robotics
The Robot Drops 85 Percent of the Scene Tokens
Training-free pruning and attention bias keep action-critical regions while improving robustness under object-placement shifts.
Summary
Training-free pruning and attention bias keep action-critical regions while improving robustness under object-placement shifts.
T3R first uses segmentation to remove irrelevant visual patches, then biases attention toward action-critical regions and language instructions. On LIBERO-Spatial it removed 85 percent of scene-camera tokens while essentially matching unpruned success, and improved success by as much as seven percentage points under out-of-distribution object placement. Profiling showed up to a 3.3-times language-model forward-pass speedup as resolution increased. Results across three vision-language-action backbones favor multi-camera settings; they do not establish general robot safety.
Why it matters
Training-free pruning and attention bias keep action-critical regions while improving robustness under object-placement shifts.
Limits and context
- Results across three vision-language-action backbones favor multi-camera settings; they do not establish general robot safety.
Key claims
Training-free pruning and attention bias keep action-critical regions while improving robustness under object-placement shifts.
Qualification: Results across three vision-language-action backbones favor multi-camera settings; they do not establish general robot safety.
Evidence: source-2026-11-08-013
Sources
- arXiv preprint 2611.08413, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.