GenHand: generalised human grasp kinematic retargeting

Abstract Kinematic retargeting is vital for teleoperation and imitation learning, enabling the conversion of human gestures into robotic end-effector actions, particularly in object manipulation tasks requiring robust, human-like grasping. Existing methods primarily focus on replicating hand shapes, especially fingertip locations, but often suffer from reduced grasp quality when adapting to grippers with different morphologies. Thus, we proposed GenHand to address these issues by offering a kinematic retargeting algorithm that generates human-like grasps for different grippers by optimising for force closure and kinematic similarity. The proposed method provides differentiable solutions to generate physically plausible grasping for various grippers based on input human grasping. Extensive evaluations show that GenHand achieves lower net wrench residuals and improved surface contact consistency, and outperforms the key-vector based baseline by 39.8% in simulated experiments across four grippers and twenty objects, while maintaining comparable grasp similarity.