully spin an air powered engine or is it now water powered so lets build this tank i could have opted for an easy choice of using a plastic drinks bottle as these are very cheap and easy to source and they can hold a lot of air pressure but i want something a bit more fancy made out of clear acrylic but as ive never made an acrylic pressure chamber before i need to do some calculations for example a millimeter acrylic tube will require two plates on the end to hold the air inside and at just . bar the force pushing out on these two end plates is about newtons or equivalent to supporting a kilogram weight so i used fusion cad software to design the acrylic pressure chamber and i chose to use millimeter thick sheets as the end plates after running a pressure simulation the thick end plates were probably a bit overkill but it allows me to cut this lip all the way around to give more surface area for gluing ive only ever cut plastic on my cnc router once before and it was the clear perspex type stuff you can buy at local stores for shed windows and im not sure if it was my cutting settings or the melting temperature of the plastic but it was a nightmare to cut so i crossed my fingers and attempted the acrylic lid which to my surprise cut really easily once finished it perfectly fit inside of the tube and it was now time to drill some holes and thread tap them for later adding some pneumatic fittings i then sanded down the gluing area on the tube and began mixing some epoxy resin which probably isnt the best glue option for acrylic but its what i had laying around and its pretty strong i also made sure to use plenty of it to fill all the gaps between the inner wall of the tube and the lip of the lid which got a bit messy but worked well in the end then we can fit a tube inlet on the bottom and a d printed stand to give enough clearance from the ground and on the top lid i can fit a pressure gauge and another tube outlet for the air supply to the engine the next step is to run a long tube all the way from the ground floor up meters into my loft where ill mount the water tank okay this is the first pressure test ive got five liters of water here and the container with the tube running downstairs gonna prop my phone up the problem also that i cant see if theres any leaks downstairs until i go down the stairs so im just gonna have to hope that this doesnt leak lets start pouring waters running down lets hope i dont just find a massive puddle downstairs so whilst we wait for the water to drain down let me tell you about the sponsor of this video nordvpn a vpn is great for internet security and privacy as it can hide your ip address as well as encrypt your data but nordvpn is more than just a vpn as they now offer thread protection which will guide your device from malicious websites malware and trackers even if youre not connected to a vpn but if you do want to use it as a vpn check out how easy this is to use lets say for example i want to watch shawshank redemption on netflix unfortunately it doesnt exist here on the uk netflix but i can simply open up nordvpn scroll over to the united states click here quick connect and as you can see its there at the top click it click play and you can watch the whole movie in full hd due to their high speed connection for the price of a coffee the advantages that nordvpn gives you is an amazing deal and for a limited time only viewers of this channel can sign up for two years and get an extra month for free by clicking the link in the description down below plus nordvpn offers a day moneyback guarantee so why not give it a go still going down i might need to go check out whats going on downstairs this is making me a bit nervous lets go check out whats happening downstairs please dont be a league please dont be a leak oh no hang on hang on hang on hang on hang on theres no leaks and weve got five bar right i suppose i should go back upstairs and keep pouring nice i think that might be the water level seems to be staying still now so lets go check what the pressure is downstairs the pressure downstairs was about . bar or about . psi which is very close to the . bar that i originally calculated but is this really enough pressure to run an air powered engine well if you havent seen my air powered engine videos before then youll have no idea what im talking about so lets have a quick recap in november of i d printed my first air engine that was based off the old air hogs engines and although it did run it was extremely inefficient so in april of i built a new engine that used a square piston and a cam and pushrod design which was more reliable than my first engine but still quite inefficient as it was very difficult to seal the piston to the cylinder without lots of friction then two and a half years later in november of i came up with a completely new design that removed the need for an airtight piston by using a diaphragm and a new inlet valve that removes the need for any push rod valves and after refining the design way too many times it outperformed the previous best air engine by a factor of five and could also run at pressures as low as . bar or about psi but will it work with my new pressure tank not quite and this reveals a problem that i hadnt previously considered not only do i need an engine to work at . bar but i also need to consider the flow rate of the water into the tank if i release all of the pressurized air the tank takes about a minute and seconds to fill back up to full pressure again and its too late to increase the diameter of the tubing as that would require new holes to be drilled for larger fittings so although id hoped that the engine would run at . bar which is not that much lower than the . bar that it previously worked at in reality it needs to run at much lower pressure so the first modification i made was to increase the diameter of the diaphragm within the engine as this will increase the area that the air can push against hopefully allowing lower pressures to apply more force to the piston and to my surprise it actually worked really well Applause at least down to about . bar when the engine would stop and i would have to wait for the pressure to build back up again the obvious next step would be to build an engine with an even larger diaphragm to run at lower pressure but theres a small problem because the design of the cylinder requires this cone section to allow for the diaphragm to expand it means the piston has very little support when it reaches the top of the stroke and making the diaphragm larger would mean increasing the unsupported distance for the piston which means the piston needs to be longer and so does the cylinder which will probably just increase friction also the crank case and crankshaft needs to be longer to avoid the propeller from hitting the cylinder so instead of completely redesigning the engine i decided to try tweaking a few things on the current engine the first being the ball valve that lets the air into the engine when d printing the cylinder head the printer often leaves small blobs of plastic as it moves on to the next layer and this means the oring used to seal the ball valve doesnt fit perfectly which isnt an issue at much higher pressures that i normally run the engines at due to the extra force pushing the oring into any small gaps but at these lower pressures if the ball doesnt perfectly sit on the oring itll leak enough air to drop the pressure in the tank by almost half a bar to fix this i used my soldering heat gun to soften the plastic inside of the cylinder head and then push the oring and ball into position to mould the plastic around it which achieved a much better seal i then made sure everything in the engine was running as smooth as possible by sanding down any tight spots in the cylinder and aligning the diaphragm as best as i could and the engine now sounds much smoother but its spinning way faster than it needs to be and running out of air so i took the piston out and shortened the small pin that opens the inlet valve which will hopefully reduce the amount of air that enters the engine per rotation almost like lowering the throttle this reduced the pressure at a much slower rate and the tank started to hold a constant pressure at just over . bar which made me incredibly excited no dont stop i thought we had a continuously running one then so i emptied the tank carried the water back upstairs and crossed my fingers for another run no i then realized that clicking noise you can hear was actually a huge amount of play in the crankshaft which i think throws the piston off center when its unsupported by the cylinder so i remove this plate with a small spacer mounted the engine on a solid stand and attempted another full tank run the engine was off to a smooth start and took around seconds to stabilize the tank pressure at around . bar then its about minutes and seconds into the run the engine suddenly dropped in rpm as the pressure reduced but somehow the lower rpm allowed the tank to fill back up to pressure and the engine rpm increased again then at minutes and seconds into the run the engine nearly came to a complete stop multiple times before switching direction which again allowed the pressure in the tank to rise at this point there is so little air left in the tank that each breath of air that the engine takes will reduce the pressure by a huge amount so it struggles to keep the pressure constant and as the water reached the top of the tank i cut it off early to prevent the engine from drowning itself Music we did it we actually did it oh no dont get water in the engine all them i cant believe that actually worked not only was that the lowest pressure any of my engines have run at but it was also the longest run time ever had out of an engine with such a small amount of air with a total run time of four minutes and seconds running just on the pressure created from a water tank in my roof i hope you found this interesting and uh ill see you the next one thanks for watching

write a youtube script

this is a watt drone motor it weighs just grams and this is a watt ebike motor it weighs kilograms but is it possible to drive an ebike with a motor this small this video is sponsored by onshape more on them later now when i say these are both watt motors im using the data provided by the motor manufacturers so to be precise the drone motor is rated to watts and the ebike motor is probably rounded to the nearest kilowatts but theres more to the specification than just the power for example this tractor has the same horsepower as a nissan gtr but the nissan has a top speed of miles per hour and the tractor can barely reach and this is very similar to the motors we are comparing the ebike has a top speed of miles per hour and has a wheel diameter of inches this means that at top speed the ebike motor spins at about rpm but the drone motor can spin up to nearly rpm this causes a bit of an issue because if we were to spin a inch bike wheel at rpm the bike would be going faster than the sr blackbird so we need to gear it down a little to spin the wheel at a more reasonable speed an tooth bike sprocket has a diameter of roughly centimeters and to get a ratio of to which is what we roughly need to convert the rpm down to rpm means we would need a rear sprocket with a diameter of roughly meters which obviously isnt possible but instead we can use a series of belts and pulleys to bring the rpm down on the motor side before attaching the chain for this specific application im going to use a to pulley ratio straight off of the motor and then another to ratio to the output shaft this means the output sprocket will spin just under rpm and because my previous super capacitor bike project required a large rear sprocket too when combined with the pulleys we should have a final wheel rpm of around which matches the ebike perfectly so lets start designing these pulleys using onshape which is a cloud native cad system if youre familiar with using cad softwares already onshape is very intuitive to use with its labeled tool menu and clear parts list and also includes industry leading manufacturing specific features for sheet metal frame based design configurations and detailed drawings but unlike many other cad softwares onshape works in a browser and is accessible across all operating systems as well as ios or android devices and because its cloudbased you no longer need an expensive computer to run it and all your work is backed up on the cloud so you wont lose anything if your computer freezes or crashes it also allows you to collaborate with teammates or vendors on the same project at the same time which is why engineering teams at companies like formlabs dhl and trek choose to use onshape and if your company uses a cad system i highly recommend you try onshape at onshape.pro forward slash tomstanton or click the link in the description down below the next step is to make the mounting plate for the motor and pulleys which i cut from aluminium on my cnc router this could then be clamped to the bike frame using a couple of d printed brackets and we can then start mounting the components starting with the output pulley which ive attached to sprocket to using a few bolts this can then be attached to the main mount with the shoulder bolts which is essentially a precision shaft bolt that the bearings will spin on i can then fit the bike chain and tension it accordingly by sliding the mount along the frame so with the belt fitted the center pulley can then be attached in a very similar way and can be tensioned using a small idler bearing in this slot and finally we can mount the drone motor and see how well it turns how does that look with the hardware complete its time to fit the electronics thatll power the motor as well as a small arduino board to convert the analog throttle signal into a digital signal that the drone motor controller is designed for perfect man look how good wiring that is here we go Music that has some power though like i can feel the wheel theres some talk in there first test hey look at that Music just moaning oh dear oh the chains coming off just look at it its ridiculous how small that little motor is so baby oh no oh dead i can smell it it doesnt smell good i think we need a slightly bigger motor oh yeah that is yep you cook an egg on that smells but it smells really bad doesnt it so we obviously need a new motor with some more torque but also one that will fit the current motor mount so i found this motor which is also designed for drones and is only slightly larger than the previous motor however the difference is that it only spins to rpm instead of rpm in theory with half the amount of max rpm the motor should output double the amount of torque for the same power now i say in theory because the thousand watt ratings on these motors are actually for the electrical power and not the actual physical output power so the only way to find out is to give it a test i also move the motor and pulleys to a different part of the bike to shorten the chain as well as stopping it from rubbing on the frame which should hopefully reduce the chance of it coming off its plagiary i think thats quite a bit more power now lets see how this goes good luck sir thank you oh shes going for it Music but the problem is the chains i need to tighten the chain lets have a look its really loose because the chains so loose it hits his pulley you see there it hits the pulley so its actually like worn if you look at the belt here it seems like one right there you can see it just focus on obviously didnt design it to be mounted here you need one of those like little little bobbins there yeah hold it up just sit up there like that i reckon tension the chain a bit more though yeah be right for another run oh my god that was top speed that was it looked quick it was uh my watch was saying just over miles an hour really its pretty good thats really good just for a little baby do i have a go maybe warm is it warm uh yeah i just about can hold it without or degrees or degrees or so yeah i mean thats still safe temperature for a motor yeah right pedal to get going uh you can use the motor you have to pedal just do a little bit oh this is crazy thats fantastic its all right isnt it loads more talk than i expected i thought it would take you ages to get up to speed but just shoot straight off thats actually really impressive is it hot oh yeah thats warming is it really hot uh its pretty warm dont grab it like i did yeah so would you uh ever buy a dronepowered ebike maybe so im sure youre all wondering how does this compare to the ebike motor well lets see with the drag race Music oh thats really hot ow oh wait ive i think i might have melted the coils yeah thats the absolute limit of that motor and the hub motor is feels cold same temperature as the bike frame absolutely freezing so to answer the question yes you can power an ebike with a tiny drone motor but would i recommend it probably not thanks for watching

