Reproduce “GraspGF: Learning Score-based Grasping Primitive for Human-assisting Dexterous Grasping” on OpenResearch

About this paper

Tianhao Wu, Mingdong Wu, Jiyao Zhang, Yunchong Gan, Hao Dong

Researchers at Peking University developed GraspGF, a two-stage framework for human-assisting dexterous grasping that combines a score-based primitive policy with a reinforcement learning residual policy. This approach achieves higher success and generalization rates in simulation and a 66% real-world success rate by adapting to diverse objects and human user intentions.

Key insights

• Decomposing the human-assisting grasping problem into a static 'how-to-grasp' primitive and a dynamic 'when/how-fast-to-grasp' residual policy enables more efficient and generalizable learning.
• Score-based generative models can effectively learn robust, low-level control primitives from static synthetic grasp data, reducing the need for extensive human demonstrations.
• The combination of a score-based primitive and an RL-based residual policy allows for adaptive, user-centric control that accounts for both object geometry and real-time user intentions and trajectories.

Summary by alphaXiv.