Overview
- CW-Net is a concept-classifier module inserted into the middle of a machine-learning planner that constrains the planner to use its labeled concepts when choosing a trajectory.
- The researchers trained the model on roughly 130 million labeled driving scenes so it can detect high-level concepts such as “approaching stopped vehicle” or “close to cyclist.”
- CW-Net produces clear, real-time concept labels alongside vehicle trajectories and does not degrade the planner’s driving performance.
- In simulation using recorded Las Vegas scenes and on a Motional robotaxi on a private test track, the explanations helped safety drivers better predict vehicle actions and exposed specific failures, for example showing a stop came from emergency braking rather than correct cyclist detection.
- The work gives engineers immediate diagnostic signals to fix models and could boost situational awareness and calibrated trust, but authors say broader concept coverage, more real-world testing, and further validation are still needed.