Diffusion Policy

Diffusion Policy

Diffusion Policy is the official implementation of "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion" (RSS 2023), a method that represents a robot's action-generation policy as a conditional denoising diffusion process. Instead of directly regressing actions, it iteratively refines action sequences, which captures multimodal behavior and yields notably more robust manipulation policies.

Key features

• Denoising-diffusion policy that predicts action sequences (receding-horizon control) from visual observations
• Handles multimodal demonstrations and high-dimensional action spaces where naive behavior cloning struggles
• CNN- and Transformer-based backbones, plus image and low-dimensional state variants
• Reproducible training and evaluation on standard manipulation benchmarks (Robomimic, Push-T, and more)
• Config-driven pipeline (Hydra) with checkpoints for both simulation and real-robot experiments

Widely adopted in the robot-learning community, Diffusion Policy has become a strong baseline for imitation-learning-based visuomotor control.

Curated mirror of the open-source Diffusion Policy (MIT). Get it from the source.