# Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

**Authors:** Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, Shuran Song  
**arXiv:** [2303.04137](https://arxiv.org/abs/2303.04137)  
**Date:** 2023-03-07  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2303.04137v5/

## Key points

- The paper introduces a novel approach called Diffusion Policy for generating robot behavior by representing visuomotor policy as a conditional denoising diffusion process.
- Diffusion Policy outperforms existing state-of-the-art robot learning methods with an average improvement of 46.9% across 12 tasks from 4 robot manipulation benchmarks.
- Key innovation includes learning the gradient of the action-distribution score function and iteratively optimizing through stochastic Langevin dynamics steps, offering advantages like handling multimodal action distributions and training stability.
- Technical contributions include incorporating receding horizon control, visual conditioning, and time-series diffusion transformer to leverage diffusion models for visuomotor policy learning on physical robots.
- The authors aim to inspire new policy learning techniques using diffusion models and provide access to extended journal version, code, data, and training details at diffusion-policy.cs.columbia.edu.

## Summary

The paper "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion" introduces a novel approach to generating robot behavior by representing a robot's visuomotor policy as a conditional denoising diffusion process. The study benchmarks the Diffusion Policy across 12 different tasks from 4 distinct robot manipulation benchmarks and demonstrates its superior performance compared to existing state-of-the-art robot learning methods with an average improvement of 46.9%. The key innovation of the Diffusion Policy lies in its ability to learn the gradient of the action-distribution score function and iteratively optimize with respect to this gradient field during inference through a series of stochastic Langevin dynamics steps. This formulation offers significant advantages for robot policies, including effectively handling multimodal action distributions, suitability for high-dimensional action spaces, and remarkable training stability. To fully leverage the potential of diffusion models for visuomotor policy learning on physical robots, the paper presents several crucial technical contributions such as incorporating receding horizon control, visual conditioning, and the time-series diffusion transformer. The authors aim to inspire a new generation of policy learning techniques that can harness the powerful generative modeling capabilities of diffusion models. Furthermore, an extended journal version of the original RSS2023 paper is available for further exploration. The code, data, and training details related to this research are publicly accessible at diffusion-policy.cs.columbia.edu. This comprehensive study not only advances the field of robotics but also paves the way for future advancements in policy learning techniques using diffusion models.