Diffusion Policy represents a visuomotor policy as a conditional denoising diffusion process over robot actions, so the policy can represent several valid ways of doing the same thing.
Findings
- Benchmarked on 12 tasks from 4 robot manipulation benchmarks, with an average improvement of 46.9% over the best earlier robot learning methods.
Limits
- Abstract-level reading only.
Relevance
A common baseline downstream: OpenVLA reports beating it by 20.4% when fine-tuned on new settings, and in RoboTwin 2.0 it falls from 28.0% on clean tasks to 0.6% under randomisation. That collapse is the open question for robust robotics: what does a multimodal action distribution do when the observation itself is off-distribution? Available in LeRobot.
Links
- arXiv
- Subfield: Robot learning