write a youtube script

the v osprey is one of my favorite aircraft due to its capability of hovering like a helicopter and then transitioning into ford flight as a fixedwing plane allowing it to land almost anywhere in fact ive built a number of vertical takeoff and landing aircraft over the past years that resemble the osprey with the distinct wingtip motors that rotate as the aircraft transitions but its been a while since my last veto project so in this video im going to take you on a journey of how i design and build each and every component of a radio controlled veto aircraft ive chosen to manufacture the fuselage out of glass reinforced plastic as its lightweight and quite durable in the event of a hard landing or crash only issue is these sheets are only one millimeter thick but with the addition of some d printed parts the rigidity can be greatly increased the rear section of the fuselage will be just a straight carbon fiber tube as theres no need to mount anything rear of the wing apart from the rudder and elevator which makes the aircraft look more like a model helicopter than a plane the next step is to add the main wing spas which use the same carbon fiber tubes as the tail and these are clamped together with some d printed components and some more fiberglass sheets for added strength which will give the wing an upwards dihedral angle of degrees then the wing spark can slot into the fuselage and be fixed in position with a few bolts making a strong yet lightweight frame this is important for this style of veto aircraft as when it is hovering the full weight of the plane is held up by the wingtips unlike a normal plane where the wing has a distributed amount of lift along its span speaking of lifting from the wingtips lets move on to the motor mounts in the past ive experimented with several different mechanisms to tilt the motors such as my first dual servo setup where one large servo would tilt for transition and a smaller servo would tilt for hover control which worked pretty well but was also quite heavy so on my second veto i literally bolted the motor straight to a single servo which did both the transition tilt and hover control my third veto also used a single servo but transferred the rotation through a pair of gears to take a bit of load off the servo bearings and finally my fourth veto which used a single servo to tilt both motors via a shaft through the wing which was quite lightweight but wasnt ideal as the motors couldnt tilt independently so control during a hover was limited because this aircraft only has two motors its like trying to balance on a twolegged stall so for the aircraft to control the pitch axis in a hover it must tilt these motors back and forth this works by creating an offset between the thrust line and the center of mass causing the aircraft to rotate and the faster this motor can rotate means the more responsive the pitch axis will be and therefore once gyro controlled the aircraft will be more stable so this tilt mechanism is designed to be as lightweight as possible with the servo fixed to the wingspart to reduce the mass of the rotating portion hopefully making it as responsive as mechanically possible so i d printed all the tilt mechanism components and glued together some of the parts that couldnt be printed as one piece then the motor can be bolted in position as well as the servo which i can attach a large d printed gear to for tilt control these two halves then slot together and can be slid onto the wingspa clamping the servo portion of the mount with a bolt to stop it twisting and there we have it a lightweight and simple motor tilt mechanism and the motor and servo wires can be neatly passed through the wingspa now as you can probably imagine the speed at which the motors and tilt mechanisms need to operate at to keep the plane stable during a hover needs to be pretty fast so in the same way that the drone requires gyros and accelerometers to compute the motor speeds this aircraft needs them too usually these gyros and accelerometers are packaged in something called a flight controller which takes all the data from the sensors and combines it with a transmitted control signal to tell the drone to move in a certain direction but unfortunately there arent any offtheshelf options for this type of vertical takeoff and landing aircraft so were going to need something a bit more custom fortunately a veto capable flight controller has been developed by a guy called nicholas reem as part of his masters research and hes posted a bunch of videos on youtube of all of the various projects hes used the flight controller in how his flight controller works is it uses a small processor board called a teensy which is combined with a gyro and accelerometer board for sensing the orientation of the aircraft these two electronic boards are all that is needed to build the flight controller and it runs its custom code called dreamflight now ive done a bunch of arduino coding in the past and fortunately his flight controller also runs arduino code so i found it very intuitive to use and honestly i am very impressed because all the code is well laid out and easy to read it almost feels like the possibilities of aircraft configurations are endless so im definitely excited to use his code in this project and ill most likely be using it in many future projects to come to simplify the wiring of the flight controller i mounted the components to a prototype board which made it easier to connect all the tinsey outputs to connectors for the servos and motors and with a few lines of test code the gyro stabilization was working well the next step is to solder in the electronic speed controller for the motors which ive chosen to use a in speed controller these are designed to be used for drones as they have four motor outputs now my aircraft only has two motor outputs but surprisingly its more compact than having two separate motor controllers then all the signal wires can be connected to the flight controller as well as a receiver and a five volt regulator to power the servos and it was time to write some code this is the main mixing code i wrote for this specific aircraft and if youre interested in learning how it works its very similar to the mixing code i wrote for my previous vtol which ill post a link to in the description down below once all the motor and server directions were configured i mounted some propellers to check how well it corrected for any external movements which i dont recommend trying as its not exactly safe but it looks promising time for a first test hover thats true it actually flies first try and its about to tip down with rain so i better go inside whilst i waited for the rain to clear i had another read through the code in the flight controller and noticed i had it set to the wrong flight mode so i changed the flight mode and fortunately the sun came out for the next test and this time it was far more stable and felt so natural to fly that i ended up doing some pirouettes and some figure of eight and then landed before i was able to crash it now all we need to do is add a wing and tail to this thing and see how it flies as a plane as you can imagine the wing on a veto aircraft is very important as unlike a conventional plane which uses a runway to take off and land a veto aircraft has to do it all whilst already in the air this means it has to accelerate from zero airspeed up to cruise speed as smooth as possible and also decelerate from crew speed down to a hover as smooth as possible fortunately there is a website called airfoiltools.com which lists loads of different error foils as well as a lift and drag data which helps me decide which one i need after lots of comparing aerofoils ive decided to use a because its coefficient of lift versus angle of attack graph is very linear between negative and positive degrees angle of attack and because im going to mount the wing with a positive angle of degrees the aircraft will be able to pitch up or down by six degrees in either direction without producing negative lift or stalling as an example here is a less desirable wing where the coefficient of lift drops off rapidly below zero degrees angle of attack so now the aerofoil shape has been chosen whats the best way to build the wing well in the past i have d printed wing ribs and folded foam board around them to create a rough error full shape but this isnt usually that accurate because the foam board starts to deform around the d printed ribs so instead im going to d print the whole wing out of a material called lightweight pla this material foams up when it is d printed and weighs as much as a regular d print the only issue with this material is the expansion caused by the foaming makes it very difficult for the printer to create nice prints let me show you an example here is an airfoil printing with zero internal structure and as you can see the printer just follows a continuous outline of the aerofoil profile however say that we want to add an internal rib to increase its strength the printer has to stop and lift the nozzle to move to the internal rib this leaves behind some excess material which has pushed out the nozzle by the foaming process and is usually fixed by pulling the plastic back out of the printer to stop it extruding but this doesnt work with the lightweight pla to stop this the wing needs to be designed to allow the printer to draw a continuous line throughout the whole printing process this means to add an internal rib to the wing we actually need to cut a slot like this this makes the rib part of the external profile and removes the need for the printer to lift the nozzle which not only removes any extra material left behind by the retraction process but also reduces the printing time now of course a single rib is not strong enough to support the whole wing so ive spent a lot of time configuring the design to produce this where the internal ribs produce a grid pattern and each rib has a hole down the center to reduce weight and this whole wing can be produced as one continuous line by the printer resulting in a fast printing lightweight wing once all the sections of the wing were printed i started gluing them all together using the wingspa to line them up each wing half is centimeters in length and weighs about grams each which isnt too bad considering its a very precise aerofoil or at least a lot better than i could make out of other materials these wings can then slide onto the wing spas and then lock into the fuselage like so making for a very modular design so if the wings were to get damaged and needed replacing i can print a whole new wing in a day or two and just swap them out you may notice this wing doesnt have any control surfaces or flaps and this is because im hoping to control the aircraft using a few different mixing methods so lets go over how this aircraft will control itself in both a hover and forward flight starting with the pitch axis as previously mentioned the pitch control during a hover works by tilting the motors back and forth and this worked well from initial test flights but youll also notice the rear elevator is moving too this wont help pitch the aircraft when hovering in zero wind but if there is a slight headwind or the aircraft is moving slowly forward it will help assist the motor tilts for roll control the motors work in the same way as they would on a conventional drone where one speeds up and the other slows down which is very responsive thanks to the technology in drone motor controllers for the yaw control the motor tilts are used again but this time in opposite directions producing a forward force on one wing tip and a backwards force on the other and as you can imagine combining all these movements simultaneously is the reason why the flight controller is required when transitioning the control movements fade into forward flight mode where the only thing that doesnt change is the rear elevator movement so to pitch the aircraft the motors are now locked in position to only produce thrust but the rear elevator continues to work like it would on a conventional plane but unlike a conventional plane the roll control will actually use the motor tilt instead of ailerons on the wings this is an unusual method to control the role of an aircraft but in theory it should work and might actually have some benefits as the roll control isnt relative to the oncoming airflow like it would be if it had ailerons potentially meaning the aircraft will still have full roll control during a stall then for your control in forward flight the aircraft will use differential thrust to push one wingtip forward and the other wing back as there isnt a moving control surface on the rudder this means in theory we can control all axes of the aircraft in both hover and forward flight with only five moving parts including the motors keeping the aircraft mechanically simple and lightweight i then built a basic tail section out of some foam board as it doesnt require a specific airfoil for the speeds at which this aircraft will operate at and mounted the elevator servo inside of the rudder this can then slide onto the tail boom and were ready for another test it was very windy on this day but i quickly noticed that one tilt mechanism was shaking far more than the other which was concerning as the electronics should move each servo at the same rate however i was desperate to test fly it so took it out to the field anyway despite being thrown around in the wind i was able to keep control of it whilst cruising around in hover mode then i tried to transition into forward flight and quickly realized i had almost zero roll control so quickly transition back to hover mode and landed it safely when i returned home i realized one of the tilt servos wasnt moving as fast as the other and was getting quite warm the reason for this was my mistake as when i assembled the tilt pivot i didnt realize how much resistance it had which puts a lot of unwanted load on the servo to make sure this doesnt happen again i replaced the servos with much larger and faster servos and freed up any resistance in the mechanism so now even the weight of the motor is enough to move the pivot down also because of the higher torque of these new servos i can run a different gear ratio which allows the motors to tilt further and therefore hopefully give me more roll control during forward flight so another quick test hover in the garden was successful and both servos were moving at the same rate though the aircraft did make a pretty bad vibrating noise which was due to excessive play in the tilt pivot however this can be easily fixed with a small clamp on the wingspar to stop it moving little did i know there were other larger issues about to show up but we can deal with them later lets take it for a proper test flight Music flicking the transition switch on my controller was almost seamless as the aircraft accelerated away from me and with the new servos the roll control worked really well they honestly flew like a regular fourchannel radio controlled aircraft but one thing i hadnt anticipated is the low drag of this aircraft makes it very difficult to slow down for a hover fortunately in hover mode i can just fly it back towards me like a helicopter the next outbound transition was a little sketchy as i have the transition speed set to rotate the motors in about three seconds which is too fast and causes the nose to pitch down and also i need to practice the inbound transitions as pitching the nose up to bleed off speed results in a rapid loss of lift when the wing stalls so whats going on with the ripples in the wings the lightweight pla that the wings are printed from starts to melt at about degrees celsius and is d printed at about degrees celsius however pla can start to warp at just degrees now i didnt think it would warp so quickly especially here in the uk but it probably doesnt help that its a dark matte color the easiest solution to this is to print the wing out of a white material but say i left it in my car on a hot day it could also cause the wing to melt when not in direct sunlight so i need to redesign the wing this new wing has t

write a youtube script

an electric motor uses a combination of coils and magnets to produce mechanical movement using electrical energy but what if instead of mounting these coils in a circle to rotate a shaft we mount them in a linear line to launch some magnets this concept isnt a new idea and has been used on a number of roller coasters to accelerate passengers to high speed often referred to as linear synchronous motors and in fact quite recently a form of linear motor has been implemented onto an aircraft carrier instead of the conventional steampowered catapults now im not entirely sure on the specifics of these launch systems so lets run through how i plan to build mine if we position a magnet near a coil of wire and apply a current to the coil in the correct direction it will create a magnetic field and launch the permanent magnet forwards at least temporarily as the magnet will be attracted to the center of the coil to prevent the magnet from bouncing back the power to the coil must be disconnected at the exact point the magnet passes the center of the coil this needs to happen in a split second which obviously cant be controlled manually so we need some electronics to help for my design im going to use a magnetic sensor called a hall effect sensor this outputs either a one or a zero depending on the nearby magnetic field so currently it would output a zero then when the magnet gets close it will output a one which all gets sent to an external control board this information can then be used to control a mosfet which is essentially an electronic switch that can turn on and off really fast by applying a small voltage through the blue wire this applies current to the coil and the magnet is pulled to the right then as the magnet gets close to the center of the coil the sensor tells the control board to turn off the mosfet and the magnet is free to continue moving towards the right Music so lets build some electromagnets by winding this wire into a coil and by applying some current through the wire we can create a magnetic field though in its current form its not a great electromagnet but by simply adding some ferromagnetic metal to its core in this case just a hardened steel bolt we can greatly increase its magnetic strength then we can use a d printed bracket to mount it to an aluminium extrusion which will act as a guide rail for whatever we choose to launch here is the hall effect sensor that detects a magnetic sled and sends a signal to the control board this will then control the mosfet on the other side of the rail which is hidden underneath this custom aluminium heatsink to keep it cool when dealing with high current because when this thing launches its basically shorting a bank of super capacitors through the coils which can exceed over amps of current per coil pair and considering these coils only use . millimeter diameter wire anything over about amps of current can get pretty toasty so lets arm the control board and see if it launches fortunately because the sled clears the coils in a split second the coils dont have to endure high current for very long but i still programmed in a safety time limit to put the launcher into a standby mode in case the sled gets stuck keeping the coils turned on and well melting them i guess now unlike a slingshot i cant just increase the distance between the electromagnets and the sled as the magnetic field greatly decreases with more distance therefore to get this sled moving faster we need to increase the number of coils along the guide rail so the sled has always been pulled forward by a nearby magnetic field so lets take this working launcher and multiply it by say which means i have a long week ahead of winding coils which i was able to average about six per hour then cnc cutting heatsinks to keep the mosfets cool and lots and lots of soldering which as you can see turned out to be a bit of a wiring nightmare fun fact though with over turns of wire per electromagnet theres coils on this thing which was well worth the effort each of these coil pairs works exactly the same as the first cool pair thats demonstrated earlier there is a magnetic sensor in between each coil which detects a sled and tells the control board inside of here to apply power from this supercapacitor bank so to launch the sled all i need to do is arm the system and press the green button but this wouldnt be a proper youtube video without launching some household products his eyes are just gonna hit the right are you ready insta cook it in the candle three two one Music two one thats gonna be good that is gonna be good but unlike lots of other magnetic launchers here on youtube having an open guide rail allows for some more interesting launch items what are you making sam concord mark ii concord mark ii put it on the launcher Applause oh wow look at the wings wow that picked up really quick that looks so good Music Music theres something rather satisfying seeing a sled accelerate down a guide rail with nothing touching it other than the friction of the rail all propelled through magnetic attraction but aside from this probably being the most complex and expensive paper plane launcher lets talk about its performance it accelerates the gram magnetic sled to about meters per second which might not seem very fast but it achieves this in just over a tenth of a second thats equivalent to about gs of acceleration which is quite a lot for a paper plane my next plan is to experiment with different magnet configurations so not only do the coils attract the sled but theyll also repel it hopefully doubling the acceleration and increasing its speed so stay tuned for the next video also have you ever realized my youtube banner image is actually computer generated thats right this image isnt a real photo and it wouldnt have been possible without a high spec computer or lots of time which brings me onto this video sponsor micro center micro center stock all the parts youll need to build your own pc okay you might not need to create a d generated youtube banner like i did but even running most cad softwares can be quite demanding and once youve finished designing a part micro center can also hook you up with a d printer from large professional grade d printers to small beginnerfriendly models so you can pick one that matches your needs and stop by their shop and start printing the same day they also have a large selection of arduino related electronics which if you watch my videos youll know that i love to use arduino code to control stuff such as the exact project youve just watched computers for designing stuff arduino electronics and d printers are basically what all my projects are based around so head over to microcenter to find the parts you need via the link in the description as micro center wants to offer new customers a free gigabyte usb flash drive and microsd card thank you to microcenter for sponsoring this video and thank you very much for watching this video thanks to all of my supporters over on patreon.com for making these projects possible and its probably not a good idea to rub your head on the on the trigger like that well well not that end either anyway our see you in the next video

