Diffusion Policy
Visuomotor policy learning via conditional denoising diffusion. Columbia University.
Overview
Diffusion Policy represents robot behavior as a conditional denoising diffusion process. It handles multimodal action distributions, high-dimensional action spaces, and exhibits strong training stability. Average +46.9% improvement over prior methods across 15 manipulation tasks.
Architecture
- Receding horizon control
- Visual conditioning
- Time-series diffusion transformers
- IJRR 2024
Official Links
- diffusion-policy.cs.columbia.edu — Project site
- github.com/columbia-ai-robotics/diffusion_policy — Code, data, Colab
Citation
IJRR 2024. See the project site for BibTeX.