Reproduce “FoveaVLA: Foveation Actions Replace the Wrist Camera in VLAs” 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
Bingao Chen, Lihan Zha, C. Karen Liu, Guy Tevet
FoveaVLA lets a robot select close-up crops from fixed external camera images, while retaining full views. At each policy query it predicts an action chunk and boxes for regions to inspect next, using RoboInter annotations for training. In four real-world suites, FoveaVLA-Single achieves 70.8% success versus 52.1% for wrist-equipped π₀.₅ across 48 trials, a result limited to this evaluation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.