Researchers have created a system called CW-Net that allows self-driving cars to display the reasoning behind their driving decisions in real-time.

Researchers from Motional and MIT’s Computer Science and Artificial Intelligence Laboratory have developed a system designed to address the "black-box" nature of autonomous vehicle AI. Published in Nature, the project introduces the Concept-Wrapper Network (CW-Net), which translates a vehicle's internal neural network calculations into human-readable concepts. The system aims to provide transparency by showing exactly which factors influence driving decisions as they occur.

How CW-Net Functions

CW-Net converts internal logic into specific concepts, such as "Approaching Stopped Vehicle" or "Close to Cyclist," which can be displayed on a dashboard. Unlike systems that generate natural-language explanations after an event, CW-Net is integrated into the vehicle's decision-making process. This ensures the explanations are causally faithful to the actions taken by the car, rather than being post-hoc guesses.

Testing and Real-World Application

The team tested the system on public roads and private tracks in Las Vegas using an autonomous vehicle with a safety operator. During testing, the system identified that the vehicle was "hallucinating" a stopped vehicle due to training data patterns, and revealed that a safety backup system—rather than the primary planner—was responsible for braking near a cyclist. While adding explainability can impact performance, researchers found the difference in driving capability compared to leading algorithms was less than one percent.