project bluebird
NATS Partnership with the University of Exeter and the Alan Turing Institute
UK airspace is undergoing a momentous transformation. Passenger numbers are expected to increase by 50% by 2050, and new categories of aircraft are expected to proliferate, from drones to electric vertical takeoff and landing vehicles (eVTOLs). At the same time, the aviation sector faces increasing pressure to reduce emissions and operate more sustainably. This concentration of power is putting an unprecedented strain on an air traffic management system that has remained largely unchanged for more than half a century.
Project Bluebird, a £13.7m EPSRC-funded collaboration, was set up to tackle this challenge head-on. Bringing together more than 90 experts from NATS, the Alan Turing Institute and the University of Exeter, the project is developing an AI-enabled air traffic control system designed to set new global benchmarks for automation in safety-critical environments. Live trials are scheduled to begin in the UK in spring 2026 and will be a major milestone in the evolution of airspace management.
At the heart of this project is a series of innovations that redefine what is possible in air traffic management. The central result is the creation of a new AI controller combined with the world’s first rigorous framework for evaluating performance against human standards. This combination of technical capability and trusted evaluation addresses two of the biggest barriers to automation adoption in aviation: performance and trust.
The foundation of this work is a high-fidelity, probabilistic digital twin of the real airspace environment. Unlike traditional simulators, this digital twin integrates real-world operational procedures with live or historical traffic data. Its probabilistic design takes into account uncertainties such as changes in weather or variations in aircraft arrival times, allowing the system to test AI behavior in both routine operations and rare, high-stress edge cases.
Designed for flexibility and scalability, digital twins can run in the cloud or locally, support multiple concurrent users, and replicate the layout and visualization tools of a real operating room. This makes it a powerful platform not only for research, but also for training, evaluation, and operational prototyping.
This system is supported by several technological advances. A physics-based machine learning model of aircraft performance combines the interpretability of physics-based modeling with the adaptability of data-driven techniques, delivering a 40% increase in accuracy compared to current industry standards. The team also developed a pipeline that uses large language models to automatically generate training scenarios and uses graph neural networks to predict demand for controller tasks.
The AI agent team has considered a wide range of approaches, from reinforcement learning and Monte Carlo tree search to constrained optimization and graph-based techniques. They have also developed a rules-based agent designed for future systemized airspace. With a focus on reliability and explainability, agents can now view plans directly within the digital twin and air traffic controllers can provide feedback through structured workshops.
The LiveShadow trial, planned for 2026, will be the first real-world test of an AI agent working alongside a human controller in an active environment.
Perhaps the most important innovation of this project is the world’s first AI air traffic controller competency assessment framework. For the first time, AI agents can now be evaluated using the same criteria applied to human trainee controllers. NATS training airspace and scenarios have been integrated into a digital twin to enable a breakthrough test in March 2025 in which NATS instructors will conduct over 60 hours of human loop evaluation. The AI agent achieved passing scores in three of the four competencies assessed, and the team is aiming for a full pass by mid-2026.
In 2026, Bluebird plans to release an open source version of its digital twin with an agent training framework and benchmarking agents. This dramatically lowers the barrier to entry for researchers around the world, allowing them to focus on applying cutting-edge AI technology without having to first master the complexities of air traffic control. The framework will also support competition at major international AI conferences and establish the UK as a leader in global aviation innovation.
Modernization in this safety-critical environment has historically been constrained by slow, expensive, and disconnected innovation pathways. Project Bluebird was also built on the recognition that neither academia nor industry can solve air traffic control challenges alone. Cutting-edge AI research often lacks the operational context needed for real-world deployment, and operational expertise alone cannot deliver the breakthroughs needed to manage the airspace of the future.
By overcoming cultural, institutional, and geographic barriers, Project Bluebird achieved results that no partner could have achieved alone.
Bluebird bridges this gap by bringing together the complementary strengths of NATS’ operational knowledge, the Alan Turing Institute’s leadership in data science and AI, and the University of Exeter’s modeling and simulation expertise.
Collaboration is built into the structure of the project. The study is structured around three research themes, each co-led by academics and industry leaders. This ensures a balance between academic priorities such as innovation, experimentation and publication with NATS’ need for practical, secure and operationally sound solutions.
Reconciling different cultures and time scales was difficult. Academia values novelty and peer-reviewed artifacts, while industry prioritizes reliability, safety, and implementation. Through collaborative leadership and open negotiation, these differences became strengths and produced rigorous and directly applicable research.
Geographical distance posed another challenge, as the team was spread across the UK. Projects are tightly integrated with regular face-to-face workshops and a clear governance structure.
Collaboration is characterized by a commitment to sharing domain knowledge. All participants, including academics, engineers, and data scientists, received on-site training at NATS’ technical school, the same facility used for air traffic controller trainees. This common ground ensures that research findings are real, relevant and readily testable, avoiding the disconnects that often undermine research industry partnerships.
The result is a true community that transcends academia and industry. By overcoming cultural, institutional, and geographic barriers, Project Bluebird achieved results that no partner could have achieved alone.
The impact of this project is already being felt across industry, regulation, research and public engagement. At its core is BluebirdDT, a software suite that integrates high-fidelity digital twins with benchmark AI agents, visualization environments, and training and evaluation platforms. For the first time, the UK has a safe and realistic testbed to develop and validate new automation concepts before they go into production.
Incorporating BluebirdDT within NATS’ innovation incubator reduced prototype delivery time and costs by 80%. The secure cloud-based instance is already in use and is central to NATS’ recruitment process, supporting the assessment of thousands of trainee controller applicants each year. Plans are underway to expand its use in airspace design and controller training, including support for the UK Airspace Design Strategy and a proposed third runway at Heathrow Airport.
The system’s machine-learned trajectory generator predicts the aircraft’s path with very high accuracy. Independent research suggests this feature can increase capacity by up to 40% without compromising safety. Accurate predictions also enable greener routing, such as avoiding persistent contrails, which account for nearly half of aviation’s climate impact. This provides a reliable path to reducing the sector’s environmental footprint.
Bluebird also shapes the regulatory environment. By providing a functional platform for exploring automation concepts, NATS can provide civil aviation authorities with tested insights rather than theoretical proposals. It supports the evolution of the regulatory framework for AI in safety-critical systems and provides a model for how regulators, service providers and operators can work together to prepare for the next generation of aviation automation.
Skills development is also a huge accomplishment. More than 100 people have contributed directly to this project, including 15 PhD internships and 12 PhD opportunities, dedicatedly supporting neurodiverse researchers. The University of Exeter’s new PhD training center, established in partnership with NATS, will create a continuing pipeline of AI and ATC experts.
The open source release of BluebirdDT is scheduled for April 2026, in parallel with a large-scale machine learning competition aimed at connecting the global research community with real-world air traffic control challenges.
