基于强化学习的空间机器人在轨装配柔顺控制方法

To address challenges such as vibration of assembly components, dynamic coupling between robots and structures in on-orbit assembly by space robots, as well as difficulties in parameter tuning, and suboptimal control performance in existing methods, a model-data hybrid driving approach is proposed that integrates impedance control with deep reinforcement learning to enable efficient learning of assembly strategies. Firstly, a modular on-orbit assembly scenario for a segmented space telescope is established, and the dynamic coupling between the free-floating space robot and the assembly components is analyzed. The modular assembly task is then formulated as a Markov decision process. Subsequently, joint impedance control of the space robot is introduced as a prior model, and a deep reinforcement learning-based assembly strategy learning method is developed to address the dynamic coupling effects between space robot and assembly components. Finally, the proximal policy optimization (PPO) algorithm is employed to learn the assembly strategy. To facilitate rapid validation of the proposed assembly strategy learning method, a parallelized training and testing environment for on-orbit assembly by space robots is constructed using Isaac Gym. Simulations and analyses demonstrate the proposed method’s effectiveness in improving compliant control performance and robustness against uncertainty.