Current visual world models treat robot manipulation as a stack of discrete frames — frame 1, frame 2, frame 3 — predicting each next instant and rolling forward step by step. But physical motion happens between snapshots, not at them. Working with UC Berkeley's BAIR group, AIR proposes ODEWorld, a continuous-time embodied world model that learns the direction and speed at which the world changes in physical time, rather than solving a frame-by-frame "what comes next" fill-in-the-blank. The approach enables efficient generation, arbitrary frame interpolation between any two moments, and even backward rollout from a future state.
https://air.tsinghua.edu.cn/info/1007/2535.htm