Beyond Safety Logic: The Case for AI-Autonomous Runway Protection

ANALYSIS / TECHNOLOGY SIDEBAR — March 23, 2026

Technology Analysis  ·  Surface Safety  ·  Artificial Intelligence

The LaGuardia collision forces a fundamental question: after three decades of incremental sensor investment, is the time overdue to remove the human controller from the loop as the last line of runway defense?

The architecture of runway incursion prevention in the United States has not fundamentally changed since the early 1990s. Every improvement since the 1991 LAX disaster — ASDE-3, AMASS, ASDE-X Safety Logic, Runway Status Lights — has followed the same design philosophy: give the controller better information and hope the controller acts on it. Sunday night at LaGuardia, that philosophy may have failed again in the worst possible way. If the ASDE-X Safety Logic alarmed and the controller could not act — overloaded, task-saturated, simultaneously managing two emergencies — then the system did exactly what it was designed to do and still produced two dead pilots. That is not a sensor failure. It is an architectural one.

In virtually every other safety-critical domain that has undergone AI transformation in the past decade — autonomous vehicles, industrial process control, air-to-air collision avoidance (TCAS), even modern fighter aircraft departure protection systems — the design principle has shifted from informing the human to autonomously intervening when the human is not fast enough or situationally aware enough to prevent a collision. The runway environment is one of the last holdouts of the inform-and-hope model. The LaGuardia accident is a powerful argument that this holdout is no longer defensible.

What ASDE-X Safety Logic Actually Does — and Doesn't Do

ASDE-X Safety Logic (AXSL) is a rule-based algorithm that uses fused surveillance data from radar, multilateration, and ADS-B to project the future positions of tracked targets and flag geometric conflicts. It generates visual and aural alerts to the tower controller. It does not transmit anything directly to aircraft. It does not illuminate lights independently (that is RWSL's role). It does not halt vehicle movement. It does not transmit a stop command. It alerts a human and then waits.

The inherent latency of this human-in-the-loop design is the central vulnerability. From the moment Safety Logic detects a conflict geometry to the moment an alert is acknowledged, processed, communicated, and acted upon, multiple seconds elapse. A CRJ-900 decelerating on landing rollout at even 30 mph travels 44 feet per second. The window between a Safety Logic alert and an unavoidable collision can be measured in seconds — a window that a controller managing two simultaneous emergencies cannot reliably exploit.

An alert that fires and is not acted upon is operationally indistinguishable from no alert at all. The system informed the human. The human could not respond. The architecture accepted that outcome as designed.

Additionally, Safety Logic's rule-based conflict detection logic carries a fundamental tension that engineers on the original AMASS program understood well: the system must be tuned aggressively enough to catch real conflicts but conservatively enough to avoid flooding controllers with nuisance alerts from ARFF vehicles, fuel trucks, and service equipment that routinely operate near runway thresholds. If that threshold was set too conservatively at LGA — suppressing alerts for slow-moving ground vehicles near taxiway intersections as a noise reduction measure — a real conflict would have generated no alert at all. This is not a hypothetical failure mode. It is a known, documented challenge of rule-based conflict detection in complex surface environments.

The Gap Map: Where Current Technology Falls Short

Layer	Current Capability	Critical Gap	Status
Surface Radar	Skin-track of all surface targets; feeds Safety Logic	Passive sensor only; no autonomous action authority	Deployed
ASDE-X Safety Logic	Rule-based conflict detection; aural/visual alert to controller	Human-in-the-loop; tuning tension between sensitivity and nuisance; no direct vehicle intervention	Gap
Runway Status Lights	Automated pavement lights warn vehicle operators at hold bars; independent of ATC clearance	Only warns at static hold bar locations; cannot override a vehicle already in motion mid-runway; driver override possible	Gap
Vehicle ADS-B	Optional; major airports have partial coverage; provides positive-ID tracking	Not mandated; ground vehicles at many airports including ARFF units may not carry ADS-B; no cooperative ID guarantee	Gap
Cockpit Awareness	ADS-B In equipped aircraft can display surface traffic; CDTI available	CRJ-900 at landing rollout speed has near-zero avoidance window; pilot awareness does not prevent collision	Gap
AI Predictive Conflict Engine	Research stage (GNN-LSTM, STGNN approaches active in academic literature); SESAR/FAA exploratory work underway	Not deployed operationally at any U.S. airport; no certified system exists; certification pathway undefined	Not Deployed
Autonomous Intervention Authority	None in current architecture	No system has authority to halt ground vehicles, transmit direct stop commands, or initiate go-arounds without ATC action	Does Not Exist

What AI Could Change — and the Research Already Underway

The academic literature on AI-driven airport surface safety has advanced substantially in the past three years. Researchers at multiple institutions have demonstrated that Graph Neural Network / Long Short-Term Memory (GNN-LSTM) hybrid architectures can predict surface conflict probability at taxiway intersections with meaningful lead time by learning the spatiotemporal patterns of surface traffic from historical ADS-B and radar data. A 2025 study published in IET Intelligent Transport Systems applied a GNN-LSTM model to Shenzhen Bao'an International Airport's surface traffic, demonstrating the ability to predict potential conflicts at hot-spot intersections ahead of the conflict event. A 2024 Transportation Research Part C study applied spatial-temporal graph convolutional networks to airport surface movement prediction, fusing ADS-B and airport operations data to model surface traffic state. Research published in MDPI Aerospace demonstrated that a graph-based deep learning ATC system, simulated over 24 hours, avoided 100 percent of potential collisions and prevented 89.8 percent of potential conflicts in simulation.

These are not mature operational systems. They are research demonstrations with important limitations — they operate on historical data, they have not been hardened against sensor noise and adversarial conditions, they have not been certified to any aviation standard, and the path from simulation result to FAA deployment is measured in years and hundreds of millions of dollars. But the underlying signal from the research community is clear: AI systems that predict conflict trajectories and act with latencies far below human reaction time are technically achievable in the surface domain.

Meanwhile, the FAA's own Surface Awareness Initiative (SAI) — accelerated after the 2023 runway incursion wave — deploys simplified ADS-B-based surface surveillance at airports that lack ASDE-X, with the specific design decision to initially omit conflict alerting in order to speed deployment. The FAA's Runway Safety Alerting Subgroup has explicitly recommended expansion of ground vehicle ADS-B transponder equipage as a runway safety priority, recognizing that a vehicle without cooperative surveillance is a target the system cannot reliably track by radar skin-return alone in all conditions. These are steps in the right direction. They are not sufficient.

A Proposed Architecture: Five Layers of Autonomous Protection

What would a genuinely AI-augmented runway protection architecture look like? Drawing on both current operational systems and near-term research capabilities, a coherent next-generation design would comprise five layers — each with autonomous action authority, not just advisory output to a human controller:

1. Mandatory ADS-B Equipage on All Airside Vehicles Vehicle ADS-B is currently voluntary and unevenly deployed. The ARFF truck at the center of Sunday's collision may or may not have been broadcasting a cooperative track. Mandatory ADS-B Out on all vehicles operating in the airport movement area — ARFF, fuel, baggage, maintenance — eliminates the reliance on radar skin-tracks as the primary detection mechanism for ground vehicles and provides positive identification data to every downstream safety system. This is a regulatory mandate, not a technology development problem. Advisory Circular AC 150/5220-26 already provides the framework; it needs to become a requirement at Part 139 certificated airports.
2. Replace Rule-Based Safety Logic with AI-Trained Conflict Prediction The fixed geometric rule-sets of ASDE-X Safety Logic should be replaced or augmented with a machine-learning conflict prediction engine trained on the full corpus of historical surface movement data — including near-misses, incursion events, and normal operations — at each specific airport. A GNN or transformer-based model can learn the nuanced difference between an ARFF vehicle crossing a runway on a routine non-emergency basis (low conflict probability) and an ARFF vehicle crossing on an emergency dispatch while an aircraft is on final rollout (high conflict probability). This contextual awareness is exactly what rule-based systems cannot achieve without generating catastrophic nuisance alert rates. The result is dramatically better sensitivity without increased false alarm burden — directly addressing the alert-suppression problem that may have contributed to Sunday's accident.
3. Extend RWSL to In-Motion Vehicle Stop Authority Runway Status Lights warn vehicle operators at static hold bars. Once a vehicle has been cleared and is in motion on the runway, RWSL provides no additional alert. The system needs a second mode: dynamic in-motion conflict lighting that illuminates directly ahead of a vehicle already crossing an active runway when a landing or departing aircraft is detected on a collision course. This does not require new pavement infrastructure — it requires extending the RWSL conflict logic to trigger lights that are already installed beyond the hold-bar position. Combined with a radio stop command issued autonomously by the conflict detection system — not waiting for a controller to recognize and verbalize the stop order — this creates a multi-channel intervention that does not depend on a single human's situational awareness.
4. Aircraft Cockpit Autonomous Go-Around Authority at Low Altitude A landing aircraft within the final few hundred feet of the runway threshold, with a conflict detected on the surface by both ASDE-X and the AI prediction engine, should receive an autonomous go-around initiation signal through existing datalink architecture — not just an advisory, but a commanded go-around that activates thrust and flight director guidance without requiring the crew to first identify the conflict visually. Below approximately 200 feet AGL, at landing rollout speeds, a crew has near-zero reaction margin. The TCAS architecture already demonstrates that autonomous avoidance commands in safety-critical windows are operationally accepted, certified, and effective. The same principle needs to be extended to the runway surface domain at minimum separation thresholds.
5. AI Workload Monitoring and Automatic Position Splitting The concurrent-emergency scenario that appears to have preceded Sunday's collision — a single controller simultaneously managing an active landing and a declared emergency on the opposite side of the airport — is a known, predictable high-risk state. An AI workload monitoring system that continuously tracks controller task load, number of simultaneous active operations, and declared emergencies could automatically trigger mandatory position split alerts and supervisory notifications when workload thresholds are exceeded, independent of whether the controller recognizes the overload condition. Research on AI-based cognitive human-machine interfaces for ATC is active in the SESAR and FAA research programs; deployment is overdue.

