Reproduce “LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence” 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
Jiajun Cao, Kai Tang, Bohua Zhang, Zezhong Qian, Ying Li, Xiaozhu Ju, Xiaobao Wei, Jiaming Liu and 13 more
LAW uses semantic subtasks to coordinate planning, action selection, and completion checks in long-horizon manipulation. Its world-action model proposes action chunks, a value expert ranks them, and a state expert checks progress asynchronously. It averages 94.21% success across 50 RoboTwin tasks, versus 91.83% for Fast-WAM, and completes three real-robot tasks in 65% of trials; those tests do not establish general transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.