Action chunking
- In French:
- Découpage en séquences d’actions
Definition
An imitation learning technique in which the model predicts not one action at a time but a short sequence of actions to be carried out in one go. It limits the accumulation of errors that makes policies learned from demonstrations drift. It was introduced in 2023 with the ACT algorithm (Action Chunking with Transformers), trained on demonstrations collected through teleoperation.
Checked on against the references below.