Abstract

Project Bluebird is a collaboration between NATS (the UK's leading Air Traffic Control provider), the Alan Turing Institute, and the University of Exeter. The goals of the project are: to build a probabilistic Digital Twin of UK airspace; to develop and train AI agents to perform Air Traffic Control (ATC) in that Digital Twin, and to investigate and promote the safety and trustworthiness of ATC tools.
We are currently three years into this five-year project, and are at the stage where the Digital Twin can replay historical scenarios, or generate artificial ones, and where humans, AI agents, or a mixture of both, can control aircraft, which then respond realistically. The "sandbox" nature of the digital twin is ideal for prototyping and testing different types of agent (e.g. rules-based, reinforcement learning, optimization), and the more advanced of these are able to deal with complex scenarios and high volumes of air traffic, safely and efficiently.
Extensive and continuing feedback from real Air Traffic Controllers and Instructors has been invaluable for refining the behaviours of the agents, and also in improving the usefulness of the frontend and the information displayed to users. This close collaboration will continue for the remainder of the project, hopefully resulting in a tool that will promote human-AI collaboration in this complex and safety-critical service.