Explainable RL takes on air traffic control, one saliency map at a time

Air traffic controllers may one day get help from an AI that explains its reasoning. A new paper posted to arXiv on July 24, 2026, applies explainable reinforcement learning to a simplified air traffic control environment, training an RL agent to reroute flights around no-fly zones. The researchers then use a saliency map to highlight which input features, such as distance to a zone or approach angle, most influenced each decision, a form of transparency with counterparts in other high-stakes domains, such as agentic finance tracking.

The authors argue that building trust in AI is essential for high-stakes domains like aviation, and that explainability is a key part of that trust. This work is a first step; the environment is basic and the explainability technique is preliminary, but it demonstrates how RL-based routing can be made transparent. The effort parallels developments in neuro-symbolic reasoning and RL-driven brain-computer interfaces, where interpretability also takes center stage. More broadly, RL systems require careful reward design to avoid instability, a challenge explored in recent reward collapse research.