Reproduce “From World Models to World Action Models: Rethinking Next-State Prediction” 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
Tingyu Yuan, Ziming Ji, Biaoliang Guan, Wen Ye, Wenrui Tian, Zhaopeng Gu, Feihong Zhang, Xu Yang and 4 more
CF-WAM trains one shared world-action model to predict one of four descriptions of the same future—RGB, semantics, geometry, or interaction labels—selected across updates, while masking robot-specific actions in human data. On RoboCasa-GR1, it reaches 82.50% average success across 24 tasks, versus 75.17% for WALA; this benchmark average does not establish universal per-task or real-world superiority.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.