This special episode of the Turing Podcast revisits Project Bluebird – the project that featured on our very first ever episode on the podcast in 2020! Host Amelia Jabry is joined by Ben Carvell, an aviation researcher leading AI agent design on Project Bluebird, and Andrew “Andy” Pace, a London terminal air traffic control officer (ATCO), who describes ATCO work, sectorisation, radar-based operations, and the still-analogue reliance on voice. They examine the growing complexity of air traffic control as traffic returns to, and surpasses, pre-COVID levels and the added challenges from drones, new aircraft types, and emissions constraints, stressing human cognitive limits in a safety-critical environment. Bluebird is a multi-year partnership between the Alan Turing Institute, NATS, and the University of Exeter with three themes: building a high-fidelity digital twin, developing virtual controller agents using methods from rules to optimisation and reinforcement learning, and addressing trust/explainability. Join Amelia, Andy and Ben as they discuss the challenges and opportunities of building an AI system to assist ATCOs in one of the most safety-critical industries in the world. 

00:00 The Hidden Chessboard at the heart of Air Traffic Control

00:37 Meet Ben Carvell and Andy Pace

03:20 What Air Traffic Control Officers (ATCOs) Do

06:20 Safety Rules for aviation and the moral responsibility of ATCOs

08:00 Introducing Project Bluebird

08:52 3 Main themes of Bluebird’s research: Digital Twins, AI Agents for Air Traffic Control, and Trust

10:30 The legal and moral responsibility of Air Traffic Control

12:20 The Science: The Physics Informed ML (PML) behind Bluebird. Including dealing with variability and the unknown.

14:55 Rules Optimisation versus Reinforcement Learning (RL)

15:50 A blast from the past

17:40 A collaboration with Johns Hopkins University tackling the issue of explainability in RL introducing ‘Relevant traffic’ 

20:36 TRUST: Safety Minima vs Collisions in Airspace – what is ‘non-negotiable’?

23:40 Interrogating the AI’s logic and maintaining trust

24:00 A curriculum for machines? The AI’s and ATCOs final exam

26:00 Bringing an ATCO onto project Bluebird – tool support for ATCOs

30:00 Intuition vs learning – the ‘mystical ATCO scan’ 

32:30 The ever-crowding airspace and increasing challenges to ATCOs

33:30 They history of Bluebird: 2020 versus now

35:40 How ATCOs really feel about Bluebird

36:02 Introducing Falcon - Are humans the ultimate guarantor of safety, or is our cognitive limit the ceiling for system efficiency?

40:15 Move 37 and the potential for Discovery

43:20 Collaboration between institutions and the real outputs Bluebird has made (the game!)

44:22 Wrap Up and Find Out More