DSML Reading Club #6: Diffusion Policy for robot control

At this seminar, we will discuss “Diffusion Policy: Visuomotor Policy Learning via Action Diffusion” (https://arxiv.org/pdf/2303.04137v4).

The paper presents Diffusion Policy, a new approach to generating robot behavior in which a visuomotor policy is modeled as a conditional denoising diffusion process. It consistently outperforms existing methods by 46.9% on average. The model learns the gradient of the action density function and optimizes behavior during inference through stochastic Langevin dynamics. It handles multimodal distributions and high-dimensional actions well while maintaining stable training. Key technical choices include receding-horizon control, visual conditioning, and a time-series transformer. The work opens a path toward new diffusion-based policy-learning methods.

• Saturday, August 9, 11:00 Kazakhstan time
• Add to calendar
• Google Meet
• Speaker: James Yamazaki
• Host: Yelaman Abdullin