Reproduce “WEVO: World Model-Enabled VLA Optimization for Whole-Body Loco-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

Qianzhong Chen, Muyuan Lin, Yuyin Sun, Mac Schwager, Cheng-Hao Kuo, Roberto Martín-Martín

WEVO trains a humanoid vision-language-action policy with simulation demonstrations, then improves it through reinforcement learning in a video world model aligned with autonomous robot experience. The model predicts camera views from motion. Across three Unitree G1 tasks, video-world-model learning reaches mean success of 75.0%, versus 23.3% for behavior cloning and 48.3% for real-robot reinforcement learning; results cover these tasks only.

Summary by alphaXiv.

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