Hierarchical Policy Learning for Humanoid Robots Whole-Body Dexterous Manipulation

Abstract

Humanoid robots equipped with dexterous hands hold immense promise for performing complex space sampling missions. However, achieving dexterous whole-body manipulation involving sequential locomotion, grasping, and manipulation remains challenging, primarily due to the high-dimensional complexity of interactions between humanoid robots and objects. To address this issue, we propose a novel hierarchical policy-learning framework designed explicitly for whole-body humanoid dexterous manipulation tasks. The proposed approach systematically decomposes the entire task into three sequential subtasks: (i) locomotion for approaching the target object, (ii) grasping with high-degree-of-freedom dexterous hands, and (iii) loco-manipulation to transport the grasped object along a specified trajectory. We first train decoupled skill priors separately for the robot’s body and hand to capture a diverse set of motion skills, and subsequently integrate these skill priors through a high-level policy to achieve coordinated whole-body interaction. Simulation experiments demonstrate that our method effectively completes the entire humanoid robot and object interaction task, significantly enhancing the robot’s task efficiency and autonomy. The proposed method provides a new perspective on applying humanoid robots to challenging whole-body dexterous manipulation tasks, such as sample collection during space missions.