Capture of Free-Floating Targets Using End-to-End Reinforcement Learning: Theoretical Foundations and Ground Demonstrations

Abstract

This doctoral research addresses the challenge of autonomously capturing free-floating, tumbling space debris using robotic manipulators for debris removal in space. Space debris poses a growing risk to near-Earth operations and satellite services. The dissertation explores autonomous grasping of non-cooperative debris in microgravity through Deep Reinforcement Learning (DRL) and advanced robotic manipulation. A high-fidelity simulation environment is developed using the PyBullet physics engine with domain randomization to train and evaluate DRL agents under realistic dynamics. The agents use hierarchical control strategies, combining curriculum-based learning for path planning and tactile sensor feedback for force regulation to minimize disturbances to the target.

Simulation studies of multiple DRL algorithms, including Soft Actor-Critic, show success rates over 90% in grasping tasks with varying debris shapes and conditions. A Hardware-in-the-Loop testbed at York University, with dual Fanuc robots and active gravity compensation, further tests these policies, while additional experiments at the University of Luxembourg’s Zero-G Lab confirm the system’s adaptability in microgravity. Results demonstrate that the agents can align and capture free-floating targets without destabilizing forces. Hardware tests verify high grasp stability even with spinning debris, showcasing the potential for autonomous space debris removal. This research advances space robotics by integrating DRL methodologies with experimental validation, enabling safer, more efficient future debris removal missions.