write a youtube script

Applause syringes can be used in many types of scientific projects what I find interesting is if you seal off the end cap and pull the piston back you can create a vacuum inside of the syringe now with some water inside and some lubrication you get a pretty strong springing effect as the piston shoots back up to the top now the other day I realized that if you mount the top part of the syringe on something which would allow it to move freely upwards pull the piston down and release you get a pretty strong effect launching these syringe straight up into the air and possibly making a dent in the roof but how about we expand on this and try it with a slightly larger syringe so to test this syringe Im going to have to launch it outside because I dont really fancy putting a hole in the roof of my shed and also be interesting to see how high this thing goes so Im going to have to make some kind of launch stand for it because I cant easily take my voice outside okay so Ive got the large syringe set up in its launch pad its essentially just a block of wood to hold this string in tension and a small peg below two screws for the release mechanism now Im got the highspeed camera running over there so Im going to have to pull the release from over there run across over there and hopefully this launches up into the air okay large syringe rocket test in three two one that didnt work as expected there okay so it seemed as though the speed of that piston moving upwards has just taken a whole top off of that sory landed there this okay so Ive now put millimeters of air inside of the piston before pulling it back so its not going to have as much of a vacuum as previously and but hopefully thatll prevent the top of the syringe from exploding in three two one oh maybe I had too much air in there before I continue with the testing lets look at how these syringes launch into the air there are three main things occurring which cause the syringe to launch the first is the atmospheric pressure pushing upwards on the piston due to the vacuum inside of the cylinder this tiny milliliter syringe experiences . newtons of upwards force or the equivalent to one point eight kilograms required to pull the piston down the second is a conversion of this potential energy into kinetic energy as the piston accelerates to the top of the syringe reaching a top speed of about meters per second or forty four point eight miles per hour in under six thousandths of a second and finally the momentum transfer from the piston to the syringe as the piston and syringe are almost exactly . grams each the final velocity of the piston and syringe combined is halved and launches the whole thing skyward at meters per second to maximize this end velocity after the momentum transfer we need a low mass syringe and a high mass piston however too heavy of a piston will reduce the speed at which it hits the top of the cylinder so there is a balance to be struck right so Ive now got a gram piston and Ive removed all the air in the hope that with the weighted piston theres less kinetic energy hitting the top of the I suppose is the same kinetic energy just lower velocity in threetwoone while the weighted piston launch the syringe a bit higher its still not very impressive I think in order to improve on this syringe rocket we need something slightly larger something like this to turn this acrylic tube into a giant syringe I had to d print a bulkhead that will hold an oring at one end this will be secured in position using three bolts threaded through the acrylic tube but before installing the bulkhead I need to install the piston which was a dense d printed part with mounting space for LED tie weights in case it wasnt heavy enough once the top bulkhead was bolted in position the piston can be pulled back to create a vacuum and we now have a milliliter syringe or nearly pint i then added some fins and printed a mountain to hold the retracted piston before launch okay so this rocket is now ready for a test theres actually a slight leak in the oring system when I pull it back so what I need to do is I need to load it up really quick take it outside and then launch it as quick as possible because the more air that leaks into the vacuum obviously the less force that will push the piston upwards so what Ive got to do is Ive got to so we need to hold this like this on the launch pad put it next to my vise and get this funny handle Ive made out of a peg to grip onto this string which attached to the piston and then I need to very carefully pull this back as soon as it gets all the way back I need to put this pin through the rope that holds a piston then before only I need to quickly run outside and give it a test here we go and Oh yep that was terribly disappointing the piston accelerated upwards as expected but stopped short to the bulkhead due to air leaking into the cylinder before launch this meant the piston couldnt gather anywhere near the required speed for this thing to launch to a slightly more impressive altitude so I decided to launch it from a table with the idea that I could pull the piston down and instantly release it reducing the amount of air entering the cylinder but it still wasnt great the piston reached an upwards velocity of meters per second and due to the high mass of the acrylic tube the end velocity of the piston and cylinder was just four point eight meters per second it seems like there are multiple factors that need to be fixed one the air leaking around the orings is reducing the strength of the vacuum to the friction of the orings sliding up the cylinder is preventing the piston from reaching a high end velocity and three the weight of the acrylic tube is just too high making for a low end velocity after the momentum transfer lets move on to plan B now this is known as a vacuum cannon you may have seen videos of people previously launching pingpong balls at really high speeds using a tube and pulling a vacuum inside of that tube now essentially how it works is the top here is sealed off with just a basic packing tape and there is a projectile in here this blue part here and there is a piston now by pulling on this rope I can pull the piston downwards and instead of holding the piston at the bottom and releasing it launching the tube upwards I actually pull the piston out the bottom of the tube now because this whole cylinder is a vacuum inside the air from the atmospheric pressure rushes in and pushes his blue projectile up a really high speed breaking through the packing tape and hopefully launching into the sky so am lets see if this actually works okay here we go vacuum cannon test number one in three two one well that went high not majorly high okay vacuum cannon test number two in three two one so this vacuum cannon essentially eliminates all of the previous issues there are now two orings on the piston to reduce any air leaking in the projectile doesnt require an oring so the friction is reduced and there isnt any momentum transfer resulting in a projectile velocity of meters per second or nearly miles per hour Applause now while this cannon may sound impressive the pop noise is nothing but the packing tape bursting and there is no difference in projectile speed from this to this and considering it exits the tube at miles per hour it doesnt go very high due to a lack of stability I want to explore this vacuum cannon system further in the future but its getting a bit offtrack from the syringe rockets so there is one more thing I want to build and if you own a d printer youll be able to build one too but first Id like to mention the sponsor for this project Skillshare school share is an online learning community with thousands of classes covering dozens of creative and entrepreneurial skills a Premium Membership gives you unlimited access so you can join the classes and communities that are just right for you I highly recommend the advanced aerial videography class as it runs through a short film shoot from start to end and in the same way a regular camera is so much more than just getting a single shot the drone can be used to capture all types of unique shots Skillshare is also incredibly affordable with an annual subscription working out at less than a month Skua is sponsoring this project by offering a month free trial if you sign up via the link in the description below so if you want to check out more about skill share and help support my projects go check out the link in the description down below so this is a d printed syringe rocket launcher I know this project is on a full circle back to the small ten milliliter syringes but they just work so well with the low friction piston combined with the mass of the piston being very close to the mass of the cylinder they just they they launch so well and by pressing down this small tab they they launch quite violently too now if you want to build this syringe rocket launcher at home the download files for it will be down in the description below so if you have a d print to yourself or you know of a d printing service nearby you can build your own syringe rocket launcher now Ive tried different types of syringe brands this is a slightly more expensive version because it has a screw on cap and these worked really well theyre really easy to set up but this is also a cheaper version and these also fit this launcher and work quite well what I do is I suck up some water into the syringe which helps to get rid of any air bubbles inside of the syringe as well as acting as some lubrication I then use a hot glue gun to inject hot glue into the nozzle of the syringe and pull the piston back slightly to suck the hot glue in this creates a pretty good seal and it works perfect for these Rockets as well as the hot glue on the end acting like a bit of a buffer if you hit anything hard it sort of protects the nozzle end of the syringe to load the syringe onto the launcher you just place a syringe at the top and the handles of the fringe should fit there and then you pull the plunger down hook the handle into the tab and then you just press this button to release now obviously safety precautions dont load this thing with your head above it these things do go quite fast and most syringes have a pointy end on them so obviously be careful with that but if you enjoyed this project itd be great if you could leave a thumbs up down below if youre new to my channel and want to see more odd projects similar to this then please click Subscribe down below and also a massive massive thank you to all of my supporters over on patreon comm I honestly could have put the time and effort into these projects without your support so thanks once again for that thanks again for watching and Ill see in the next video goodbye

write a youtube script

in the early s British engineer by the name Barnes Wallis was experimenting with skipping marbles across a water tank in his garden this eventually led to one of the most famous engineering marvels of world war ii where a barrel shaped explosive was skipped across the surface of a lake to destroy enemy dams commonly known as the dam Busters now instead of this video being a historical documentary why dont we try to replicate this on a slightly smaller scale this video sponsored by Kiwi Co more on that later a d printed this prototype barrel with the plan that the slot will contain a CNC cut aluminium disc this would allow the weight of the bow to be adjusted as well as its rotational inertia when bonds Wallis tested the barrels he discovered that adding backspin was essential to getting him to skip the spinning created gyroscopic stability and kept the barrels pointing in the correct orientation but I soon decided I didnt really want to see and see cut a bunch of aluminium discs that would most likely end up at the bottom of a lake so I decided to print this barrel with threaded countersunk holes this way I can use large bolts to increase the mass of the barrel which saves a lot of time and cost speaking of time this took and a half hours to d print and Im going to need a few of them so the final design will be cut from this PVC pipe and a quick tip to make a quick and easy straight cut through a tube wrap packing tape around until the ends meet without creasing the tape then work your way around using a hacksaw and the cuts should be pretty straight then I can use these d printed end caps to mount the bolts too as well as a center axle to spin the barrel youll probably notice from this test footage the barrel looks more like a sphere than the cylinder so Ive d printed these curved sections that clamp around the cylindrical barrel however the actual spherical barrel is proved difficult to manufacture at the time and therefore these straight cylindrical barrels were eventually used in the raid but Id still like to test both now we need to make the release mechanism the actual release mechanism Barnes Wallis designed sandwiched a barrel between two arms which would pivot outwards slightly and drop the barrel but before release it was spun up using an electric motor via a belt and pulley system shadow types are Brian from the YouTube channel real engineering for letting me use his animation so a d printed a replica mount to hold the barrel but instead of using a belt and pulley system to spin the barrel I use a small brush this motor directly on the barrel axle for simplicity so one arm contains a brushless motor and that other contains a bearing each with a cone that meets well with the barrel but in order to keep the barrel in position the arms needs to be held in words and this is done using a string system and a short piece of elastic attached to a servo this causes the arms to be sprung loaded and creates a tight clamping force against the sides of the barrel then when it is ready to release the servo simply lets go of the elastic so the release mechanism works well but theres a slight problem the barrel sink now because I dont want to be littering these barrels all over the bottom of a lake and I also dont want to have to make loads of them for testing I need them to float I then realized a really simple fix is to get a ziploc bag and fill it with air then once sealed I can put it inside of the barrel and the barrel will float no matter how much water leaks in so now we need something to drop it from so Ive teamed up to wave at James from Project air the YouTube channel where he builds things like this this is a large scale Lancaster bomber and you built this out of foam board right yep its all homemade bit of a buzz job recycle some electronics in there now were going to throw it over a lake and see if Tom sir bouncing barrel works lets just do a controls check quickly thats left thats right thats up thats down that goes that way that goes out yeah all right chaps yep as ready as well ever be right oh god its got a lot of up oh wow that looks amazing I think the battery cables come out its flying though look at that how she looks really good yeah so this is a lot of foam and a lot of look how slow its going as well its got theres a lot of motors in the air at one time no idea what the stall speed is oh yes nice one yeah oh my god the circle nice format we decided it would be best to test the curve balls first as they might be a bit more forgiving in terms of skipping across the water so with all the cameras rolling and the surprisingly large crowd watching we were somewhat ready Im ready okay unsurprisingly Jamess awesome homemade lancaster handled the weight of the barrel and its release mechanism with each good yep okay ready Tom okay down coming down then we just pause right here Barnes Wallis calculated that in order to skip the barrel successfully across the water the planes we need to fly at meters per second ground speed and drop the barrels from an altitude of meters this would result in the boughs hitting the water at about a degree angle some quick calculations estimate that this plane is traveling at . meters per second at an altitude of . meters which should result in the barrel hitting the water at degrees now this may seem like a rookie mistake but Barnes Wallis actually made the same mistake when initially testing his barrels it was in time to land the plane which wasnt exactly easy seeing as there was a wall and a lake on one side and a very steep drop on the other so its a little damaged isnt it James yes but dont worry so we fixed the plane and got it back in the air this time around my mechanism failed to spin the barrel up fast enough and I told James to fly back around but before he was able to do another pass the barrel came loose from the mechanism completely my fault and very frustrating as the landings werent easy so every flight counted fortunately James landed it beautifully and it was all good friend of the flight yep this time with the straight cylindrical barrel just like what was used in the actual rate Im getting nervous again scary all this work in the air I want the video to be good so I hope youre enjoying this yep good okay James flew the plane round with a long runup to allow the barrel to spin up slowly lower the altitude was also gaining speed and I did that was great yes lets take a look at that again from the high speed the plane was traveling at an estimated fifteen point six meters per second and dropped the barrel from an altitude of about one point three four meters which was really impressive flying from James causing the barrel to impact the water at an angle of degrees which is quite a bit steeper than Barnes Wallis originally aimed for but to still managed an initial skip distance of five point five meters with a total of three skips spanning eight point four five meters however it wasnt over there the mission still had to be completed which was made significantly harder by this rogue propeller this planes done its mission and then it just it didnt make it home boys like I was so nervous it was very low actually three times I think yeah quite a few times it was then time to collect the barrels with Jamess homemade hydrofoil boat whilst I made some modifications so the barrel dropped mechanism its amount of the underside of a large drone I just built the week before for future heavy lift projects how would you feel about flying it Tom whilst Ill show some of the drone tests Id like to mention the sponsor for this video Kiwi Co Kiwi Co delivers monthly projects designed to make learning about science arts math and more fun and accessible they have seven different crates for kids aged zero to sixteen plus and they sent me one of their tinkerer craze which is aimed at ages nine and above now if you follow my project here on YouTube youll know why I couldnt have been happier to receive this each crate contains detailed instructions and an educational magazine along with the project supplies I enjoyed assembling this kit and appreciated that it can be assembled without the use of glue or other outside tools once assembled I added a few batteries and bolts as a counterweight though this could be anything even rocks from the garden I then loaded it and it launched a supplied pingpong ball first try I was also pleased to read the educational magazine which covered the history of trebuchet DS the basic physics behind levers and even how changing variables such as arm length and counterweight mass will affect the projectiles range Kirikou a subscriptionbased service which deliver their crates monthly so there are many of other different projects to build so if you want to find out more about QE code go to kiwi cocom forward slash tom stanton and receive your first monthly subscription crate for free so thanks the Kiwi Co for sponsoring this project now back to the barrel skipping yes so I was able to get a few good skips using the drone but it suffered in terms of forward speed especially when accelerating in such a short distance it may have helped if I had flown lower but I didnt really want to risk crashing it I still cant believe how lucky we were to have that one successful test if that propeller had fallen off literally a minute earlier we wouldnt have had that a successful test and the whole project would have just flopped I travelled quite far north to visit James and he put a load of effort into building the plane so there was a lot riding on that final flight we didnt know it was going to be the final flight but thats how it ended also the weather that day was just perfect low winds clear skies the day before and the day after were really windy and really wet so I think we were just lucky all round I want to say a massive thanks to James for building Lancaster model he did a great job in terms of making it look scale and also in it like it flew really well and unfortunately I wasnt able to fly it due to the unchanged or disassembly but James did a great job of piloting it also thanks to James for coming up with the idea to do this project because if it wasnt for James he wouldnt be watching this video right now Ill be posting a link to Jamess channel in the description down below so go check out some of his videos hes put together a twopart miniseries of him building the plane and also his side of the testing so if youre interested in find out more then go and subscribe to James in the description down below I also want to say thanks to Matt from the YouTube channel DIY perks for coming along he helped film a lot of the shots James and I had our hands full all day with James obviously flying the plane and I was launching the plane and controlling the release mechanism so the rest of the cameras were just on tripods and Matt was able to film the tracking shots across the surface of the lake so thanks to Matt for helping with that thanks once again for watching if you enjoyed this video be great if you could leave a thumbs up down below if were new to my channel and want to see more projects similar to this then click subscribe down below a massive thank you to all of my supporters over on patreon comm you make these projects possible and I honestly couldnt do these projects without your support so thanks once again thanks for us again for watching and Ill see you in the next video goodbye Music I did those graves it actually did Music

