Allan Grosvenor, MSBAI; Abdul Wahab, MSBAI; Kyrylo Bohachov, MSBAI; Anton Zemlyansky, MSBAI; Ryland Adams, MSBAI; Dwyer Deighan, MSBAI
Keywords: autonomous systems, spacecraft control, anomaly detection, hierarchical AI, neuro-symbolic architecture, reinforcement learning, space situational awareness, space domain awareness, maneuver prediction, space traffic management
Abstract:
The rapid proliferation of spacecraft in Earth’s orbital environment has heightened the demand for autonomous systems capable of advanced maneuvering and reliable decision-making under uncertainty. Conventional AI approaches often struggle with the dynamic, noisy, and resource-constrained conditions of space operations, limiting their effectiveness for mission-critical applications such as collision avoidance and anomaly detection. This paper presents a novel hierarchical neuro-symbolic architecture designed to enhance spacecraft autonomy and maneuverability, addressing these challenges through a synergistic integration of structured reasoning and advanced machine learning techniques.
Our proposed system is structured across three specialized layers. The planning agent employs large language models and vision-language models to devise high-level mission strategies, adapting to evolving objectives and environmental contexts. The grounding agent utilizes graph neural networks and Decision Transformers to translate these strategies into actionable commands, ensuring alignment between abstract plans and operational realities. The control agent, driven by reinforcement learning with Proximal Policy Optimization, executes precise maneuvers in real time, optimizing performance in dynamic settings. A standout feature of our approach is the incorporation of Joint Embedding Predictive Architecture (JEPA), which aligns latent embeddings of current observations with predicted future states. This enables robust representation learning, excelling at filtering noise and detecting anomalies in spacecraft behavior.
To address the computational limitations inherent in spaceflight, our methodology leverages GPU-accelerated reinforcement learning, achieving a 1000-fold speedup compared to traditional training methods. This efficiency is further enhanced by simulation-based virtual environments and leadership computing resources, allowing extensive testing and adaptation with limited real-world data. Ensemble methods and meta-learning are integrated to bolster decision-making robustness, ensuring adaptability across varied mission profiles.
Preliminary evaluations in simulated space domain awareness tasks demonstrate exceptional performance, with our system achieving 94 percent accuracy in detecting satellite maneuvers and 98 percent accuracy in classification tasks. Applied to spacecraft control, we anticipate similar gains in maneuver precision and anomaly detection, surpassing existing benchmarks. These advancements promise to enhance autonomy by enabling spacecraft to navigate complex trajectories, respond to unexpected events, and maintain operational safety with minimal human intervention.
The relevance to Space Situational Awareness and Space Domain Awareness is profound. Accurate maneuver prediction supports proactive space traffic management, reducing collision risks in increasingly crowded orbits. Enhanced anomaly detection capabilities enable early identification of malfunctions or adversarial actions, bolstering mission security. The scalability of our architecture ensures it can process multi-source data streams, aligning with the Department of Defense need for reliable AI in critical operations applications.
This work advances the frontier of autonomous spacecraft control by merging symbolic reasoning with machine learning, delivering a system that thrives in the unpredictable space environment. By improving maneuverability, autonomy, and situational awareness, our methodology contributes directly to the sustainability and security of the space domain, offering a transformative tool for next-generation space operations.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications