Rolling-WAM spreads video-action denoising across cycles for 4.5x faster robot replanning

USC-PSI-Lab · hf · 2026-09-29

World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation, but completing the joint video-action denoising at each replanning cycle causes latency that limits closed-loop responsiveness.

USC-PSI-Lab's Rolling-WAM maintains a sliding window of video-action chunks at staggered noise levels. Each step, a rolling noise schedule fully denoises the imminent action chunk for execution while partially refining farther-future chunks; as new camera observations arrive, retained future chunks continue denoising. This distributes compute over time while carrying an evolving visual-action context across chunk boundaries.

Evaluations on LIBERO, RoboTwin, and a real Unitree G1 humanoid show competitive manipulation performance, with a 4.5x steady-state replanning speedup over standard joint WAMs by avoiding full-horizon denoising from scratch.

Original post →

More from Embodied

Embodied channel →