Directly skill-transferred DMP-based trajectory planning for robot soft-capture of space tumbling targets

Abstract

The trajectory planning of space robot for capturing space tumbling targets still faces challenging technical problems due to the moving targets, which requires the robot to precisely pinpoint the target in a reliable, safe way. Hard-capture strategy that only cares about the precise pinpoint of the final capturing states cannot ensure a reliable capture due to its lack of a thoughtful approaching strategy. To this end, this paper proposes a soft-capture planning method based on the imitation learning theory for the space robot, considering the robotic approaching strategy by imitating the target’s motion. The most frequently used imitation learning algorithm, named dynamical movement primitives (DMPs), are employed to plan both the positions and orientations of the robot end-effectors in Cartesian space to capture grapple fixtures fixed on the target, in which the pose DMPs are specially formulated to capture such a moving target. To imitate the target’s motion, the soft-capture approaching strategy directly generalizes the motion skill from the target to the robot by learning the weight vectors of the pose DMPs for the robot from the demonstrated target’s pose trajectories. Moreover, the final capturing time, poses, and velocities/angular velocities are considerately optimized to ensure good robot kinematic manipulability. Numerical simulations implemented on a dual-arm space robot validate the effectiveness and generalization of the proposed soft-capture planning method, which shows a longer imitated tracking process for about 4 sec before capturing than that of the hard-capture strategy for less than 1 s The longer imitated tracking process leads to more time for the robot capturing adaptations, which is more reliable for the space capturing tasks.