Diffusion Policy

Key Insight

Diffusion policy represents the action distribution of a robot as a denoising diffusion model conditioned on current observations, allowing it to excel at complex, multi-step manipulation tasks. Unlike standard behavior cloning which uses deterministic networks that fail when demonstrations show multiple valid paths, diffusion policies handle multimodal distributions naturally by gradually refining random noise into smooth action trajectories. This iterative generation process ensures that the robot makes a clear, decisive choice (such as passing an obstacle on the left or the right) instead of outputting a hazardous average of all demonstrations.