The Certification Challenge

Every one of the above proposals confronts the same obstacle: FAA certification of safety-critical AI systems for operational use in the National Airspace System is a process without a clear established pathway. Unlike rule-based deterministic systems, AI models trained on historical data present challenges of explainability, distribution shift, and adversarial robustness that existing DO-178C software certification standards do not address. The FAA's AI in Aviation Framework, published in 2023, acknowledges the gap. RTCA SC-228 is developing guidance. EASA has advanced further with its AI roadmap. None of this moves at the speed of the accidents occurring in the meantime. The regulatory framework needs to accelerate in parallel with the technical development, not sequentially after it.

The Precedent the System Already Has

It is worth remembering that this argument has been made before and won. TCAS — the Traffic Collision Avoidance System — was resisted for years on the grounds that autonomous avoidance commands would conflict with ATC authority and create unpredictable interactions in complex airspace. The Überlingen mid-air collision in 2002, in which a crew followed an ATC instruction rather than the TCAS Resolution Advisory, was the tragic evidence point that finally settled the debate. TCAS authority now supersedes ATC in its activation window. Pilots are trained to follow the RA first and notify ATC second. The system works.

The runway surface domain needs its TCAS moment — not another accident to catalyze it, but the institutional will to build and certify autonomous intervention systems before the next one. Sunday night at LaGuardia demonstrated that informing a controller is not the same as protecting an aircraft. After 35 years of building better information systems, it is time to build a system that does not wait for the information to be acted upon.

Sources

1. FAA. "Airport Surface Detection Equipment, Model X (ASDE-X)." faa.gov/air_traffic/technology/asde-x
2. FAA. "Runway Status Lights Questions and Answers." faa.gov/air_traffic/technology/rwsl/faqs
3. FAA. "ADS-B Vehicle Transmitter Maps." (Vehicle ADS-B not mandated.) faa.gov/air_traffic/technology/adsb/vehicle_transmitter_maps
4. FAA Runway Safety Alerting Subgroup. "RS4B – Expansion of Ground Vehicle ADS-B Transponder Equipage." FAA ARC Report. faa.gov/media/88956
5. FAA. "FAA to Install New Runway Safety Technology." (Surface Awareness Initiative.) faa.gov/newsroom/faa-install-new-runway-safety-technology
6. NBAA. "Emerging Airport Technology Aims to Reduce Runway Incursions." June 2025. nbaa.org
7. Saab. "FAA Surface Safety Solutions." (ASDE-X and RWSL program history.) saab.com/products/faa-surface-safety-solutions
8. Defense Daily. "Technology to Reduce Runway Incursions." (ASDE-X Safety Logic architecture.) defensedaily.com
9. Yuan et al. "Prediction of Airport Surface Potential Conflict Based on GNN-LSTM." IET Intelligent Transport Systems, 2025. doi:10.1049/itr2.12611
10. ScienceDirect. "Airport Surface Movement Prediction and Safety Assessment with Spatial-Temporal Graph Convolutional Neural Network." Transportation Research Part C, 2022/2024. doi:10.1016/j.trc.2022
11. Pérez-Castán et al. "Deep Learning in Air Traffic Management (ATM): A Survey." Aerospace (MDPI), Vol. 10, No. 4, 2023. doi:10.3390/aerospace10040358
12. ResearchGate / AI in ATC review. "Artificial Intelligence in Air Traffic Control: Advancing Safety, Efficiency, and Automation." February 2025. researchgate.net/publication/388960570
13. von der Burg & Sharpanskykh. "Multi-Agent Planning for Autonomous Airport Surface Movement Operations." SESAR Innovation Days 2023. (15% taxi-time reduction result.) Cited in: yenra.com/ai20/air-traffic-control-optimization
14. ScienceDirect. "AI4ATM: A Review on how Artificial Intelligence Paves the Way Towards Autonomous Air Traffic Management." June 2025. doi:10.1016/j.jatm.2025
15. MIT Lincoln Laboratory. "Operational Evaluation of Runway Status Lights." Eggert et al. (RWSL genesis and ASDE-X dependency.) ll.mit.edu
16. SKYbrary. "Runway Status Lights (RWSL)." (RWSL architecture and coverage.) skybrary.aero/articles/runway-status-lights-rwsl
17. Aviation File. "Autonomous Runway Incursion Warning System (ARIWS)." (64% reduction figure, FAA evaluation data.) aviationfile.com

Aviation Week & Space Technology  ·  Technology Analysis Sidebar  ·  March 23, 2026  ·  Views expressed in analysis sections represent editorial assessment based on available technical literature and are not NTSB findings  ·  Investigation ongoing