Topic 08 of 17

Imitation & diffusion policies in robotics

Imitation learning trains robot policies from demonstrations rather than reward signals, and diffusion policies, which generate action sequences the way image models generate pictures, made it work at scale by handling multimodal demonstrations gracefully. This family powers most current manipulation results.

arXiv query: cat:cs.RO AND (abs:"imitation learning" OR abs:"diffusion policy" OR abs:"behavior cloning")