Rolling-WAM: World Action Models with Rolling Imagination
Rolling-WAM speeds robotic replanning 4.5x by spreading joint denoising across successive cycles.
Rolling-WAM is a world-action-model formulation for robotic manipulation that spreads joint video-action denoising across replanning cycles instead of restarting the full horizon each time. A sliding window holds chunks at staggered noise levels, fully denoising the next action while only partially refining farther-future chunks as new camera observations arrive. On LIBERO, RoboTwin, and a real Unitree G1 humanoid, it matches competitive manipulation performance and reports a 4.5x steady-state replanning speedup over standard joint WAMs.
- Rolling-WAM keeps video-action chunks at staggered noise levels in a sliding window.
- Each step fully denoises the imminent action and partially refines future chunks.
- Evaluated on LIBERO, RoboTwin, and a real Unitree G1 humanoid.
- Steady-state replanning is 4.5 times faster than standard joint world action models.
Full article153 words · extracted from arxiv.org · click to collapse
World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action updates and limiting closed-loop responsiveness. We present Rolling-WAM, a formulation that distributes joint denoising across successive replanning cycles. Our method maintains a sliding window of video-action chunks at staggered noise levels. At each step, a rolling noise schedule fully denoises the imminent action chunk for execution, while partially refining farther-future chunks. As the window advances with new camera observations, the retained future chunks continue their denoising process. This distributes the computational cost over time while carrying an evolving visual-action context across chunk boundaries. Evaluations on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid show that Rolling-WAM achieves competitive manipulation performance. By removing the need to denoise the entire prediction horizon from scratch, it delivers a 4.5x steady-state replanning speedup over standard joint WAMs.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.30247