PITTSBURGH — Ask anyone where the geographical center of autonomous car development is, and most will give the same answer: Silicon Valley. But 2,500 miles to the east, there’s another focal point for autonomous drive technology development in the industrial-era heart of Pittsburgh.
Robotics experts may not be surprised. One of the top teams in the series of DARPA Grand Challenge and Urban Challenge competitions that kick-started the recent groundswell in autonomous car development came out of Pittsburgh’s Carnegie Mellon University. Home to one of the world’s foremost robotics programs, Carnegie Mellon has become one of the top pipelines for robotics research and talent that forms the core of many of the leading self-driving car developers.
Today, three of those companies — Uber Advanced Technologies Group, Aurora Innovation and Argo AI — have established offices blocks from one another, forming a “Robocar Row” in an old industrial area here along the Allegheny River. The concentration of robotics, artificial intelligence, software, sensor and electrical engineering talent in these few square miles is among the highest in the country.
On Aurora Island
Aurora Innovation is the youngest of the three companies and came out of “stealth mode” only this year. Its freshly redesigned offices are decorated in a restrained navy blue, giving it a modern, high-tech feel that’s free from the quirky indulgences of so many high-tech startups. A monastic calm hangs over the mostly empty office and its adjoining garage workshop, as engineers work silently on electronic components and screens filled with code.
Aurora’s famous founders, former head of the Google self-driving car program Chris Urmson and former head of Tesla’s Autopilot program Sterling Anderson, are at the firm’s Bay Area offices, and Chief Technical Officer Drew Bagnell, who heads up work at the Pittsburgh office, is lurking somewhere in the background. But four engineers are on hand to give an inside look at the day-to-day development of Aurora’s self-driving software stack: technical program manager Dima Kislovskiy, robotics engineer Clint Liddick, perception engineer Davis King and software engineer Ethan Eade.
All four explain that the hushed office reflects the way work gets done at Aurora.
“At a high level, our days are periods of intense individual concentration and work interspersed with frenzied communication,” Eade said.
Aurora manages this balance by designating the time before lunch as Maker Mornings, Liddick explains.
“You are free to not respond to email or Slack. If you want to take all of the morning and just concentrate on the hard problem and work by yourself with headphones, that’s respected,” Liddick said.
For many of Aurora’s engineers, this work time is spent “turning ideas about how to solve some subpart of this big problem space into a working software machine,” Eade said.
After lunch, the office livens up as a more collaborative phase of work begins. “The communication is broad,” Eade said. “When we do it best, it involves a surprisingly large fraction of the company at one time.”
Because each part of the self-driving stack is such a thorny problem, and yet each part needs to be tightly integrated, Aurora puts a strong emphasis on cross-team collaboration to prevent the formation of fiefdoms, round out individual skill sets and keep everyone pulling together.
A company software review infrastructure forces engineers to check one another’s code every time they submit a new piece of the software. Comparing the system to open-source projects, King says the cross-company review policy “keeps everyone in the loop, keeps everyone engaged. Everybody knows what’s going on because you’re constantly pulled into reviewing people’s code. It seems like a trivial thing, but big things are built out of tons of trivial things.”
That collaboration ensures a shared sense of purpose as employees develop technology with real-world, potentially life-or-death consequences.
“Safety is a process, and it’s an attitude, and it’s a willingness to hear reservations, objections, thoughts from anywhere at any time and to treat them with all the attention they deserve,” King said.
One of the company’s maxims is that building a self-driving system is about turning science problems into engineering problems, and as Kislovskiy puts it, “Engineering is fundamentally a human problem.” The belief that a combination of brilliant talent and collaborative culture is the only way forward not only creates what Eade calls an “unironic Aurora spirit,” but also a sense that the company can more or less ignore the ups and downs of its competitors and even perceptions of self-driving technology more broadly.
“We’re sort of on Aurora Island,” Kislovskiy said. “There’s too much noise; there’s too much commotion; there’s too much distraction; and we think we know what we’re doing. … One of the things we try to do is not only not say anything externally, but try to keep to our own game, play it the way we want to and really stay super focused.”
Aurora Innovation systems engineer Prasad Venkiteswaran, left, and software engineer Erick Fuentes test a Volkswagen Golf equipped with self-driving technology. Aurora puts strong emphasis on cross-team collaboration. Photo credit: GETTY PHOTOS
Uber lays low
A short walk from Aurora, past the hip shops and restaurants on Smallman Street and under the trestle bridge that runs over 33rd Street, is the original heavyweight of Pittsburgh’s self-driving car scene: Uber Advanced Technologies Group. Inside its sprawling riverside building, the black and white decor feels on-brand and boasts more of the traditional signifiers of startup life. A large, buffet-style eating area complete with a fridge full of kombucha looks through giant plate-glass windows into a quiet workshop full of idle Volvo XC90s. Nearby, a large, stepped amphitheater for all-hands meetings looms over a pair of pingpong tables.
