Dex-One2Many: Learning Dexterous Manipulation from a Single Human Demonstration
A single human video is turned into scene-graph guidance for RL, then transferred to a real multi-fingered hand.
Dex-One2Many abstracts the demonstration into sequential scene graphs instead of copying exact motions. Those graphs constrain object relations, letting simulation sample varied resets and provide dense stage rewards. The authors say this keeps exploration guided while allowing new object poses, goal poses, and grasps. Across five tasks, it beat baselines by 6.5% on seen configurations and 71% on unseen ones. HF Daily Papers' note
Dex-One2Many abstracts the demonstration into sequential scene graphs instead of copying exact motions. Those graphs constrain object relations, letting simulation sample varied resets and provide dense stage rewards. The authors say this keeps exploration guided while allowing new object poses, goal poses, and grasps. Across five tasks, it beat baselines by 6.5% on seen configurations and 71% on unseen ones. HF Daily Papers' note
score 5