Daniel Huterer Prats, University of Colorado Boulder; Hanspeter Schaub, University of Colorado Boulder
https://doi.org/10.64861/WVYB5020
Keywords: Space Situational Awareness, Space Domain Awareness, Reinforcement Learning, Autonomous Satellite Operations, Space-to-Space Surveillance, Resident Space Objects, Multi-Orbit Imaging, Onboard Autonomy
Abstract:
Growth of the resident space object population increases the sensing rate and responsiveness required for space domain awareness. A space-based optical sensor can observe targets from changing geometries without atmospheric interference, but its scheduler must coordinate access, illumination, attitude motion, onboard resources, and intermittent communication. This paper presents a reinforcement-learning scheduler for persistent space-to-space imaging and catalog maintenance by a low-Earth-orbit observer. The formulation introduces variable-duration imaging and downlink actions, repeated observations after a cooldown period, online high-interest target promotion, and targets in low, medium, and geosynchronous Earth orbit. The scheduler adapts a target-wise attention policy that encodes a changing candidate set separately from spacecraft and mission context information. The computational study trains policies using different mixtures of immediate image-acquisition reward and delayed value delivered to the ground. Two policy families, trained with either 100 low-Earth-orbit targets or a 50/30/20 low-, medium-, and geosynchronous-orbit catalog, are evaluated on identical 100-target mixed-regime scenarios over 100 Monte Carlo seeds. Mixed-regime training improves delivered ground value at every reward weight and shows best ground-value performance in the ? =0.1?0.3 regime. At ? =0.1, the increase over low-Earth-orbit training is 5.55% and the policy offers a practical operating balance, imaging 95.07±0.43 unique targets and revisiting 89.34±0.61 targets per episode. A 200-target catalog-size ablation evaluates zero-shot scale transfer and response to online priority promotions. Within the same evaluation, high-interest and super-high-interest objects promoted at the episode midpoint reach 89.1% successful-image coverage compared with 71.4% for an equal-sized random control group. These results show that representative training geometry and a balanced imagingdownlink reward are important for learning efficient closed-loop, priority-aware policies that acquire, revisit, and deliver imagery while managing spacecraft resources. The policy also responds online to target priority promotions, redirecting both imaging and useful downlink actions toward newly elevated targets without requiring a revised offline schedule.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications