MIT and Motional built a method that explains why a self-driving car acts as it does. Called CW-Net, it plugs into the middle of the car's existing planning model. It labels the scene with plain concepts such as "approaching stopped vehicle" or "close to cyclist". The final planning step is forced to use those labels, so the stated reason matches the real one. Explanations are output alongside the driving path in real time. Driving performance was not changed by the addition. On a private test track, a safety driver predicted the car's behaviour more accurately. One case showed the car stopping only because emergency braking triggered, not because it saw a cyclist. A larger simulation study with non-experts using Las Vegas road footage found similar gains. The researchers want to cover more concepts next.
What changed
Self-driving planners were black boxes, giving no reason for sudden braking.
What it unlocks
Reading a self-driving car's stated reason for each move in real time.
- 130 million labeled driving scenes
What you need to act on it
- access to a vehicle's machine-learning planner architecture
Sources