MIT, Motional Unveil CW-Net Using 130 Million Scenes to Explain Self-Driving AI Decisions
Updated
Updated · MIT News · Sep 2
MIT, Motional Unveil CW-Net Using 130 Million Scenes to Explain Self-Driving AI Decisions
1 articles · Updated · MIT News · Sep 2
Summary
CW-Net, published in Nature, adds a concept-classifier module to self-driving planners that explains decisions in real time with labels such as “approaching stopped vehicle” while preserving driving performance.
130 million labeled driving scenes trained the system to map a model’s internal reasoning into human-readable concepts and force the planner to use those concepts, aiming to keep explanations causally faithful rather than misleading.
Private-track tests on a Motional robotaxi showed safety drivers predicted vehicle behavior more accurately, including spotting a case where a stop came from emergency braking—not cyclist detection—and could have ended in a collision.
Las Vegas-based simulation studies with nonexpert users found similar gains, suggesting the tool could help both passengers anticipate mistakes and engineers debug autonomous-vehicle AI in safety-critical settings.