Reproduce “ForeAct: Geometry-Guided Action Learning from World Representations” 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
Wei Lu, Jiaxi Zhang, Yunheng Wang, Taowen Wang, Yuetong Fang, Zhenhui Zhang, Shuning Zhang, Zizhao Yuan and 3 more
ForeAct trains a pretrained video model to support robot actions by adding geometry tokens supervised to predict next-observation relative depth, while retaining a single-pass action path that needs no depth input at deployment. On LIBERO, it reaches 96.6% average success, above its 94.7% no-depth-supervision ablation, but below reported Fast-WAM and Cosmos Policy averages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.