Putting the AI in air traffic control

The UK’s leading air traffic control provider, NATS, has teamed up with the Turing to explore how the industry can evolve with machine learning

Last updated
Friday 17 Jan 2020


Flight movements above the UK are at a historic high and demand is forecast to increase by up to 40% over the next 10 years. As a result, the technological advancement of air traffic control is a key challenge to providing the robust and resilient transport infrastructure the nation needs for the future. So, it’s perfect timing for the Turing to be in collaboration with NATS (formerly National Air Traffic Services), the UK’s leading air traffic control provider that manages all the UK’s upper airspace, as well as many key UK airports.

"Collaborating with NATS gives us a unique insight into real-world challenges in a new area for machine learning research," says Evelina Gabasova, a Principal Research Data Scientist on the Turing’s Research Engineering Group, who leads the research. The project is part of the Turing’s Data-Centric Engineering (DCE) programme. Its aim is to find out if machine learning can be usefully applied to air traffic control, to both understand the current limitations of the technology, and to begin to pave the way towards automation. In the nearer term the research will also provide critical insights which could contribute to the development of new tools and decision aids for air traffic controllers.

"Collaborating with NATS gives us a unique insight into real-world challenges."

Evelina Gabasova, Principal Research Data Scientist, Turing

Air traffic control (ATC) is enormously complex. Controllers modify the trajectory of each aircraft in order to maintain a minimum distance between them at all times – to avoid “loss of separation” (LoS) between aircraft – while getting them safely to their destinations. Meanwhile, they must also maintain a range of backup plans to ensure safety in the face of the unexpected. Then there’s other considerations like the orderly transfer of planes between sectors, fuel efficiency and environmental improvements, landing sequence optimisation, and more.

To optimise these elements requires looking into each flight’s potential future, predicting possible conflicts and issuing timely instructions. And there are many sources of uncertainty, from aircraft mass to individual pilot behaviour, airline preferences and variations in weather conditions. Any decision-making system applied in air traffic control has to account for these uncertainties while optimising for the other objectives. In other words, this is an exceptionally complex arena which offers rich opportunities for high-impact developments in machine learning.

Watch NATS's data visualisation showing air traffic over the UK in one day.

How did it start?

“The Data-Centric Engineering programme has a range of ongoing major partnerships with multinational companies, regulated industries and UK utilities,” says the programme’s director, Professor Mark Girolami, “so it was natural when NATS first approached me to explore whether a partnership with DCE at Turing could advance their research and development strategies to deliver the next generation of air traffic control services.”

NATS wanted to explore a question: what is the state of the art in machine learning (ML), and what is its potential for air traffic control? “NATS values our openness, willingness to understand the domain and its inherent challenges,” says Gabasova. “They also appreciated our fresh thinking, in the sense that our approaches are not those traditionally taken at NATS.”

There is a dearth of useful ML research in the scientific literature with regard to air traffic control, and one reason is that researchers tend to simplify the issues to such a degree that the research becomes useless in practical terms. For example, they might splice together ATC algorithms from code previously used to direct ground-based robots, which results in ‘solutions’ based mostly on aircraft headings. “This is potentially an unrealistic treatment for ATC, as it omits the more critical element of vertical control,” says Gabasova. So why do some researchers do this? “It's because it’s easier, and you don't have to engage with the industry to build a true understanding of the challenge! When you have a hammer, everything looks like a nail.”

Air traffic controllers at NATS Swanwick Centre
Air traffic controllers at NATS Swanwick Centre managing the London area (Image: NATS)

Turing researchers go deeper: from the outset of this 14-month collaboration, which officially kicked off in October 2018, the research team spent a great deal of time building their understanding of the challenges faced by the air traffic controllers at NATS. “We are trying to solve the challenges that exist in practice, rather than inventing a problem that's easy to solve with an off-the-shelf algorithm,” says Gabasova. This means frequent meetings with NATS, both virtual and in person. Turing researchers were also put into NATS training simulators, to give them a taste of the job.

That’s a human element, but what does an artificial agent ‘care’ about? “From a data science perspective, it's all about defining what actions an AI agent can perform – such as issuing an instruction to change an aircraft’s speed, altitude or heading – and how to find out if a situation is good or bad, expressed as a number. But to get to that point, we have to understand the challenge very well.”

What happened?

The collaboration, which also included the University of Edinburgh's high-performance computer centre EPCC, expended considerable effort building knowledge on both sides, with the researchers getting thoroughly acquainted with the nature of ATC work, and NATS learning how state-of-the-art data-wrangling worked. To this end, NATS were involved with several of the Turing’s Data Study Groups, in which industry and commercial groups each bring a data challenge to the Turing, to be tackled by some of the country’s sharpest data scientists. A big impact from this work was on NATS’s trajectory prediction capability. Realistic trajectory prediction is a crucial aspect of an ATC simulator, and the results of this work will ultimately be built into NATS’s own simulator, currently in development.

For the Turing, the main part of the collaboration has been focused on the development of an open-source experimentation platform for automated air traffic control agents. Based on an open-source ATC simulator called BlueSky, and with a user-friendly interface, this platform allows the outcomes of different algorithmic approaches to be explored using automated AI agents.

The experimental platform, which is also being integrated with NATS’s own simulator, will next be put to use evaluating different ATC algorithms for real-time decision making, with ML techniques ranging from optimisation approaches to reinforcement learning. The optimisation approach essentially treats the ATC environment as a geometric problem to solve: the algorithm makes a plan and executes it.

Bringing in reinforcements

The reinforcement-learning approach may be more promising. Here, an AI agent would have an overview of the ATC environment (i.e. the information available to an air traffic controller) that it would consider, looking for potential safety problems and so on, and then issue commands to its various aircraft. It then receives information about the new state of the ATC environment, and receives "rewards" if the situation is improved, or penalties for, say, creating conflicts between aircraft or issuing too many instructions. Constant updates and feedback will teach an AI what approach works best to meet given goals, hopefully providing valuable insights for human air traffic controllers. 

In collaboration with NATS, the Turing team is working towards turning criteria that are used in human controller’s training into quantitative benchmarks for the ML system.

"We can explore the potential of AI for our industry and it just would not have been possible without the fantastic team at the Turing."

Benjamin Carvell, Senior Researcher at NATS

The experimental platform built by the Turing can generate ATC scenarios automatically, and can scale the complexity. Gabasova says: “We can go from two aircraft – which would be simple for a human air traffic controller to manage – to enormously complicated situations beyond the limit of human-decision making. We want to explore the limits of AI agents.” And, ultimately, turn these insights into practical tools.

“The Turing is unique in having both deep research strength and a large Research Engineering Group that is capable of turning research outcomes into deployable software systems,” says Girolami. “The relationship between NATS and the Turing’s DCE programme is an outstanding example of how research teams, working with research-software developers and an industry partner, can achieve so much more.”

“This collaboration goes right to the core of air traffic control,” says Benjamin Carvell, Senior Researcher at NATS. “Together we have built the research infrastructure we need to truly explore the potential of AI for our industry, and it just would not have been possible without working so closely with the fantastic team at the Turing.

“We have shared many insights on both sides over the last year, and now look forward to the next stage – creating agents which can learn to safely balance the many facets of air traffic control, leveraging new data sources from the real world, and starting to explore the boundaries of what is possible.”

Air traffic controllers
The research is aiming to solve challenges that exist in practice, using data from real world ATC scenarios (Image: NATS)

What does the future hold?

The close collaboration should ultimately contribute to new and better tools and decision aids for air traffic controllers in the critical role they play as guardians of the air, as well as laying the groundwork for future investigations into the broader potential of automation for the industry. The results of this project will guide future work in this area, expanding the capacity and safety of the UK’s airspace.

It has also given the Turing’s Research Engineering group valuable experience of building collaborations in a complex and critical industry. And there should be more to come, says Katrina Payne, the Turing’s Partnerships Development Lead: “We’re in discussion with NATS about a more strategic collaboration, looking to work at scale and build on the strong foundations of this first project.”

“This will accelerate science not just for NATS and for the Turing, but for the wider machine learning community."

Evelina Gabasova, Principal Research Data Scientist, Turing

The Turing is committed to open-source research wherever possible, as exemplified in our Turing Way philosophy. As a result, the benefits of this ongoing work could ripple out across the globe, says Gabasova: “Because we’re making the entire experimental platform open source, other research teams will be able to write their own air-traffic control agents, evaluated on our ATC scenarios, and publish them.”

Such competitions for AI agents are very popular, and can be found in diverse arenas, from computer games to protein folding, which is important in the discovery of new medicines. “This will accelerate science not just for NATS and for the Turing, but for the wider machine learning and reinforcement-learning communities,” says Gabasova.

For now, though, the main goal of the collaboration is simply to enable AI to spread its wings in commercial aviation, where there is huge potential for beneficial impact. Watch this airspace.

PDF Summary

NATS impact story preview