Reproduce “$μ_0$: A Scalable 3D Interaction-Trace World Model” 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

Seungjae Lee, Yoonkyo Jung, Jusuk Lee, Jonghun Shin, Amir Hossein Shahidzadeh, Yao-Chih Lee, H. Jin Kim, Jia-Bin Huang and 1 more

The authors introduce μ₀, a model that learns from video to predict 3D motion traces of selected physical points, then supplies those traces to a separately trained action expert. The world model uses no action-labelled pretraining, but the expert uses robot demonstrations. On three UR3 tasks, μ₀ averages 91.7% success versus 73.3% without traces; this evidence covers one setup.

Summary by alphaXiv.

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