Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

Framed robot visuomotor action generation as a conditional diffusion process over continuous action trajectories, outperforming traditional Gaussian mixture and deterministic policies on multimodal manipulation tasks.

Serves as qualified theoretical and systems foundation for task-continuous-control, task-robotics.

Assumptions

• Markovian state dynamics and stationary reward functions hold in target evaluation environments

Limitations

• Sample efficiency, exploration stability, and real-world sim-to-real transfer gaps require specialized tuning