Theia: Distilling Diverse Vision Foundation Models for Robot Learning