Trier Mortlock, Lawrence Livermore National Laboratory; Nathan Golovich, Lawrence Livermore National Laboratory; Alex Geringer-Sameth, Lawrence Livermore National Laboratory; Robert Armstrong, Lawrence Livermore National Laboratory; Kshitij Bhardwaj, Lawrence Livermore National Laboratory; Peter McGill, Lawrence Livermore National Laboratory
https://doi.org/10.64861/GNGM8143
Keywords: Autonomous tasking, space situational awareness, reinforcement learning, object tracking, sensor scheduling
Abstract:
Current space situational awareness (SSA) sensing networks face significant challenges tracking an exponentially growing number of Resident Space Objects (RSOs), particularly in the increasingly congested geostationary orbit (GEO) regime. Meeting this challenge requires persistent, autonomous monitoring capabilities enabled by constellations of space-based sensors operating in and around GEO. However, autonomous tasking of GEO-based sensor constellations presents a complex optimization problem as algorithms must balance tracking known objects against discovering new ones, all while conforming to computational constraints and enabling real-time decision-making in dynamic scenarios. Traditional sensor scheduling approaches, including greedy and heuristic methods, can scale poorly as the number of tracked objects and sensors increases. They may also have limited ability to coordinate sensing observations, discover unknown objects, and exploit recurring patterns in RSO distributions while adapting to new scenarios. Reinforcement learning (RL) offers a promising alternative by enabling a constellation of sensors to learn optimal policies through interaction with the environment, naturally handling long-horizon credit assignment and non-linear observation spaces. In this paper, we formulate the GEO-to-GEO sensor tasking problem as a constrained optimization problem with competing objectives, including sensing coverage, revisit frequency, and characterization quality. We develop a multi-agent reinforcement learning (MARL) framework autonomous tasking of coordinated GEO-based sensors. The framework is designed to balance the number of tracked and newly discovered objects, observation frequency, and characterization quality as represented by estimation uncertainty. We describe an evaluation methodology based on simulated orbital dynamics and optical sensing models, including comparisons with optimization-based and other learning-based tasking approaches. The study also considers scalability across constellation configurations, sensor characteristics, and RSO populations, together with computational requirements relevant to onboard and edge processing. This work establishes a framework for investigating the use of multi-agent reinforcement learning for scalable, coordinated tasking of GEO sensor networks in complex sensing environments. _x000D_
_x000D_
This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. The document number is LLNL-ABS-2016204. _x000D_
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications