Hand2Robot: Object-Centric 4D Hand Reconstruction from Human Demonstrations for Robotics

Hand2Robot is a framework that reconstructs 4D hand motion from human demonstrations to provide robot action supervision. It extracts interaction clips, reconstructs object geometry, and optimizes hand-object interactions for robot policy post-training. Experiments on DexYCB and HO3D-v3 show improved object pose accuracy and interaction quality, with real-robot tasks demonstrating effective data supplementation for policy training.