Self-driving vehicles depend on deep-learning models to interpret their surroundings and decide what to do next. Yet these models can behave unexpectedly, leaving safety drivers and engineers unsure why a vehicle suddenly stopped or chose a particular path. Researchers at MIT and autonomous vehicle technology company Motional have developed a system that makes these decisions easier for humans to understand, tells MIT News.
Called the Concept-Wrapper Network, or CW-Net, the method translates an autonomous vehicle’s internal reasoning into recognizable concepts such as approaching a stopped vehicle or being close to a cyclist. Machine-learning planners normally operate as black boxes, processing information from cameras and lidar before generating a trajectory. CW-Net is inserted into this planning architecture and identifies concepts that directly influence the vehicle’s final decision.
The researchers trained CW-Net using 130 million examples of scenes captured by self-driving vehicles. The system produces understandable explanations in real time while preserving the performance of the original deep-learning model. This causal connection is important because explanations that merely sound plausible could give drivers a false understanding of why a vehicle behaved in a particular way.
Testing on a Motional robotaxi demonstrated the system’s potential. In one case, a safety driver believed the vehicle stopped because it detected a cyclist. CW-Net revealed that the planner had actually failed to properly detect the cyclist. The vehicle stopped only after its emergency braking system activated. Knowing the real cause could help safety drivers intervene earlier and engineers correct underlying problems.
Online simulation studies using driving situations recorded in Las Vegas produced similar results. Participants who received CW-Net explanations became significantly better at predicting autonomous vehicle behavior. The researchers believe the approach could improve transparency, situational awareness, and appropriate trust in self-driving systems while also supporting engineers developing AI for other safety-critical applications.
