Trajectory planning of free-floating space manipulator based on reinforcement learning and smooth parameterization

Abstract

Space manipulators provide an important means for performing in-orbit servicing tasks, but their motion planning becomes challenging when the spacecraft base is free-floating. In this case, joint motion not only drives the end-effector but also induces base motion, so the end-effector response cannot be obtained from conventional fixed-base kinematic relations. As a result, evaluating a candidate joint trajectory generally requires propagation of the coupled system dynamics, and online planning may become computationally expensive. This paper develops a learning-based trajectory generation method for free-floating space manipulators by combining smooth trajectory parameterization with reinforcement learning. The kinematic coupling between the manipulator and the spacecraft base is described through the generalized Jacobian matrix, from which an ordinary differential equation is established to propagate the system state. To obtain compact and smooth joint commands, fifth-order Bézier curves are used to represent the joint-angle histories, leaving only a small number of curve coefficients to be determined. The original planning task is therefore reformulated as the search for suitable Bézier parameters under the coupled ODE model. A DDPG-based policy is trained to infer these parameters directly from the desired end-effector position, avoiding repeated online optimization after training. Numerical tests show that the trained policy can produce feasible joint trajectories for target positions not included in the training stage. For 10,000 randomly sampled test cases, all control parameters are generated within 0.12 s, demonstrating the efficiency of the proposed approach for fast trajectory planning of free-floating space manipulators.