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:
The rapid growth of resident space objects (RSOs) across low, medium, and geosynchronous Earth orbits continues to challenge existing Space Situational Awareness (SSA) and Space Domain Awareness (SDA) architectures. Unlike ground-based systems that are traditionally used for this purpose, space-based SSA sensors are generally unaffected by atmospheric turbulence, weather, and horizon masking effects [1]. Their continuous motion provides a dynamic field of regard that is much larger than a single ground sensor. This comes with improved revisit characteristics, and they can achieve higher sensitivity for detecting smaller or dimmer RSOs [2]. These advantages motivate increased interest in autonomous space-based surveillance architectures capable of persistent monitoring and catalog maintenance._x000D_
_x000D_
Prior RL-based autonomy efforts for space-based sensor tasking demonstrate the feasibility of learning imaging policies for LEO platforms observing GEO targets [3], however, these approaches do not incorporate spacecraft safety constraints, onboard resource management, or imaging–downlink tradeoffs. References 4 and 5 apply more mature RL methods to Earth-observing satellite scheduling under predictable target motion and event-based tasking. However, RSOs exhibit diverse and rapidly evolving relative trajectories that demand fundamentally different decision-making strategies compared to stationary or slow-moving ground targets. Reinforcement learning is particularly well-suited to this problem because it can learn complex scheduling policies that balance competing objectives including imaging opportunities, resource constraints, downlink windows, and safety margins in highly dynamic environments where traditional heuristics struggle to predict achievable slew and dwell times._x000D_
_x000D_
While prior work demonstrates the basic feasibility of autonomous space-to-space imaging using RL for LEO targets, significant gaps remain in extending these capabilities toward realistic multi-orbit operations, closed-loop catalog maintenance, and operationally meaningful performance metrics. This paper presents a substantial extension of previously published work at AMOS [6], expanding both the problem formulation and the autonomy architecture to better reflect real-world space surveillance operations._x000D_
_x000D_
Building on the prior AMOS contribution [6], this work advances the original single-objective, fixed-duration LEO-to-LEO imaging formulation into a multi-objective scheduling framework that explicitly distinguishes between onboard imaging reward and downlinked, ground-usable imagery. This separation reflects operational realities in which images only contribute to SSA value once successfully transmitted and delivered to operators. The resulting formulation enables the RL agent to reason not only about when and which RSOs to image, but also about when imaging should be deferred in favor of downlinking or resource recovery to maximize long-term mission utility. Similar imaging–downlink tradeoffs have recently been explored by the authors in a related multi-orbit context [7], showing that significant behavior and performance differences can be obtained based on reward-balancing._x000D_
_x000D_
This work introduces three major extensions beyond prior approaches. First, whereas previous work [7] trained agents on LEO-to-LEO scenarios and applied them to multi-orbit regimes, this paper trains agents directly on LEO-to-any-orbit environments, including LEO, MEO, and GEO targets, capturing large variations in relative velocity, lighting geometry, access duration, and pointing agility requirements during training. Second, the fixed imaging duration assumption used in all prior formulations [6, 7] is relaxed in favor of variable imaging durations up to a defined maximum, enabling rapid imaging actions immediately upon completion of slews when geometry permits. This allows the agent to adapt imaging time based on relative motion, target priority, onboard resource state, and expected downstream utility. This new capability is expected to yield significant performance gains and highlights a major challenge of heuristic approaches that struggle to estimate slew times and visibility windows effectively. Third, this work introduces repeated re-imaging of RSOs as a step toward closed-loop catalog maintenance. Rather than treating targets as one-time objectives, the agent is incentivized to revisit high-interest RSOs over time, maintaining spatial-temporal coverage and enabling persistent characterization, a capability not addressed in prior RL-based space-based imaging work [3, 6, 7]. This reframing aligns autonomous space-based surveillance with operational SSA needs, where maintaining accurate and up-to-date catalogs is as critical as initial imaging. Moreover, because this work employs an online RL approach, it enables new high-interest RSOs to be added to the catalog at any time and immediately incorporated into imaging decisions without requiring time-intensive offline replanning of the observation sequence._x000D_
_x000D_
The RL agent operates within a high-fidelity and flight-proven Basilisk-based simulation environment using the BSK-RL framework. Compared to prior work, pointing constraints are tightened to reflect conservative optical safety margins and realistic agility limits. Training studies demonstrate that increased batch sizes and extended training durations yield substantial improvements in policy performance, particularly in multi-orbit scenarios dominated by sparse but high-value imaging opportunities._x000D_
_x000D_
Overall, this paper demonstrates that reinforcement learning provides a scalable and adaptable framework for autonomous space-to-space surveillance across multiple orbital regimes. By incorporating multi-objective reward structures, variable imaging durations, and closed-loop re-imaging, the proposed approach moves significantly closer to operationally relevant SDA architectures and supports the growing role of onboard autonomy in future space surveillance missions._x000D_
_x000D_
References:_x000D_
_x000D_
[1] M. R. Ackermann, R. R. Kiziah, P. C. Zimmer, J. T. McGraw, and D. D. Cox. A systematic examination of ground-based and space-based approaches to optical detection and tracking of satellites. In 31st Space Symposium, Colorado Springs, CO, April 2015. URL: https://www.spacesymposium.org/wp-content/uploads/2017/10/M.Ackermann_31st_Space_Symposium_Tech_Track_paper.pdf_x000D_
_x000D_
[2] B. Lal, A. Balakrishnan, B. M. Caldwell, R. S. Buenconsejo, and S. A. Carioscia. Global trends in space situational awareness (SSA) and space traffic management (STM). Technical Report D-9074, Log H 18-000179, Institute for Defense Analyses Science and Technology Policy Institute, Washington, DC, April 2018. URL: https://apps.dtic.mil/sti/pdfs/AD1123106.pdf_x000D_
_x000D_
[3] P. M. Siew, D. Jang, T. G. Roberts, and R. Linares, “Space-Based Sensor Tasking Using Deep Reinforcement Learning,” Journal of the Astronautical Sciences, vol. 69, no. 6, pp. 1855–1892, Dec. 2022._x000D_
_x000D_
[4] L. Q. Mantovani, Y. Nagano, and H. Schaub, “Reinforcement Learning for Satellite Autonomy under Different Cloud Coverage Probability Observations.”_x000D_
_x000D_
[5] M. Stephenson and H. Schaub, “Reinforcement Learning for Earth-Observing Satellite Autonomy with Event-Based Task Intervals.”_x000D_
_x000D_
[6] D. Huterer Prats, H. Schaub, and C. Wheeler, “Reinforcement Learning for Space-to-Space Surveillance: Autonomous Scheduling for Resident Space Object Imaging,” AMOS Conference._x000D_
_x000D_
[7] D. Huterer Prats and H. Schaub, “Autonomous Space-Based Imaging: Reinforcement Learning Scheduling with Imaging–Downlink Tradeoffs Across Orbital Regimes,” AAS-26-052, AAS/AIAA Space Flight Mechanics and GN&C Conference, 2026._x000D_
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications