Project Bluebird The mission to create a digital twin of UK airspace
Continue reading...
Our industry has a long history of pioneering the latest technologies to extend the limits of human performance. From radio telephony and radar to flight data processing and satellite surveillance. Technology now allows previously unimaginable numbers of aircraft to safely fly all over the world every day. So, with that track record in mind, can we expect Artificial Intelligence – heralded just about everywhere as the next big technology revolution - to have a role in air traffic management?
That’s exactly the question being posed by researchers and data scientists in the UK as part of the ground-breaking Project Bluebird.
The Project Bluebird team
A collaboration between the UK air traffic service NATS, the Alan Turing Institute - the UK’s leading body for data science and AI - and Exeter University and the University of Cambridge, with government funding from UK Research and Innovation, Bluebird aims to create a ‘digital twin’ of UK airspace.
What is a digital twin?
“A digital twin is a digital model of a physical object, person or process that is specialised and calibrated to that particular entity. Digital twins can help an organisation realistically simulate potentially thousands of situations, ask “what if” questions and ultimately allow it to make better decisions.”
Richard Everson is a Professor of Machine Learning, a Turing Fellow, and the Academic lead for AI agent design on Bluebird.
Within this virtual sky, AI controllers will be developed and honed, with a view to transforming how real controllers are trained, how airspace change is planned and how the next generation of ATM tools are produced. Now two years into a five-year programme, it is already making startling progress.
But what is it about AI that makes it so uniquely suitable to the task? Why use AI at all? “Artificial Intelligence is really what you use when nothing else will work,” laughs Ben Carvell, AI Design Lead for the project at NATS. “When something is so complex it makes it almost impossible to apply ‘rules-based’ programs, that’s when you turn to Artificial Intelligence.”
Ben Carvell is the Industrial Lead for AI Agent Design on Project Bluebird, a collaboration between NATS, The Alan Turing Institute, Exeter University and the University of Cambridge looking at applications of AI to Tactical Air Traffic Control.
He has worked as a researcher at NATS for 10 years across a range of automation projects, and is passionate about emerging technologies and their potential to transform the future of ATC.
It seems that air traffic control is so complicated, with so much uncertainty and so many variables, that it would be next to impossible to build a traditional computer program - one that relies on the logic-based rules of ‘if x = y then do z’ - sophisticated enough to replicate what a controller does every day. “The sheer variety they face, plus the need to find solutions to non-standard scenarios – be it bad weather, a closed runway or an aircraft emergency –means you need something smarter, something that’s able to learn and adapt and deal with that uncertainty. You need AI and Project Bluebird is our effort to do just that.”
Artificial Intelligence is really what you use when nothing else will work
Ben Carvell, NATS AI Design Lead
In which case, is the aim to replace controllers altogether? Will AI soon be taking over the skies above the UK? “We’re a very long way from AI being able to replace a human controller, but there is a clear role for it to help us improve the safety and efficiency of UK airspace and to support human controllers with tools that help them with decision making” says Ben. The focus then is on helping NATS better understand how efficient its airspace is, to support the training of new and experienced controllers, and to test new airspace designs and procedures. “The beauty of it,” gleams Ben, “is that using our digital twin, we can run tens of thousands of highly accurate, realistic simulations in hours. We can learn things in days it might have taken months or even years to do beforehand.”
We can learn things in days it might have taken months or even years to do beforehand.
Ben Carvell, NATS AI Design Lead
Bluebird is built upon a foundation of 10 years of data, capturing almost every single commercial aircraft movement in UK airspace. It’s an incredibly rich data set to feed into the digital twin and provides a vividly real playground in which the project’s AI ‘agents’ can go about simulating the role of a controller. “What we are trying to do is understand the whole task of what a controller does, so it needs to be as realistic as possible if this can really have value.”
Putting the Artificial Intelligence through its paces
So far, the team has succeeded in creating a digital replica of two UK airspace sectors in which the agents are put through their paces. In practice, this has taken the form of 3 months’ worth of simulations , complete with radar display and flight information strips, the whole works. Ben says: “What we’ve done so far is set up the agent to work on one controller workstation, with a real controller working right next to them. The agent’s performance is then assessed by an instructor like they would be in our training college. We want to see what behaviours the AI shows, how it copes with uncertainty, whether it can ‘fail safely’ when something goes wrong.”
And the real-life controllers? "They loved being part of it."
Support from training and operations is the lifeblood of projects like this, so we were really happy to see how willing people were to engage with the research.
NATS Principal Investigator, Richard Cannon
After three months of simulations, what results have the team seen so far? “We’re marking the agents against the standard we set for our basic course trainees , so while we’re not setting any records, we have seen definite improvement over the course of the simulations. That’s exactly what we’d hope for.”
But in one particularly intriguing twist, it seems not all AI agents are created equal. The Bluebird team has actually programmed multiple agents to have their own way of tackling problems. “We’ve given them different attributes to see how they perform because that can teach us different things. One agent ‘plays’ each simulation like a video game . It plays, gets a mark and then tries again. It can do this 10,000 times, learning each time and refining its approach.” Obviously, that’s not how any real controller would behave, but NATS hopes the approach will enable the AI to better deal with uncertainty and be resilient to sudden change, something all controllers have to cope with.
AI taking control in the simulators
For Ben one of the most fulfilling elements of the project has been seeing industry and academia working together. “Working with Turing, Exeter and Cambridge means we’re approaching this whole project with real academic vigour. Our work will be peer reviewed and we know we’re working the absolute best in the business . For them, it also means they’re working on a project with that has a genuine, real-world application.” NATS hopes all the project will help it best direct future investments in airspace design and technology. The freedom to test and prove the veracity of new tools in an environment that exactly mirrors real-world conditions will reduce the time and cost of developing and introducing these things in what is a safety critical, 24/7 environment.
Working with Turing, Exeter and Cambridge means we’re approaching this whole project with real academic vigour. Our work will be peer reviewed and we know we’re working the absolute best in the business.
Ben Carvell, NATS AI Design Lead
So, with three years still to run and hugely impressive results already achieved, what’s next for Bluebird? “Ultimately” Ben says, “we want to be the first AI system in the world to control a sector of airspace in shadow mode.” That will mean setting up an AI agent to control traffic as it presents in reality, while an actual controller does the job for real. “The results of that will be fascinating.” They certainly will.
What does The Turing Institute hope to get out of the project?
As the national institute for data science and AI, the Turing is engaging in programmes and projects that have impact and innovation. Project Bluebird is a great example of a large scale data-driven project with real challenges that need innovative theoretical and practical solutions. By bringing together industry and academia, AI, data science and software engineering will exemplify the Turing’s role in the national AI ecosystem.
What are the biggest challenges the team as encountered in creating the digital twin?
Digital twins are effective when they are spacialised and calibrated to a particular domain. An exciting challenge for the Bluebird digital twin has been to bring the specialised expertise about aircraft and air traffic control into the software embodiment of the digital twin. Other significant challenges, which bring huge opportunities, are the fantastic amount of data on air traffic that NATS has which informs and lets us calibrate a probabilistic digital twin, and of course the software engineering to make a fast, scalable twin covering many sectors.
Why use AI for a project at all for a project like this? Is there something unique about ATC?
Chatting with the general public at the British Science Festival showed that many people think that AI is already commonplace in ATC. AI has been shown to be extremely effective in limited scenarios, such as games, and air traffic control is a fantastic opportunity to investigate the feasibility of using AI in a much more general, uncertain and unpredictable situation. While machines could have the advantage of unflagging vigilance, the safety critical nature of ATC brings exciting challenges for understanding how AI might work with people and how safety principles can be effectively embodied in AI.
Richard Everson is a Professor of Machine Learning, a Turing Fellow, and the Academic lead for AI agent design on Bluebird.