humanoidsdata.com

Search

Search datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Models & learning

Imitation learning

Imitation learning is a family of methods that learns a policy from examples of expert behaviour rather than specifying every control rule by hand. In robotics, demonstrations pair observations or states with actions, trajectories or inferred objectives. Behaviour cloning is one imitation-learning method; interactive and inverse approaches address different supervision and distribution-shift problems.

Also known as: learning from demonstration, LfD, learning by demonstration

Updated

What is learned from an expert

Imitation learning uses demonstrated behaviour to learn how a robot should act. The supervision can be direct state-action examples, comparisons between behaviours or an objective inferred from demonstrations. The published survey An Algorithmic Perspective on Imitation Learning distinguishes direct replication, including behaviour cloning, from methods that infer the expert’s underlying objective.

For robots, a demonstration may come from teleoperation, kinesthetic teaching, an autonomous expert or retargeted human motion. A video alone does not contain the robot’s motor command, so human demonstrations need correspondence, action inference or retargeting before they can supervise low-level control.

Behaviour cloning is one approach

Behaviour cloning treats expert observation-action pairs as supervised training data. It is simple and widely used, but the learned policy can enter states absent from the expert dataset after making small errors. Interactive imitation methods collect supervision in states reached by the learner.

DAgger is a foundational example: it repeatedly runs the current policy, asks the expert for actions on visited states and adds those labelled states to the training set. This targets the mismatch between the expert’s state distribution and the learner’s own rollouts.

Data considerations for humanoids

Useful humanoid demonstrations align camera data, proprioception, contact signals and actions on one timeline. They also identify the robot, reference frames, action units and controller mode. Coverage of corrections, balance disturbances and failed attempts matters when recovery is part of the task. Demonstrations show what the collector attempted under particular conditions; they do not by themselves prove that a learned policy will be safe or reliable outside that distribution.

Sources