Towards AI-based Air Traffic Control

Air traffic, despite the recent dip due to Covid, is expected to grow 30-40% year on year. With the potential inclusion of UAVs (Unmanned Aerial Vehicles) into controlled airspace over the next decade, it is anticipated that the congestion levels in airspace will increase 10 fold. This paper presents an AI-based approach to air traffic control, with the aim of alleviating the load and improving the efficiency of human agents (air traffic controllers). One of the primary goals of air traffic control is to safely navigate an aircraft through controlled airspace using real-time control actions – such as changes to speed, heading (direction of travel) and altitude of an aircraft. The safety critical nature of this environment calls for precise explanations (why take an action) and counterfactual (why not take an action) explanations, real-time responsiveness, the ability to present succinct actions to a human agent, while simultaneously optimizing for air traffic delays, fuel burn rates, and weather conditions. This paper presents algorithms and a system architecture for anticipating separation losses (conflicts in airspace) and a lattice-based search space exploration AI planner to recommend actions to avoid such conflicts. The key contributions of the paper include: (i) fast detection (prediction) of conflicts in a controlled airspace, and (ii) fast lattice space exploration based AI solver to produce a set of feasible resolutions for the detected conflicts. Additionally, this paper discusses how to weight the different resolutions and how future work on optimisation techniques could improve the efficiency of the algorithm and address various known limitations of the current approach from both technical and human-agent perspective. The evaluations are conducted against an air traffic simulator, Narsim, showing the ability to avoid separation losses, while minimizing the number of actions even at 3 x normal capacity.