Extracting Learning Signal from Suboptimal Robot Data with Ambient Diffusion

giannis_daras · x · 2026-07-30

Ambient Diffusion Policy introduces a method for imitation learning using suboptimal data in robotics. The research shows that Diffusion Policy learns different data features at varying noise levels: global task structure at high noise and local motion refinement at low noise. By restricting suboptimal data usage to specific noise levels through a simple dataloader modification, the algorithm extracts meaningful signals while ignoring harmful features.

Related event: Ambient Diffusion Wins Dual Awards for Training Robots with Suboptimal Data(2 posts)→

Original post →

More from Embodied

Embodied channel →