Researchers from MIT and autonomous vehicle technology company Motional developed a method that translates the internal reasoning of deep-learning driving planners into understandable concepts, such as “approaching stopped vehicle” or “close to cyclist.” The team reported in Nature on September 2, 2026, that the approach, called the Concept-Wrapper Network (CW-Net), describes an autonomous vehicle's decisions without altering its driving performance. The researchers said they trained the module on 130 million labeled examples of driving scenes.

According to the report, a road test on a private track with a safety driver in a Motional robotaxi found that CW-Net explanations helped the driver better predict vehicle behavior in surprising situations. In one case, the vehicle stopped near a cyclist; the explanations indicated the model had not been properly configured to detect the cyclist and had stopped through emergency braking. Larger online simulation studies using real driving situations captured in Las Vegas yielded similar results on participants' ability to predict vehicle behavior, the researchers said.

The authors described the experiments as early evidence and suggested future work could extend the module to more concepts and explore other training and design techniques. MIT professor Julie Shah said the work shows how explanations can support a person's understanding of a system's behavior and could be used in engineering and development.