Transformer-Based Reinforcement Learning for Autonomous Orbital Collision Avoidance in Partially Observable Environments

Researchers have developed a new AI system for satellites to autonomously avoid collisions, even when sensor data is noisy or incomplete. It uses a transformer-based architecture, similar to advanced language models, to better interpret uncertain information over time. This approach is more reliable than traditional methods in partially observable environments. The framework includes a configurable simulator and a state estimator to handle the uncertainty of relative motion in orbit.