Glossary
Diffusion Policy
A behavior-cloning policy that predicts short action trajectories via iterative denoising — the dominant imitation-learning baseline in modern robotics.
Diffusion Policy is an imitation-learning approach where the policy predicts a short sequence of future actions by iteratively denoising, the same way image diffusion models generate images. The model is conditioned on the recent observation history and outputs a probability distribution over action trajectories.
Two things make it work well:
- Predicting a trajectory smooths out the policy and handles multimodal demonstrations gracefully (multiple valid ways to do a task).
- The denoising formulation lets the model express uncertainty over action sequences instead of collapsing to a mean.
As of 2026, Diffusion Policy or a close variant is the default baseline for new manipulation work.