Self-driving technology can make complex decisions in seconds, but understanding why a vehicle chooses a particular action is not always easy.

Researchers have developed a new approach that gives autonomous driving systems clearer, human-readable explanations for their decisions.

The Problem With Hidden Decisions

Modern autonomous vehicles increasingly depend on deep-learning systems to interpret surroundings and select driving actions. These systems can process large amounts of information quickly, but their internal decision-making can be difficult for people to follow.

That creates a practical challenge for a human supervising an automated vehicle. If the car slows down, stops, or chooses an unexpected path, knowing why that decision happened can help the person understand what the vehicle may do next. The research focuses on this communication gap. Instead of simply showing what the vehicle is doing, the researchers wanted to provide information about the factors influencing its decisions.

A New Way to Explain AI

The researchers introduced a system called the Concept-Wrapper Network, or CW-Net. It works alongside an existing machine-learning driving planner and translates aspects of the planner's reasoning into recognizable concepts. These concepts can describe situations such as approaching a stopped vehicle or being close to another road user. Rather than requiring people to interpret complex technical information, the system presents the reasoning through concepts that are easier to understand.

An important feature is that these concepts are directly connected to the decision-making process. The researchers designed the system so that the concepts help determine which driving action is selected, rather than simply generating an explanation after the decision has already been made.

Driving Performance Stayed Stable

Adding explanations would have limited value if it significantly changed how the vehicle drove. The researchers therefore compared the original planner with the modified system. Testing showed that driving performance remained effectively equivalent, with differences of less than 1% across the evaluated measures. The concept-classification component achieved a mean accuracy of 54%, while recall reached 77%.

Testing With Human Drivers

The team then moved beyond computer simulations and tested the technology with a real autonomous vehicle. Safety drivers observed the vehicle during trials and received information generated by CW-Net while the system operated.

Several unexpected situations showed why this could matter. In one example, the vehicle repeatedly stopped near parked vehicles. The driver's initial assumption about the reason for stopping turned out to be different from the factor identified by the explanation system. When the vehicle moved farther away from the parked cars, the relevant concept became less prominent and the vehicle started moving again.

Another test involved the vehicle stopping near a traffic cone. The explanation indicated that the planner was responding as though a stopped vehicle were present. Removing the cone did not eliminate the behavior, providing additional evidence about what was influencing the planner's decision.

Explanations Improved Understanding

The researchers also examined whether these explanations actually changed how people understood the vehicle. Participants were shown driving scenarios and asked to explain the vehicle's behavior and predict what it might do under different circumstances.

The results showed improvements among both experienced participants and people without specialized expertise. In the study, 8 of 9 experts and 27 of 30 non-experts showed improved mental models after receiving the explanations. The research also found a connection between better understanding and improved predictions.

Real-World Situational Awareness

The researchers extended the testing to public-road driving scenarios and conducted a larger online study involving 100 participants. The results indicated that explanations improved several aspects of situational awareness when the vehicle behaved unexpectedly.

Participants performed better at identifying what the vehicle was responding to, understanding why it acted in a certain way, and predicting what it might do next. Importantly, the explanations did not significantly reduce situational awareness during more predictable situations.

Potential Beyond Self-Driving Cars

The value of this research may extend beyond autonomous vehicles. Other systems that work alongside people can also make decisions that are difficult to interpret, particularly when they rely on complex machine-learning models. The researchers suggest that similar approaches could potentially support autonomous drones, robotic systems, and other applications where people need to understand automated decisions quickly.

Explainable AI could become an important part of making autonomous technology easier for people to understand. For self-driving technology, that added understanding could help create a clearer relationship between automated decisions and human supervision.