Fault Resilience of Reinforcement-Learning-Based Autonomous Satellite Task Scheduling

This paper investigates the resilience of deep reinforcement learning for autonomous satellite task scheduling under reaction wheel fault conditions. The scheduling problem for an agile Earth-observing satellite is formulated as a partially observable decision-making framework that captures realistic operational constraints, including power management, momentum control, and imaging opportunities. Four representative fault modes are considered: complete loss of control, power limitation, increased friction, and encoder measurement faults. To improve robustness, three training strategies are evaluated: nominal training, curriculum learning, and domain randomization. Policies are trained within a high-fidelity spacecraft simulation environment and evaluated across varying target densities and fault severities. Performance is assessed in terms of cumulative mission reward and mission safety, defined by successful completion without battery depletion or reaction wheel saturation. Results show that policies trained only under nominal conditions exhibit gradual performance degradation as fault severity increases, whereas curriculum learning and domain randomization mitigate this degradation while maintaining comparable performance in the nominal environment. Domain randomization provides the most consistent performance across fault conditions and outperforms a heuristic scheduling approach in most fault scenarios, demonstrating the potential of deep reinforcement learning for resilient autonomous satellite task scheduling.