write a youtube script

in the early s British engineer by the name Barnes Wallis was experimenting with skipping marbles across a water tank in his garden this eventually led to one of the most famous engineering marvels of world war ii where a barrel shaped explosive was skipped across the surface of a lake to destroy enemy dams commonly known as the dam Busters now instead of this video being a historical documentary why dont we try to replicate this on a slightly smaller scale this video sponsored by Kiwi Co more on that later a d printed this prototype barrel with the plan that the slot will contain a CNC cut aluminium disc this would allow the weight of the bow to be adjusted as well as its rotational inertia when bonds Wallis tested the barrels he discovered that adding backspin was essential to getting him to skip the spinning created gyroscopic stability and kept the barrels pointing in the correct orientation but I soon decided I didnt really want to see and see cut a bunch of aluminium discs that would most likely end up at the bottom of a lake so I decided to print this barrel with threaded countersunk holes this way I can use large bolts to increase the mass of the barrel which saves a lot of time and cost speaking of time this took and a half hours to d print and Im going to need a few of them so the final design will be cut from this PVC pipe and a quick tip to make a quick and easy straight cut through a tube wrap packing tape around until the ends meet without creasing the tape then work your way around using a hacksaw and the cuts should be pretty straight then I can use these d printed end caps to mount the bolts too as well as a center axle to spin the barrel youll probably notice from this test footage the barrel looks more like a sphere than the cylinder so Ive d printed these curved sections that clamp around the cylindrical barrel however the actual spherical barrel is proved difficult to manufacture at the time and therefore these straight cylindrical barrels were eventually used in the raid but Id still like to test both now we need to make the release mechanism the actual release mechanism Barnes Wallis designed sandwiched a barrel between two arms which would pivot outwards slightly and drop the barrel but before release it was spun up using an electric motor via a belt and pulley system shadow types are Brian from the YouTube channel real engineering for letting me use his animation so a d printed a replica mount to hold the barrel but instead of using a belt and pulley system to spin the barrel I use a small brush this motor directly on the barrel axle for simplicity so one arm contains a brushless motor and that other contains a bearing each with a cone that meets well with the barrel but in order to keep the barrel in position the arms needs to be held in words and this is done using a string system and a short piece of elastic attached to a servo this causes the arms to be sprung loaded and creates a tight clamping force against the sides of the barrel then when it is ready to release the servo simply lets go of the elastic so the release mechanism works well but theres a slight problem the barrel sink now because I dont want to be littering these barrels all over the bottom of a lake and I also dont want to have to make loads of them for testing I need them to float I then realized a really simple fix is to get a ziploc bag and fill it with air then once sealed I can put it inside of the barrel and the barrel will float no matter how much water leaks in so now we need something to drop it from so Ive teamed up to wave at James from Project air the YouTube channel where he builds things like this this is a large scale Lancaster bomber and you built this out of foam board right yep its all homemade bit of a buzz job recycle some electronics in there now were going to throw it over a lake and see if Tom sir bouncing barrel works lets just do a controls check quickly thats left thats right thats up thats down that goes that way that goes out yeah all right chaps yep as ready as well ever be right oh god its got a lot of up oh wow that looks amazing I think the battery cables come out its flying though look at that how she looks really good yeah so this is a lot of foam and a lot of look how slow its going as well its got theres a lot of motors in the air at one time no idea what the stall speed is oh yes nice one yeah oh my god the circle nice format we decided it would be best to test the curve balls first as they might be a bit more forgiving in terms of skipping across the water so with all the cameras rolling and the surprisingly large crowd watching we were somewhat ready Im ready okay unsurprisingly Jamess awesome homemade lancaster handled the weight of the barrel and its release mechanism with each good yep okay ready Tom okay down coming down then we just pause right here Barnes Wallis calculated that in order to skip the barrel successfully across the water the planes we need to fly at meters per second ground speed and drop the barrels from an altitude of meters this would result in the boughs hitting the water at about a degree angle some quick calculations estimate that this plane is traveling at . meters per second at an altitude of . meters which should result in the barrel hitting the water at degrees now this may seem like a rookie mistake but Barnes Wallis actually made the same mistake when initially testing his barrels it was in time to land the plane which wasnt exactly easy seeing as there was a wall and a lake on one side and a very steep drop on the other so its a little damaged isnt it James yes but dont worry so we fixed the plane and got it back in the air this time around my mechanism failed to spin the barrel up fast enough and I told James to fly back around but before he was able to do another pass the barrel came loose from the mechanism completely my fault and very frustrating as the landings werent easy so every flight counted fortunately James landed it beautifully and it was all good friend of the flight yep this time with the straight cylindrical barrel just like what was used in the actual rate Im getting nervous again scary all this work in the air I want the video to be good so I hope youre enjoying this yep good okay James flew the plane round with a long runup to allow the barrel to spin up slowly lower the altitude was also gaining speed and I did that was great yes lets take a look at that again from the high speed the plane was traveling at an estimated fifteen point six meters per second and dropped the barrel from an altitude of about one point three four meters which was really impressive flying from James causing the barrel to impact the water at an angle of degrees which is quite a bit steeper than Barnes Wallis originally aimed for but to still managed an initial skip distance of five point five meters with a total of three skips spanning eight point four five meters however it wasnt over there the mission still had to be completed which was made significantly harder by this rogue propeller this planes done its mission and then it just it didnt make it home boys like I was so nervous it was very low actually three times I think yeah quite a few times it was then time to collect the barrels with Jamess homemade hydrofoil boat whilst I made some modifications so the barrel dropped mechanism its amount of the underside of a large drone I just built the week before for future heavy lift projects how would you feel about flying it Tom whilst Ill show some of the drone tests Id like to mention the sponsor for this video Kiwi Co Kiwi Co delivers monthly projects designed to make learning about science arts math and more fun and accessible they have seven different crates for kids aged zero to sixteen plus and they sent me one of their tinkerer craze which is aimed at ages nine and above now if you follow my project here on YouTube youll know why I couldnt have been happier to receive this each crate contains detailed instructions and an educational magazine along with the project supplies I enjoyed assembling this kit and appreciated that it can be assembled without the use of glue or other outside tools once assembled I added a few batteries and bolts as a counterweight though this could be anything even rocks from the garden I then loaded it and it launched a supplied pingpong ball first try I was also pleased to read the educational magazine which covered the history of trebuchet DS the basic physics behind levers and even how changing variables such as arm length and counterweight mass will affect the projectiles range Kirikou a subscriptionbased service which deliver their crates monthly so there are many of other different projects to build so if you want to find out more about QE code go to kiwi cocom forward slash tom stanton and receive your first monthly subscription crate for free so thanks the Kiwi Co for sponsoring this project now back to the barrel skipping yes so I was able to get a few good skips using the drone but it suffered in terms of forward speed especially when accelerating in such a short distance it may have helped if I had flown lower but I didnt really want to risk crashing it I still cant believe how lucky we were to have that one successful test if that propeller had fallen off literally a minute earlier we wouldnt have had that a successful test and the whole project would have just flopped I travelled quite far north to visit James and he put a load of effort into building the plane so there was a lot riding on that final flight we didnt know it was going to be the final flight but thats how it ended also the weather that day was just perfect low winds clear skies the day before and the day after were really windy and really wet so I think we were just lucky all round I want to say a massive thanks to James for building Lancaster model he did a great job in terms of making it look scale and also in it like it flew really well and unfortunately I wasnt able to fly it due to the unchanged or disassembly but James did a great job of piloting it also thanks to James for coming up with the idea to do this project because if it wasnt for James he wouldnt be watching this video right now Ill be posting a link to Jamess channel in the description down below so go check out some of his videos hes put together a twopart miniseries of him building the plane and also his side of the testing so if youre interested in find out more then go and subscribe to James in the description down below I also want to say thanks to Matt from the YouTube channel DIY perks for coming along he helped film a lot of the shots James and I had our hands full all day with James obviously flying the plane and I was launching the plane and controlling the release mechanism so the rest of the cameras were just on tripods and Matt was able to film the tracking shots across the surface of the lake so thanks to Matt for helping with that thanks once again for watching if you enjoyed this video be great if you could leave a thumbs up down below if were new to my channel and want to see more projects similar to this then click subscribe down below a massive thank you to all of my supporters over on patreon comm you make these projects possible and I honestly couldnt do these projects without your support so thanks once again thanks for us again for watching and Ill see you in the next video goodbye Music I did those graves it actually did Music

write a youtube script

I have come to the United States of America or Los Angeles specifically William Osmonds house we need to hang out with this guy and talk about nerdy stuff now Ive run into a bit of a problem my phone is dead well it has a low battery and all of the plug sockets here in America dont seem to work so I need to find an alternate solution this video is sponsored by Skillshare more on that later so I need to find a USBC phone charger plug hey William can I borrow your phone charger since no one else owned a phone charger abuse I took to Amazon to find what I needed and came across his handcrank generator with the USB output which should be perfect for charging my phone right so the handcrank generator has arrived from amazon by turning this handle lights have a small red LED which I assume is showing that the USB is outputting power lets hook these all up hours and minutes to fully charged Im turning this up probably one and a half revolutions per second and I really dont want to be doing that for the next hours so lets come up with a better solution Peter so this is basically the same crank generator that Ive bought off Amazon yeah but geared so that you can hang a big weight off of it and extract the energy from the way maybe I should convert my hand crank generator into one of these yeah charging a phone from gravity would it disassembling the handcrank generator revealed a gear system a brushed motor and a small electronic board with the USB port I then modelled all the parts into CAD to design a custom casing which could be manufactured on Williams d printer once printed I can attach the pulley to the output gear and fit the remaining components inside whats poppin Tom whats purple how does it work so this is the gravity pod phone charger basically a hand crank generator modified into a d printed housing gear system inside here its brushed motor which then powers this USB port here now Ive designed it to work with this beaded rope system so I can hang a weight from it drop it off the edge of this ledge and we should be able to charge my phone okay ready yeah how long does a is it four hours why is this Jack McGee the bat your invention sucks time spending anything you made yeah Wow how many times do I need to lower this to fully charge your phone well continuously for four hours so if that took how long did that take like so would be sick so you have to do that six or times I think we need to make this thing lower slower okay so like more weight so in a pulley system that pulls it slower exactly so in order to slow the descent rate Ive added two extra pulleys this means the weight will descend three times slower for the same rpm of the generator and since the USB wont output more than . amps at volts even when dropping really fast its best to prolong the descent time as much as possible without the current dropping too low so lets try lowering a larger weight from a greater height I am setting up the gravity phone charger to try off the edge of this high ledge its kind of scary standing on a chair where youre gonna take the potential energy from this coke bottle and put it into my phone I wring the coke weight were gonna run out beep oh no whats the problem I need to winch it with the way back up so we found the fundamental issue with this gravity powered charger and that is that I have to win yeah no no no no we can try to train the cat to get in the basket write it down and then climb up the stairs get back into the basket right again lets try the first test of the cat in the box ready yes a totally word beer not this was a manager Jimmy was not in the bag at any point in time whatsoever right Jimmy one last idea to test this charger currently dead if I press the power button it doesnt turn on were gonna see how many times we need to lower this coke bottle maybe charge one percent in battery and take one photo in the comments down below take a guess how many times you reckon this is going to take and once were doing this were gonna hit a word from our sponsor Skillshare is an online learning community with thousands of classes covering dozens of creative and entrepreneurial skills a premium membership gives you unlimited access so you can join the classes and communities that are just right for you seeing as my goal is to take a gravity powered photo Id like to recommend this course on phone photography each topic has an easy to watch short video explaining very simple solutions to achieving impressive looking photos Skoosh it is also incredibly affordable with an annual subscription working out at less than per month Skillshare is sponsoring this project by offering two months free trial if you sign up via the link in the description so if you want to check out more about skill share and help support my projects go check out the link in the description down below coming to the bottom of the drop thats three drops thats me the phone lets see if we can turn it on huh its got into like super power mode let the brightness is dying oh I know weve got weve got something or a take the stupid selfie yeah but we took a gravitypowered selfie if youre doing this video leave a thumbs up down below if you do too much I dont want to see more crazy projects click subscribe down below a huge thanks to all of my patreon supporters for making these ridiculous video as possible and yeah thanks once again for watching thanks to Peter for filming this and I see in the next video

write a youtube script

Music this is my homemade electric bike it isnt powered by a hub motor or addon kit but instead powered by a large drone motor its taken me three years to get to this point and I want to share what it took to build an electric bike from scratch so grab a drink sit back and lets read one three years at the time I was studying at university and my housemate Dannie built an electric skateboard using these hobbygrade brushless motors combined with speed controllers designed to radiocontrolled cars now I personally prefer two large wheels over four small wheels and one did this setup can be modified for a bike so I purchased a cheap bike from Amazon with a large motor that can be attached to the rear of the frame using a d printed mount this motor would then drive the rear wheel via a belt and pulley system where the large rear pulley was also d printed and cable tied to the wheel spokes this drivetrain worked well but the cheap speed controller had a slight issue the throttle input controlled the speed of the motor and not the power output which makes sense as it is called a speed controller but the issue is that it will do anything to maintain the speed in but you give it so when climbing a hill if the motor cant maintain the speed that you want it to the speed controller will dump a load of current into the clothes and toast the motor a little after this I didnt want to spend too much money on a project that might not work but I also didnt want to risk ease in the same speed controller and burn out another motor so I purchased a V ESC or vest this speed controller was twice the price of the previous one but I had currentlimiting capabilities meaning the motor was essentially never burn out if I set the limit correctly also I didnt want to buy another motor and therefore ended up rewinding the motor by hand which is why the coils in the original electric bike video looked different to the original motor coils and it was this rewound motor and Vesak setup that powered the bike in my first electric bike video back in but there were some fundamental flaws of the design the first being the pulley ratio was far too high meaning the acceleration wasnt great on the plus side at a top speed of miles per hour I think d printed a new larger pulley which produced more torque at the rear wheel this would also reduce the top speed and this began my quest for more power so I built this ridiculous setup yep thats two motors attached to a bag rack above the rear wheel these two motors will link together via a single belt to drive the rear wheel but it never rode more than a couple hundred meters due to an unforeseen issue as this motor applied talked to the belt it would increase attention here but also decrease attention here meaning the reduced tension around the second pulley caused the belt to skip rendering the second motor useless and then made this mount where theres only one output pulley driving the main belt and a secondary small about connecting the two motors this setup works surprisingly well and had lots of pound Music but the biggest problem that they set up was not being able to tension the smaller belt joining two motors together at the outer case of the motors would spin also surely afterwards one of the two speed controllers died fortunately I was contacted by a company called tramper boards who are working with Benjamin Vedder on a new speed controller called the vests six this new speed controller could handle a lot more current meaning I could use one large single motor rather than two smaller motors and get a similar power output I also invested in some lithiumion cells to build a volts amp hour battery pack and built apart d printed part aluminium motor mount which was pretty similar to the version one electric bike but with around three times the power its also had for dinner two breaking which was controlled by a thumb lever on my left handlebar regen charge and coming up to this but after lots of testing it only increased the range by about . and get the power on oh yeah its got plenty of torque I ran this same set up to the following eight months and rode around miles with minimal issues other than replacing the belt once in fact I ended up replacing the tires and brake pads more often which led me to eventually upgrade the bike which required a whole new motor mount design and to do this I imported photos from multiple camera angles into my CAD software to produce a d model of the bike and design this bracket where the motor could be mounted behind the seat post and cut a prototype from plywood thinking difficult part of design the rear pulley as there was very little space between the wheel spokes and the brake caliper the plan was to mount the pulley to the brake disc as this was a pretty solid mounting point and then Ben some aluminium arms to an angle where they would just miss the brake caliper which was a very tight fit but work nonetheless then with a total of six spokes and a d printed toothed pulley it was time to work on the motor mount again which I realized needed a full redesign as I wasnt comfortable with this long arm potentially twisting into the rear will this new design utilized the water bottle mounting bolts and a clamp around the seat post at the rear which kept the motor protected inside of the frame but there was a problem the pedals hit the motor so I cut the end off and also mounted the belts and pulleys which as you can see there are two belts instead of just one this is to gain a larger pulley ratio between the motor and rear wheel where this belt produces a twotoone step down and the larger belt produces a to step down meaning a total motor to wheel ratio of to or . to but dont tell my math teacher this produced so much torque at the rear wheel that the pulley exploded instantly upon testing due to a weak design so I beefed up the spoke mounting holes and it held up a lot better but then I was having motor issues it seemed it wasnt possible to run the motor so close to the aluminium mount and yes aluminium isnt magnetic but if I drop this magnet down this aluminium tube you can see theres clearly some resistance I wont go into detail how it works but it was causing so much resistance on the motor that the motor lost sync with the speed controller and for a short period tried spinning the wrong way causing the Bell to skip and make that horrible noise so I had to widen the slot a little also the shaft was slipping in the motor due to my genius idea of cutting the end off so I had to make this plate to hold it in place and then d printed a cover for the motor and the bike was ready for a test ride and this was the last electric bike video I uploaded to YouTube but a lot has changed since then the dual pulley setup produce a lot of noise and resistance which reduced the efficiency and range so I flipped the most around I made a new mount which connected the motor output directly to the main drive belt and completely redesigned the large rear pulley to be manufactured from a single piece of aluminium which removed the need for the singular spokes and were spaced away from the brake using a few aluminium discs I also programmed an Arduino board together serial data from the speed controller and display all the important information on an LCD screen if you own a vest and want to make a similar display Ill post a link in the description for more info and this was a final design for the next year and a half when I decided to upgrade the motor the diameter of this new motor is larger than the previous motor and its got magnets instead of which I had no idea if it would improve the performance but it was worth a try the only problem is this motor lacked an output shaft long enough to mount a pulley as its designed for mounting propellers too so I had to make my own shaft which would bolt to the motor housing passed through all the bearings and attached to a pulley like so this could then be bolted to the new aluminium motor mount with the other mount plate being cut from three millimeter carbon fiber to avoid the magnetic motor issues again and this fit nicely inside of the frame but spawn testing there was a noticeable vibration coming from the motor this was due to my amateur lathe turning skills where I machined the shafts slightly too small and there was playing the bearings so this is my work around a millimeter shoulder bolt these boats have a shaft section which is machined perfectly to fit within bearings but to turn it into a usable motor shaft I filed down the head to lock into an aluminium bracket that I cut on my CNC router I then sprayed it black and bolted it to the motor and also cut me threaded section off then all I had to do was replace the bearings in the motor for millimeter inside diameter rather than millimeter and now I have a motor shaft with zero at play to complement the new motor I decided to build a new battery pack using these cells to do this I had to cycle charge every cell to measure their capacity which took just over a week entering all the values into a website called second life storage will calculate which cells should be put into parallel groups and I could then spot well the negative of one cell to the positive of a cell from another group until I had six pairs and I had to repeat this to make a pack of six in series and in parallel then I added some power wise and a balanced lead to keep each cell balanced I choose not to run a battery management system on my electric bikes as I prefer to keep the batteries as small and simple as possible and I use my charger to manage the cell balancing so thats a volt amp hour lithiumion battery pack and I need another one of them to white in series for volts but it requires a different cell layout hes a slowmo video of the spot welder at frames per second it shouldnt always spark like that as I held the contacts at an angle which looks pretty cool but it also destroys the welding tips this new layout battery pack was the same volts amp hours but was slightly angled to miss the frame which when sealed in heat shrink was a very snug fit and then d printed a guard to stuck out the same distance as a motor pulley to prevent my leg touching it and this guard had holds through to the motor to help with cooling on the other side of the frame almost the whole motor was covered to prevent dirt and water from entering but they kept some slots on the rear side for air flow which seems to prevent oncoming mud from entering the only issue is air flow carries dust and since the manufacturer of the motor design at Fort remarkable cooling performance it also sucks up a lot of dust so I plan to build a dust filter for the in their holes using some open cell foam in the future so would I recommend you build an electric bike like mine the . kilowatt motor performance is great allowing it to accelerate to miles an hour in a few seconds the range is decent with about hours of fun riding and even more if cruising but if you just want performance hub motors are the best option this is a kilowatt hub motor and its super quiet when running of s as well as having better acceleration it is quite a bit heavier at . kilograms considering my whole setup including the battery weighs just points but the hub motors will fit into pretty much any bike frame whereas my needs a very specific design to fit one big advantage of mine is it can be converted back to a regular bike in about a minute or two by simply removing the battery and taking the belt off leaving the motor and wires behind which adds about two and a half kilograms of weight I know that many of you who follow my channel are engineers and most likely wants to build an electric bike from scratch and the most common question I get asked is how do you choose a correct motor these hobbygrade brushless motors have two important specifications the size of the motor is usually labeled as a fourdigit number for example this is a seventyfour turnigy motor and I wouldnt recommend going much smaller than this for an electric bike the second specification and probably the most important is the kv rating which isnt a measurement of kilovolt but instead the velocity constant and it essentially relates to how fast the motor will spin depending on the voltage of battery that you have for example powering this kV motor with a volt battery will rotate at roughly seven and a half thousand rpm ideally for this size of motor you want to be within to kV for an electric bike the next thing to calculate is the motor to wheel ratio whether you choose to use a belt chain or geared system there needs to be a ratio between the two throughout all my ebike developments Ive realized one golden rule to getting this right and thats this calculation the kv rating multiplied by the motor to wheel ratio needs to be approximately for example my first electric bike had a kV motor with a tooth wheel pulley which outputs a number of . which is way too high and is the reason it lacked in talk my version II bike had a value of ten point four which was pretty much perfect and my latest version has a value of eleven point two which is still fine as the motor is quite a bit larger and can handle the talk interestingly enough the three kilowatt hub motor is . kV which when plugged into the equation with a one to one ratio gives a value of eight point nine so essentially this equation converts the smaller motor rpm rating to a hub motor equivalent so I hope its also vinter some of you and hopefully inspire some of you to try and build your own electric bike if you enjoyed the video if you go if you can leave a thumbs up down below if youre new to my channel I want to see other crazy projects click subscribe down below a massive thanks to all of my supporters over on patreon comm for making these projects possible thanks once again for watching and Ill see in the next video goodbye Oh

