Rolling-WAM Speeds Robot Replanning With Staggered Denoising

World Action Models, or WAMs, are a class of robot control systems that predict both what a robot should do next and what its camera will see as a result, denoising video and action data together through a diffusion process. That joint prediction helps robots plan more coherently, but according to a paper posted to arXiv, redoing the full denoising process from scratch at every replanning cycle creates significant latency, which slows down how quickly a robot can react to its environment.

A team of researchers — Yinghua Zhou, Junjie Ye, Yiqi Zhao, Hao Dong, Celina Shiyu Wang, Ruohai Ge, Tingyi Yang, Basile Van Hoorick, Gaurav Sukhatme, Vitor Guizilini, and Yue Wang — describe a fix they call Rolling-WAM. Instead of resetting the denoising process for an entire prediction window each cycle, the method keeps a sliding window of video-action "chunks," each at a different, staggered noise level. At every step, the chunk about to be executed is fully denoised and ready to act on, while chunks further in the future are only partially refined.

As the robot receives new camera input and the window slides forward, those partially refined future chunks keep getting cleaned up incrementally rather than being thrown out and recomputed. The authors describe this as spreading computation over time while preserving an evolving visual-action context that carries across chunk boundaries, so the model isn't starting from zero at each planning step.

The team evaluated Rolling-WAM on two simulated manipulation benchmarks, LIBERO and RoboTwin, as well as on a real-world Unitree G1 humanoid robot. Across these tests, the method delivered manipulation performance competitive with standard joint WAMs that denoise the entire horizon at once. The key payoff was speed: Rolling-WAM achieved a 4.5x steady-state speedup in replanning compared to conventional joint denoising, according to the paper.

The work is presented in a 10-page paper with 7 figures and 5 tables and is currently under review; the authors note a project page exists alongside the arXiv listing. For robotics teams building on diffusion-based planners, the approach suggests a practical route to closing the gap between the computational heaviness of generative world models and the tight timing demands of real-time closed-loop control.