A robot model cut the time to prepare its next action from 978 to 215 milliseconds by carrying unfinished video-and-action predictions from one planning cycle into the next, researchers report. The method, Rolling-WAM, keeps a forecast of what the robot may see alongside the movements it plans to make.
World-action models jointly predict robot movements and future video, giving a controller a visual forecast of how a plan could change the scene. A standard model rebuilds that whole forecast from noise each time it replans, even though the robot executes only the first part. That repeated work can delay the next action.
Rolling-WAM keeps a sliding window of video and action chunks at different stages of denoising, the process of gradually turning noisy guesses into a usable prediction. It finishes the nearest action chunk, leaves farther predictions partly noisy, then shifts the window after the robot acts and receives a new camera image. The retained chunks are refined with that fresh observation while a new chunk is added. The approach reuses unfinished predictions instead of discarding the entire horizon at every update.
In the researchers’ default RoboTwin 2.0 timing test, Rolling-WAM took 215 milliseconds per steady-state update on one NVIDIA A100 GPU. The standard joint world-action comparison took 978 milliseconds, and Fast-WAM, which omits future-video generation during execution, took 548 milliseconds. The test used 384-by-320 images and executed 16 actions per update. Rolling-WAM maintained an 80-action prediction window, while the baselines predicted 16 actions. The measurements excluded initialization and warm-up.
Faster updates did not come with a large drop in the reported task scores. Rolling-WAM averaged 98.1% success on the LIBERO simulation benchmark, compared with 98.5% for the joint model, and 93.3% on RoboTwin 2.0, compared with 90.6%. On a Unitree G1 humanoid, one policy tested on three manipulation tasks averaged 85% success across 20 trials per task; the joint model averaged 78.3%.
The physical test covered only doll placement, plate stacking and bead pouring, with a small number of trials. The paper also found that a longer prediction window did not always improve results: on six selected RoboTwin tasks, success peaked at five chunks and fell with eight. Forecasts that lag sudden scene changes remain a practical risk; the authors identify adaptive window management as a direction for further work.
Sources
- Yinghua Zhou et al., “Rolling-WAM: World Action Models with Rolling Imagination”, arXiv preprint, September 24, 2026. Method, benchmark comparisons, timing setup and real-robot trials.
- Lead image: Rolling-WAM teaser figure by Zhou et al., from the same paper, licensed under CC BY 4.0; image file. Diagram illustrates the method, not a robot trial.