Abstract
With the continuous advancement of marine development, underwater operational tasks are becoming increasingly diverse and complex. Addressing the limitations of traditional methods and intelligent planning—which focus solely on acquiring task skills while separating grasp planning from force planning—this paper proposes a modeling approach integrating impedance control with deep reinforcement learning. Using a five-finger humanoid underwater dexterous hand as the grasping execution platform, this method achieves collaborative decision-making between grasp planning and force control for underwater dexterous hands. First, a modular underwater dexterous grasping scenario is established. Its kinematic model and inverse solution are analyzed, and the grasping problem is modeled as a Markov decision process. Second, based on the dexterous fingertip impedance control model for simulation, a grasping strategy learning method grounded in deep reinforcement learning is constructed to address the complex control challenges posed by the high degrees of freedom of the dexterous manipulator. Finally, the Proximal Policy Optimization (PPO) algorithm is employed for grasping strategy learning. An underwater dexterous grasping parallel training and testing environment is established using the Isaac Lab simulation platform to rapidly validate the learning method. Simulation results demonstrate the proposed method’s excellent dexterous compliant control performance and strong robustness to underwater variable environments: the PPO-based impedance control scheme reduces contact force variance by 56% compared to pure position control. The average maximum contact force is suppressed to 3.26 N, representing a 60.4% reduction compared to pure position control. This study achieves the organic integration of underwater hydrodynamic compensation, adaptive impedance control, and grasping strategy learning, providing a novel and effective solution for compliant grasping control of underwater dexterous manipulators.