Anthropomorphic Grasp Motion Planning for Humanoid Robots via Learned Riemannian Metric and Dextrous Grasp Evaluator

In this work, we tackle the challenge of generating human-like grasp motions on humanoid robots—a task that remains difficult for traditional methods due to the complexity of modeling human arm joint preferences and synthesizing dexterous grasps for novel objects. Our approach, termed anthropomorphic grasp motion planning (AGMP), integrates deep learning models by leveraging on human demonstrations and annotated grasps. Specifically, we learn a Riemannian metric in the joint space from human data to capture the directional preferences of upper limb movements, and formulate a quadratic Riemannian distance function as the optimization objective. To assess the grasp quality, we propose a dexterous grasp evaluator that accepts candidate grasps and object as inputs, returning scalar costs. By exploiting the differentiable nature of these networks, we iteratively solve the problem through approximations to both the learned Riemannian distance and the grasp constraint. Simulation results demonstrate that our approach reduces trajectory errors by up to 40% compared to the minimum-jerk model. Moreover, our approach improves the success rate by 11% and significantly reduces planning time relative to existing methods. We further validate our method on a humanoid robot prototype in realworld experiments, achieving a 73.3% success rate.