All Tools
D
OtherFreeOpen Source
DIFFUSION POLICY
Visuomotor robot policies learned by denoising action trajectories
MIT
ABOUT
Imitation policies often collapse to one mode when human demos contain several valid ways to complete a contact-rich task. Diffusion Policy models the action sequence as a diffusion process so robots sample multimodal trajectories from vision or state observations instead of averaging incompatible behaviors.
INTEGRATION GUIDE
1. Train a visuomotor policy from human demos for contact-rich tabletop tasks
2. Compare state-based and vision-based diffusion policies on simulation benchmarks
3. Reproduce RSS 2023 checkpoints and logs for manipulation research
4. Generate multimodal robot action sequences instead of a single mean trajectory
TAGS
pythonroboticsimitation-learningdiffusionvisuomotormanipulationpytorch