Abstract

Action Chunking—the practice of predicting a sequence of actions and executing a prefix open-loop—has emerged as a key enabler of recent progress in imitation learning for robotic manipulation. A growing body of work has sought to understand why action chunking appears so effective, citing mitigating compounding errors, absorbing inference latency, increasing action smoothness, or consequences of partial observability such as mode switching or idling. We unify this last category of explanations under the language of "non-Markovian expert demonstrations," and argue that non-Markovian expert demonstrations, combined with policies' limited memory, are the primary factors that necessitate action chunking. We support our argument with extensive experiments on popular robotics benchmarks. Further, we show that, when policies are provided sufficient context, action chunking is no longer beneficial and the most reactive, closed-loop policies perform best. While imitation learning has seen great success using action chunking policies, our findings suggest that long-context, reactive policies are a more principled and performant paradigm.

Citation

@misc{zeng2026revisitingopenloopexecutionrobotics,
  title={Revisiting Open-Loop Execution in Robotics: Toward Reactive, Higher-Performing Policies}, 
  author={Michael Zeng and Abhinav Agarwal and Ajay Bati and Brian Lee and Siddharth Ancha and Russ Tedrake},
  year={2026},
  eprint={2608.15938},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2608.15938}
}