Rapid neural motion planning for dual space manipulators under closed-chain kinematic constraints

In the context of on-orbit service tasks involving multiple space manipulators, closed-chain kinematic constraints are a common and important class of constraints. Current Sampling-based Constrained Motion Planning (SCMP) algorithms are usually time-consuming. This paper investigates the Neural Motion Planner (NMP) based on reinforcement learning and presents an object-centered action space to achieve efficient sampling on the constraint manifold. To reduce the difficulty of training and improve the success rate of the NMP algorithm, this paper proposes a reward function with target connectivity as the termination condition and an attitude interpolation method based on rotation exponential coordinates. Finally, a comprehensive comparative analysis between the SCMP algorithms and the NMP algorithm was conducted in the cooperative transport task of dual space manipulators with fixed bases. The planning success rate and planning time of each algorithm were tested, thereby validating the effectiveness of the proposed method.