Research/ ORBITAL ROBOTICS
Learning-based Model Predictive Control for capturing tumbling debris using a space manipulator
Ruipeng Liu, Jian Guo, Eberhard K. A. Gill
Learning-based MethodsOpenAlex topic: Reinforcement Learning in Robotics · Manipulation & GraspingOpenAlex keyword: space manipulators · Guidance, Dynamics & ControlOpenAlex keyword: model predictive control
This paper proposes a motion planning approach for a robotic manipulator on a free-floating spacecraft to capture tumbling space debris. The approach combines an improved Proximal Policy Optimization (PPO) algorithm with the Expectation-Maximization (EM) algorithm to optimize high-level control strategies that guide a Model Predictive Controller (MPC) for precise motion execution. Rather than treating the task as a chasing problem, the method formulates it as a future rendezvous process, aiming to synchronize the manipulator’s end-effector with the debris’ graspable point in both position and velocity. To generate smooth capture trajectories, the EM algorithm is used to jointly optimize B-spline control points and timing, resolving the mutual dependency between them. The learned policy is designed to minimize base disturbances and ensure the manipulator performs accurate actions within tight time windows, enabling safe and effective capture even when the target’s motion approaches the manipulator’s dynamic limits.