Reproduce “Leap-VLA: High-Rate Robot Control from a Low-Rate VLM Backbone” 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

Lin Yang, Yunjie Shen, Zheyu Li, Zhenjie Tian, Yongli Wang, Dongze Huang, Weiqi Wang, Yucheng Zhu and 3 more

Leap-VLA separates slow semantic processing from fast robot control: an action expert uses cached vision-language context, fresh visual input, and cache age. It trains the expert with asynchronous data, measured cache ages, and stale-frame masking. On LIBERO-Spatial, success at a 2-second semantic refresh rises from 61.4% with naive reuse to 95.5% with the full recipe, within the tested cache-age range.

Summary by alphaXiv.

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