write a youtube script

this is my D printed Lego minifigure Im trying to match the geometry from this guy to this guy even down to the joints you could kind of think of this like a tshirt blank where I have all the geometry exactly how I want it but then I can D print in whatever color I want and create custom minifigures last week I was working on these legs I was able to design it so its a print in place assembly so this print has one piece and has moving pieces after you take the supports off however I feel like this looks too different than this and these kind of look like knee joints I then tried to design that prints the legs separately from the waist and then they snap together this worked and it was really cool but I really wanted to get print in place to work so I tried again and it works and this is my best version yet I was also able to design a print and place torso with moving arms and wrists theyre printed in a Mr Crab stance to allow for Max detail and easy supporting Im really happy I got print in place to work because if I ever want to print like a hundred of these at a time itll save me a ton of time rather than having to D print all the pieces separately and then assembling them together

write a youtube script

come on its okay youre still a winner youre the best loser that i know instead of being a loser thats somehow a winner lets actually win i dont have skill but i do have technology so lets be honest robots are better than us theyre more precise great at computation and they work long hours for no pay im going to apply these strains to pool to make a robot that will find and take shots that humans can only dream of ill also be using this robot for something that you might not expect im going to use it to play pool online physically which raises some pretty interesting ethical questions it definitely doesnt seem like its working this was kind of a random idea but turned out to be really fun this just might be the star of the whole show playing pool takes skill for now at least its also a game with many variants im going to be focusing on ball pool two players take turns trying to sink all their balls by striking a white cue ball with a stick the gameplay boils down to two steps figure out what shot you want to take and then hit the balls very precisely to execute it i want the robot to do both of these things figuring out what shot to take henceforth referred to as the big brain algorithm is going to take some explaining and ill get to that in a minute the idea behind the robot is a bit more straightforward say im trying to make this shot if i had a robot in the tip of this queue it could move the tip so that it hits in just the right way so that the ball goes in heres what were gonna have to build a camera system to watch the game and figure out whats going on a complicated algorithm to interpret this and figure out the best shot and a pool cue with a magic robotic tip that can take the shot for us this is probably the most complex project that ive done the robot q is obviously going to have to move and rotate laterally but pull is not d where you hit the ball vertically matters a lot off center hits put spin on the ball which gives you control over where it goes i want the robot to have control that means moving and rotating vertically as well fitting a mechanism that can do all these motions into the tip of a pull cue is going to be very hard its going to get fat but my goal is for it to at least remain a stick i have an initial design but i dont have a good feel for it i need to hold it i havent really talked about my glowforge but its really useful for making quick prototypes like this i use it kind of like my plasma cutter where i can very quickly make structures out of flat patterns except for in this case out of plastic and wood rather than metal im glad i quickly prototyped this because theres no way this is going to work its just too bulky and theres no way im going to be able to get this stiff enough deep down i knew that a steward platform was probably the right way to go here but ive done everything i can to suppress that thought because i know its going to be very painful to make the idea is that you arrange six telescoping rods in a very special arrangement you can completely control how this platform moves around by extending and contracting these rods i can stick a tiny one of these in the queue to move it wherever i want its very stiff so it wont deflect when i take a shot and its also really compact i wanted to avoid a stewart platform because theres a ton of finicky little parts instead of telescopic rods i have the linkages on these cranks which push them up and pull them down but its essentially the same the defining feature of this stewart platform is that its driven remotely through tension cables the servos are on one side of this broomstick and the moving end is on the other the electronics are pretty straightforward theyre mostly off the shelf microcontrollers and other components man that end got thick not really the best look there we go these platforms take a bit of math to move them around but its really not too bad say i wanted this platform to move over and rotate like this one of the actuators gets shorter two get longer and it results in it being this position if i want to calculate how long these actuators need to be i have the solution right here in front of me this is how long they need to be for it to go here conceptually all i have to do is move the platform where i want it and then measure the distance between these sets of points to tell me how long each actuator needs to be i solved it by drawing a picture but its also very easy to solve this mathematically using linear algebra that is exactly what my program is doing this is a case where the results make it look much harder than it actually is if i did everything right the platform should move wherever i want now Music wow working on the first try is not normal its pretty cool how the motions all combine isnt it this air cylinder on the tip of the queue will be used to fire the shot its connected to pressurized air when i open a valve that shoots out really fast but not all shots are done at maximum power we need a way to vary the power from gentle touch to battering ram to do this i built the power pack this tank is normally intended for car horns the tank has a small wimpy computer attached to it which people call a microcontroller the microcontroller measures the pressure with this pressure gauge if the pressure is too low itll open this valve which will connect it to my shop air and pressurize the tank if its too high itll open this valve and release the pressure and to fire i open this valve its really simple and works well this cue is really cool but its completely useless without some more information if im holding it here how does the computer know that it needs to move it over here so the shot will go in i have to know where everything is to be able to figure that out i have this camera mounted on the ceiling looking straight down it can see everything going on on the pool table it has a big problem though although this table was the cheapest one that money can buy it is not this out of square this image makes the balls look like theyre over here but in reality theyre over here this is caused by the camera being out of square with the table if i took the position of the balls from this image then im going to calculate the wrong thing programmers call this garbage in garbage out ive heard this also applies to raising children i glued these funny looking tags to all the corners of the table and theyre called fiducials theyre basically a reference point thats very easy for a computer to pick out of an image kind of like a qr code i can find these four points and then use them to straighten the image up i know where the holes and bumpers are relative to the fiducials so i know where they are now too and then if i find the balls i know where they are and the last thing to find is the queue which is why it has these two fiducials one fiducial tells me where it is two fiducials tells me what direction its pointing there is still one little problem the camera only tells me where things are laterally because its a d image i dont know how high the cue is above the table or how its tilted normally you rest the q on your hand but the height of my hand isnt controlled theres an easy fix for this its just like my hand except plastic screwed to the queue and most importantly known height i still dont know how its tilted so i added this imu to the end of the queue and imu is what you give someone when you owe them money but youre broke imus can do all kinds of things but in my case it tells me what direction is down so now i can compensate for tilt i still need to make the big brain algorithm to figure out the best shot for now all it can do is a single ball into the hole i show the shot in this really simple interface that i made and if i want to hit this shot it tells me that i need to hold the cue in this area the problem is that the ui is on the computer and im at the table i guess i could look at the computer screen but thats awkward i want to be unencumbered i want to enjoy the game of pool the way that it was meant to be played ive got an idea im going to project the user interface onto the pool table oh no one little tiny hole took out my furnace my lights my security system and my internet this must be the most important cable in my entire house all right lets see if this thing works i have an annoying problem the projector is just projecting on the table it has no sense of where anything is if i just display my ui on the projector its totally wrong cue ball is here not here if i project a grid of fiducials with the projector i can figure out where the projector can project relative to the table from there its easy to make the ui line up with anything all right we have everything we need to try it lets see what this thing can do at this point it looks like i succeeded in building exactly the opposite of what i wanted which is a pool cue that always misses even if the shot is really easy this would make a really good wife mode for some reason the robot is moving to the wrong position sometimes so what could cause this i have a camera watching the table a program which modifies the image to correct for camera distortion and then another program which extracts the positions of everything this is plugged into a hacky physics simulation which does a lot of math to decide the angle of the tip which is then converted to six actuator positions which are then realized by an untested robot so why is it going to the wrong position i have no idea i present to you a week of pain in seconds alright i think i figured it out no these plastic pulleys sometimes slip lets redesign them that is the wrong ball i found some math issues which must be causing problems you know what math is useless i made a system that records where the cue goes versus where i want and then corrects for it this has to work at this point im just systematically testing everything this is painful but eventually you run out of stuff that can be wrong i finally found the problem these six little servo motors are what control the end of the platform they were sold to me with specs of how far they can rotate but two of them dont do that which means theyre rotating a different amount than i think which makes the end either sort of wrong or really wrong depending on where im trying to move it in my defense you sort of assume that the thing you bought does what it says that it does all right let me show you what this thing can do have you ever woken up from a nightmare but youre actually still asleep and youre in another nightmare yeah it turns out theres more than one big problem so something somewhere is broken i know that but i have no idea what it is i feel like i went through everything trying to solve that servo problem i flailed around helplessly for quite a while were gonna just fast forward through all that oh steps up this is the good part now back a little bit it looks like things are in a slightly different position than my software thinks and remember garbage in garbage out its not a huge amount but its enough to mess up most shots i think this is due to some kind of optical distortion if youve ever used a wideangle camera lens youll have some idea of what im talking about if i take a wideangle view of the table and square it up these points are square but look at this edge its supposed to be straight this is a distortion caused by the wideangle lens my camera already has a calibration that should correct for this but im guessing something somewhere is wrong day of flailing and im trying to decide if this is what the pit of despair really feels like everything i do that should make the system better is making it worse because there are so many places that this can be wrong im doing a hail mary by developing my own calibration system the rough idea is i take a bunch of pictures of something that i know the shape and size of i can solve a series of equations thatll take a location in the image and directly calculate the corresponding location in the real world since this goes directly from the image to a predicted location it will skip over anything that might be broken this picture is giving me an idea of the distortion on the pool table i dont see any physical way this could be an actual optical distortion which implies that i am an idiot you know how occasionally a surgeon will accidentally leave their scalpel inside of you and then youre trying to figure out why youre sick all the time but you never thought to look for the scalpel inside your belly the same thing happened to me and i guess by me i mean my code way back in the beginning when i was developing this code i was doing a little test and i accidentally left that code dormant deep in the bowels of my system this code does a camera distortion correction and this correction is for a completely different camera than what im using this was subtly messing up my images and it could be applied in a couple of different ways and in one case it was applied twice it made it very confusing im exaggerating the effect so you can see it it was a lot more subtle five to ten millimeters of position error so this is roughly equivalent to punching myself in the face for five days straight all right moment of truth all right one Music two yes finally they went in and it wasnt random a neat thing i realized you can do is use this to guide normal pool shots too this thing is just really fun although if im going to go against my wife i needed to do more than just hit a single ball im going to get to the big brain algorithm soon i promise but theres one thing i want to try first so it should be possible to play pool with someone over the internet if i send a video feed of the pool table to someone and then i give them some controls to select a shot to take the computer can tell me where to hold the stick and then it will take their shot for them physically so its like online pool but real which is weird i went ahead and built it were just going to skip over that park so its kind of boring all right lets test it with a game with the wife first when its her turn she chooses the shot on her computer it tells me where to hold it the trigger tells the computer im ready and then she hits fire cool im even gonna beat you from upstairs i think youve just been lucky that was pretty close it doesnt seem like it went where i aimed the queue not your best shot ever this is definitely broken it seems like its doing exactly what its supposed to do we should just play when you fix it all right its working perfectly hey hows it going this concept is so weird i basically built it on a whim but it is surprisingly fun boom i really thought this would be basically the same as online pool maybe a little crappier but it really is different i was so engrossed i didnt even notice that i lost the game wait whered they ball go yeah ive actually been looking for it for a while wait well thats that im playing with some people who support me on patreon which helps me a lot so im trying to understand why this is so fun i think it has something to do with it being a real game in fact i challenge you to find an online full game with a more realistic physics engine than this one its almost like two people holding the stick at the same time which is strange oh who is this redneck yeah man you just call someone up and play a real game of pool with them nice through the internet through the internet maybe this isnt just a goofy tangent it feels like a legitimate way to connect and have fun with others oh you sunk you moved it huh nothing remember when they connected pokemon to twitch what if i did the same thing with this pool table basically me versus a crowd do you think thatd be worth doing why didnt you go for the green i because i call the shots literally i call the shots youre just youre just the meat servo to align my servo all right finally we get to the fun part the big brain algorithm the idea is to make a physics simulation based on what the camera sees this should look at the entire table and calculate the best shot for you so what is the best shot this shot may be the easiest to sink this shot is a bit harder but it leaves the ball right next to another ball which will be easy to sync what i want is a list of all the possible shots that i could take and then i can select the shot i want from that list based on risk and probability and whatever other criteria matter imagine i have a table with three balls on it the simplest way to figure out all the possibilities is to just brute force it i can make a physics simulation where i virtually take hundreds of thousands of different shots and see what happens theres a lot of reasons doing it this way is bad the biggest reason for me is that this is slow so heres what im doing i look at each ball and i calculate all the ways that it can go into holes bounce off other balls and anything else that would advance the game it makes a big fat mess i do the same thing for all the other balls now i have all the possible things that each ball could do individually the question now is how do these things relate here are a few of the many things that the red ball could do could any of these things be caused by something that another ball could do one of the possibilities for this cue ball is that it directly hits this red ball in reality this cue ball can hit this red ball anywhere in this range if we hit it right here that would send the red ball along this trajectory the white ball could also bounce off the bumper into the red ball which could also cause this shot to happen or it could make the red ball hit this blue ball that hit could make the blue ball go into this pocket if you do this systematically for every ball you end up with a dense network of all the possible outcomes this structure is called a graph specifically a digraph because its directional since everything begins at the white ball all the paths through this represent a potential shot that i could take this is a very simplified version that ive drawn the real version is just a gigantic mess if youve heard of ray tracing before what i did here is kind of a second cousin of ray tracing it looks incomprehensible but this graph is very easy for the computer to process and it has all the information we need to rank the shots for example how long is the shot or how precisely do i have to hit the ball i can ask the graph for the shortest path the longest which is almost certainly impossible the robot currently chooses shots based on a blend of distance precision and estimated speed im feeling good about my odds of winning at pool for once a lot of people will tell me that it would have been more efficient just to practice rather than building this whole system i agree but here we are although before i go up against my wife i want to talk about this videos sponsor so you know how this works the sponsor is what basically enables me to spend so much time developing insane pool cues and other crazy things if you like this video or you like me taking a minute to check out the sponsor is one of the most helpful things you can do to support me and these these projects this video sponsors micro center and you need to know about them when they asked me if i was interested in doing a sponsorship i just said where do i sign the amount of hours and dollars ive spent at micro center is just i dont even want to know how much money have i spent at micro center um lots i need a jumbo calculator for that one if you are remotely interested in technology you have to go check out one of these stores theres just nothing else like it they have everything its the only physical store i can go to and buy parts for my robots their selection of computer parts is just fantastic and somehow their prices are also really good my computer is built entirely from parts from micro center i dont know if im allowed to say this but they employ a bunch of nerds which is amazing when youre trying to figure out what components you should get for building a pc all that im really asking is for you to go and visit a micro center theres of them in the u.s and more people should know about them and before you hop in your car and peel out if you take just a second to click on the link in the description and enter your email address this lets microcenter know that stuff made here sent you this is how they decide if these videos are worth sponsoring which helps me out a ton but it also helps you out because when you do this theyll give you a coupon code which you can take to the store and theyll give you a free micro sd card and flash drive so thank you for hearing me out and thank you microcenter for supporting this video q versus wife who will win oh those things perfect theres some initial back and forth but then the robot really hit its stride Music i cant hold it still enough Music i have no shots now yep my little stripeys are blocking you and then the eight ball nice you really didnt win though the robot one it was my brain encoded in the software sounds a little creepy let me put it this way if your kid wins do you win yeah then im a winner the robot is not your kid weve talked about this getting this thing done was a bad as hard as labor oh please so hard all right i think this pool cue is awesome although there are still a lot of things that i just didnt have time to get to id like to take the pool move solver to the next level and have it look ahead multiple moves like a chess engine this is where ball spin becomes really useful and so far i pretty much just ignored it i think this is a good proof of concept and hopefully you liked it if you have any thoughts or ideas let me know id love to hear them thanks if you like this video please consider subscribing it helps me out a lot and keeps you in the loop Music you

