Event
Feedback World Model Proposed to Online-Correct Diffusion Policy Predictions
1 reports · 1 sources
Tuo An et al. propose a feedback world model that, at inference time, uses the real states observed after robot execution to online closed-loop correct predictions, without requiring additional training data or parameter updates, and introduces action-aware guidance to transform the corrected predictions into control. The method can be interpreted as a latent-space observer, with convergence guarantees given bounded asymptotically convergent dynamics residuals and stable feedback gains. The paper reports that in experiments on LIBERO-Plus, Robomimic, and real manipulation tasks, the method improved prediction accuracy and policy performance under distribution shift, reducing world model prediction error by up to 76.4%. The paper has been accepted to NeurIPS 2026.
Report timeline
- Feedback World Model Enables Precise Guidance of Diffusion Policy (NeurIPS 2026)
This paper proposes a feedback world model that uses the true states observed after robot execution to perform online closed-loop correction of predictions at inference time, without additional training data or parameter updates. The method can be interpreted as a latent-space observer and provides convergence guarantees under bounded, asymptotically convergent dynamics residuals and a stable feedback gain; it also introduces action-aware guidance to better translate corrected predictions into control.
arXiv cs.RO ·