The Day Vehicles Started Driving Themselves

March 19, 2004

By 2020, it has been estimated, as many as 10 million self-driving cars will be on the road. Though optimistic, the prediction seems plausible enough: Google is making amazing strides in its self-driving prototype, and BMW, Mercedes, and Tesla are already adding self-driving features into their autos.

Remember, if you will, that not all that long ago the very idea of a self-navigating vehicle was mostly the stuff of SciFi. It certainly seemed fairly preposterous in 2004, when I had the opportunity to witness the first-ever autonomous vehicle race, held in the Mojave Desert, by the U.S. Defense Advanced Research Projects Agency(DARPA). So it was remarkable then that the winning vehicle had managed to travel was 7.4 miles on its own, just as it is remarkable today how quickly these robot vehicles have evolved since then. Here is my account of that day.

Virginia Tech's vehicle bolts out of the starting gate, followed closely by
                  the control vehicle, which was manned by a DARPA official to watch
                  the autonomous vehicle and stop it if got into a dangerous situation

It was somewhere around Barstow, on the edge of the desert, where the robots began to take control.

As the sun rose, a fire engine-red Hummer bounded from a starting chute--fashioned from a pair of concrete bunkers--and drove off into the desert. No driver was on board. The vehicle was controlled by a set of Intel Itanium processors.

As a small crowd cheered on from the bleachers, the driverless vehicle rolled away so assuredly that one might think that a computer-driven vehicle could actually navigate itself across the Mojave desert. The Hummer, with a giant fin welded on top, gracefully guided itself down a dirt road and around the series of curves that doubled back along the path behind the staging area. Within an hour, it was a few miles from the starting line.

It was a solid start to the Defense Department's first autonomous vehicle race, called the Grand Challenge, held March 13.

The race began in the pre-dawn hour of 6:30 a.m. at the Slash X Cafe, a tavern favored by off-roaders eight miles outside Barstow, Calif. Over the next few hours, a total of 12 more self-directing, or autonomous, vehicles would set off on a journey to find their way across 142 miles to the finish line, in Primm, Nevada. The first vehicle to finish, DARPA promised, would win a $1 million prize.

DARPA's robot race got underway in the pre-dawn hours. Starting chutes were
                  fashioned from concrete bunkers. Once an official waved a green flag,
                  the vehicle had 30 seconds to get out the starting gate.

What made this event different from the usual off-road race was that once the robotic vehicles were set loose, they were entirely on their own. No remote-controlled operations were allowed.

                  Williams, who headed up Team Phantasm. Williams' team had to
                  forfeit in the last day of qualifications. A defective radar burned
                  fried his vehicle's computer in the process. Unfortunately, his
                  programmer did not make a backup of the control program for the

Nor was rebooting. Each vehicle had to find its own way. It was as if after 400 years of close adult supervision, machines were finally being allowed to take a few baby steps on their own.

"If a vehicle decides to go over a ravine, it will go over the ravine," said DARPA director Anthony Tether, in a conference beforehand.

Alas, no vehicle made it past the eight-mile marker. In fact, most didn't make it very far from their starting gates before becoming befuddled by their surroundings. A few, however, made enough headway to astonish DARPA officials.

Cognitive Tractor Pull

"It didn't matter to us if anybody completed the course," Tether said. "We wanted to spark the interest in science and engineering in this area."

As the long-term research arm of the Defense Department, DARPA has a vested interest in getting people to think about autonomous navigation. Congress has set a goal for the Defense Department to make one-third of all its ground vehicles unmanned by 2015.

Tech's vehicle being driven off the course, after its brakes locked
                  up, disqualifying it from competition.

Autonomous vehicles are nothing new. Some are already in commercial production--think robotic lawn mowers and farm tractors. But none of those are designed to navigate through a course as complicated as the Grand Challenge, a rock-strewn route punctuated with steep slopes, dry lake beds, water, mud, cattle guards and, of course, lots of sand. It was a path even humans found difficult to navigate, people who knew the land said.

In order for these vehicles to guide themselves across this terrain, they needed to have plenty of computational smarts.

The idea of machine cognition is one problem that DARPA researchers have been trying to crack for years, with limited success.


How can a complex robotic vehicle get to where it needs to go, even if that means finding its way over big rocks, downed trees, or rubble? "The vehicle needs to perceive its environment, and it needs to send the right sequence of commands to its actuators so that it goes where it should," said DARPA program manager Larry Jackel during a presentation at the annual DARPATech conference, held earlier that week in Anaheim, Calif.

Jackel had earlier overseen research for autonomous navigation, testing systems in obstacle-strewn off-road courses. He found that state-of-the-art technology could get a vehicle through woods or thick grass, but vehicles were easily confused by such things as thickets of large bushes and fallen logs. Vehicles assumed shallow puddles and tall weeds were formidable barriers and refused to go through them.

