Reproduce “Knowing When to Stop: Adaptive Action Chunking via Internal Cross-Attention Dynamics in VLAs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Runze Xu, Xiaolong Shan, Shuang Dai, Yu Wang, Jincheng Yu
The method uses a policy’s visual-language attention to choose when a robot should stop executing a predicted action prefix and replan. It measures attention entropy, where higher values mean attention is spread across input tokens, and truncates only when high entropy is sustained. On π0.5 across RoboTwin 2.0, it reports 62.9% mean success versus 61.6% for fixed chunks, with 18.11 steps, not measured time savings.
Key insights
- Cross-attention entropy correlates with future action error: after accounting for action position, the partial Pearson correlation is 0.214, and samples in the highest entropy quartile have 10.4 times the mean MSE of the remaining samples.
- The same normalized plateau criterion works with both decoupled attention in π0.5 and coupled attention in X-VLA, suggesting that the signal is not limited to one particular VLA architecture.
- Adaptive chunking can preserve long executions during well-grounded motion while shortening the control loop near uncertain or precision-sensitive stages, rather than selecting one horizon for an entire task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.