Just because Waymo settled its high profile lidar trade secrets case against Uber earlier this month, it doesn’t mean Anthony Levandowski is out of the spotlight. The U.S. Justice Department could still file criminal charges against the ex-Waymo engineer for the alleged theft of technical documents from his former employer. And then there’s the question of what Levandowski is planning to do next: Will he use his vast experience with autonomous vehicles to launch another startup—and make a comeback?
During a deposition last April, Levandowski did not want his experience and plans scrutinized. When Waymo lawyers asked him hundreds of questions, mostly about his activities at Waymo and Uber, Levandowski took the Fifth, to avoid answering questions that might incriminate him. There was, however, one project he was eager to talk about: GhostRider.
“What was your entry into the  DARPA Challenge?” asked one lawyer, referring to the Pentagon’s famous $1 million self-driving vehicle competition that kickstarted the entire industry. “The entry was called GhostRider, and it was a two-wheeled motorcycle,” replied Levandowski. “It was the first of its kind... [and] frankly, a pretty crazy idea.”
Building GhostRider cast Levandowski as a robotics wunderkind, secured his place at a follow-up DARPA Grand Challenge, and ultimately enabled him to build Google’s first self-driving car—a step that would later make him a multimillionaire. In 2007, Levandowski immortalized his role as an autonomous vehicle pioneer by donating GhostRider to the Smithsonian’s National Museum of American History in Washington, D.C.
Levandowski, who has been profiled numerous times, has discussed the project before. And yet GhostRider’s full story has never been told. In particular, it seems the riderless motorbike that launched Levandowski’s career was lucky to race in the first DARPA Grand Challenge at all.
IEEE Spectrum has pieced together GhostRider’s history from new and contemporary interviews with Levandowski, as well as records that include a 14-year-old press kit—a glossy white folder with the words “GHOSTRIDER ROBOT” on the cover—recently discovered in a box of old files by a Spectrum editor.
Levandowski first heard about DARPA’s Grand Challenge while a graduate engineering student at the University of California at Berkeley in 2002. He decided immediately he wanted to enter the race.
While brainstorming in a hot tub, Levandowski and his friend Randy Miller came up with a bunch of ideas, including riderless motorbikes and a robotic forklift. “But the idea for a motorbike got cemented when we were driving back from a Grand Challenge conference and a bunch of motorcycles went around us on the freeway,” Levandowski said in an interview last week.
The motorbike, originally a Honda XR, was light enough to be loaded into a pickup, and for Levandowski to physically pick it up when it inevitably fell over. The bike was also one of the cheapest vehicles to star in any of DARPA’s Grand Challenges, with Levandowski estimating the multi-year project at around $100,000. The money came from his own pocket and also from corporate sponsors and individual donors.
“People gave money through PayPal, ten bucks here, a hundred bucks here,” he said, adding, “It was a bit like an early Kickstarter.”
Levandowski began building the self-driving motorbike in his own garage near Berkeley, gathering together half a dozen fellow engineering students to help. The volunteers were paid in burritos and some of them even eventually moved in. The group would be called the Blue Team, a reference to the friendly team in military exercises. They named their robot bike Dexterit.
One of the members of Blue Team, Bryon Majusiak, who now builds agricultural robots, remembers getting a call from Levandowski at 11 one night to help unload the motorbike. “He was so impressed that I actually showed up and I started with him all the time,” Majusiak remembers.
The first task for the Blue Team was installing servos to control gas, clutch, and brake. A dc motor and worm gear reducer operated the handlebars. A lead-acid battery powered the electronics. Levandowski went through a couple of Honda bikes before settling on a Yamaha 125 for the first Grand Challenge. (The final GhostRider bike would be a child’s Yamaha 90 dirt bike, chosen for its automatic clutch that Levandowski called a “lifesaver.”)
Once the mechanical controls were installed, the Blue Team faced its hardest challenge. “To get a car to move down the street, you can kind of apply a little bit of accelerator and not steer, and the vehicle will do that,” Levandowski said in his deposition. “To get a motorcycle to move forward, you have to build a lot of technology beforehand to make it able to just drive in a straight line.” He added, “It turned out that the complexities and challenges of adding the balancing before you could start testing all of the other navigation and optical [systems] were hard.”