write a youtube script

so i look very silly in this machine and the reason im in here is that the longer i go without cutting my hair the younger i look if you base my age on the youtube comments on my videos im somewhere between a and year old boy and the other thing that people are saying is that i look like the reference model for the lego minifig im not really seeing the resemblance but its time to cut my hair i would rather not have someone cut my hair whos touching other peoples heads all day long so i built this robot and it is a hair cutting robot its going to cut my hair the hard way using scissors no trimmers or any easy mode like that and i think it works so i dont know because i still have all my hair i havent tested it yet and im a little bit worried because the initial trials with the dummy didnt go totally smoothly but i think ive mostly worked out those software bugs i did a quick dry run just a little bit before this just to see if everything was working without cutting and it had this other problem of ripping my hair out wow ive fixed it now the fingers clamp a lot less hard than it did before im glad i tested that and even if this thing gives me a terrible haircut i still feel like ive won how many people have a robot made haircut its basically art my prediction is that its going to give me somewhere between completely terrible and a haircut well ive put this off as long as i possibly can so this is the first haircut that its ever going to attempt so lets do it Music so how about that weather i dont know i mean i guess its warm out oh well cool you can tell from my face that i wasnt sure if it was completely wrecking my hair or not i honestly was not sure if this was gonna work because it was a hard problem so the hard part even with the robot is handling the hair the hair wants to go all over the place and what you need to do is separate out a very local chunk of hair and you dont want to have any other hair trapped in that chunk and so thats what a lot of the combing and finger work that stylists are doing its to separate out the hairs i knew in my heart that trying to build a robot that works like a stylist to separate out hair was a bad idea i still tried to do it that way initially before i go way off the deep end into the details i need to take care of just a little bit of housekeeping so im sure what i want to build next i want to hit a baseball really far like really far it just sounds like good oldfashioned fun to me im excited because i found the perfect thing to put some extra pep into my swing this is a device which shoots nails into concrete using basically bullets so thats cool im super aware of the safety considerations here im going to do this very carefully in a very safe way so no need to tell me to be safe in the comments but if you want to you can if that sounds as awesome to you as it does to me you should subscribe to follow along and see how it goes and then these videos are very time consuming and expensive to make if you want to help support making more awesome things and more videos you can check out the patreon and in exchange im giving more behind the scenes content probably the coolest thing is the discord server im having a whole lot of fun hanging out in there with people and getting advice on my projects alright thats enough of that back to the fun stuff how about they sports team i mean i dont think there really are any right now right oh well cool the first concept that i was trying to do involved two combs and do a complicated series of motions to select just the hair that i want i even designed the mechanism to do it and it was pretty cool the problem was i just couldnt find a way to do this reliably without some kind of camera to see whats going on but ultimately trying to figure out where the hair is with the camera is a nightmare of phd proportions i think to try this would be insane i told you thats an insane design its way too hard you need to pull out the essential things that it needs to do and then design a way to do them thats easier for a robot the way to go here is a vacuum its gonna make life a lot easier yeah so as hes so rudely interjected i did have some alternate ideas that at least on paper seemed a good bit easier since when does rude dove good ideas you know im really looking forward to seeing you stick your big head into this robot my head no youre the one with the bad ideas anyway i came up with another way of grabbing hair that uses a vacuum that is a whole lot easier than i was going to do before the vacuum sucks all the hair straight up and then if i slide two fingers in i get just the hair that i want and no hair from the sides the way the cutter head turned out to do this is pretty cool it only takes three servos so heres the fingers the scissors are on a little stage driven by a servo and then the scissors are actuated by another servo so putting these things together i can move the fingers grab some hair lift up and then cut unfortunately this wig hair is its like wire so the scissors cant cut it this is unfortunate because i wanted to test it on a wig but it doesnt work so im going to have to be the first real haircut this explains the sense of impending doom in the pit of my stomach do anything fun this weekend it was really sad my dog died oh well cool the other big challenge of this haircutting attachment is how do i move it to all the different locations on the head and so what i did is rather than having a three axis system i have an arm that can move in and out and up and down only and so it can reach the face and they can reach the head then i rotate the entire head so i can reach the side i just rotate the head now im touching the side and so rotating the head is hard because i have to rotate my whole body i do the equivalent thing which i rotate the entire robot around the head i always need to be pointing at the head when i cut so im cutting on the top i need the scissors to be pointing this way if im kind of on the side maybe pointing this way and so i added a whole extra rotational mechanism to the to the robot gantry to allow this one other problem that i ran into that really annoyed me because it just made things way more complicated is that depending on where youre cutting on the head you want the scissors to be going different directions so when i cut the bangs i want to be cutting across the head i dont want to be cutting vertically if i cut vertically im going to end up with a diagonal sawtooth bangs which is going to make me look like a complete idiot so i added a whole new access to my robot that allows me to rotate the cutter head this machine is made with a bunch of different techniques the first is plasma cut and folded sheet metal you can see the plasma cutter is drawing these marker lines and thats so i can bend the parts very accurately so they fit together later some of the plasma cut parts are spot welded together and you can see im using a little access port so that i can reach the spot welds inside this thing there was only one machine part and that was because i was able to d print pretty much everything this is great because theyre complicated shapes which have been very difficult to make everything is built on top of this giant bearing which is what my head goes through and then theres a simple two axis gantry i need to cut a hole in a sheet of plywood which this cnc router i just got would be perfect for but its not set up so i had to make a tool to do it using your tools to make new tools is one of lifes great pleasures anyway so it was fun and then everything rides in this stand i built which bolts to my workbench so electrically theres really not that much interesting theres four big stepper drivers which are responsible for controlling different motors theres a microcontroller its a teensy thats running the show its communicating with the computer for receiving commands and all that stuff it is a giant mess but theres nothing like a good facade to give the impression that everything is okay all right how you feeling im a mess i mean i am good thanks how are you cant beat a facade when can i cut your wifes hair its not gonna happen trust me i tried all right what if i give you a haircut that is a super set of another haircut that you want then if something goes wrong we can cut your hair down to another haircut what do you mean if something goes wrong i mean i dont plan on anything going wrong its just its a prototype and sometimes stuff happens i already like my hair but come on its important its for humanity humanity again yeah theres all these people stuck at home in quarantine and they want to look good how does getting my haircut help it just does that doesnt make any sense just put your head in the robot its that easy you just put it in and itll cut it my analysis is that youve made yourself into the man who cried wolf even if you really mean well youre not going to be able to convince where did that come from oh well cool one of the big challenges of getting this robot to work is that it needs to know where my head is in order to cut the right length if my head is here its going to cut a different length than if my head is here your head moves around a good bit and id rather not have every little movement that my head does baked into my haircut i spent way too much time trying to get this depth camera working so this is an intel realsense depth camera and what i was doing is doing facial recognition to find where my head is and then combining with the depth data to figure out where my head is in d space but there was a really big problem with this switch this camera is completely blocked by this stand a lot of the time even when its not directly in front of it it really blocks it and i could maybe deal with that but it makes the software so much more complicated i just didnt think it was worth it what i ended up doing is i put a little switch on these fingers and this allows me to move and sense when it hits my head and then i can use that to measure a bunch of different points on my head and figure out roughly where it is so say i get a little bit too comfortable in here and i kind of move over to the side it knows because it has this probe move back move back move back go ahead all right were back in business how do you tell a robot what haircut you want the way that i solved this problem is i took a d model of a head and i painted on it the haircut that i want basically lighter is longer darker shorter at first glance this looks like i want to be a balding man but it actually makes sense to the robot the other thing i have to tell the robot is what angle to cut at because i wanted to cut my bangs horizontally and i wanted to cut around my ears so this isnt telling the robot that i want a giant unibrow this time the robot cut the bangs horizontally and then cut around my ear the robot will load this d model and then knowing how big of an area the scissors cut it will compute a plan of the minimum number of locations it needs to cut at to make my hair the desired length and then as its doing that plan itll cut one and then go the next one in the next one at each location it will check the unibrow model to see if it should cut at a specific angle and if so itll do that at least thats the plan in theory i havent done it yet hopefully it works but were doing it live so heres the whole haircut from start to finish heads up i look completely insane in this time lapse i think its a weird time lapse artifact but also i was terrified of it ruining my hair so its probably also a bit of fear midway through i realized there was a bug in my code that made it cut about six times as many locations on my head as it needed to i thought it was taking a while so it took about an hour but it should only take more like minutes every time the robot goes to cut a location it probes to find where my head is at and then the length that it cuts is relative to where it probed this is the main reason it could maintain accurate hair lengths even though i was moving around so you can see the vacuum sucking a bunch of extra hair into the tube but the fingers that grab the hair have little guards on them that keep hair that they arent grabbing from being cut making this robot hit the desired location on my head was pretty hard not only does it have a bunch of joints that are somewhat challenging to compute where the heck the scissors are going to go my head is moving and then the design of the robot doesnt allow it to approach my head from all possible angles i have to find the best achievable angle that the robot can do i have a suspicion my math isnt always correct but thankfully my hair all turned out the right length all right the cut is done what did the salons first customer have to say all right two out of ten not great zero personality i can believe that top and sides are cut well nice nice but it gave me a mullet what the robot cut the hairs that it could reach really well but theres two problems the first is it cant get close enough to my ear so i have all this hair right by my ears thats really long that should be pretty easy and fast to clean up but the big problem is it gave me a robotically perfect mullet so it just couldnt get quite low enough to cut the last maybe inch of the back of my hair i had planned to have it rotate which dropped it by about an inch maybe an inch and a half but the scissors interfere with the stand and it just couldnt do it so i said i thought it was gonna give me a five dollar haircut i think that was right i think even for five dollars they dont normally give you a mullet man thank goodness i didnt prank my wife with this machine because i had to just go give her a pair of scissors and ask her to fix my hair im actually really happy it gave me the mullet theres just something deeply satisfying about a robot giving me a perfect mullet all she did was cut those off she didnt touch the top at all and overall id say its really not bad i checked some of the lengths they seem right on and id pay five dollars for this haircut i dont get another shot at cutting my hair for a month or two but i think i might try to upgrade this to cut closer to the ears to not generate perfect mullets and then maybe also add a trimmer or something like that so that it can do the neckline for me so hopefully you enjoyed this theres a fun project thanks for watching you

