Multi-Satellite Custody of Moving Ground Targets

Maintaining persistent custody of ground targets from Low Earth Orbit (LEO) is a key challenge in space-based surveillance and reconnaissance. Traditional rule-based tasking strategies struggle with scalability as the number of assets and targets grows. In this paper, we present a centralized Reinforcement Learning (RL) approach using Proximal Policy Optimization (PPO) to coordinate four LEO satellites tasked with tracking two moving ground targets. PPO was selected for its balance of sample efficiency and stability in high-dimensional spaces, making it suitable for orbital simulations. The satellites operate in fixed orbits without maneuvering, relying on tipping and cueing strategies to maintain custody. Our implementation integrates the Skyfield orbital mechanics library with stable-baselines3 to simulate the environment and train a policy. Results demonstrate that the PPO agent significantly outperforms a heuristic baseline by achieving ∼ 25−30% higher average custody duration (mean 74.5% coverage over 5000 steps vs. 41.4 and 54.9% coverage for random and greedy heuristics respectively) and 30% fewer gaps, with p < 0.01 from t-tests confirming the robustness of improvements. This work highlights the potential of centralized RL for scalable custody management in multi-satellite constellations.