Imitation learning

Imitation learning is the family of training methods where a control policy is supervised against demonstrations — human teleoperation, motion capture, or expert behavior — rather than driven by trial-and-error reward. The strict version is behavior cloning: the policy is fit to reproduce the demonstrations directly. Looser versions (DAgger, GAIL, inverse reinforcement learning) iteratively correct for the distribution drift that pure behavior cloning suffers when the robot encounters states the demonstrator didn't.

The distinction matters because imitation-learning claims routinely gloss over data scale and quality. Figure's Helix policy was trained on hundreds of hours of operator teleoperation; Physical Intelligence's π0 series uses cross-embodiment human-video plus robot data; Boston Dynamics has built parts of Atlas's manipulation stack on a mix of imitation and reinforcement. A press claim of "the robot learned from one demo" without specifying the data regimen should be read as the loose version of the technique by default.