Clearly, a lot of work still needs to be done to overcome the myriad obstacles that await autonomous vehicles in the world. So holding a race that would be open to anybody would get more people to work on the problem.

A few teams used hardened laptops for vehicle control. SciAutonics'
                  Rascal used an Itronix GoBook.

In this regard, the race proved to be a success. Tether said that staging the event cost the agency about $13 million-- assuming the $1 million purse was claimed. But he estimated that the agency was getting at least four or five times that amount in research, simply from all the teams trying to build competitive advantages into their vehicles.

Initially, more than 80 teams applied for the race when it was first announced in early 2003, though that number was whittled down to 25 finalists.

The teams themselves ranged from highly organized consortiums of companies and academia to southern California gearhead shops specializing in high performance racing automobiles. A few high school- and university-student teams even participated.

The prior Monday, DARPA started a week's worth of qualification runs--mostly to insure that the vehicles were safe--at the California Speedway in Fontana, Calif., cutting the number of contestants to 15. Two teams backed out in the last minute, citing technical difficulties.

The vehicles that made the final cut largely all had the same basic design. Most of the teams used commercial vehicles, ranging from a Hummer and four-wheel-drive pickups to the modified all-terrain-vehicles favored by off-road enthusiasts. Various commercially available actuators, hydraulic pumps and servers were rigged to operating the steering, acceleration and braking.

No steering wheel needed (CalTech).

The hardware the teams used also was largely commercial. Commercial Global Positioning System receivers provided coordinates of the vehicles in relation to their chosen route. Off-the-shelf-radar was used to provide long-range vision for the vehicles, while stereo cameras or light-detection and ranging devices, called lidar, provided more detailed information of closer objects. On the computational side, most teams used commercial computers or laptops, usually running Linux or a real-time operating system, though some teams used rugged laptops or shock-resistant hard-drives.

CalTech's vehicle was run by a team of personal computers, networked together

What was unique about each vehicle--and the ingredient that DARPA was interested in--was the software that would process the sensors' information and make decisions on how to move forward.

Warner Williams, who headed up Team Phantasm, said the vehicle control program would be simple--the software for his own modified ATV was less than a megabyte. The idea would be to "let the GPS fight with the radar for control of the vehicle."

"The secret sauce is in interpreting these sensors and figuring out what to make the vehicle do," Tether said during the wrap-up press conference. "Not only to be able to say, 'Yes, that is path over there,' but to transform that to the vehicle to go fast, put a brake on, turn left."

SciAutonics' vehicle traveled almost a mile before veering off-route

Out The Gate

If observers want to know where the cutting edge of autonomous vehicle control is in 2004, it would appear to be at about seven miles from the starting line. Even the most well-designed vehicles only got that far before malfunctioning for one reason or another.

University of Florida's Navigator trucks off down the road

It was no surprise that the red Hummer, which was overseen by a Carnegie Mellon University-led group called the Red Team, provided an exciting start to the race. Starts were staggered in five-minute intervals and DARPA officials had ordered the starting line-up based on the probability each vehicle had of completing the race, the strongest contenders departing first. Of all the teams, the Red Team appeared to put the largest investment into its vehicle, an estimated $3 million worth of parts and labor.

Needless to say, most of the launches that followed the Hummer's were not as smoothly executed. Team Caltech's white sport utility vehicle lurched out of the starting gate far more tentatively and stopped every 30 seconds or so to recalculate its course.

"I wanted to keep the truck in one piece," said team leader Richard Warren. As a result, the team programmed the vehicle to move ahead cautiously.

CanjunBot runs itself into a fence, and tries to hop over a bush Team CajunBot's all terrain vehicle, painted the color of green hot pepper sauce, ran about 20 yards before sideswiping a concrete bunker, a jam it tried to extricate itself from by attempting to roll over a bush.

Team member Chris Meaux later said that they had programmed the vehicle to understeer in order to compensate for the dirt road.

Team ENSCO's six-wheeled vehicle speedily jumped out of the starting gate and took off down the straightaway. It soon jumped up into bushes alongside the road itself and continued on its mad drive.

When it attempted to get back on the road after a few hundred yards, it flipped over.

ENSCO's vehicle came to a spectacular end, after bolting from the
                  starting gate and flipping over in some rough terrain

The vehicle may have been engineered to go a bit faster than was necessary, admitted Jacinda Clemenzi, an ENSCO team member, as the vehicle was being taken off the course by a tow-truck.

Other vehicles didn't even get out of the starting coral. The brakes of Virginia Tech's golf cart-like vehicle locked up at the start of its run.

Axion Racing's sport utility vehicle got only a few feet out of the starting gate before losing its way. It wandered around seemingly at random until it was disqualified by the judges.

The Caltech vehicle ultimately fared a little better. It lurchingly made its way through the first mile before coming to a stop off in the far distance, barely within eyesight.

