# Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

**Authors:** Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, Shuran Song  
**arXiv:** [2303.04137](https://arxiv.org/abs/2303.04137)  
**Date:** 2023-03-07  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2303.04137v4/

## Key points

- Authors introduce a novel approach to generating robot behavior using diffusion models
- Diffusion Policy represents a robot's visuomotor policy as a conditional denoising diffusion process
- Evaluation across 12 tasks shows consistent outperformance of existing methods with an average improvement of 46.9%
- Key innovation is the ability to learn the gradient of the action-distribution score function and optimize it iteratively through stochastic Langevin dynamics steps during inference
- Advantages include effectively handling multimodal action distributions and suitability for high-dimensional action spaces with remarkable training stability
- Technical contributions include receding horizon control, visual conditioning techniques, and time-series diffusion transformer
- Aim is to inspire new policy learning techniques leveraging generative modeling capabilities offered by diffusion models
- Commitment to making codebase, data sets, and detailed training procedures publicly available for reproducibility and further research

## Summary

In their paper titled "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion," authors Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song introduce a novel approach to generating robot behavior through the use of diffusion models. This approach involves representing a robot's visuomotor policy as a conditional denoising diffusion process. The proposed Diffusion Policy is evaluated across 12 different tasks from 4 distinct robot manipulation benchmarks and consistently outperforms existing state-of-the-art methods with an average improvement of 46.9%. The key innovation of this method lies in its ability to learn the gradient of the action-distribution score function and optimize it iteratively through stochastic Langevin dynamics steps during inference. This formulation offers significant advantages for robot policies such as effectively handling multimodal action distributions and being suitable for high-dimensional action spaces while maintaining remarkable training stability. To fully harness the potential of diffusion models for visuomotor policy learning on physical robots, the authors present several crucial technical contributions in their work including incorporating receding horizon control, visual conditioning techniques, and the time-series diffusion transformer. By leveraging these advancements, the authors aim to inspire a new generation of policy learning techniques that can leverage the powerful generative modeling capabilities offered by diffusion models. Furthermore, they commit to making their codebase, data sets used in experiments, and detailed training procedures publicly available to promote reproducibility and further research in the field of robot learning. Overall,this work represents a significant step forward in enhancing robot behavior generation through innovative diffusion-based approaches.