Researchers from Motional and MIT have developed a system that enables self-driving cars to explain their decisions in real time. Known as CW-Net, the technology translates the internal logic of neural networks into contexts that humans can understand. This makes it easier to see why a vehicle brakes, stops, or responds to other road users.
Tests in Las Vegas showed that the system can also uncover hidden problems, including cases where a car incorrectly identifies a stationary vehicle and brakes for the wrong reason. Remarkably, adding the explanations reduced driving performance by less than one percent. The technology could improve safety, troubleshooting, and trust in autonomous vehicles while also helping meet future regulatory requirements.