|
Diffusion models have demonstrated strong performance as imitation policies and motion planners in robot learning. However,
diffusion models generally involve hours of training black box neural networks from expert data. In this talk, I'll discuss our recent work on imposing structure on diffusion by viewing it as a modular plug-and-play architecture that can incorporate diverse
target distribution specifications, sources of refinement, and inference-time modifications. We show how to enforce constraints on diffusion planners, perform imitation learning in milliseconds without any training, and perform inference-time neural diffusion
policy editing.
|
