Egang: Data Driver

On The Cover/Top Stories
Data Driver
09.04.06

Sebastian Thrun is very clear on the future. "Forty-two thousand people in the U.S. die in traffic accidents every year," he says. "Down the road, I know cars will drive themselves. It's just right!" Robots don't get drunk, fall asleep or talk on the phone. If we turn highways into "invisible rail systems," he says, we can combine the convenience of a car with the safety of a train.

This vision has been a futurist phantasm since the 1950s, but Thrun and a car named Stanley are bringing the age of the autoroad a bit closer. Stanley is a Volkswagen Touareg built by Thrun, 39, who teaches computer science at Stanford, and his graduate students. In October 2005 Stanley drove itself to a rookie victory in the Grand Challenge, a 131-mile race across the Mojave Desert. First prize: $2 million. Race sponsor, tellingly: the U.S. Department of Defense.

In the same race a year earlier none of the 15 robotic entrants even finished. Thrun entered the race last fall largely out of frustration at these results; he had just arrived at Stanford from Carnegie Mellon's Robotics Institute. Researchers there had spent years coming up with some of the fundamental principles for building robotic vehicles. Analyzing tons of data was key: They packed cars with ever more sensors to navigate the world around them.

But Thrun knew that sometimes the data points are simply wrong. He had built a robotic tour guide for the Smithsonian National Museum of American History in 1998, and the contraption got confused when people waved at it. Thrun taught his bots to stop trusting all data equally and instead assign some probability to its utility and accuracy.

"The robot has to figure out: Is this data a wall? Could it be a kid? Could it be a shadow or a trick?" Thrun says.

Stanley sports five global-positioning antennas, six Pentium M computers, five laser range finders and a video camera. To train Stanley's brains, Thrun and his crew drove it 1,000 miles over desert terrain to reveal what conditions drivers deem critical and what they ignore. A bad bump jostling Stanley's sensors might accidentally produce data indicating a wall ahead even if the road is clear. Stanley learned to downplay such data and stay attuned to more significant obstacles.

To win the race in October Stanley took six hours and 54 minutes; Carnegie Mellon's car needed another 11 minutes to finish. Best of all, Thrun says, 5 contestants finished (of 23 total). "We all won. The robotics community won." The next Grand Challenge is in November 2007, with a 60-mile romp through a city yet to be named. Contestants will have to avoid static obstacles and, most important, one another. Thrun will be ready with a new Volkswagen Passat. "To make robots really useful," he says, "they will have to deal with traffic."

http://www.forbes.com/free_forbes/2006/0904/102.html?boxes=custom

--##--

Sphere: Related Content