Glossary
Behavior cloning
Train a model to predict the action a human would take, given the same observation.
Behavior cloning is the simplest form of imitation learning: given a dataset of (observation, action) pairs collected from human demonstrations, train a model to predict the action from the observation. That's it — a supervised learning problem.
Behavior cloning works surprisingly well when the dataset is good, but it suffers from compounding errors: a small mistake takes the robot off the data distribution, the next action is more wrong, and the trajectory diverges. Modern policies (Diffusion Policy, ACT, VLAs) are all extensions of behavior cloning with structure that reduces this problem.