Elastic-WAM spends world-model computing only where a robot task needs it

World-action models help robots plan by predicting how the scene will change, but repeated video denoising makes them slow at inference, according to a paper on arXiv. Haitian Li and colleagues note that earlier fixes use a fixed refinement budget, while the useful depth of refinement actually varies by task and by state.

Their Elastic-WAM learns a single spatiotemporal demand field that decides when to stop refining, how far each denoising step advances and which future regions to refresh. The field is trained in three supervised stages tied to the action.

In simulation benchmarks and real long-horizon manipulation, the authors report control results matching the best systems with lower inference latency.