|The PayPal Mafia of Self-Driving Cars Has Been at It a Decade|
A Pentagon-funded contest spawned many of today’s self-driving startups. A decade later, the competitors are struggling to perfect the technology.
By Dana Hull
October 30, 2017
Boss, a self-driving SUV from Carnegie Mellon University's Tartan Racing team, on display during the 10th anniversary celebrations of Tartan Racing's victory in the 2007 DARPA Urban Challenge on Saturday, Oct. 18, 2017 in Pittsburgh, Pa.
Photographer: Stephanie Strasburg/Bloomberg
The brief history of the self-driving car is parked in Pittsburgh. Robotic vehicles built by students and professors stand sentry on the campus green at Carnegie Mellon University. These cars carry scars and dents—one shows rollover damage—from participation in a legendary series of competitions organized by the Defense Advanced Research Projects Agency (Darpa), a research and development wing of the U.S. military. Many of the engineers involved in those contests went on to launch autonomous vehicle startups that are now steering the future of transportation.
A few employees from Aurora Innovation Inc.—you could tell because they all wore Aurora t-shirts—arrived on a recent Saturday morning to pay homage to their autonomous ancestors. “It’s like looking at cell phones from the 1980s,” says Clint Liddick, 27, a software engineer at the company’s Pittsburgh office. “Now I see self-driving cars every single day, outside of my own work. My three-year-old son instantly recognizes them.”
The head of Aurora, a roboticist named Chris Urmson, was the assistant professor who led the Carnegie Mellon team behind “Boss,” the self-driving Chevy Tahoe that won the Darpa Urban Challenge in November 2007, exactly 10 years ago this week. He has come back to campus over an October weekend to reminisce with veterans of the Boss team, an anniversary-inspired reunion. The obligatory event swag—gray polo shirt, black ball cap, stainless steel tumbler—features a red-outlined image of Boss and its bulky sensors.
“It’s a lot more obviously hardware-heavy,” says Urmson, pointing to a clunky laser on Boss that looks remarkably like a coffee maker. “But much of the technology that is on here is what is being used today.”
Aurora, whose sparse website promises “safe, efficient mobility,” is part of the new crop of companies whose origins can be traced back to Pentagon-funded contests a decade ago. A close-knit diaspora emerged from the Darpa competitions in much the same way early members of the so-called PayPal Mafia went on to lead or fund a long list of Silicon Valley successes. Most of the self-driving descendants have vaguely futuristic, one-word names—Argo, Nuro, Waymo, Zoox—as well as a fervent idealism about preventing traffic deaths and transforming the way people live and move. Urmson and his peers from the contest embody how far autonomous vehicles have come in a decade—and how much work remains.
Aurora co-founder and Chief Executive Officer Chris Urmson (left) leads a panel discussion at Carnegie Mellon University during the 10th anniversary celebrations.
Photographer: Stephanie Strasburg/Bloomberg
There was federal money at the inception of the self-driving vehicle, the same government largesse lurking in the origin stories of the internet, global-positioning system technology and alternative energy. Darpa has pursued the same mission since the Sputnik era: Make key investments in breakthrough technologies to promote national security.
The autonomous-vehicle challenges were designed to bring out into the world technology that had been under development for decades in labs. There was military urgency at the time. The U.S. was fighting wars in Afghanistan and Iraq, and scores of soldiers were being killed by roadside bombs. Driverless vehicles could save lives on the front lines.
An initial competition, the Darpa Grand Challenge of 2004, asked robotic cars to travel roughly 140 miles across the Mojave Desert. Carnegie Mellon’s entrant, a Hummer named “Sandstorm,” managed to travel the farthest—a whopping seven miles. At a follow-up event, in 2005, Stanford University came in first place, and Carnegie Mellon’s contenders placed second and third. Each university team included dozens of students and professors, as well as corporate sponsors.
The two early races proved that autonomous vehicles worked. But the courses were static—the cars didn't have to navigate dynamic traffic conditions or interact with each other. A final contest, the Darpa Urban Challenge, would bring the contestants back together at an abandoned Air Force base in Victorville, California. There would be a mock city this time, with moving parts.
