As self-driving cars become more common on city streets, one important question remains difficult to answer: Why do they sometimes crash?
Now, researchers at King’s College London have developed a new artificial intelligence approach that could help uncover the exact reasons behind specific accidents involving autonomous vehicles.
The technology could improve safety and help build public trust in self-driving systems.
The research was presented at the 2026 IEEE International Conference on Robotics and Automation.
Self-driving vehicles are already operating in cities around the world, including London and San Francisco. While these vehicles have the potential to reduce human driving errors, several collisions and safety incidents have raised concerns about how they make decisions and why mistakes occur.
Traditionally, engineers have relied on statistics to study failures. These methods can estimate how likely a crash is to happen under certain conditions, but they usually cannot explain the exact chain of events that caused a particular accident.
The new system takes a different approach. Instead of focusing only on probabilities, it examines a crash after it happens and works backward to identify the specific factors that led to the failure.
According to the researchers, this method uses a concept called “actual causality.” The goal is to determine not just what happened, but which events directly contributed to the crash.
This is especially important for autonomous vehicles because their decisions are influenced by huge amounts of information. At any moment, a self-driving car must monitor traffic, pedestrians, road signs, weather conditions, and the movements of nearby vehicles.
As a result, a single accident may have hundreds or even thousands of possible contributing factors. In some cases, an event that occurred several minutes earlier—or even miles before the crash—may have started a chain reaction that eventually led to a collision.
To tackle this challenge, the team created a new “responsibility-guided” search algorithm. The system rapidly examines all possible causes and identifies the most important ones without needing enormous computing power.
The researchers say their method can find explanations for accidents using far less computational effort than previous approaches.
The work builds on earlier research by the same team, which focused on identifying rare situations that could cause autonomous systems to fail. The new study goes one step further by explaining why those failures occur.
Although the researchers tested the technology using self-driving cars, they believe the approach could be useful for many other AI-powered systems. Future applications might include autonomous robots used in hospitals, care homes, and other environments where safety and reliability are critical.
The team hopes that better explanations will help engineers design safer systems and make it easier for the public to trust AI technologies.
As autonomous vehicles and robots become a larger part of everyday life, understanding why they make mistakes may be just as important as preventing those mistakes in the first place.
