ActPrune: Training-free action-guided visual token pruning

Sungwon Moon and other authors propose ActPrune, a training-free visual token pruning method for vision-language-action (VLA) models. It moves the pruning decision ahead of the visual encoder, using the action chunk already predicted at the previous control step as a geometric prior: this action chunk is available before encoding the current observation and provides the end-effector and camera motion between adjacent observations. Specifically: under a fixed third-person view, it retains image patches near the projected end-effector trajectory and semantic anchors; under the wrist view, it prunes image patches that remain unchanged after compensating for camera motion caused by the action. The authors report that on the two simulation environments LIBERO and VLABench and four real robot tasks, the method achieves up to 1.61x speedup and reduces FLOPs by 59.5%, with negligible drop in success rate.

AI 根据报道生成 · 2 days ago更新