MIT and Motional Build AI That Exposes Robotaxi Planning Errors

A robotaxi stopped near a cyclist, but not for the reason its safety driver assumed. CW-Net showed that the vehicle’s planner had selected a collision-causing path; emergency braking activated only when the car got too close. MIT and Motional’s research on the system appeared in Nature on September 2.

The Concept-Wrapper Network, or CW-Net, translates an autonomous-driving planner’s internal reasoning into readable concepts, such as “approaching stopped vehicle” and “close to cyclist.” MIT and Motional developed it to explain machine-learning-based driving decisions.

Built into the choice, not added afterward

CW-Net sits within an existing planner. It identifies high-level concepts from the vehicle’s inputs, then requires the final planning stage to use those concepts when selecting a trajectory. It produces an explanation alongside that trajectory in real time, rather than trying to describe the choice after the fact.

130 millionTraining scenes

CW-Net was trained on 130 million labeled autonomous-driving scene examples.

Less than 1%Driving-capability difference

Benchmarking reported a less-than-1% difference in driving capability after CW-Net was added.

A stop can conceal the wrong failure

In the cyclist test, the safety driver initially believed the car had recognized the cyclist. CW-Net showed that the planner was not properly configured to detect the cyclist and had chosen a path that would have caused a collision. The vehicle stopped because its emergency-braking procedure took over at close range.

The team tested CW-Net in a Motional robotaxi on a private track with a safety driver. The explanations helped safety drivers better predict surprising behavior. Online simulations using real Las Vegas driving situations also found that participants were significantly better at predicting the vehicle’s behavior with the explanations.

Motional also described repeated stops near a traffic cone. CW-Net identified the experimental planner as hallucinating a stopped vehicle ahead, a pattern the company linked to training data. Researchers plan to extend the method to more concepts and explore designs intended to improve performance and interpretability.