It’s nearly showtime for Waymo CTO Dmitri Dolgov.
Almost a decade since he and a dozen other engineers went to work on Google’s Self-Driving Car Project, Waymo, the project’s commercial successor, is months from launching a paid robotaxi service. It’s doing so in the aftermath of a fatal accident in which an Uber self-driving test vehicle killed a pedestrian in metro Phoenix, near where Waymo operates. That shook public confidence in the technology, and a recent media report suggesting even Waymo’s very advanced program has flaws didn’t help.
But ask Dolgov if the artificial intelligence behind the wheel of Waymo’s Chrysler Pacifica minivans will be ready for daily driving duties in Chandler, Arizona, and the Russian-born scientist’s confidence is unwavering: “Yes, absolutely,” he told Forbes.
(The interview took place before The Information published “Waymo’s Big Ambitions Slowed by Tech Trouble,” highlighting anecdotal problems its minivans have had and their annoyance factor for some locals. The company says Dolgov’s views are unchanged.)
Waymo launches in Phoenix with 600 vehicles but its aspirations are far bigger, with plans to add up to 82,000 Chrysler hybrid vans and electric Jaguar i-Pace SUVs to its fleet over the next few years. At the same time, CEO John Krafcik has cautioned that the “time period will be longer than you think” for automated vehicles to be available everywhere.
Expanding a Waymo-branded ride service to multiple cities in the U.S. and abroad is its next evolution. Doing that in an efficient, safe way means scaling up everything.
“At a high level, there are maybe three things that have to come together. You have to build maps and maintain them, in order to scale and deploy this technology. You have to build and design self-driving hardware, lasers, radars, cameras, computers, and then you have to build and be able to deploy the full software stack,” he said. “In all of the three main pillars of our system, we’re at a point where we are industrializing the whole thing.”
Maps might seem like a redundant effort since as an Alphabet company it has access to Google Maps, Street View and even Google Earth. But those don’t work for autonomous vehicles.
“Our maps are different, much higher resolution, much higher fidelity. But Google is a company that figured out how to build maps at huge scale, so we’ve been building off of that work,” he said. “We still need to send our cars out there to collect the data. On a more fundamental level, it’s the know-how of how to build stuff at scale. There’s a lot of stuff that happens at the back end. Data collection is just a part of it. You need the infrastructure, you need the pipelines for crunching all of the data to produce your maps.”
Similarly, advanced hardware Waymo uses, particularly laser lidar sensors that generate 3-D point-cloud maps of a vehicle’s surroundings, are expensive and still relatively hard to come by.
“It’s not a mature industry. You can’t just go and buy something off the shelf. You can’t go to a supplier and say `build me 100,000 of these things,’” Dolgov said.
Waymo won’t detail what it costs to load up each of its vehicles with multiple cameras, radar and long- and short-range lidar sensors and a computer that are all designed in-house. Krafcik has said that Waymo cut the cost of its priciest lidar unit by 90%. More such cost-reduction efforts continue, though Dolgov would only say, “this is something we’ve been working on for years.”
As an Alphabet company, Waymo also has access to the network of global suppliers that make high-tech hardware for Google.
“Now it’s turning that, all those designs, all those partnerships and then going to the next scale.”
With a top-shelf education and professional background, Dolgov might have been sent from central casting to lead Waymo’s tech team after the turnover in 2016 when several fellow stars, including Chris Urmson and Anthony Levandowski, opted to seek new opportunities after receiving highly lucrative Google bonuses.
But Dolgov stuck around to see Google’s R&D transform into a commercial service under CEO Krafcik. A lean, fit 40-year-old, he has the no-nonsense bearing of a veteran engineer or career military officer. He holds a Masters degree from the Moscow Institute of Physics and Technology, a Ph.D. in computer science from the University of Michigan and did postdoctoral research at Stanford. Oh, and he’s credited as an inventor on more than 90 patents.
While at Stanford Dolgov joined Sebastian Thrun’s robotic vehicle team for the DARPA Urban Challenge, the self-driving car movement’s Big Bang. After Stanford, he worked as a senior scientist with Toyota Research Institute before Thrun brought him to the Google Self-Driving Car Project in 2009.
Dolgov speaks about Waymo’s technology and autonomous vehicles with the confidence of someone who’s worked in the field for most of his adult life. And with Waymo now past 9 million miles of on-road testing and billions more miles logged in computer simulation, there’s a basis for that confidence. Structured testing at “The Castle,” a Waymo facility in central California where it recreates tricky real-world driving scenarios, like merging into busy traffic, making left turns or handling pedestrians and objects that unexpectedly cross a vehicle’s path.
Nonstop on-road, virtual and structured testing is vital to keep training Waymo’s AI to safely handle all the situations human drivers confront, with improvements continuously shared with every vehicle in its network to ensure uniform performance.
Still, The Information said Waymo minivans continue to struggle with basic driving maneuvers, including making turns at the T-intersection near its vehicle depot in Chandler, unprotected left turns and merging into busy traffic. The story included complaints from locals about the vehicles, many of whom work at businesses near Waymo’s depot and no doubt confront them frequently.
Circumstances described in the story were a surprise to city officials and other locals, however.
“We haven’t received complaints at City Council meetings or similar public forums from Chandler residents about Waymo or autonomous vehicle testing in Chandler,” Matt Burdick, a spokesman for the city, told Forbes.
“I can’t think of any complaint we’ve received about” Waymo vans, said Detective Seth Tyler, a Chandler Police Department spokesman.
“It’s probably fair to expect that no matter how far along Waymo is, there will be some rough spots for them when they actually launch,” said Mike Ramsey, an industry analyst for Gartner who closely follows Waymo and developments in autonomous vehicle technology. “That said, no company has done more real-world testing and my expectation is that the launch will be limited in size and scope – kind of like a soft opening for a restaurant.”
A driverless test-ride Waymo provided for Forbes a few months ago in Chandler was smooth and uneventful. It was also too short to draw meaningful conclusions beyond the obvious: It handles familiar routes with no issues.
Still, the minivan safely navigated a different T-intersection, contending with a bicyclist approaching from the left and oncoming traffic to the right. The vehicle made a quick left turn to avoid both that was surprising because it was much the way an alert human driver might have handled it.
Waymo defends its Chandler operations, and a spokesman said it’s made significant improvements in how its vehicles handle left turns in front of the Chandler depot.
“Waymo was founded on a mission to make our roads safer, and that’s why we built a cautious and defensive driver,” the company said in a statement. “The way to responsibly deploy our fully driverless technology is to robustly test and validate in a geo-fenced territory that grows over time. Anything less than that undermines safety and the promise of this technology.”
Dolgov acknowledged that left turns, merges, lane changes, construction zones and many other road conditions are challenging for its self-driving vehicles (Waymo even blogs about some of them).
“Things that are sometimes hard for humans are also hard for self-driving systems,” he said. “In narrow streets, you have to negotiate this bottleneck. Who goes first? Do I go into this pocket and let you through? Those things are hard.”
“Merges, do I go ahead of you, do I go behind you? There’s a lot of signaling, there’s feedback of your behavior and my behavior. It gets into this problem of deeply understanding the intent of other actors in the world, making predictions about how they are going to behave and then reacting in a way that is socially acceptable,” Dolgov said.
“Good drivers are good at that and good self-driving vehicles are good at that.”
Mark Twain may or may not have quipped, “reports of my death are greatly exaggerated.” With autonomous vehicle tech currently said to be in what Gartner’s Ramsey calls the Trough of Disillusionment people working in that field may share Twain’s sentiment. None more so than Dolgov.