At one point in early 2003, after struggling to make progress, Levandowski remembers telling his team that if they could not get Dexterit to travel a mile [1.6 kilometers] before the next Sunday, he was going to abandon the project.
But then they had a breakthrough, an elegant solution to the twin problems of balancing and turning. Instead of shifting weight on the bike to balance it like a human rider, the Blue Team realized that steering the Yamaha slightly in the direction it was tipping would create a force to balance gravity.
“Counter-steering creates centripetal acceleration, which causes a torque in the other direction,” Levandowski told me. “You then balance that back and forth to keep going straight.”
Turning was accomplished by allowing the motorcycle to lean into a curve while keeping the handlebars straight. In a video from 2003 or 2004, an early version of the bike can be seen balancing itself while stationary in a driveway. In a later video, a more advanced version drives in circles on a lawn.
While the concept sounds simple, it took months of work for the Blue Team to achieve a smooth, human-like ride. Levandowski said they flipped the motorcycle hundreds of times.
A technical document for the 2004 competition reveals that the motorbike’s primary sensors were optical, with a range of about 40 meters. A pair of forward-facing monochrome webcams scanned for moving obstacles, while a single color camera was used to detect the road itself. An on-board computer using an AMD Athlon 64-bit CPU and 512 MB of RAM was able to process only about one frame every 4 seconds.
The cameras were mounted on a gimbaled gyro-stabilized mount above the front wheel, and another gyro and an inertial measurement unit provided orientation and acceleration data. Optical encoders tracked steering angle and speed, and a GPS unit fixed its location.
The plans was to take a dataset of waypoints—to be provided by DARPA shortly before the race—and process it with a custom application to “increase the density of waypoints to provide a tighter route for the vehicle to follow.” The bike would rely on that data to guide itself through the desert, while at the same time using its cameras to detect rocks, pits, and other vehicles.
“It was basically a bunch of people putting stuff together and seeing what worked,” said Levandowski. He had hoped to get a 77 GHz mechanically scanning radar on board but could never get it powered up. “I got shocked like 10 times in one day trying to work out where the short was in the device,” he remembers. “We never got it to work—and we never tried lasers.”
Many components were supplied as in-kind sponsorship from manufacturers and retailers. Hobby Engineering, an internet seller of robotics kits, got involved when Levandowski called its founder, Al Margolis, at 3 am one morning. “[Levandowski] was dumbfounded to have the call answered,” Margolis wrote in a press release at the time of the Challenge. “He wasn’t really expecting an answer, let alone actual help, but he was desperate enough to try.”
A few hours later, the two met at the Hobby Engineering showroom, where Margolis provided some microcontroller chips and debugging assistance to help Levandowski meet a demonstration deadline.
No one on the team worked harder than Levandowski himself, remembers Randy Miller, now an engineer and developer: “He would work through the night and go two or three days without sleep, just grinding away.”
As the Challenge approached, Levandowski gave more thought to what might happen if the Blue Team actually made it through a week-long qualification event in California to the high-profile (and potentially lucrative) competition itself. Levandowski changed Dexterit’s dorky name to the more evocative GhostRider and put together a press kit containing a profile of himself (in which he calls Bill Gates his hero), a slide deck, and letters from sponsors like Agilent and Crossbow Technology. Early in 2004, he also set up a company called Robotic Infantry Inc. to “explore the possible military and commercial applications of this technology.”
According to the slide deck, Robotic Infantry’s next steps would be to model GhostRider’s components in 3D, build an electric-powered carbon fiber model, then refine its sensors and improve its software. By 2007, Levandowski hoped to have a road-worthy product and be ready to “consider technology acquisitions” and “investigate venture capital potential/IPO.”
The Blue Team worked on GhostRider until the last minute, only adding a remote emergency deactivation feature required by DARPA days before the competition. “Like a lot of other people at the Grand Challenge, it was very difficult to get [GhostRider] working in time,” Levandowski said. “We prioritized making the thing work before making it stop. Stopping wasn’t really a problem for us: If the system didn’t work, it just fell over.”
On 8 March 2004, 25 teams arrived at the California Speedway near Los Angeles for the qualification event. DARPA had marked out a 2.2-kilometer course with obstacles that the vehicles would meet on the real race, such as dirt hills, ditches, cattle grids, and a sand pit. The vehicles had to pass a safety inspection and were expected to complete at least two runs of the course.
