Reproduce “Learning What Actions Change: Action Effect Matching for Dynamics-Faithful 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

John Won, Yonghoon Dong, Jinwoo Shin

AFFECT trains world models to make generated futures respond to robot action changes, using perturbed actions and recorded outcomes during play. It matches feature-level action effects without repeated runs from identical starting states. On DROID with DreamDojo, policy-success correlation rises from 0.468 to 0.790, while long-horizon FVD falls from 807 to 622. The method is limited to effects its features capture.

Summary by alphaXiv.

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