Diffusion policy

Diffusion policy is a neural-network architecture for robot control that generates joint commands by iteratively denoising from a noisy initial state — adapting diffusion-model techniques from image generation. Particularly effective for multi-modal manipulation policies where a deterministic policy would average inconsistent demonstrations into incoherent behavior.

The distinction matters because diffusion policy was one of the breakout architectures of 2023–2024 for learned manipulation. Physical Intelligence's π0 and π0.5 series use diffusion policy as the action decoder; the original Diffusion Policy paper (Toyota Research / Stanford) showed strong gains over earlier behavior-cloning approaches. The architecture has become a default building block for VLA-style models, and the multi-modality property makes it natural for imitation-learning data with inconsistent demonstrations.