Robotic Vehicles Race, but Innovation Wins

The New York Times
September 14, 2005
By JOHN MARKOFF

FLORENCE, Ariz. - Cresting a hill on a gravel road at a brisk 20 miles an hour, a driverless, computer-controlled Volkswagen Touareg plunges smartly into a swale. When its laser guidance system spots an overhanging limb, it lurches violently left and right before abruptly swerving off the road.

With their robotic Touareg, known as Stanley, impaled in the brush, the two passengers - Sebastian Thrun and Michael Montemerlo, both Stanford computer scientists - pull off their crash helmets and scramble out to untangle the machine.

A quick survey reveals that the sport utility vehicle is covered with debris, but the bug-eyed laser, radar and optical vision system on top of the vehicle is undamaged. So Stanley and its passengers continue on their way, over 50 miles of dirt road through a cactus-covered landscape, in the final weeks of preparation for the second round of the Pentagon's great race.

It has been almost 18 months since the Pentagon's research arm, the Defense Advanced Research Projects Agency, first attracted a motley array of autonomous vehicles with a prize of $1 million for the first to complete a 142-mile desert course from Barstow, Calif., to Las Vegas. The most successful robot, developed by a Carnegie Mellon University team, managed all of seven miles.

With the next running scheduled for Oct. 8 - and this time a $2 million purse for the winner among 43 entries - it is clear that many of the participants have made vast progress. For some researchers, it is an indication of a significant transformation in what has been largely a science fiction fantasy.

"Computers are starting to sprout legs and move around in the environment," said Andy Rubin, a Silicon Valley technologist and a financial backer of this year's Stanford Racing Team, which produced Stanley. Mr. Rubin, who tinkers with robots himself, was the co-founder of Danger Inc., which created the Sidekick hand-held.

The Pentagon agency, known as Darpa, struck upon the idea of a race - calling it the Grand Challenge - as a way to stimulate innovations useful in battlefield applications like unmanned logistics vehicles.

For the two Stanford scientists, however, the Grand Challenge is about something larger. "The military are interested in more potent weapons, and by itself that's a bad answer," said Mr. Thrun, a roboticist and director of the Stanford Artificial Intelligence Laboratory. His broader goal is to advance robotics as a science and explore applications ranging from aids for the elderly to basic advances in intelligent computerized systems.

Several years ago, when Mr. Thrun was a professor at Carnegie Mellon and Mr. Montemerlo was a graduate student, they helped develop a prototype of a mobile robotic companion for the home that used natural-language voice commands and was able to provide useful information taken from the Internet like weather and television schedules.

There are a myriad of other possible applications for their software, which can reason about the immediate environment; distinguish sky from ground, road and trees; and make lightning-quick decisions.

Already in the automotive industry, intelligent cruise control has become more adept at automatically maintaining the spacing between cars, and intelligent lane-change and collision-avoidance software is close to being a reality. Robots are routinely used in manufacturing, and in Japan a three-foot-tall "house sitter" robot that can recognize 10,000 words and 10 different faces will go on sale in September, offered by Mitsubishi Heavy Industries.

In the Darpa contest, though, the proving ground is not the home but the desert. And several of the contestants, who range from garage hackers to teams from giant automotive and aerospace corporations, say this year's course is expected to be even more difficult than last year's.

The exact course will be secret until just hours before the event begins, but Darpa officials are said to believe that the original test was too much an exercise in automatically following global positioning satellite "bread crumbs" - the data points outlining the route that are given to the contestants shortly before the race begins.

So this year the course is likely to include unexpected man-made obstacles and other hurdles that would be trivial for a human driver, but vexing for the computer-controlled navigational systems that are at the heart of the technical challenge the Pentagon has laid out.

Despite the added complexity, there is a widespread expectation among robotics researchers that this time the course will be completed.

The machines, many of which wandered seemingly randomly in the desert last year, have benefited from more than a year's experience as well as a significant rush in improvement in every aspect of robotic vehicle technology. And on a hot August day in the desert here, it was clear that the field of artificial intelligence has made significant strides.

The increasing power of the technology was evident during the testing of the Stanford Racing Team's robotic Touareg, which looks unexceptional from the outside except for a festoon of sensors and the slogan "Drivers Not Required" on its side, a play on Volkswagen's "Drivers Needed" slogan.

Stanley was able to complete a 47-mile dirt-road course here - strewn with potholes, tight turns, puddles and lined with boulders, foot-high berms and cactuses - with only two "incidents," which in Mr. Thrun's scientific vernacular is when his robot does something unplanned, like leaving the road.

When their Touareg swerved abruptly in a roadside thicket, the team was quite certain why.

The previous evening Mr. Thrun had persuaded Mr. Montemerlo to remove an irritating software module, which forced the car to brake rapidly after swerving to avoid an obstacle. Without the module, at speed the Touareg fishtailed on the desert road and plunged into the brush before Mr. Thrun, sitting in the driver's seat with his hand on a large red "E-Stop" button, could react.

Back inside the Touareg, Mr. Montemerlo, seated in the rear seat with a laptop computer that is networked to the seven mobile Intel Pentium processors that comprise Stanley's control logic, fiddles with the software and reinserts the problematic code. Now the vehicle will behave more cautiously, although the hard braking will be a little uncomfortable for its human passengers.

[After fixing two software bugs, the Stanford team managed to put Stanley through the entire test course on Sept. 7 without crashing.]

In the actual race, of course, there will be no passengers along for the ride. The teams will be able to follow a short distance behind, but will have no communication with their vehicles.

For the two researchers, who have been leading a team of 60 developers from Stanford and Volkswagen, the hiccup is all part of the process of trying to create machines that can mimic what human drivers do effortlessly.

The challenge is heightened by the obvious rivalry that the two scientists feel with their alma mater, Carnegie Mellon. This year, the Carnegie Mellon Red Team - led by the roboticist William L. Whittaker, known to all as Red - is testing two robotic vehicles, Sandstorm and H1ghlander, in the Nevada desert.

With an array of sponsors including Caterpillar, Intel, Boeing, Harris, Google, and Hummer's manufacturer, AM General, Mr. Whittaker's team is once again the favorite.

For decades Mr. Whittaker has been one of the most passionate advocates of robotic vehicles. Despite being bitterly disappointed last year, when Sandstorm edged off the course after almost completing the most difficult section of the route, he is confident that more than one team will succeed this year.

"I would love it if the high school kids won this year," he said, in a reference to a team from Palos Verdes High School in California, which is backed by Boeing, Northrop Grumman, Analog Devices, Goodyear and others.

Whether or not one of the vehicles arrives at the finish line this year, Mr. Whittaker says the credibility problems that have dogged the field are largely in the past.

Of the event, which will begin this year near a rough-and-tumble bar south of Barstow, he said, "I don't know whether it's going to be more like Lindbergh landing in Paris or more like Woodstock."