Rolling-WAM: World Action Models with Rolling Imagination
Rolling-WAM speeds up robotic replanning by spreading video-action denoising across successive control cycles.
The method keeps a sliding window of future video-action chunks at different noise levels, fully resolving only the next action while continuing to refine later ones. That lets the model reuse evolving future context as new camera observations arrive instead of restarting the whole horizon each time. The authors report competitive manipulation results on LIBERO, RoboTwin, and a real Unitree G1 humanoid, with a 4.5x steady-state replanning speedup over standard joint WAMs. ArXiv · AI/CL/LG's note
The method keeps a sliding window of future video-action chunks at different noise levels, fully resolving only the next action while continuing to refine later ones. That lets the model reuse evolving future context as new camera observations arrive instead of restarting the whole horizon each time. The authors report competitive manipulation results on LIBERO, RoboTwin, and a real Unitree G1 humanoid, with a 4.5x steady-state replanning speedup over standard joint WAMs. ArXiv · AI/CL/LG's note
score 5