“We had dummies that looked like people walk in front of the car, and the sensors had to show that they could recognize that it was a person and stop,” says Anthony Tether, who ran Darpa from 2001 until his retirement in 2009. Only 11 of 89 teams made it through to the final round.
More than 2,000 people attended the final event on Nov. 3, 2007. Hundreds of Darpa personnel managed the course from a command center, while large teams of university students and faculty wired on adrenaline and anxiety mingled with auto industry sponsors and observers. Larry Page, the co-founder of Google, flew in on his corporate jet. A JumboTron was installed to broadcast the six-hour event. (This video footage from YouTube gives a good sense of what it looked like.)
Tether now jokes about the similarity between the Urban Challenge and Woodstock: They’re two events that improbably large numbers of people claim to have participated in, and both invoke wild stories that have become part of communal lore.
Boss almost failed to get out of the starting gate, a story Tether recounts over a reunion lunch between panel discussions. It turned out that the JumboTron was interfering with its GPS signals, a snafu that crazed the already frantic Carnegie Mellon team. Tether finally killed the power on the massive video screen, and Boss came back to life.
The teams didn’t have a bird’s eye view of the entire course and spent much of the morning wondering if their robotic cars would survive. “They didn’t want anyone cheating. You had no visibility. For six hours, you had no idea what was going on. I was sick to my stomach the whole time. It was like please, please, please come back,” says Bryan Salesky, the software lead for the Carnegie Mellon team. He later joined Google and co-founded Argo AI LLC, in which Ford has committed to invest up to $1 billion.
Former Darpa Director Anthony Tether at the reunion on Oct. 18, 2017.
Photographer: Stephanie Strasburg/Bloomberg
The old Air Force base was a hive of activity. In addition to 11 autonomous contenders, roughly 30 vehicles driven by humans simulated traffic. By mid-morning, about half the field had been disqualified.
“There was this big traffic jam, and Carnegie Mellon’s vehicle did a three-way turn all by itself. My God, it was tremendous,” says Tether. “Boss worked its way out of a major traffic jam better than most humans would.” Even the head of DARPA had trouble persuading himself there was no one inside the car.
The winners were announced the next morning, after Darpa officials spent hours poring over reams of technical data. Virginia Tech came in third, Stanford placed second and Carnegie Mellon won bragging rights, $2 million and an enormous bronze eagle.
“The first two Grand Challenges were our Kitty Hawk moments,” says Tether, referring to the place in North Carolina where the Wright Brothers experimented with flight. “But the Urban Challenge was like Charles Lindbergh flying to Paris. When he did that, aviation became a real thing and people said, ‘We can make a business out of this.’ Once we showed it could be done, the Google people came in with the money.”
Sandstorm, a self-driving race car from Carnegie Mellon University's Tartan Racing team.
Photographer: Stephanie Strasburg/Bloomberg
Before Google’s money, however, came the burst of the housing bubble, the 2008 recession and the near-destruction of the U.S. auto industry, which cut research funding in its struggle for survival. General Motors Co.—the lead corporate sponsor of the Carnegie Mellon team—filed for bankruptcy protection in 2009. But Google Inc. had plenty of available resources, and Page, who studied computer science at Stanford, had gotten to know Sebastian Thrun, one of the leaders of the Stanford team.
“In 2009, Larry was very eager to find a new home for innovation within Google, and the very first project was the self-driving car project,” says Thrun. Page told him to build a dream team, and Thrun’s first hires were all alumni of the Darpa Urban Challenge: Mike Montemerlo, Chris Urmson, Anthony Levandowski, Dirk Haehnel and Dmitri Dolgov.
It wasn’t that auto makers didn’t see the potential, but Detroit’s timeline was too timid. “The auto industry CEOs at the time saw this as technology that would evolve, maybe in the 2030 to 2040 timeframe,” says Larry Burns, GM’s vice president of research and planning at the time of the Darpa contest. He now consults for Waymo LLC, the self-driving company spun out of Google last year. “By stepping up when they did,” he says of Google’s co-founders, “they pulled this industry forward by at least a decade.”
