We’re releasing a demonstration of our research into using AI systems to improve airport surface safety by monitoring voice-based clearances and instructions.
Last month we released Y4, our de novo multimodal aviation model which delivers industry-leading performance at understanding pilot-controller voice communications. This new level of performance and increasing reliability opens the door to research on how to help prevent runway incursions in real-time by detecting potential conflicts between human instructions and the reality on the airport surface. The data shown in this demo is a live view of our real-time analysis at LaGuardia (LGA) and Atlanta (ATL).
America's National Airspace System remains the safest in the world, with commercial carriers transporting more than the global population without a fatal crash between 2009 and early 2025.
Still, the margin of safety on the ground is narrowing. Recent data reveals a significant uptick in runway incursions, defined as the “incorrect presence” of an aircraft, vehicle or person on a runway. Annually, U.S. airports report over 1,500 such incursions. Some airports, such as Dallas/Fort Worth and Boston Logan, experience up to two runway incursions per month on average.
While runway incursions remain an area of concern in the U.S., critical steps are being taken to improve safety on the airport surface. Surface Awareness Initiative (SAI) systems are currently being rolled out at some 200 U.S. airports. At larger airports ASDE-X with Safety Logic works to detect conflicts based on aircraft/vehicle positions and trajectories.
Ultimately, however, a surface safety system must extend beyond the capabilities of today’s technology to deliver the most comprehensive protection. In the current paradigm, systems essentially infer where an aircraft may be in the future by extrapolating speed and direction from its current location. This functionality is important, but incomplete: it cannot account for the human decisions that do not appear in surveillance data.
Has a given aircraft approaching the runway been instructed to hold short? An aircraft just lined up to depart a runway that intersects another active runway — has this aircraft been cleared for takeoff, or merely instructed to "line up and wait?"
In short, a truly comprehensive system must not just understand what is happening; it must also understand why. In most cases, this data is available in just one place: the voice communications of the system participants.
This research is an important step toward a future where AI systems work in tandem with experienced professionals to improve the safety and capacity of the National Airspace System. The flying public stands to benefit from the broad, persistent, and high-bandwidth capabilities of specialized AI systems, teamed 1:1 with the remarkable judgment and reliability of aviation professionals.