Overview
- Dmitri Dolgov reiterated at Y Combinator this week that camera-only systems can match human drivers for some tasks but run out of capability long before they reach the near-perfect reliability needed for Level 4 driverless service.
- Dolgov showed test examples where LiDAR or radar detected hazards that cameras missed, including a Phoenix dust storm where LiDAR picked out a pedestrian while camera images were nearly unusable.
- Waymo runs a three‑modal perception stack with separate encoders for cameras, LiDAR and radar whose outputs are fused into a single scene; the company says that redundancy is essential to climb the ‘exponential ladder of nines’ toward very high safety levels.
- Regulatory and performance contrast is sharpening the debate: Waymo reports hundreds of millions of paid-rider miles and lower serious-injury crash rates, while Tesla’s camera-only robotaxi program shows far fewer driverless miles and NHTSA has opened an engineering analysis that identified nine crashes tied to potential visibility-degradation detection failures.
- Falling LiDAR prices and NHTSA scrutiny of camera-only systems could shift industry economics and regulation, which may speed adoption of multi-sensor stacks and affect when and how robotaxi services expand for riders.