Behavior Cloning

Learn robot control policies by directly imitating expert demonstrations, treating the problem as supervised learning: map observations to actions. Behavior cloning frames robot learning as a supervised problem: given a dataset of expert demonstrations (observation → action pairs), train a neural network to predict the expert action from the observation. While conceptually simple, practical BC requires addressing distribution shift (the policy encounters states not in the training data), action stochasticity, and demonstration quality. Modern BC for robotics uses large vision-language backbones (RT-2), diffusion-based action prediction, and data augmentation to handle multimodal action distributions.