Enhancing Space Manipulator Fault Tolerance for In-Orbit Servicing through Meta-Reinforcement Learning

This study develops a data-driven motion planning framework for a 7-DoF robotic manipulator to enhance fault tolerance against joint failures during the pre-capture phase of In-Orbit Servicing missions. The system is trained using meta-Reinforcement Learning (meta-RL), exposing the agent to a wide range of randomized scenarios, including stochastic joint-locking. This process achieves two main goals: it equips the agent with strong generalization capabilities for conditions beyond its training domain, and it bypasses traditional inverse kinematics by learning a direct mapping from sensor data to joint-space commands. The meta-learning framework is shown to significantly improve the system’s ability to autonomously manage unforeseen failure events. In non-critical scenarios, the agent successfully learns to repurpose redundant joints to complete its task. However, the system cannot recover from failures that cause an unrecoverable reduction of the manipulator’s workspace. The results demonstrate that this approach leads to more predictable autonomous behavior during failures and, most importantly, shows how some unexpected anomalies may be overcome by simply reframing them as an operational constraint, before critical operations begin.