AI Improves Air Corridor Conflict Resolution with Degraded Surveillance

Researchers have developed a Multi-Agent Reinforcement Learning (MARL) framework, utilizing Deep Q-Networks, to enable decentralized conflict resolution for various aircraft types in structured three-dimensional air corridors. This system is specifically designed to operate safely even when surveillance information is compromised by noise, delays, incompleteness, or temporary unavailability, a critical challenge for Advanced Air Mobility (AAM) operations. Separate policies were trained for small unmanned aerial vehicles (UAVs) and electric vertical takeoff and landing (eVTOL) aircraft, using local observations and a comprehensive 14-action space. The simulation environment incorporated realistic factors such as aircraft-specific dynamics, energy consumption, corridor constraints, observation noise, communication delays, information dropout, wind disturbances, and actuator/model uncertainties. The trained policies were evaluated across 90 combinations of traffic density and separation thresholds, showing that while loss-of-separation frequency increased with density, most events were resolved quickly. Agents maintained their course approximately 79% of the time under safe conditions, with turning, speed control, and vertical maneuvers being key actions during conflicts. The study identified Pareto-optimal configurations, highlighting trade-offs between safety and corridor capacity, and supports simulation-based evaluation of safer AAM conflict-resolution strategies.