Team CalTech's vehicle

Caltech's team leader Richard Warren, who was standing along the course just outside the starting gates with other spectators, was trying to figure out why his vehicle wasn't moving. He talked by cell phone and walkie-talkie with other team members, who were equally at a loss.

Each vehicle was closely followed by a control vehicle manned with DARPA personnel, who, through remote control, could pause or shut off the autonomous vehicle if it endangered the safety of humans or wildlife.

But none of the vehicle's team members accompanied the DARPA control vehicles, so information teams could get about their vehicles was scanty at best.

CalTech team leader Richard Warren, his team's vehicle off in the distance,
                     beyond his help. "You just want to find out what is wrong and then go out there and fix it," Warren said, squinting in the distance. Slowly word came back that the vehicle had jumped over a fence, wrapping barbed wire around its rear axle.

The tenacious machine slowly kept on, plowing through a number of bushes before jumping back over the fence and onto the road again. Then it stopped. And for the remainder of the race, the vehicle sat in distance, not moving. It was busy "thinking," as the announcer of the event put it.

A few teams surprised the officials and fellow racers with their designs. The Digital Auto Design team, with a modified Toyota Tundra four-wheel-drive, consisted of only two people. Yet their entry nearly equaled the much better-funded Red Team in terms of performance, admitted Jay Gowdy, the technical adviser for the Red Team.

Another team that surprised many was the Golem Group, which was led by Richard Mason. Mason largely financed the team himself, using winnings from an appearance on "Jeopardy." He scavenged volunteers from around the Southern California area.

Golem Team's dark house vehicle, a Ford F-150, which defied expectations by traveling over 5 miles
                  after a weak qualification

"This Golem Group impressed me in particular. They were a truly independent team," said Tim Grayson, the DARPA deputy program manager who followed Golem's vehicle. The vehicle, a black Ford F-150 pickup truck, struggled through the final qualifications, with the team working hard after each run to tweak its performance. As a result, they were scheduled to start second to last.

"When we flipped the switch this morning and sent it off, we hoped they'd make at least a good show of it. But they kept going," Grayson said.

The truck got 5.2 miles before it got stuck on a steep incline. What waylaid the truck was a simple mechanical limitation, said team member Mari Beth Mason. The team had set the vehicle servo to depress the accelerator no more than halfway down. So when the truck encountered a hill that needed more throttle to surmount, it simply stalled.

The Blue Team's self-driving motorcycle proved to be a limited success.
                  Here it is disabled, a few feet from the starting gate

The Bitter End

By 11:30 a.m., the race was declared over. Even the best vehicles had been judged as disabled.

The day's favorite, Carnegie Mellon's Hummer, drove for 7.4 miles before dying a smoky death. The vehicle experienced "what is known in robotics as a system failure," said Gowdy, meaning it experienced a series of minor failures that cascaded into entire system inoperability.

The chain of problems started with the truck knocking out its front fender, losing its sensors. That left only the less-accurate global positioning satellite coordinate to navigate by. As a result, the Hummer soon veered off course, tearing up its underside and eventually wedging itself on top of a large rock. Trying to free itself, it then tore the rubber off its tires, which landed near the engine and caught fire, Gowdy said.

Another favorite that came to a grinding halt was TerraMax. Here was a vehicle for which difficult terrain would certainly be no challenge: a 5-ton tactical military truck manufactured by Oshkosh Truck Corp., of Oshkosh, Wis. OshKosh had teamed with Ohio State University to enter the competition. Observers at the qualifications remarked on how easily the lime green vehicle could push entire concrete bunkers out of its way.

Team TerraMax's 5-ton truck tools down the road, sans driver

Nonetheless, information overload brought the beast to a halt a mere 1.2 miles from the starting line, said TerraMax team member Don Verhoff. When the vehicle veered slightly off the path, the information from its sensors overwhelmed the operating system.

The truck stopped to reboot its computer, but when it reawakened, it couldn't place where it was on the route map. So it backed up to find its bearings. Even when it backed across the pathway, it still couldn't place itself.

Despite the fact that no vehicle completed more than a twentieth of the course, DARPA was happy with the results. The performances were "extremely impressive," said Thomas Strat, DARPA deputy program manager for the event.

He noted that after the first five miles the course got considerably more difficult, with steep narrow slopes and a switchback to navigate through. A total of four teams' vehicles made it into this rough terrain--The Red Team, SciAutonics, Digital Auto Drive and the Golem Group.

One day, promised Tether, entire military conveys of trucks filled with supplies will find their way to the front lines, perhaps using some of the technologies pioneered by this event.

Judging from today's race however, it might be a while before the machines take to the roadways themselves.

The event was "a good day for the tow-truck drivers," as one desertbilly said.

Team CajunBot takes their vehicle out for a joyride after the race

This story originally appeared in Government Computer News