search

lucidrains / diffusion-policy

Implementation of Diffusion Policy, Toyota Research's supposed breakthrough in leveraging DDPMs for learning policies for real-world Robotics
MIT License
89 stars 1 forks source link
artificial-intelligence attention-mechanisms deep-learning denoising-diffusion robotics transformers

readme

Diffusion Policy (wip)

Implementation of Diffusion Policy, Toyota Research's supposed breakthrough in leveraging DDPMs for learning policies for real-world Robotics

What seemed to have happened is that a research group at Columbia adapted the popular SOTA text-to-image models (complete with denoising diffusion with cross attention conditioning) to policy generation (predicting robot actions conditioned on observations). Toyota research then validated this at a certain scale for imitation learning with real world robotic demonstrations. It is hard to know how much of a breakthrough this is given corporate press is prone to exaggerations, but let me try to get a clean implementation out, just in the case that it is.

The great thing is, if this really works, all the advances being made in text-to-image space can translate to robotics. Yes, this includes stuff like dreambooth.

Discord

Todo

Citations

@article{Chi2023DiffusionPV,
    title   = {Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
    author  = {Cheng Chi and Siyuan Feng and Yilun Du and Zhenjia Xu and Eric A. Cousineau and Benjamin Burchfiel and Shuran Song},
    journal = {ArXiv},
    year    = {2023},
    volume  = {abs/2303.04137},
    url     = {https://api.semanticscholar.org/CorpusID:257378658}
}
@article{Sauer2023AdversarialDD,
    title   = {Adversarial Diffusion Distillation},
    author  = {Axel Sauer and Dominik Lorenz and A. Blattmann and Robin Rombach},
    journal = {ArXiv},
    year    = {2023},
    volume  = {abs/2311.17042},
    url     = {https://api.semanticscholar.org/CorpusID:265466173}
}