Dynamic Grasping for Free-Flying Space Robots via Deep Lagrangian Network Assisted Reinforcement Learning

Capturing moving targets is crucial in various space missions. However, the base and manipulator coupling of free-flying space robots (FFSRs) imposes challenges to traditional analytical modeling and control approaches. On the other hand, model-free reinforcement learning faces long training cycles, difficulty in parameter tuning, and model inaccuracy. This paper proposes a novel model-based reinforcement learning framework grounded in the Deep Lagrangian Neural Network. This framework employs deep Lagrangian networks to train a high-fidelity system dynamics, with which FFSRs are capable of generating a significant volume of virtual experiences based on current states, enhancing the iterative policy optimization process in reinforcement learning algorithms, thereby substantially improving training speed and control accuracy in moving targets grasping. Simulations conducted on an FFSR with 12-degree-of-freedom (6 for the manipulator and 6 for the base) validate the superior performance of the proposed framework.