World2Rules Learns Patterns of Unsafe Aircraft Behavior and Explains Potential Risks Before Disaster Strikes

The Breakdown: 

• World2Rules learns patterns of risky aircraft behavior from real airport operations and incident data.
• It explains its warnings in simple terms that are easy for humans to understand.
• World2Rules is designed to work alongside existing prediction systems to enhance aircraft safety.

* * *

When managing airport traffic, small errors can lead to catastrophe. 

Just a few months ago, an air traffic controller told an Air Canada jet that had just landed at New York’s JFK Airport to wait before crossing a runway to avoid an EVA Air jet that was landing. The Air Canada crew acknowledged the instructions but kept moving while the EVA jet was traveling toward them at high speed. 

Fortunately, an alert controller transmitted, “Stop, stop, stop, stop!” in time. The Air Canada plane stopped, the EVA plane zoomed by and disaster was averted — but it was close.

Near misses like the one at JFK inspired a group from the AirLab in Carnegie Mellon University’s Robotics Institute (RI) to create World2Rules, an AI system that learns interpretable safety rules from data to analyze, verify and explain potential collision scenarios. 

By learning from both everyday airport activity and documented safety violations, the system builds a clear picture of what “normal” and “unsafe” behaviors look like. When it detects a potential violation, Word2Rules does more than just raise an alert. It identifies which safety rule is being broken and explains why the situation is risky, showing how the scenario matches known patterns of danger rather than issuing a vague warning.

“The overall idea we’ve been working on with this project is to see how we can improve safety in the aviation domain or other safety-critical domains,” said Jack Wang, an RI master’s student. “As shown on the news, runway incursions have been happening. Sometimes they’re minor, but sometimes they can be quite catastrophic.”

Wang is passionate about aviation safety and flight. He joined the CMU Flying Club as a first-year student and later taught a Student College course to help students get the ground instruction they need to pursue a pilot’s license.

The World2Rules team wanted to design an AI system that could not only recognize when aircraft were on a dangerous path, but could also predict potential collisions early enough to give pilots and controllers critical extra moments to react.

To do so, the AirLab and the Bot Intelligence Group jointly developed the Amelia-42 dataset. The set contains two years of Federal Aviation Administration (FAA) airport surface movement data from 42 U.S. airports. It includes massive amounts of information, tracking aircraft and vehicle movement across runways and taxiways. To process the large amount of information, they used the Bridges-2 supercomputer at the Pittsburgh Supercomputing Center. 

“The data we collected includes both normal airport operations and crash and incident reports,” said Jay Patrikar, a recent RI graduate who worked on World2Rules and was also a founder of the CMU Flying Club. “That data helps our system distinguish between normal and unsafe situations. We not only want to understand that a crash is happening, but also want to predict if a crash will happen in the future.” 

World2Rules is designed to plug into a broader collision-prediction pipeline. It learns explicit safety rules from the Amelia dataset, recognizing patterns that lead to unsafe situations, such as aircraft occupying the same runway at the same time. It then applies those rules to aircraft trajectories, flagging when a future scenario would violate them. Instead of simply signaling risk, the system can then point to the specific rule being broken and explain why the behavior is dangerous in terms humans can understand.  

“In practice, this ideally would mean air traffic controllers or automated systems could get earlier, clearer warnings of potential dangers,” Wang said.

To make sense of all that data, World2Rules combines two types of AI approaches, neural and symbolic. The neural side picks up on patterns buried in the airport data. The symbolic side turns those patterns into clear, logical rules that humans can read. By pairing pattern recognition with rule-based reasoning, the system can both identify risky situations and explain them in a structured way.

“Beyond aviation, World2Rules could also be used in other areas where safety is critical,” said Sebastian Scherer, an associate research professor in the RI and head of the AirLab. “The system can be adapted to different environments by teaching it the relevant rules and behaviors for that domain. Once that information is defined, the same core technology can learn and monitor safety risks without needing to be redesigned.”

The team reported their results at the NASA Formal Methods Symposium in Los Angeles earlier this month. To learn more, read the preprint of the article.

For More Information: Aaron Aupperlee | 412-268-9068 | aaupperlee@cmu.edu

Share This Story!