Autonomous Task Rescheduling in a Heterogeneous LEO Satellite Constellation Using Reinforcement Learning

Earth observation satellites play a critical role in global monitoring, environmental analysis, and disaster response. However, the rapid growth of low Earth orbit (LEO) satellite constellations introduces new operational challenges, including communication disruptions, contested orbital environments, limited onboard resources, and increased scheduling complexity. Traditional satellite management systems rely on centralized, human-operated task planning, which lacks the agility and resilience needed to respond to failures, intermittent connectivity, or dynamic task demands. This work presents a decentralized autonomy framework for heterogeneous LEO constellations, integrating high-fidelity Basilisk spacecraft dynamics with a reinforcement learning environment to evaluate task allocation strategies over the Continental United States (CONUS). We compare three decision-making approaches, rule-based queue scheduling, a myopic greedy heuristic, and a learned Proximal Policy Optimization (PPO) algorithm, under identical task distributions, resource models, and single-satellite failure scenarios. Results show that while baseline methods sustain basic task execution, they lack priority awareness, coordination, and resilience following the loss of a high-capability satellite. In contrast, the PPO agent consistently prioritized higher-value observations, learned to anticipate resource usage, and demonstrated promising retasking behavior after failure.