AI at work: mastering the airspace
By Aaron Karp and Ben Iannotta|November 2024
U.K. researchers are in the midst of Project Bluebird, a multiyear effort to create competing versions of artificial intelligence agents to see if one could in theory be capable of autonomously controlling air traffic. Aaron Karp and Ben Iannotta spoke to researchers about their plans to conduct live trials in the safety of a digital twin of the airspace.
NATS, the quasi-private company that controls much of the United Kingdom’s airspace, expects jet travel to continue increasing annually by 2% for the foreseeable future. While that’s happening, the U.K. has pledged to meet net-zero carbon emissions by 2050 through its Jet Zero initiative, and that means aircraft must fly the most efficient routes possible. Meanwhile, in urban areas, a host of eVTOLs — electric vertical takeoff and landing aircraft — should be ready for trials as air taxis in a couple years. NATS must be ready for them to share the airspace.
Put simply: The U.K. has laid out “targets and expectations” for air traffic control of these complicated skies, says NATS researcher Richard Cannon. He is principal investigator for Project Bluebird, a NATS-led effort to apply artificial intelligence, machine learning and digital twinning to the problem of more efficiently managing the airspace under its control.
Various versions of AI agents, sometimes called “digital controllers,” are in development under the project. Researchers want to find out how well the best of them performs compared to human air traffic controllers. The twin could serve as a foundation for developing ever more sophisticated agents.
“Once you’ve got that digital twin tapestry, once you’ve got that little ecosystem in place, actually testing and validating these agents can take a matter of months from start to finish,” Cannon says.
But for now, NATS and its Bluebird partners, the Alan Turing Institute in London and the University of Exeter, have in mind accomplishing a specific milestone: a 2026 experiment in which an AI agent will be fed real-time information about the location and trajectories of aircraft and other factors available to human controllers. The agent will send instructions to digitally represented aircraft within a digital twin of the airspace. The performance of the agent will then be compared to that of the human controllers on the operations floor in a building known as Swanwick, shorthand for the London Area Control Centre. The controllers there send instructions to aircraft flying over England, Wales, Northern Ireland and Scotland — in short, anywhere within most of the nation’s airspace.
“It’s really busy,” Cannon says.
In these “shadow mode live trials,” the twin must “accurately model the flight paths of aircraft,” explains Nick Pepper, a research associate for Project Bluebird at the Turing Institute. That means displaying the actual locations of the planes as they are known to the controllers, while also accurately predicting where the aircraft will be.
“The idea is to see how far we can get in building an independent controller that can function autonomously,” the Turing Institute said in a summary of the experiment provided to us. But because an autonomously functioning digital controller “would rightly take a very long time to regulate and get into operations, (even if we were able to perfectly automate the entire task which is unlikely), the plan is to then step back to exploring more realistic short to medium term use cases,” Turing explained.
A digital controller could, in the medium term, serve as a “decision support” tool for human controllers, says Marc Thomas, a software engineer and the digital twin lead at NATS. But, adds Thomas, having in mind that an autonomously functioning agent puts no ceiling on the work, the hope is that regulators will start contemplating how and whether they would regulate such digital controllers or agents.
“We want to have an eye on pushing the regulation for higher levels of automation, otherwise, is that going to happen?” Thomas says.
Whether a digital controller could ever replace a human one remains an open question that won’t be answered in 2026.
Safety has been top of mind for those planning the live trials. The whole process of attaining live flight information must meet the U.K.’s cybersecurity standards for critical national infrastructure.
“I have to make sure that I have secure egress from the operations room,” Cannon says.
Setting up the hardware and software is already well underway and is scheduled to be fully in place by Easter 2025. The result will be a strictly one-way flow of flight data: “There’s no way in,” Cannon says, and there will be “firewall provisions.”
The flight data will flow into servers in the operations building and to the nearby NATS corporate center, where it will be uploaded into a commercial cloud infrastructure. There, NATS has already established a “sandbox” environment for researchers.
“It’s like a play zone where we’ve got our academic team, we’ve got our postdocs, our Turing people, my team leading on various things,” Cannon says.
The researchers have developed multiple AI agents and aren’t sure yet which one they’ll use in 2026. Ten million NATS historical flight records were made available for developing the agents and to calibrate the digital twin of the airspace.
“So some people come in and develop and want to use quite a lot of that data,” Cannon explains, “and some people can find leaner methods and use maybe a million flights or maybe even half a million flights.”
As for the digital twin, “You can play it like a computer game,” says Thomas — a version of it, anyway, one with simplified scenarios amenable to control with a keyboard and mouse. “It exists for public engagement (e.g. British Science Festival), where the full scenarios and input instructions would be too complicated for the general public,” according to Turing.
Preparatory trials with historic flight records began last year and will culminate with the 2026 live trials. These first trials involved the most advanced Bluebird AI agent so far, plus air traffic control officers and collegiate trainers. The AI agent controlled historic air traffic over the London Middle sector, 20,000-30,000 feet above London, while at the same time air traffic control officers directed historic traffic in the next layer up of airspace, the London Upper Sector that’s 30,000-60,000 feet. The human controllers then commented on the performance of the AI agent.
Turing describes the results as “promising.” Feedback from the air traffic control officers has been especially valuable, because it has helped developers make the twin more realistic.
“Sometimes you show it to the air traffic controllers and they say, ‘Well, that’s not really realistic. You know, the aircraft wouldn’t have come in there, and that sector is in the wrong shape,’” Thomas says.
In the live trials, success will depend in large part on how well the digital twin reflects the real world of air traffic control. Human controllers live in a world of incomplete information: They don’t know the precise strength of the winds aloft, including tailwinds. They don’t know exactly how fast a pilot will climb or descend, or the actual cruise speed that will be attained, or a host of other factors. The competing agents can’t operate in “some synthetic, rarefied, simplified world that we’ve given it,” Cannon explains. “What these agents are actually doing is decision making with uncertainty.”
The digital twin will therefore have to be probabilistic, and it’s all cutting edge, Cannon says.
“We’re building most of what we do from foundational principles, where there’s not much you’re taking on off the shelf for this. We are having to build a lot from the start. That’s why we use the Turing Institute. That’s why we use Exeter” — and sometimes other universities — “because we’re building these kind of algorithms for the first time.”
Less exciting, but perhaps no less important, is the role AI agents and the digital twin could one day play in empowering controllers to operate strategically, even though that is not currently a central goal of the project.
Today, most air traffic control decisions “are made quite tactically,” Thomas says. “You’re really literally looking at the traffic that’s in the air coming your way.”
Even if an AI agent has no authority to guide aircraft, machine learning capabilities could empower it to instantly process thousands of scenarios, giving human controllers a much more advanced look at likely — and unlikely — traffic flows, he explains. The agent could conduct “strategic planning at a higher level on an hours, days and sometimes weeks sort of timeline, in terms of how much traffic is going to go through” the sector, he adds.
For the U.K., the stakes can be seen in the number of flights. Air traffic controllers in the U.K. process as many as 8,000 aircraft daily. That number is expected to keep rising, according to NATS, putting more pressure on the controller workforce, which may not grow sufficiently enough to keep up with demand. The answer, NATS believes, could lie in having AI agents in the mix — in a precise fashion to be decided.