Only the 15 fastest and most capable vehicles would be allowed to enter the Grand Challenge proper, and GhostRider was facing well-funded teams from the likes of Stanford and Carnegie Mellon universities, Caltech, and the Oshkosh Truck Corporation.
“On the first year, we actually did manage to qualify,” Levandowski said in his deposition last year. “And out of the 109 teams or so that applied, you know, 14 or so qualified, and we were one of them.”
But how, exactly, GhostRider ended up in the finals is a bit hazy. The Blue Team was due to make its first run on 9 March, and its unique two-wheeled design attracted a large crowd of spectators. Sadly, the Los Angeles Times reported that GhostRider fell over that day after traveling only about 4.5 meters. The next morning, GhostRider was up again but, once more, failed to make a run, according to a DARPA news release. In fact, Levandowski told DARPA that the Blue Team had terminated its attempt. GhostRider was not alone—many teams were struggling to get their vehicles moving without bumping into obstacles or veering off course.
On March 11, the final day of qualification, a total of 38 last-minute attempts were made to complete the course. The Blue Team was not among them. By the end of the qualification event, a total of seven vehicles had managed to complete at least one circuit of the course, and eight more had made what DARPA considered partial completions. That neatly added up to the 15 teams that DARPA wanted for the race itself on March 13.
Fourteen of those fifteen teams duly made it onto the race-day roster. But one, Rover Systems, an innovative off-road robot with a low center of gravity and four-wheel steering, was not on the list, despite having completed parts of the course on two occasions. Instead, DARPA selected GhostRider, which had only made a little progress on the first day.
Levandowski’s explanation is that “we earned our spot” in the race. “We had showed that we were able to get the motorcycle out of the gate, make a turn, go straight and then crash into a fence,” he said. GhostRider probably also benefited from Levandowski driving it in circles in a parking lot, outside the official qualification runs.
Whether it was that impromptu demonstration or Levandowski’s knack for showmanship, GhostRider would get to compete for the $1 million prize in a challenging 230-kilometer off-road race between Barstow, Calif., and Las Vegas, Nev.
Levandowski was realistic about GhostRider’s prospects. “There was never a chance we were going to win, as we didn’t even have enough gas to make it the whole way,” he said. “But we were hoping to make out of the gate, past the first turns, and then perhaps crash into some tree 300 feet away.”
But when the starting gun fired, GhostRider could not even match its personal best of 4.5 meters. Instead of buzzing off into the desert, the bike simply fell down. Its Grand Challenge was over in a matter of seconds.
“It’s really difficult when the bike is trying to balance itself and you’re walking it—it’s fighting you,” Levandowski told me. “So we turned the stabilization off while walking the bike to start line and then forgot to flip the switch to turn the balancing on again. It was highly embarrassing.”
Embarrassing or not, GhostRider’s participation in the original Grand Challenge marked out the 23-year-old Levandowski as a name to watch. Because no vehicles completed the Challenge that year, DARPA invited all the finalists back to compete in a follow-up race in 2005, this time with a $2 million prize.
GhostRider once again failed to finish the qualification course, although this time DARPA passed it over for the main event. The race was won by Stanley, an autonomous Volkswagen Touareg SUV built by a Stanford University team led by Sebastian Thrun.
Thrun took a shine to Levandowski, and after the race gave him a tour of Stanford’s lab. In 2006, Thrun invited Levandowski to help out with a camera mapping project called VueTool that caught the eye of Google, which hired the entire VueTool team the next year to develop its own StreetView system. Within a couple of years, Levandowski and Thrun were building Google’s first self-driving car, and the rest is history.
As to what he plans to do next, Levandowski is not ready to share any details. But it probably won’t involve motorcycles. He laughs when I ask him whether he has ever given any thought to revisiting the idea of a riderless bike. “I don’t think it’s actually a great robotics vehicle platform,” he said. “It’s a great training tool and you can totally get it to work but the added complexity isn’t worth it, as you can’t carry cargo very easily.”
“It would be fun to remake GhostRider for pure enjoyment,” he added, “but it’s not significantly relevant to self-driving cars.”