Behavior cloning

A form of imitation learning that trains a policy to predict demonstrated actions from observations.

The training target is the demonstrator’s action. Inconsistent controls, poor timing or unmarked interventions can weaken that target. A policy may also encounter states absent from the demonstrations; strong performance on recorded examples alone does not establish reliable autonomous behavior.

In practice

A recovery movement is recorded with its intervention context, so it is not mistaken for the normal continuation of a successful task.

Further reading

• LeRobot: learning from interventions