num_train_frames: 500_010
batch_size: 32
batch_length: 32
agent.imag_reward_fn: video_text_reward
eval_modality: task_imag