write a youtube script

this is not a normal golf club obviously it can do two very special things the first is that its all the golf clubs in one say i wanted this to be an eight iron all i do is turn this knob boom now its an eight iron maybe i want it to be a two iron turn the knob now its two iron and thats really cool but thats not even the coolest thing that this club does i have a problem where i cant swing the golf club consistently at all id really rather say to the club i would like the ball to go maybe yards and then when i swing the club it will detect whats going on and then correct the head of the club in real time to make my shot go closer to the desired distance so im going to show you how i made this how it works and of course put it to the test ill also be challenging my wife to a shipping competition so is this not a prank its a golf club how could it possibly be prank this project totally kicked my engineering butt it took two major hardware iterations to get it right the first club that i built was actually hydraulic which sounds totally awesome but also doesnt work for reasons that ill get into so in this design i focus on adjusting the loft angle of the club which is the angle of the face striking the ball what im doing is im sensing where the club is going as well as the speed and adjusting the loft angle so that the ball goes basically the right distance i also made this i guess apparatus youd call it it goes on to the upper shaft of the golf club near the grip and allows me to rotate the entire face of the club this should allow me to correct for slice and hook i do plan to integrate this in a club ill probably make a future video about that although in the meantime im really interested in the idea of human robot augmentation not just for being sports i think the next frontier for me is music i want to see if i can make myself good at violin and take the fun out of that as well so if that sounds interesting you should subscribe also if you have any other good ideas about human robot augmentation leave a comment id love to hear about them and maybe ill build them getting the face of the club to move to a different angle at least in principle is not very hard certainly nowhere near as difficult as the backboard that i built in my previous project its small and light and the distance that i have to go is not very far but it turns out its very hard to do and thats because of golf when the ball strikes the face of the club it is insanely violent its like walmart at am on black friday they only have three flat screen tvs but theyre its bad it can generate up to pounds of force as a point of reference my milling machine is pounds its crazy so this is a little hobby servo its used for rc planes and cars and stuff like that this little arm moves left and right and theyre very handy i could attach this to the face of the golf club and move it around to get it to go where i want problem is imagine i then take a hammer and i smash this arm its gonna destroy this so my first attempt to deal with these giant forces was building a hydraulic golf club which isnt as insane as it sounds maybe a little bit i needed a way to move the face angle of the club quickly and also in a way thats very stiff and can take a beating i designed and made two little hydraulic pumps that can pump fluid through these hoses this is obviously not yet miniaturized for the golf club id probably be wearing this on my back the hydraulic fluid goes through these tubes and actuates two hydraulic cylinders that i built into the club so unfortunately this club has a pretty fatal flaw which is that it breaks apart every time you use it when i originally designed this system i was planning on using proper hydraulic line so when i ran the numbers i saw that i would basically need two of these engines to force the fluid through the hoses fast enough imagine trying to suck a gallon milkshake through a straw in seconds youd have to be able to suck pretty darn hard in order for that to happen but if you make the straw bigger it becomes easier to do that so to make the system work i bumped the size of the hoses up and this is where i committed a terrible sin that totally perverted the design but i knew what i was doing i honestly even liked it these hoses are stretchy theyre very stretchy if you pressurize them they get longer and thats a problem for a hydraulic system where im generating high pressures and i need it to not move when i swing the club and hit a ball i dont feel the impact of that ball with the club until the ball has already left the club and is flying away and this is due to the speed of sound when i strike a golf ball with a golf club it generates a pressure wave in the material which as far as physics are concerned is basically sound it goes at the speed of sound up the golf club in microseconds the sound will go to maybe about here and this is what i was hoping would make my hydraulic system work even though the hoses were stretchy the shock wave from hitting the ball with this club would generate pressure in the fluid which would also go at the speed of sound and it should only go not too far up the hose before the ball is gone theres only this much hose that can stretch out and i was hoping that would mean the face would move a minimal amount and that would be acceptable the problem as far as i can tell is that that wave does propagate up the hoses the system is very springy so all that energy goes up and then it shoots right back down and launches the face of the club off to try to deal with this i did get some proper hydraulic hose with the right inside diameter but this is just way too heavy i cant put this on a golf club i decided just to scrap this which hurt me to do is very sad i really like this club a lot its very cool looking it just isnt good enough all right i think i have a conceptual way for how to make an actuator that can survive the forces of golfing without immediately breaking but i am having a heck of a time trying to fit into a club this is probably iteration five or six of this design and this one almost kind of works but it still has some pretty fatal flaws i think i just keep chugging i think im close Music Music okay so heres what i came up with im using a servo a really fast one to actuate the face of the golf club and i know i said before that servo motors cant work but this one works because im doing a very sneaky trick im rotating the face of the golf club by moving this rotational cam here the geometry is such that it does not matter how hard i hit the face of this golf club with the golf ball zero force is going to be transferred to the servo and let me explain why this red triangle represents the cam thats moved left and right by the servo this blue represents the matching cam thats thats rotated when this moves in and out so if this red cam slides out the blue cam slides down and vice versa this is how it converts linear motion into rotational motion the reason that my arrangement works is that this angle is very shallow i can push on the top of this green wedge with any amount of force and it will not be able to push this wedge out basically what happens is as this angle gets shallow the contact forces between the two surfaces increases which increases the friction also when i push down on the top of this wedge portion of the force is going sideways trying to eject this wedge from underneath of it theres an angle the force to eject the wedge is less than the frictional force and it does not matter how much force i put on the top it will not eject a neat little parallel is this is why screws dont unscrew if you think about it a screw is just a wedge thats wrapped around a shaft if the thread is shallow enough it doesnt matter how hard i push or pull on the screw it wont unthread so the other cool thing about this design is that it was fully qualified using almost entirely d printed parts i did this because i didnt want to spend a bunch of time machining a lot of parts that dont work at all so in fact i normally wouldnt even have attempted this but im using a very special material called durable its designed for high impact and high wear applications which is great for a golf club i did eventually break it so it is cracked here and that happened when i plowed the golf club into the ground at what felt like mach but well never know the speed for years i ran the engineering teams that developed these printers and i hated this material because it is really difficult to get it printing reliably and so i just had a blind hatred for it but now that im using it its quite good so after qualifying the parts with printed parts i went back and machined these although i can use the club with the printed parts they do flex a good bit more than metal does and that absorbs a lot of energy from the swing you can you can actually feel a pretty big difference im still using d printed addons for the stuff that doesnt have to be metal and that way i dont have to machine it the cam parts are still printed and thats basically because they would be a real nightmare to machine and with my setup theres nothing really inherently bulky about this design which is pretty cool if i made these out of say steel this whole mechanism could be way smaller than it is here the same thing is true for this servo theres no reason it needs to be this big and all the wires could be run up the shaft all the electronics would be on could be on one small pcb kind of in the top of the handle with a small battery so i finally have mechanics that i think can work i just have to get the software and the electronics working so i was just working on bringing up the electronics for this and youre not going to believe what i just did all right pins just straighten up a little bit thank goodness im a mature adult who can handle this in a very mature way i promise thats exactly how it went down so this board is the imu its what ill be using to figure out where the club is in space and how fast its going and all that stuff i just ripped off these cancellations which are what i was going to be using to get data off this board and this board is very expensive its about so im not going to be buying another one theres usb on this as well theoretically i should be able to talk to this over usb and get the data off that way unfortunately i suck at usb like i said im not buying another one so i guess im writing a usb driver oh man its like am on sunday morning ive been banging my head against this stupid usb thing for like hours now and i should have just bought another imu this is a mistake the reason that it has been working is because i had a one instead of a four anyway it works now and im going to sleep all right i got some sleep lets talk about the software what the software needs to do is determine the speed that the club head is going to be going when it impacts the ball as well as the angle of the club shaft relative to the ground its not a good assumption that the club shaft is vertical i need to know the angle of the club shaft so the club face is the right angle relative to the world and the ground in gravity so that the ball will go in the right direction if you look at a club swing the club starts at a standstill it accelerates throughout the stroke and it will sometimes even drop speed before it hits the ball when i start the swing the club isnt going very fast as i accelerate it the speed gets faster and faster the reason this is tricky is that i need to rotate the face of the club to whatever angle it needs to go before it strikes the ball and the speed is changing all the time so what im trying to do is predict the speed that the face of the club will be going relative to the ball earlier so i did as i just kind of brute forced this problem i collected a ton of data of me swinging each swing gives me a curve of the velocity im going versus the angle of the club so this might be the start of the stroke up here and then this is the velocity thats going when it strikes the ball swings that start shallower like a chip might look like this because the club is starting at a lower angle im at zero velocity here you can imagine collecting lots of these and theres noise so what i do is i take all these curves and i compute the average swing trajectory for each starting position this one might correspond to this position and so on then when the club senses me swinging i look up the closest curve so lets imagine that im taking a swing and ive matched to this swing profile the tricky thing is that depending on if im swinging the club harder or softer i might get a profile like this or i might get a profile like this i also might get something totally different in which case this method breaks down i have some ideas for how i would deal with that but i was trying to keep it simpler so imagine i start taking a swing and its going to follow this trajectory as the data starts to come in from the imu i see what my velocity profile looks like i then take the ideal velocity curve and i scale it so that it matches this one it doesnt perfectly predict the result because theyre different trajectories although the shapes do seem to be relatively consistent and then i look forward to the end and assume thats the speed that im going to be going this seems to work reasonably for my testing its tuned to me obviously if i was trying to make this a product i wouldnt use this method but it was a quick way to to get it done so the other problem that i mentioned is i dont know what angle the club is going to be when it strikes the ball which of these is it going to be this has a very big impact on where the ball goes this problem is a tricky one and i actually thought of a few very complicated schemes to try to predict what the angle would be so i just decided to sidestep the whole problem by assuming that the swing starts at the ball so i just assume that if the club is like this when the swing starts and then the swing comes up and then comes back down its gonna hit at the same angle thats not a totally safe assumption it seems okay for now thats pretty much all the interesting things about this club i did pack all the electronics into this way too small enclosure so this thing has two main operating modes theres distance select where i enter the distance that i would like the ball to go and whenever its in this mode its continuously listening to the imu and its looking for a swing it will start adjusting the face of the club to wherever it thinks it needs to be so the ball will go the correct distance and the other mode is basically this acting like a normal club so i have the ability to select whatever club i want with this i think the notable thing here is it does go up to i mean it has to right i guess that would be a sand wedge even just this mode here is pretty cool the fact that i have a single club thats all the clubs is neat but i think the robotic aspect is cooler thats pretty much it i think all thats left is to put this thing through its paces ive challenged my life to a friendly shipping challenge and what were gonna do is take three shots each and see who can place them most consistently downrange short distance so no strength advantage anything like that and were going to see with the help of this if it levels the playing field so is this another prank its a golf club how could it possibly be a prank this club stinks maybe you stink i just played holes of a par with this club it survived and i have the results with this club i shot in which still isnt great and with the regular clubs i shot a and that sounds like a meaningful difference but im not sure that all that had to do with this club at least not directly when im using this club i feel like a better golfer and i was just hitting the ball better with it to really know i need to collect more data i think what i need to do is shoot like balls with them without this club and then measure the distribution but id really like to get the slice and hook correction in before i do that because i only want to do that once hopefully well get back to it and have a super duper club in the near f

write a youtube script

ive made two basketball hoops that seriously improve your game theyre fun but they both have this fatal flaw the ball doesnt go in if you entirely miss the hoop and i mean come on its . i shouldnt have to be good at anything so the problem is that the ball is over here and the hoop is over here im building a hoop that always goes to where the ball is going so that even if you miss entirely its going to go in even if you shoot blindfolded heck blindfolded backwards in the dark doesnt matter and like my previous hoop theres unlimited opportunities to torment my wife all right heres my game plan im going to point this connect which is a d sensing camera at this giant wall im going to write software that will track whenever i throw a ball and it will figure out where the ball is going to hit on the wall im going to build a basketball hoop that i can move wherever the ball is going to go and its going to do all that between the time the ball leaves my hand and strikes the wall i know i say this every time but this was a really hard project it just did not want to work everything that could break broke it just really was a lot harder than i was anticipating oh my gosh Music how i intend this all to come together will be a lot easier to explain with the actual hoop so lets get it made there are tons of plasma cut and folded sheet metal parts on this machine its just so easy and fast thats half of the rail for the xaxis these are how i make one long carbon fiber rod out of two all the wheels and pulleys are printed on a form three these are composite pulleys so the part that needs to be strong is machined and then the rest is d printed french cleat is awesome you can rearrange your shop in just a few minutes turmockr coming in handy normally two rails this far apart would be what you do when you hate yourself but my mounts are springloaded so alignment isnt very critical it is not supposed to do that or that the system was having a tough day this thing is gigantic it looked so much smaller on the computer its also kind of a weird design like my dog is designed for one thing and one thing only speed the design is optimized so the minimum amount of stuff moves and its as light as possible ill be making an extremely light composite backboard and hoop which will be mounted on this ridiculously over engineered super light frame which is on these incredibly light carbon fiber tubes which are driven by belts from these extremely heavy but stationary motors heres a really quick demo to give you an intuitive feel for why this matters i built this little cart it has bearings so very low friction its gonna be pulled by a constant force by this belt which is attached to a bucket with weights in it heres the car with five pounds on it and heres the car with pounds on it this is why lighter is so much better for this system this little cart is also the worlds best banana peel hey come check this out oh no watch out these motors im using have incredible torque which means they can yank on these belts super hard and super fast which is going to accelerate the hoop like a bullet and things are probably going to get pretty violent im really worried about breaking a belt especially this one that i spliced together like a total noob i just couldnt get one long enough theres another interesting complication on this robot which is tilting the hoop and you might be wondering why do i need to tilt the hoop i can move it anywhere on the wall shouldnt i be able to direct any shot in i actually cant theres a variety of shots that wont go in no matter where i place the hoop if i throw the ball really hard at the hoop its not going to fall down into the hoop if i want the line drive to go into the hoop i need to tilt it down so that it directs the ball down into the hoop getting the system to tilt is responsible for a lot of these ugly belts the reason i have all these belts is to keep my pants up otherwise theyll fall down and what in the world am i reading he asked me for help with his script what was he thinking the real reason i have all these belts is because i need the motors to be stationary theyre way too heavy to move around quickly so im transferring all their power with belts to move the hoop i have one continuous belt this is the really long one thats just waiting to explode in my face its interesting because the final position of the cart is a combination of the position of both motors for example if i want to move the cart to the right i rotate both motors in opposite directions the backboard has two belts attached to it one at the top and one at the bottom to make it tilt you pull on one of them and push on the other the tilt motor follows this tshaped path and connects to the second motor up top the tricky thing about this arrangement is that the angle of the hoop is coupled to the position of the cart if i pull the car up the hoop points down and if i pull the car down the hoop points up and if i take things too far stuffs gonna break if i move the tilt motor it tilts and if i move the position motors it also tilts so what i have to do is drive them together in the same direction and it wont tilt if i want it to tilt ill basically drive the belts different amounts and then thatll give me a resulting tilt having all three motor positions coupled in this way is a bit annoying but it allows me to build the system in a very light way figuring out how to rotate the motors isnt as bad as you might think you can write it all out as a pretty simple system of linear equations and then you can write that as a matrix and solve it directly without any equation manipulation if youre wondering why you might want to know linear algebra this is one of the very useful applications of it theres some other pretty cool things about this design every pulley and idler is d printed i didnt buy any most of the construction is sheet metal which is really fast to make the sliding frame is stiffened by these tensioned steel cables they make it several thousand times stiffer than it would be without them i have these spring loaded followers so i dont have to line the rails precisely and a bunch of stuff im dom time to get into the ball tracking software works in a similar way to my previous hoop if you want to see how it works you can check out the previous video i did rewrite it to fix some of the bigger issues and it did take me several days of straight programming too bad it makes for boring video all right its got my software on here its draft one gonna give it just a five millimeter move to make sure everythings working lets try again oh what the heck the belt was a bit loose because i was afraid of breaking my terrible splice and that let it jump off the pulley all right everythings back together lets try the five millimeter move again Music exactly what i was afraid is gonna happen just happened the belt ripped apart where i stitched it together hopefully i can make one that will survive otherwise totally hosed all right it only took me seven tries im getting pretty good at this i was in integration hell fixing all the software and mechanical system issues for a long time for days there were a lot of problems the wounds are still fresh so were gonna just kind of gloss over this were especially not gonna talk about how the cart would follow an lshape rather than a diagonal line and how i walked over eight miles in my own house going up and down the stairs trying to fix it its done and thats all that matters it took quite a while to get here but lets just move it to see if it works looks good enough to me im going to sleep it is really late i want to test the system out for real throwing balls but theres this tricky problem which is that the connect which is the d camera its mounted up here its looking out at the hoop in the room but it doesnt actually know which way is up or how its angled or where the wall is or where the hoop is all it tells me is that this point in the picture is this far from the camera to predict where the ball is going to go and where to move the hoop i have to know where the connect is relative to everything to deal with this issue i made a calibration program it takes an image from the connect this represents how far everything it sees is from the camera so what you do is you just paint where the wall is so you do the same thing for the floor and then you select where the center of the backboard base is so now the hoop knows where the connect is relative to the wall and everythings gonna be very accurate i wrote this program to avoid measuring everything by hand which probably would have taken about two hours maybe this took a whole day so the big question is if this is worth it or not oh come on is that even a question all right its time to test it out it says it hit there which it did so why no moving i think i hear it moving its moving really slowly all right my calibration thought that a millimeter was a meter and that the world was upside down and a few other things that happen when you write code like this but i think i have it all sorted now sweet how cool is that seeing this thing intercept the ball is like the best feeling ever but its only half the system i still need a backboard that can tilt with a rim on it ive been testing without it because the way the belts interact i didnt want to be crashing all the time with the hoop so ive just put off making it lets get that made and then take it to the next level im making it a super light fiberglass and foam i machine the backboard core on my tormach r it makes it really easy to get these inserts in the right spots Music this backboard turned out awesome its foam core with fiberglass reinforcement that makes it really really stiff but light the reason i made this aluminum rim is that the stock one is incredibly heavy steel and this is about four to five times lighter than the backboard plus rim if that isnt the coolest thing youve ever seen wow this thing is so much fun ive gotta have my wife try this out hey wife all right here to be pranked why would i prank you yeah is it not supposed to be like that oh yeah what Music all right give it a shot that definitely had to be a fluke Music ah it all makes sense now so its doing the cross product yeah thats basically what does it i thought she liked programming you know what even losing this hoop is a pleasure now i can really make it rain the soup is so cool it is really satisfying to see it working for a while there i thought it wasnt going to work sometimes it just does completely wrong thing it sometimes thinks my head is a ball i need to look into the code because my optimizer keeps trying to switch the shot rather than bounce it off the backboard which isnt the best strategy because its more sensitive to error in my estimated trajectory in other words it hits the rim backwards blindfolded in the dark Music for my next video im planning to make the next version of this bat i really want to break the record and i have a metal bat which should help me withstand the stresses and i talked a lot about powering it with three shells then i realized i have four shells so why not four one more should be really good twice as powerful as before so thats the plan if you like what i do you should consider subscribing it helps me out and it also helps you out because youll get notified when i have new videos like this back the other way you can help out is by checking out any sponsors i have on the videos if i have any so i was looking at my head and it reminded me of the sponsor todays video keeps you may think my haircut robots your only risk of hair loss but youd be wrong im not going bald me either well the stats would really beg to differ two out of the three of us will probably have some hair loss by the time were . what then what are we gonna do the best thing to do is get ahead of the problem while you still have hair arent the treatments super expensive i spend all my money on tools well youre in luck because keeps offers generic versions of the only two fda approved hair loss treatments so even if you spent all your money on tools you can still afford it its also super easy to get you can visit a doctor online and theyll mail it directly to your house so theres this old saying that an ounce of prevention is worth a pound of hair because i think thats how it goes my point is that its a lot easier to prevent hair loss when you still have hair and it can take four to six months for keeps to start showing results so you should act now while you still got some hairs left on your head so if youre ready to take action to prevent hair loss go to keeps.com stuffmade here or click on the link in description to get off your first order thats keeps dot com slash stuff made here

