Reproduce “Reducing Temporal Redundancy for Efficient Vision-Language-Action Inference” 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
Yuzhou Wu, Yuxin Zheng, Muchun Niu, Yishan Yang, Tianhao Liu, hanwen kang, Jiajian Jing, Linfeng Zhang and 1 more
The authors speed vision-language-action inference by reusing visual representations across similar frames and compressing a ten-step flow-matching action solver into two steps. On LIBERO with π0.5, the method reduces inference time from 286.9 ms to 121.2 ms, while mean success rate changes from 94.4% to 93.8%. This supports faster inference with nearly maintained success in that evaluated setting, not performance elsewhere.
Key insights
- Visual perception in VLA models contains substantial temporal redundancy, as evidenced by high cosine similarity (mostly above 0.98) between visual tokens in adjacent frames.
- Action generation via flow matching exhibits low-rank dynamics, where velocity evolution is largely driven by a few dominant components, allowing for significant step compression.
- The action generation module, particularly iterative sampling, constitutes the primary computational bottleneck in VLA pipelines, making its optimization critical for overall efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.