If there are fewer self-driving cars to be seen on Pittsburgh’s streets than there were a year ago, that’s because Uber’s massive autonomous drive development program is grounded pending an external safety review following a fatal crash in Tempe, Ariz., in March.
Uber welcomed a visit from Automotive News but didn’t make any of its staff available for interviews, saying that its public relations efforts are on hold until the safety reviewis published. The review appears to have had an effect on the company’s direction. Issues that didn’t seem to be major Uber priorities before, such as designing vehicles accessible for disabled people and establishing safety standards for autonomous vehicles, seem to be of deep interest to the company now.
There is a different vibe at ATG than at its intracity competitors, despite the fact that many of Aurora’s and Argo’s employees previously worked on autonomous drive technology at the ride-hailing giant. The staff is bigger, and the office space is more full than at the other two companies. The waiting room is busier, security seems tighter, and there’s more of a slick, corporate sheen to everything.
Upstairs from the expansive common area, large workspaces house teams of coders developing software and electrical engineers tinkering on desks strewn with sensors, circuit boards and other electronic components. There’s little evidence of the intense discussions that Aurora’s tight-knit and much smaller team describes, but at the moment, ATG’s vehicles aren’t on the road gathering data and identifying edge cases.
Salesky: “About transparency”
Down Railroad Street from ATG is a brand-new building, where Argo AI occupies the top two floors. Like Aurora, Argo is a relative newcomer. Emerging from an elevator whose fresh paint gives off “new startup smell,” visitors are greeted by decor that occupies a middle ground between Aurora’s spare minimalism and Uber ATG’s branded, startup-playground feel. There’s a grownup maturity, yet Nerf guns litter the engineering workspaces and standing desks.
Argo is open in ways the other companies aren’t.
“We’re all about transparency,” CEO Bryan Salesky said.
For instance, Automotive News received one of the first on-the-record rides in one of Argo’s test cars. It also was introduced to each major team in the technology stack, with each demonstrating its work on mapping, perception, prediction and route planning.
Argo has deep ties to Pittsburgh and its robotics community. Salesky and his two top lieutenants, President Peter Rander and Vice President of Robotics Brett Browning, have all put in considerable amounts of time at Carnegie Mellon’s National Robotics Engineering Center: Salesky for five years, Browning for nearly 13 years and Rander for nearly 15. Their grounding in robotics gives them a perspective on self-driving systems that balances hardware and software, creating a different attitude compared with a software whiz who may have moved into self-driving cars because it was the next big thing.
Instead of hyping the potential of the technology, Argo’s leadership appears to be looking for concrete commercial opportunities.
“When I first joined Carnegie Mellon, I have distinct memories of this faculty meeting where everyone was like, what’s the killer app, like what’s the thing where robotics will actually get beyond academia?” Browning said. “No one had an answer at the time. … There’d been a whole bunch of these different applications that were like wonderful creations of technology, but they hadn’t quite hit that intersection of the technology being good enough and an application that was actually financially viable.”
Self-driving cars are the commercial opportunity now vaulting robotics aficionados out of academia and applied robotics and into the high-tech spotlight. Ford Motor Co.’s billion-dollar investment got Argo off the ground. Now, in addition to trying to develop a financially viable self-driving car in time for the automaker’s planned 2021 deployment in Miami, the Argo team is talking to “a long list of suitors.”
“What’s neat about this space right now is we’re anticipating a new industry,” said Rander. “There’s all this money coming into it, but it’s a new industry. It’s not a traditional core auto industry that people have learned and refined and honed and understand the dynamics. It’s like, ‘Well, gee, I don’t know X, so that’s interesting suddenly. Oh, this does this.’ As you’re figuring it out, suddenly you’re like, ‘Well, maybe we aren’t competing.’ ”
Going through Argo’s technology stack gives even a lay observer a sense of the dizzying technological challenge that self-driving cars represent.
Tasks from mapping to perception to prediction to planning are fiendishly difficult and fraught with a million ways to make a mistake, but understanding all of them individually — let alone how to bring them all together into a single functioning system — takes a staggering level of experience and focus.
“There’s no way any one person could possibly get into every single detail,” Browning said. “But you’ve actually got to get into a lot of details because that weakest link is the problem.
“It could be this really subtle little decision that seems so innocuous that some engineer in the team made one day that actually is the problem that we now have to fix. And that means part of the job is not just building the technology, but really building the team that has the right kind of mindset to check, double-check, scrutinize every decision we’re making.”