Reproduce “EvoScene-VLA: Evolving Scene Beliefs Inside the Action Decoder for Chunked Robot Control” 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
Chushan Zhang, Ruihan Lu, Jinguang Tong, Xuesong Li, Yikai Wang, Hongdong Li
EvoScene-VLA carries a compact scene state forward while generating action chunks, predicting scene changes that the next camera observation can correct. It combines the prior with multi-view images and jointly generates actions and future scene states. Across 31 RoboTwin tasks, it averages 88.8% in Clean and 87.9% in Rand; recurrence alone did not help without scene supervision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.