Berkeley & Polimi Uncover Why Action Chunking Improves Robotic Imitation Learning

canondetortugas · x · 2026-08-07

In robotic imitation learning, Action Chunking—predicting and executing multiple actions—is widely used, but a precise understanding of why it improves performance has been limited.

Researchers from Politecnico di Milano and UC Berkeley conducted rigorous experiments showing that existing hypotheses (temporal consistency, horizon reduction, representation learning) fail to explain its success. The actual reasons are:

Greater non-Markovian expressivity and reduced compounding error.
Implicit ensembling: by learning diverse temporal relationships, action-chunked policies behave like a model ensemble, enhancing robustness and generalization.

Based on these insights, the team demonstrated how to match action chunking's performance without it, and proposed a new policy class that further amplifies the ensembling benefits.

Original post →

Berkeley & Polimi Uncover Why Action Chunking Improves Robotic Imitation Learning

More from Embodied