Conceptual Use and Architecture of LLM-Agents in Satellite Operations

Abstract

Modern satellite operations face increasing complexity due to growing constellation sizes, mission sophistication, and the need for rapid response to dynamic events. This research presents a system implementation and evaluation demonstrating the integration LLM-based agentic AI systems within the NASA NOS3 simulator. Leveraging locally hosted LLMs, Retrieval-Augmented Generation (RAG), and agent frameworks like LlamaIndex, we demonstrate a system where AI agents assist human operators. These agents monitor satellite telemetry, interpret system states against operational norms, diagnose simulated anomalies using knowledge derived from procedural documentation, propose corrective actions with clear rationale, and execute approved commands via the ground system interface under human oversight. The system architecture incorporates persistent time-series data archival using TimescaleDB, facilitating post-simulation analysis and learning. We detail the architecture, its integration with standard NOS3 components (in particular with cFS as the flight software and YAMCS as ground software), the roles of specific agents (Telemetry Monitoring, Reporting, Procedure Execution), and illustrate the complete workflow through a simulated solar array degradation scenario. This work highlights the potential for agentic AI to reduce operator workload, improve situational awareness, accelerate response times, and enhance the safety and efficiency of satellite operations, particularly demonstrating feasibility using simulation environments and documentation as primary knowledge sources early in a mission lifecycle before extensive flight data is available.