KeyGen: Unsupervised Keypoint based Object-Centric Representations for Category-Level Policy Generalization

KeyGen extracts canonicalized semantic 3D keypoints from point clouds to provide structured object-centric representations for robotic manipulation. By conditioning a visuomotor diffusion policy on these keypoints and object geometry, the framework enables policy generalization across diverse, unseen object instances within the same category.