Autonomous Task Scheduling for Earth-Observing Satellites Tracking Moving Targets With Low Maneuverability

Agile Earth-observing satellites are essential for monitoring dynamic global events such as ships, hurricanes, wildfires, and floods. Scheduling imaging tasks for moving targets is challenging, particularly when future target locations are uncertain. This study investigates autonomous satellite task scheduling using deep reinforcement learning (DRL), explicitly accounting for probabilistic target position estimates that evolve over time. Unlike prior studies assuming perfect observability and fixed imaging locations, this approach incorporates growing location uncertainty and models the required search time for successful imaging based on the estimated target location and its variance. The scheduling problem is formulated as a partially observable semi-Markov decision process (POsMDP) to capture limited state observability and variable-duration actions. Policies are trained and evaluated within BSK-RL, a high-fidelity simulation framework integrating the Basilisk spacecraft simulator with the Gymnasium RL interface. Two key factors are explored: (1) the effect of training environments and observation spaces on policy performance, and (2) the influence of reward function design on balancing target priority and uncertainty reduction. Results show that policies leveraging uncertainty observations outperform those without, and that reward formulations balancing target priority and uncertainty allow effective trade-offs, improving overall mission performance.