Standard World Action Models (WAMs) jointly denoise video and action predictions from scratch at each replanning cycle, causing significant latency.
Rolling-WAM instead maintains a sliding window of video-action chunks at staggered noise levels. At each step, only the imminent action chunk is fully denoised for execution, while future chunks are partially refined. As the window advances with new observations, retained chunks continue their denoising process, distributing computation across cycles.
On LIBERO, RoboTwin, and a real-world Unitree G1 humanoid, Rolling-WAM achieves competitive manipulation performance (98.1% success on LIBERO, 93.3% on RoboTwin) while delivering a 4.5x steady-state replanning speedup over standard joint WAMs. This reduces the computational bottleneck of closed-loop control with visual imagination.