Reproduce “$Σ$-LeWM: Learning World Models via S-Information for Robot Control” 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

Xinrui Zu, Michael Przystupa, Zhao Yang, Shujian Yu, Kevin Sebastian Luck

Σ-LeWM learns a latent world model for robot control, evaluated on simulated peg seating in a gear hub. During training, it predicts transitions forward, backward, and the action connecting consecutive states; the auxiliary predictors are discarded for test-time planning. On simulated Gear-Assembly, it achieves 74.9% strict success versus 10.9% for LeWM, with no physical-robot demonstration.

Summary by alphaXiv.

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