Motional and MIT AI explains self-driving car decisions
ainews ·
Researchers from Motional and MIT have developed a system called CW-Net that allows self-driving cars to explain their decisions in real time. The system, described in Nature, converts a neural network's internal logic into human-readable concepts such as 'Approaching Stopped Vehicle' or 'Close to Cyclist.' Unlike post-hoc explanations, it is designed to be causally faithful, meaning driving actions are directly based on these concepts. The team tested it in an autonomous vehicle around Las Vegas with a safety operator and identified two key issues. In one case, the car kept stopping near a cone because the planning system was hallucinating a stopped vehicle, not because of the cone itself. In another, the vehicle's braking for a cyclist actually came from a backup safety system, not the experimental planner. Benchmarking showed the explainability features cost less than one percent in driving performance. The authors suggest such tools could become a baseline requirement as regulators ask for more transparency in autonomous vehicles.
ainews