AI System Uses Decision Transformers to Handle Satellite Anomalies With Sparse Data

A researcher developing an autonomous satellite operations AI encountered a critical limitation when a simulated telemetry blackout exposed the system's inability to act under extreme data scarcity. The study focuses on satellite anomaly response, where missing or contradictory sensor data, long sequential dependencies, and severe cost asymmetry make standard reinforcement learning approaches unreliable. The researcher found that Decision Transformers, which frame reinforcement learning as a sequence modeling problem, offer a promising foundation but still struggle with data sparsity and sudden distribution shifts during anomalies. To address these gaps, the work explores human-aligned adaptations that incorporate expert operator judgment into the decision-making pipeline. The goal is to build AI systems capable of making high-stakes satellite decisions with minimal data while remaining aligned with experienced human oversight.