Reproduce “Action-aware Dynamic Pruning for Efficient Vision-Language-Action Manipulation” 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

Xiaohuan Pei, Yuxing Chen, Siyu Xu, Yunke Wang, Yuheng Shi, Chang Xu

ADP reduces visual-token computation in vision-language-action robots by pruning image patches according to instruction relevance and recent end-effector motion, restoring full vision when motion indicates delicate control. On LIBERO, OpenVLA-OFT with parallel action decoding reaches 96.3% average success at 50% token retention, versus 97.1% for the baseline, with 1.23× LLM-side speedup; the result is not an end-to-end episode-time claim.

Key insights

  • Visual redundancy in robotic manipulation tasks is action-aware and varies dynamically with the robot's motion, allowing for adaptive pruning that balances efficiency and precision.
  • Early Transformer layers (e.g., Layer 0) provide more stable and globally discriminative text-to-vision attention signals for robust visual token importance scoring than deeper layers.
  • A dynamic, state-aware pruning strategy that considers robot actions is crucial for efficiency gains without sacrificing accuracy, significantly outperforming static or periodically switching pruning methods.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.