Reproduce “Meituan-Robotics-0: A Vision-Language-Action Foundation Model for Desktop 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
Meituan Robotics Team, Caicheng Wang, Chen Zhao, Chengpeng Jin, Duo Xu, Fengjiao Chen, Guanglu Wan, Heqing Yang and 30 more
MR0 combines varied vision-and-robot data with continuous control and human corrective training through Rollback, which pairs replaced actions with corrected continuations. Its task-level memory stores events beyond current observations. Under matched HG-DAgger, it averages 79% success rate and 91% progress score across three desktop tasks, versus π₀.₅’s 69% and 85%; these results are limited to those tasks and setup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.