Waymo CEO: camera-only self-driving systems can't achieve full autonomy safely

Full text

Camera-only systems can only “approximately match” a human driver


Multi-sensing setup offers more than just a failsafe

Waymo co-CEO backs the use of Lidar and radar over camera-only systems

Waymo co-CEO Dmitri Dolgov has spoken out about the technology required to run a fully autonomous robotaxi service , addressing the subject of camera-only systems.

While he didn’t mention Tesla directly in his in-depth talk at Y Combinator’s Startup School, the presentation directly addressed vision-only camera setups, like those Elon Musk and Tesla have been pursuing.

“Humans, of course, can drive with just eyes, so there’s that proof of existence,” he said, according to Electrek .

“If the goal were to just approximately match human performance or to build an assist product, that’s a very reasonable way to go.”

But “weak sensing” inevitably leads to a safety curve that flattens out “way too early,” he added, highlighting the many cases where a camera system can come unstuck.

Low light or poor visibility, dust storms and objects physically blocking the lens can bring a camera-based system to a grinding halt in seconds.

Dolgov then gave an example of how Waymo’s combination of Lidar, radar and high-definition cameras allowed one stricken vehicle to safely navigate itself back to a depot when a tree branch became caught in its wipers and obscured the camera lens.

While Dolgov doesn’t completely dismiss camera-based systems, stating that they can easily support a driver with highly autonomous assistance systems, he admits they will struggle to replace a human entirely.

The issue is that fully autonomous robotaxis need to be as close to perfect as it’s possible to be.

A mounting record of accidents will likely put off customers for life and kill the technology before it has time to gain a foothold.

Many take the multi-modal path

(Image credit: Hyundai/Motional) Having ridden in numerous autonomous vehicles over the years and listened to those that have built them, I am fully in Dmitri Dolgov's camp.

While a camera-based system might help keep costs down, fitting autonomous vehicles with Lidar, radar, and other sensors feels like a natural and reasonable step to make them as safe as possible.

A visit to Motional’s HQ in Las Vegas earlier this year illustrated the need for multiple sensing technologies when you completely remove the driver from the picture.

Its robotaxis, which are now bespoke versions of the Hyundai Ioniq 5 (complete with radar and Lidar), have driven over two million autonomous miles with zero at-fault incidents.

Like Waymo, it too believes a multimodal approach that fuses various technologies is the safest way forward.

Laura Major, the company’s CEO, said at the time that rival startups were able to get to a “90% performance” very quickly, often relying solely on high-definition cameras.

“These offer really smooth rides in most situations,” Major remarked. “The problem is when you’re going to remove a driver and have true driverless operation, that’s not enough. You can’t perform well in 90% of the cases,” she added.

Admittedly, I haven't ridden in one of Tesla's driverless Robotaxis, but the numbers don't lie.

According to Electrek, Tesla's own data confirms a crash rate three times worse than humans even with a monitor at the wheel, while Waymo's numbers suggest it is 17 times safer after covering some 220 million rider-only miles.

Tesla has accumulated just 380,000 driverless miles over the past year since launching its robotaxi service.