Researchers have introduced Rolling-WAM, a new formulation designed to enhance robotic manipulation by addressing latency issues inherent in current World Action Models (WAMs). Standard WAMs couple action generation with future visual prediction, but the process of denoising video and actions at each replanning cycle creates substantial delays. These delays hinder real-time action updates and limit the responsiveness of robots operating in dynamic environments. Rolling-WAM distributes this denoising computation across multiple replanning cycles, effectively reducing latency and improving operational speed.
The core innovation of Rolling-WAM lies in its method of managing a sliding window of video-action chunks. These chunks are maintained at staggered noise levels. At each replanning step, the system fully denoises the immediate action chunk intended for execution. Simultaneously, farther-future chunks undergo partial refinement. As new camera observations become available, the window advances, and the previously processed future chunks continue their denoising process. This distribution of computational load over time allows for faster replanning cycles while maintaining an evolving visual-action context across segment boundaries.
Evaluations of Rolling-WAM on simulation benchmarks, including LIBERO and RoboTwin, as well as real-world manipulation tasks using a Unitree G1 humanoid robot, demonstrate its effectiveness. The method achieves competitive manipulation performance, matching or exceeding that of existing state-of-the-art WAMs. Notably, Rolling-WAM achieves a 4.5x steady-state replanning speedup compared to standard joint WAMs by eliminating the need to denoise the entire prediction horizon from scratch at each cycle. The system utilizes only two denoising steps per replanning cycle, suggesting that the co-refinement of current and future predictions can lead to both faster replanning and improved manipulation outcomes. Qualitative assessments show that the imagined robot motion and task progression closely align with observations, indicating the visual context maintained during rolling execution remains accurate.