Portland, Ore. Slow but sure took the $2 million purse in the second Darpa Grand Challenge. Using what its developers called a "tortoise strategy," an autonomous Volkswagen Touareg named Stanley exploited artificial-intelligence techniques to cover 132 miles of Nevada desert in 6 hours and 53 minutes, for an average of just over 19 miles per hour.
"Stanley won because of our tortoise strategy, where we concentrated on perfecting our navigational software rather than on driving fast," said Gary Bradski, a principal engineer and machine-learning expert at Intel Corp. who worked with the Stanford University team that designed Stanley.
Second and third in the Oct. 8 race, sponsored by the Defense Advanced Research Projects Agency, went to a pair of Hummers from Carnegie Mellon University. Sandstorm trailed the leader by 11 minutes and H1ghlander by 21 minutes. Of the field of 23 autonomous ground vehicles that started the race before 2,000 spectators, only five crossed the finish line, and one of those exceeded Darpa's 10-hour time limit.
Still, the results vastly improved on last year's Grand Challenge, when every entrant either stalled or crashed within seven miles.
"Just last year, there were people who said it couldn't be done," said Stanford professor Sebastian Thrun, the leader of the winning team and director of the Stanford Artificial Intelligence Laboratory. "Now we know that autonomous vehicles are possible, and that knowledge is all it takes for others to succeed, too."
Last year's starting field all tried to outpace one another with horsepower. This year, the top three winners swore off horsepower in favor of reliability, ruggedization and smarts. "Our focus was reliability from the start, and in the end it was reliability that won the race," Bradski said.
With two entries, Carnegie Mellon tried a tortoise-and-hare approach. H1ghlan-der, the "hare," had been favored to win, but ran into mechanical difficulties. "The pace we set for H1ghlander should have had it finishing 30 minutes ahead of [tortoise] Sandstorm. Unfortunately, H1ghlander had mechanical trouble that slowed down its pace," said professor William Whittaker, the CMU team leader.
"The Carnegie Mellon team had higher-performing vehicles, but in the end the extra performance didn't help," said Intel's Bradski, who spent a year's sabbatical working on Stanley. He also enlisted the help of Intel engineers Adrian Kaehler, Bob Davies and Ara Nefian.
The brains of all three winning vehicles were identical, having been donated to Stanford and Carnegie Mellon by Intel, which permitted separate engineers to work with the rival teams. The computers used were six Pentium M processors, which were low power enough to run off the alternator in two of the three vehicles (H1ghlander had an auxiliary power generator supplied by Caterpillar Inc.). In-stead of laptops, the six Pentium M's were packaged as blades in a ruggedized platform designed to be earthquake proof (no spinning hard disks and a spike-resistant power supply). Each winning vehicle used one Intel 5091 chassis and six Intel MPCBL5525 processor blades.