Espurna
EN
LearningIntermediateimitation learningbehaviour cloningteleoperationLeRobotACTaction chunkSO-101

Imitation learning

In 30 secondsA person teleoperates the robot, observation-action pairs are recorded and a policy is trained to copy them. It fails in states it has never seen; the fix is more varied demos and predicting several actions in a row.

Also available in: Català · Castellano

The idea

Sometimes it is easier to show than to write a reward. A person teleoperates the robot (for example with the SO-101 leader arm) and observation → action pairs are recorded. Training a policy to copy them is supervised learning: behaviour cloning.

The classic problem

If the robot drifts into a state it never saw in the demos, it doesn’t know what to do and the errors pile up.

Usual fixes:

  • More demos, and more varied ones.
  • Predicting action chunks (several actions in a row at once), as ACT does in LeRobot.

The LeRobot pipeline

Teleoperate with the leader → record episodes with cameras → train a policy → run it on the follower.

In the demo

Drag a demo from the start to the goal and run the policy. From a random start far from the demos, it wanders aimlessly; add demos there and it works.

Interactive demo

Related concepts

Learning

Reinforcement learning

Nobody tells the robot how to do it, only whether it did well: the agent tries actions, gets a reward and learns a policy that maximises it. In simulation, thousands of copies train at once.

Learning

Vision-language-action models (VLA)

A VLA starts from a model that already links images and text and adds an action output: it takes camera images, an instruction and the robot's state, and returns a chunk of movements.