Rolling-WAM: World Action Models with Rolling Imagination
Yinghua Zhou, Junjie Ye, Yiqi Zhao, Hao Dong, Celina Shiyu Wang, Ruohai Ge, Tingyi Yang, Basile Van Hoorick and 3 others
University of Southern California · Brown University · Fudan University · Toyota Research Institute
Published September 24, 2026 · on the feed September 29, 2026 · ▲ 3 · arXiv 2609.30247
The 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 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.
The Brief
1:52A short podcast summary of the paper. Written and voiced by AI.
Transcript
Rolling-WAM. World Action Models with Rolling Imagination. By Yinghua Zhou and ten others, from the University of Southern California, Brown University, and two other institutions. Published on the twenty-fourth of September. Its code repository is a placeholder: a readme, and nothing to run yet.
World Action Models, or WAMs, generate a robot's actions together with a prediction of what it will see next, and use both for manipulation. The trouble is timing. Every time the robot replans, the model has to run the whole joint video-action denoising process to completion. That takes long enough to delay action updates, and a robot that updates its actions late cannot respond well inside a closed loop.
Rolling-WAM spreads the denoising out over time. Instead of finishing it within one replanning cycle, it distributes the work across successive cycles, so each cycle only has to do part of it.
The model keeps a sliding window of video-action chunks, each held at a different noise level. At every step a rolling noise schedule fully denoises the chunk that is about to be executed, and only partially refines the chunks further in the future. When the window moves forward with a new camera observation, those future chunks are kept, and their denoising simply continues. Two things follow. The computational cost is distributed across time instead of landing all at once, and an evolving visual and action context is carried across chunk boundaries rather than rebuilt.
The method is evaluated on LIBERO, on RoboTwin, and on a real-world Unitree G1 humanoid, where Rolling-WAM achieves competitive manipulation performance. Because it no longer denoises the entire prediction horizon from scratch, it replans four point five times faster in steady state than standard joint WAMs.
The future the model imagines is no longer thrown away at each replan. It is kept half-finished and carried forward, and that reuse is where the speedup comes from.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A robot that thinks too long acts too late. That is the wall here.
Your robot's policy imagines the video of what it will see next and generates its actions alongside, both by denoising from noise.
The road ahead
- 1 · The problem ~7 min A robot policy that imagines future video and actions by denoising must finish all that denoising before every replan, so its actions arrive late.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Spread the denoising across replanning cycles: keep a window of chunks at different noise levels, finish only the next chunk, and keep refining the rest.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Keep staggered chunks, finish the front one, execute it, slide the window with a new observation, and keep denoising what remains.
- 6 · The experiment ~6 min You’ll guess: By spreading denoising across replanning cycles, how much faster did Rolling-WAM replan in steady state than standard joint World Action Models?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.