robotics
Cross-Attention Told the Robot When to Look Again
A training-free rule ends an action chunk when observation grounding disperses into a sustained entropy plateau.
Summary
A training-free rule ends an action chunk when observation grounding disperses into a sustained entropy plateau.
Long open-loop chunks can drift as the world changes, while short chunks repeatedly pay inference cost. The method watches cross-attention weights already computed inside the action expert and truncates execution when their entropy remains high. Evaluations across two VLA families, simulation suites and three real manipulation tasks improved average success over fixed and adaptive baselines with negligible added overhead.
Why it matters
A training-free rule ends an action chunk when observation grounding disperses into a sustained entropy plateau.
Limits and context
No additional limitation was separately recorded.
Key claims
A training-free rule ends an action chunk when observation grounding disperses into a sustained entropy plateau.
Evidence: source-2026-09-02-013
Sources
- arXiv preprint 2609.00908arXiv · primary research
Corrections
No corrections have been recorded for this story.