Abstract
Diffusion policies have become a strong option for robotic manipulation under stochastic disturbances because they can represent multimodal action distributions. In standard use, inference relies on a full reverse-time diffusion (denoising) process before an action is applied, while the physical plant is assumed comparatively slow so that plant dynamics and denoising dynamics are effectively decoupled.
This line of work studies the opposite regime: partial denoising—running the reverse diffusion only partway before commanding the robot—so that the plant evolves in parallel with ongoing denoising on the computer. The plant and the diffusion process are then coupled, which raises basic closed-loop questions that do not arise in the usual “frozen plant” picture.
We develop theoretical bounds on closed-loop stability for such coupled systems and outline a framework aimed at faster imitation learning when full denoising at every step is too costly for real-time use. The analysis also highlights how properties of the demonstrations (for example, their variance) enter a stability-oriented view of the controller.