Rolling-WAM: Accelerating Robotic World Action Models via Rolling Imagination
Rolling-WAM：利用滾動想像實現高效控制的機器人世界動作模型
Traditional World Action Models (WAMs) couple action generation with future visual prediction, but denoising the entire horizon at every step creates major latency. Rolling-WAM introduces "Rolling Imagination," maintaining a sliding window of video-action chunks at staggered noise levels. Instead of denoising from scratch, it fully denoises only the imminent action chunk for execution, while partially refining farther-future chunks. As the window slides with new observation inputs, the future chunks undergo further denoising. Tested on LIBERO, RoboTwin, and a physical Unitree G1 humanoid robot, it delivers competitive manipulation performance with a 4.5x steady-state replanning speedup.
Key points
Temporal Denoising Distribution
Distributes the heavy joint video-action denoising process over successive replanning cycles rather than starting from scratch.
Rolling Noise Schedule
Maintains a sliding window with staggered noise levels, fully denoising the immediate chunk while progressively refining future steps.
4.5x Replanning Speedup
Achieves a 4.5x steady-state speedup over standard joint WAMs, significantly improving real-time responsiveness.
Physical Humanoid Deployment
Validated successfully on simulated benchmarks and a physical Unitree G1 humanoid robot platform.
How it works
| Standard WAM | Rolling-WAM (Ours) | |
|---|---|---|
| Denoising Method | 每次規劃皆需從頭完整去噪 | 滑動視窗內分階段交錯去噪 |
| Replanning Latency | 高延遲（難以實現閉環即時控制） | 低延遲（達 4.5 倍穩態加速） |
| Future Horizon | 捨棄前次預測，全部重新生成 | 保留未來區塊並在後續步驟中持續精煉 |
Why it matters
While World Action Models (WAMs) show great promise for embodied AI, high computation latency has restricted their real-time deployment. Rolling-WAM proves that we do not have to sacrifice visual imagination for reaction speed. By breaking the denoising latency bottleneck, this approach paves the way for practical, closed-loop control on physical robots requiring agile and high-frequency updates.
Who it affects
- AI Developer
- AI Researcher
- Student & Learner
How to use it
- 1Real-time control and manipulation for humanoid robots
- 2High-frequency closed-loop decision making using diffusion models
Limitations & caveats
- Relies heavily on continuous sliding-window context; sudden, drastic environmental changes might affect the partially denoised future chunks.
- Despite the speedup, the overall computational footprint of diffusion-based world models is still higher than simpler behavioral cloning models.
Related
AD-WM: Action-Discriminative World Models for Counterfactual MPC
AD-WM：專為反事實預測控制設計的動作辨識世界模型
AD-WM enhances latent world models by learning to distinguish alternative actions from the same state, significantly improving counterfactual control and zero-shot transfer in robotics.
RAPID: Robot Agentic Programming from Demonstrations
RAPID：只需單次視覺示範，AI 代理即可自動生成與優化機器人操控程式
RAPID is a framework that automatically infers task specifications and environments from a single visual demonstration, using an agentic loop to program and refine generalized robot skills.
Coding Agents Outperform Hand-Engineered Planners in Task and Motion Planning
程式碼生成 Agent 突破機器人規劃瓶頸：自動編寫高泛化性的 TAMP 解決方案
Researchers demonstrate that coding agents can interactively synthesize highly generalizable programs for task and motion planning (TAMP), achieving up to 95% success and outperforming hand-engineered baselines.