Rolling-WAM: World Action Models with Rolling Imagination for Faster Robot Manipulation

Forum topic · 小凯 · 2026-09-28

Summary

Rolling-WAM is a new approach to World Action Models (WAMs) for robotic manipulation that reduces the latency of joint video-action denoising. Standard WAMs couple action generation with future visual prediction, but completing the full joint denoising process at every replanning cycle introduces substantial delay, limiting closed-loop responsiveness. Rolling-WAM distributes joint denoising across successive replanning cycles using 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, retained chunks continue denoising. This spreads computation over time and carries evolving visual-action context across chunk boundaries. Evaluations on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid robot show competitive manipulation performance and a 4.5x steady-state replanning speedup over standard joint WAMs. Paper: arXiv 2609.30247.

Overview

Research area: Computer Vision (CV) Authors: Yinghua Zhou, Junjie Ye, Yiqi Zhao, Hao Dong, Celina Shiyu Wang, Ruohai Ge, Tingyi Yang, Basile Van Hoorick, Gaurav Sukhatme, Vitor Guizilini, Yue Wang arXiv: 2609.30247

Abstract

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 evolving visual-action context across chunk boundaries.

Key Results

Evaluated on LIBERO, RoboTwin, and a real-world Unitree G1 humanoid robot, achieving competitive manipulation performance.
Since there is no need to denoise the entire prediction horizon from scratch, Rolling-WAM achieves a 4.5x steady-state replanning speedup compared to standard joint WAMs.