Reproduce “Exploring Temporal Context-Aware Physical Grounding in Latent World Models” 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

Yilei Man, Yang Chen, Jingcai Guo

G-JEPA improves image-based robot planning by combining temporal physical-state supervision with learned action proposals and model-based search. A shared causal S4D module grounds observed and predicted latent sequences during training; a frozen world model scores learned and Gaussian candidates. On 75-step PushT-Distance tasks, it reaches 55.21% success, versus 47.53% without proposals and 16.54% for LeWM; at 25 steps, proposals reduce success.

Summary by alphaXiv.

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