Diffusion Policy

Diffusion Policy (Chi et al., RSS 2023, extended in IJRR 2024) represents a robot's visuomotor policy as a conditional denoising diffusion process: instead of directly predicting actions, it progressively denoises from noise into action sequences via stochastic Langevin dynamics. The formulation gracefully handles multimodal action distributions (push left or right are both valid), suits high-dimensional action spaces, and trains with unusual stability.

Benchmarked across 15 tasks from 4 robot manipulation benchmarks, it consistently beat state-of-the-art robot learning methods by an average of 46.9%. With receding-horizon control, visual conditioning, and a time-series diffusion transformer, plus open code and data, it became a default method in robot learning — directly influencing pi-zero's flow matching and Octo's diffusion decoder, and shipping as a standard policy in Hugging Face's LeRobot.