write a youtube script

ive been wanting to hit a ball further than the baseball distance record for quite a while in another video i made this explosive bat which had all the right ingredients to do it it would just constantly break when i pushed it to higher powers i was only able to reliably hit a ball about feet which is a major league home run but its not setting any records thats just unacceptable allow me to introduce explosive bat v and explosive bat v which is a double barrel caliber baseball bat its powered by caliber rifle blanks i dont really see reason why this wouldnt be regulation this is the type of thing where you do it once and they make a rule just for you this bat is insane and for the first time were gonna get to see some awesome slomo of how it works and finally firing the wooden bat with three shells so feet thats the longest verified home run feet is all the way out of the stadium ive also decided that it only counts if i swing the bat for real going into this i was thinking i was gonna easy but as i got into the math it turns out feet is really hard the forces involved are insane this is what happens to my wooden back when you try to fire it with three shells more on that later im right up against the edge of whats possible and just in making this thing i crossed over that edge over and over and over Music but through those fires and flames emerged something really cool im not sure what to call it though explosive bat v is such a mouthful lets just call it the . i like how that sounds after my last bat video about million people said i should do three shells so i know its disappointing everyone wanted me to do three shells but i didnt i did four which is one more so let me break down my plan for beating the record so for step one im gonna hit the ball really really hard thats about it im gonna need a much stronger bat oh this is perfect so like i said this bat is gonna be powered by caliber blanks these are normally used for driving long nails into solid concrete and steel i took this opportunity to totally rethink the bat ive been obsessed with the idea of giving the bat just the tiniest tap and still hitting a home run i call it the four gun salute the entire end of the bat moves and it has four individual firing mechanisms which actuate this four bar linkage the force of the ball hitting the bat will fire all the shells my new bat is so much easier to load than my old one you just open the breech load in your shells up to four which is pretty scary actually and check this out oh yeah i am so excited about this all right lets put it to the test i have my stand which lets me drop a ball onto a bat so i can try it out without putting myself in danger it was this exact moment i realized i majorly screwed up this bat is so much worse than my last one it could probably break my wrists which is a little scary and check this out im so not excited about this bat its too bad this thing is a complete piece of garbage because theres a lot of cool stuff to talk about if it had worked im especially not going to get into the benefits of four separate firing mechanisms or the iterations of the firing pins before i could finally get it to fire reliably theres a really basic physics problem for this to get launched up really fast and hard there has to be an equal and opposite reaction forcing the handle down this also generates a twisting force which is what snapped the welds and because my hands are also bolted to the bat theyll also get twisted and snapped thats a problem i should have done more engineering analysis but i just wanted it too much and i thought maybe for once id be able to full nature but it never works good night sweet prince i started an entirely new design from the ground up this is the way i should have done it from the beginning and the basic idea is that im going to take a slice out of an aluminum bat and put this firing mechanism in there and what will happen is when a baseball hits the top this piece moves it will fire the shells and then fly out super fast launching the baseball so heres what goes on inside the bat if we peel back the outer layer of the steel its kind of complicated looking theres an assembly that can move like this on this assembly theres two firing pins here and here and theres some blank cartridges down here so when a ball comes in and hits this plate this mechanism will move down and these firing pins will set off these cartridges which will release a ton of gas which will shoot down this tube and then up these two tubes thatll shoot these up super fast because the pressure is really really high and that will send the ball flying when this flying piston gets to about here these rods will come into contact with these blue springs theyre super super stiff springs and then theyll start to compress them which will slow this piston down so that it doesnt shoot out im just going to give you a little taste of what it was like to make this it was pretty intense a lot of machining almost everything was machined which means lots of tools some of which gave me some really annoying problems this is the most complicated part the chamber the pistons are a very precise fit as is the seal on the bottom that connects the chambers probably of the time on this project was machining parts it just feels morally wrong to be doing this to a bat kind of felt good too though this is how i mount everything to the bat with these welded on caps i had some heat treated parts because they need to be very hard so they dont deform theres the firing pins hopefully you can now go make my bat except dont i think it goes without saying but dont try this at home i know i sometimes pretend to be careless for laughs but this is really serious business i do an enormous amount of precautions from engineering analysis to testing to wearing body armor and its no joke doing this could kill you so just live vicariously through me and dont do it alright lets give it a shot Applause so this is exactly what i was afraid was going to happen it blew the piston out of the bat and it also broke the test stand i found the piece of the test stand hey what was that noise what noise this is a good point to talk about the hard parts of this design there were two really big really important problems to get right the first and most important was how do i keep this thing from exploding in my face from the extreme pressures of detonating these cartridges this is particularly important because i plan on holding this bat surviving these insane pressures is one of the reasons i have two smaller bores instead of one big one you can think about the pressure and the bores pushing out in every direction and the worst thing that the pressure does is it tries to split this in half so you think of all the pressure acting along this surface it adds up to trying to pull the whole thing apart so the material is having to pull itself together to keep it from getting ripped apart and as this hole gets bigger the force is trying to rip the material in half also get bigger and the only way i can make it not go pop as i do that is to make these walls thicker and thicker so i want the holes to be smaller but the problem is as the holes get smaller the force pushing the pistons out also gets smaller so its a very delicate balancing act to go as small as i can while still getting the acceleration and force to launch the ball that i need by the way this phenomena of ripping the material apart is called hoop stress and its the main thing that sets the thickness of propane tanks and pipes and stuff like that having two bores also keeps things from twisting and it distributes the force so the pushing plate can be wider without bending the second big challenge with this bat is that the moving end will be going up to miles per hour if it works and my preference is to not shoot it out with the ball so i have to stop it with stop rods which compress these really stiff springs at the end of travel and the force generated by this is unreal so for my bat the forces are going to be up to pounds just top the piston and its all gonna go straight into these stop rods and springs thats three tons thats almost the combined weight of both my cars and thats normally what you apply to a steel rod when you snap it in half i think these stop rods broke because i used a higher carbon steel which hardened when i welded it i made some with better steel lets try them looks good lets try two shells welcome to hell let me make things lighter which will reduce the force Music that made things worse lets try forged rods oops i didnt realize my dog was down here also they didnt work lets try different springs Music hey something different we didnt just break the rods i also broke the back ive got better welds better springs high grade steel indestructible stops oh finally this thing has so much power i love it heres the knife edge that im walking i need these pistons to accelerate for as long as possible the longer they accelerate the faster they go as soon as they hit the springs they start slowing down and that sets the top speed i want the springs to be as short as possible so i get the maximum travel before i begin braking but if the springs are too short then my deceleration will be very harsh which will snap the stop rods this is why the ball goes faster when the stop rods break theres no springs to slow it down so it gets to accelerate longer have you ever wondered how to touch your wrist with all fingers at the same time its really easy just have the back go off while youre pushing like this heres what i came up with to keep my fingers safe ive got the pusher if the system blows theres no appendages in the path and theres no loops around my fingers to deglove them if you dont know what degloving is dont look it up even more important is this clipon safety amongst friends thats referred to as the safety boy it clips onto the pistons and completely blocks the firing pins it stays in the bat until its ready to be used that way if you drop the bat or something nothing happens this is how i keep my fingers pointing the right way so i think its time for me to test this thing and im a bit worried about the kick i mean just look at how its slamming the wood but i dont know how im gonna get more information than i already have so im just gonna go for it im not scared i mean seriously what do i have to be scared of look im not scared why would i be scared its totally fine its a little cold down here Music oh wow this bat feels awesome theres very little kick actually it looks like im getting about to miles per hour off the tee with one shell which is a real improvement over the previous bat all right lets see how far we can hit outside looks like im getting about feet off one cartridge see how two does that was a terrible hit and that was really loud sound like a gun that sound was the dying breath of the breach the breech is a piece of steel that sits behind the shells and it keeps them from popping out of the chamber when you fire them the pressure of firing them wants to shoot them out really hard and i didnt feel like properly designing this part so i did the old engineering trick of just making it out of thick steel and that usually works this is the first time thats ever failed me so to fit any more steel in here im gonna have to redesign and remake this top end i hate doing things the right way but theres no way its gonna survive at three and four shells if i dont do this all right new design for the top piece made and machined it lets me have a solid half inch steel breech i did my homework this time and this should definitely probably not bend it is now a.m and this is the only time i can attempt a or foot hit at this field because i dont want to worry about foot foul balls and people so this totally sucks i take one hit and the bat broke it was also the wimpiest hit ever it went like feet i think you see the ball right there oh man the damage is worse than i thought so heres what i think happened a ball came in and just barely hit the corner or edge or something which was enough to fire the system as the system started moving it just kind of pushed the ball to the side and then it was free to shoot out like a rocket this made it slam into the stop springs much harder so they were pulling down super hard this way which bent it into this very annoying potato chip shape i would really strongly prefer it not bend in half like this whenever i have a bad hit so that means i get to redesign and rebuild the top end again i think you know the drill at this point ive made so many of these overall ive made more than custom machine parts for these bats i hope i can get this working because it would be a huge waste of time although lets be real its a waste of time anyway all right were going all the way to four shelves just to see what happens one shell piece of cake is always heres two looking good i havent done three yet because it keeps breaking but its surviving many times were all this is fantastic Music now four oh my gosh the power is unbelievable but unfortunately yeah it broke but just look at that kick oh man i really want to fire this four shells shattered my grade . volts which is good that means im getting a lot more force than i thought i was where did the screw heads go found them man those must have gone flying and then i had kind of a big brain moment which is im putting all this effort into making these stop rods not break but the stop rods cant break if there are no stop rods right all right its time to pull out all the stops this bypasses the muffler and makes it so much more satisfying to use a good friend is letting me use his property and this is really important because the whole no stop rod situation id prefer the piston not go flying feet so ive tied it to this bucket with kevlar line its like a little parachute heres how far the ball goes without any charges so we did a little horse and round with one shell and then we gave it three that was pretty good yeah you can say that again that was pretty good so theyre claiming the ball went all the way over the trees into the neighbors yard which is incredible there it is feet but it only counts if i swing it so time to get suited up and give it a whack foul ball the foul ball was actually my piston flying into the woods so you know what that means i get to remake it feet is incredible and it totally obliterated babe ruths record and it made me realize im really close to breaking another record there are these guys down in alabama i think some guy named dusty or something they made this bat spinning machine that was hitting the ball they say feet i think i can beat that if i can tune the system the problem im having is i cant see whats going on and what weve been planning this for weeks dude what are you talking about lets go do it Music this is like holding a ferrari well well lets combine our powers and uh lets just do awesome things so were using destins high speed camera to figure out whats actually going on and then use that knowledge to do even more awesome things Laughter oh yeah like cheap chinese fireworks is what that smells like all right the first real hit really cool you can see the stop springs bouncing and this is not how i expect the ball to interact with the bat its more like launching it so a normal collision is like microseconds or something this is three three and a half milliseconds of contact its an eternity there are organisms that lived and died in that amount of time the thing i was most excited to see was four shells destroying everything that was really cool but where was the destruction oh there it is destroyed the stop springs i know where to get some destruction lets hit the wooden bat three two one wait should we give another countdown oh my gosh dude come on dude stuck all right oh golly its like by a thread three two one okay one okay Music i am so glad i didnt try swinging this bat with three shells thats the first time you fired three in the back really thats why i didnt fire with me in the back so i had a good feeling this is what was gonna happen this video was already too long for me to get into the weeds of all the stuff we saw and everything we learned but a lot of cool stuff destin is going to be making a video going more into the weeds that im just glossing over here so you should check it out on smart everyday if you are one of the three people who doesnt know what that is hes diving into the details of what happens when a ball hits the bat its not really a bounce its weird when i first met destin it just did not compute i was like how can someone be this nice and cool somethings off here what is he really up to but it turns out thats just how he really is so go check out his stuff and youre not gonna regret it the last thing i really wanted to see is four shells with no stops whoa dude thats pretty cool my people have called me holy cow were out here trying to beat babe ruths record and destins record by hand it is difficult because i am not very good at swinging a bat which the high speed makes extremely clear i think the technical term for this is loft have you heard of that term no it stands for watch ill show you right here which stands for lack of freaking talent oh yeah oh my goodness im terrified of this thing im too much yeah because i know the guy that made it because you know its gonna crush your record oh is that what were doing were in the smack talk area now yeah youre pretty tough when you get body armor i see why youd be afraid of it we got a couple alright hits then we got this hit which flew forever and then this quad was about milliseconds from getting completely pulverized which would have been awesome the control system on these quads is just unbelievable how this didnt crash is just spectacular so this hit looked amazing it was flying over the trees it looked at least as good as the other footer unfortunately we dont have a camera behind me because destin was too busy trying to crash my quad and dealing with a high speed and i was too busy trying not to die we did see where it went the only problem is it landed in this field this is like finding a spherical needle and the worlds largest haystack i searched with the quad i searched on foot i searched for hours and i thought for sure this was it thats some stupid mushroom the bat was still working i would destroy this mushroom been out here for three four hours this is the ball thats just not meant to be found im sure i broke babe ruths record because it went over the road which is right at the record line but did i beat dustin i just dont know and its bugging me you can come to alabama would you consider that we can fire the uh supersonic baseball cannon at it well yeah duh what were you i mean like what were you even talking about i mean just think about it mad batter this this plus supersonic baseball cannon all at the same time yeah yeah we should do that what happened we italicized it oh dude so we have the letter pi now yeah are we done unless we can find the other one i just realized that this would be so easy to turn into a golf club let me know if you think thatd be worth doing these videos are very costly to make and take an enormous amount of time if you enjoyed this video please consider subscribing it really helps me out and it also helps you out because youll be in the loop when i post new stuff if youd like to support these videos directly you can check out my patreon and the final way you can help out is by checking out any video sponsors Music heres the part where i tell you about this videos sponsor kiwico i think a lot of you are going to actually really like this its a good thing the number one question i get asked is where did i learn to do all the things required to build my projects and my answer kind of surprises people because i think they expect me to tell them what college i went to the truth is a lot of what i know how to do most of it i learned outside of school and in fact it started when i was about four years old or so my dad is awesome and he gave me all kinds of opportunities to get my hands dirty and build things and make things for example when i was five he gave me this kit which is just some components and a board and a soldering iron and he said go to town make it and then he left me alone in the garage i i think he remembers a bit more supervision but nah and these were incredibly formative experiences that gave me confidence and abilities to then make more and more complicated things as i got older i was even making videos about them this is my energy efficient house proposal from i think the third grade i think my video skills have gotten better but you can be the judge kiwi co is a service that sends you a project crate every month with everything that you need to make it its like a steady drip of selfdirected handson learning so im about to tell you may shock you i have a kid a daughter to be exact shes still a little bit young for the hydraulic claw but i think she could be there any day what do you think will happen if i put the big syringe in instead of the little one exactly maybe shes not into mechanical engineering thats fine theres eight different lines for different age groups and interests not only are the instructions topnotch but theres an entire magazine made for every single project that breaks down whats going on it draws connections to the real world and then theres a bunch of different projects like building a hydraulic elevator the fact that you have all the pieces to build other projects is so cool i attached my welder to the end of this arm and i almost weld it awesome can you imagine the amount of skill you would develop from a couple years of building projects like this every month and the world needs more scientists and engineers and just people who know how to get their hands dirty and make and fix things i can tell you from experience the best thing you can do to expand your kids mind and get them comfortable with technology is to give them handson learning with kits like these so whether its for your kids a relative or honestly yourself give the gift of knowledge and bask in the joy of knowing you gave a gift thats going to have long term value and leave a mark im super grateful to have been exposed to this type of stuff growing up go to kiwico.com stuff made here and youll get off your first crate and dont forget the holiday season is this close and these make really good gifts not only do they not rot the brain it keeps giving all year long so thats pretty cool if you do this you can also feel good because it helps support me and the channel and it enables me to build increasingly bizarre and insane things so thank you for that you