Reproduce “Motion-guided Contrastive Regularization For 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

Yiwen Chen, Qiqi Hou, Jisoo Jeong, Yinhao Zhu, Huaizu Jiang, Amir Said

The authors introduce a motion-guided objective for JEPA-style world models that predicts action-driven changes in visual representations. It compares predicted changes with the next observation and six other moments from the same episode, using optical-flow similarity to guide the comparisons. Planning success rises from 52% to 78% on Cube-RB and from 14% to 54% on LIBERO-Object, under sampled-goal position criteria.

Summary by alphaXiv.

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