The Camera-Versus-Lidar Debate Is Really About the Last Few Nines

Waymo co-CEO Dmitri Dolgov has offered a clear explanation for why Waymo combines cameras, lidar and radar instead of relying on cameras alone.

Dolgov acknowledges that cameras can support a capable driver-assistance system and may approximately match human perception. His argument is that this is not enough for a vehicle expected to operate without a human driver.

According to Dolgov, camera-only systems eventually reach a safety ceiling. They can improve quickly at first, but their progress flattens before reaching the reliability required for full autonomy.

He describes that challenge as an “exponential ladder of nines.” Reaching 90 or 99 percent reliability is much easier than adding each additional nine. A demonstration may tolerate occasional failures. A vehicle carrying passengers with nobody behind the wheel cannot.

Waymo uses three sensor types because each observes the world differently.

Cameras provide color and high-resolution images, but their performance deteriorates in darkness and glare. Lidar directly measures the three-dimensional structure of the environment. Radar can operate through fog, rain and snow while measuring movement through Doppler.

Dolgov says these sensors are not merely backups. Their separate observations are combined into a view of the environment that no individual sensor can provide.

The examples in the article make that argument compelling. In a Phoenix dust storm, a camera could barely see while lidar detected a pedestrian. Lidar also identified people entering a dark roadway when the camera image offered little useful detail. When a tree branch physically obstructed one of Waymo’s sensors, the vehicle used its remaining sensors to return safely to its depot.

But the article does not completely resolve the most important practical question.

How much additional equipment is justified by the improvement in safety, and can lidar and radar become inexpensive enough to install across tens of millions of ordinary vehicles?

Dolgov’s response is that Waymo has reduced its hardware costs substantially across six generations. He warns that dismissing lidar based on its current price means relying on a number with a short shelf life.

That is reasonable, but it is not yet an answer. The article does not provide a detailed cost history, a current per-vehicle sensor price or a projection showing when Waymo’s approach becomes competitive with camera-only hardware at mass-market scale.

This leaves two separate questions:

Can cameras alone reach the reliability required to remove the driver?

Can a richer collection of sensors become affordable enough to deploy broadly?

Waymo presents a strong technical case that cameras have limitations that software cannot always overcome. A lens cannot infer what it physically cannot see through darkness, glare, weather or an obstruction.

The remaining competition may be less about whether lidar works and more about which happens first: camera systems overcoming their safety ceiling, or additional sensors becoming cheap enough that excluding them no longer makes economic sense.

Until then, neither impressive demonstrations nor present-day sensor prices settle the debate.

Read “Waymo CEO explains why Tesla’s camera-only self-driving falls short.”