At last count, Waymo has at least eight employees who were involved in the Darpa challenges, including Mike Montemerlo and Dolgov, who is the vice president of engineering. Other Darpa alumni got fabulously wealthy, thanks to Google stock options, and headed off to start their own companies.
“A lot of millionaires have come out of this shop,” says William “Red” Whittaker, a longtime robotics professor at Carnegie Mellon. “When they were students, they couldn't rub two nickels together to buy breakfast.”
As the industry has matured, rivalries have broken into the clubby network of friends from the Darpa event. Waymo sued Uber Technologies Inc. earlier this year, alleging that the ride-hailing giant stole trade secrets concerning autonomous technology. Although he's not named as a defendant, at the center of the lawsuit is Anthony Levandowski, a member of the Stanford team 10 years ago. He joined Google’s self-driving car project early on, like so many others, and then left to start autonomous-trucking startup Ottomotto LLC, which was quickly snatched up by Uber. Levandowski is expected to invoke the Fifth Amendment right to avoid testifying; his lawyer did not respond to a request for comment. The trial is scheduled to start on Dec. 4.
Competition for talent is also ferocious, with contest veterans vying against each other for recruits, funding and partnerships. In California alone, more than 40 companies, including Apple Inc., Bayerische Motoren Werke AG and Samsung Group, have been issued autonomous-vehicle testing permits from the Department of Motor Vehicles. But ultimately, everyone shares the same goal.
“We all run into each other, and it’s pretty friendly, even if you work at a competitor,” says Jesse Levinson, a member of Stanford's team who in 2014 co-founded Zoox Inc., a startup that aims to make a fully robotic taxi. “We all want this technology to exist for society.”
The final Darpa race in 2007 had a total budget of about $30 million. A report this month by Jefferies Group LLC found that the total market for autonomous-vehicle technology is now at $100 billion—and that figure excludes the value of the actual cars involved.
The hype cycle has reached a fever pitch. Lyft Inc. just raised $1 billion. GM's Cruise plans to test a fleet of self-driving electric Chevy Bolts in Manhattan next year. Waymo is running public trials in Phoenix. Elon Musk claims that all Tesla Inc. cars being built at the company's factory have the hardware needed to be fully autonomous and has vowed to demonstrate a cross-country trip without touching the steering wheel. Every day brings fresh rumors of partnerships among automakers, suppliers and startups developing autonomous-technology platforms.
A detail of the sensor array on Boss.
Photographer: Stephanie Strasburg
For members of the Darpa diaspora, who have been in the trenches longer than most, much work still lies ahead. There have been enormous advances in a form of artificial intelligence known as machine learning, wherein computers can act without being explicitly programmed, and the cost of sensor hardware is coming down. But the business models behind autonomous driving remain unclear. Will we subscribe to cars instead of owning them? Will passengers pay by the mile or the amount of time? How can you prove automated systems are truly safer than human drivers? Is a car trained on the streets of San Francisco going to perform well in Beijing?
Training robotic cars to replicate the human eye and brain using multiple sensors—cameras, LiDAR, radar—takes an astonishing amount of data and time. “We’re getting very close,” says Montemerlo of Waymo. “There’s more that needs to be done, but there’s a world of difference between where we are now and where things were 10 years ago.”
The engineers who were there at the beginning now routinely mix heady optimism with hard-eyed assessments of the remaining challenges. Salesky, who built the software in Boss, dismisses anyone who thinks self-driving vehicles will be commonplace within a few years. Urmson is famous for predicting that self-driving cars would arrive before his son gets his driver's license. His son is now 14.
“It’s just like anything in life,” says Urmson. “It’s comparatively easy to get to the point where it mostly works, but to have it work all the time—and under a variety of circumstances that you have to face—is where it gets really hard.”
Urmson was already taking reporters on autonomous demonstration drives when he was at Google back in 2010, three years after Boss blew minds making a three-point turn by itself. “And seven years later,” he admits, “we're still working on it.”