Yan Huang

Fixing world-action models to a single next-state representation fundamentally bottlenecks policy learning; supervising across dynamically sampled geometric, semantic, and visual projections of the same future allows robots to directly absorb dynamics from heterogeneous human video.

Today's video world models excel at hallucinating plausible-looking frames, but their ability to actually preserve initial object identities drops by over 40% as scene complexity scales beyond 40 instances.