Motion planning of space manipulator for capturing a tumbling non-cooperative target using reinforcement learning

Abstract

Space manipulators will play a substantial role in future space missions such as On-Orbit Servicing (OOS), Active Debris Removal (ADR), manufacturing, etc. However, motion planning of the free-floating base is highly complicated

due to the coupling motion between the base and the manipulator. Also, those missions are facing new challenges since targets are usually non-cooperative and tumbling. In this thesis, docking for a non-cooperative rotating target after proximity operation has been examined with a free-floating two-finger gripper system. First, two spacecraft systems, single-arm and dual-arm spacecraft, have been designed. Generalized Jacobian Matrix (GJM) is applied to describe their coupling kinematic motion. Second, a state-of-the-art DDPG algorithm is adopted for motion planning. It has an advantage over model-based algorithms because of computation time and robustness in a dynamic environment. It does not require an uncertainty boundary. This thesis suggests to consider the capture angle in motion planning, not only the distance between the end-effector and the target. Lastly, we trained the motion planning algorithm in the OpenAI gym custom environment for simulation.