Skip to main content

CU proves it is up to the DARPA challenge, making final six

A car driven by an artificial intelligence (AI) that can navigate city streets, obey traffic laws, pass, merge and avoid other vehicles, reroute around blocked streets and above all, not hit anything, sounds almost impossible. But six university-affiliated teams, including one from Cornell, met such rigorous standards at the 2007 DARPA Urban Challenge, Nov. 3, in Victorville, Calif.

Team Cornell's "Skynet," a converted Chevy Tahoe named for the AI in the Terminator movies, was one of only 11 vehicles out 35 initial entries selected for the final test, where cars carried out three simulated military-supply missions in an urban setting. Five of those were eliminated during the first mission.

All six remaining cars performed amazingly, completing about 55 miles on city streets, merging with and passing each other as well as cars driven by actual people. But in the end, said Mark Campbell, associate professor of mechanical and aerospace engineering and one of the team's faculty advisers, it came down to time. "Our first mission ran quite well, but we had some throttle problems, which slowed us down the last two missions," he explained.

Carnegie Mellon, Stanford and Virginia Tech universities took first, second and third place, with prize money of $2 million, $1 million and $500,000. DARPA did not rank the remaining teams. The other finishers along with Cornell were from the Massachusetts Institute of Technology and a partnership of the University of Pennsylvania and Lehigh University.

The MIT and Cornell cars survived a collision as the MIT car attempted to pass Skynet and then turned into it. Skynet was moving in a start-and-stop manner, and it was difficult to determine fault, Campbell said.

But Campbell and two other faculty advisers found the Cornell team's achievement outstanding, noting that most of the other teams worked closely with well-financed industrial partners and had substantial faculty and senior researcher technical leadership.

"It was certainly a student-led project," said Ephrahim Garcia, associate professor of mechanical and aerospace engineering and another adviser. "We added some discipline in terms of software development, testing and some elements of design," he said. "Many of the key team members have been working on vehicle automation since the DARPA Grand Challenge in 2005, and it was the dedication of returning team members that made this success possible. Many students postponed graduation, taking a semester off, and one even quit his high-paying job in industry to come back and work for the team for little money. They are a zealous group of students who routinely worked 16 hour days, seven days a week. It has been a pleasure to watch then mature as engineers."

Dan Huttenlocher, the John P. and Rilla Neafsey Professor of Computing, Information Science and Business, agreed. "We took the view that we were advisers, not team leaders," he said. "This is one of the reasons I love being a faculty member at Cornell. We have students who are willing to take something like this on and drag faculty members into it."

The students also built the autonomous elements of the car pretty much from scratch, Campbell said, including designing and building a drive-by-wire system and a sensor network, in contrast to several other entrants who used off-the-shelf designs or custom designs from industrial partners. Also, he said, "we had a good [software] solution. There were some elements only we were doing in terms of tracking other vehicles." The software, he explained, fused the output of lasers, radar and vision to do "probabilistic tracking." The system not only pays attention to where other vehicles are, but tries to predict where they may be going and can deal with such things as a vehicle that is momentarily hidden from sight.

The Cornell AI drew compliments from the professional stunt drivers hired to drive the approximately 50 other cars that provided realistic traffic during the missions. "Several of them said our car drove more like a human than a robot," Huttenlocher recalled. "Very smooth and very predictable."

All this can still be improved, Campbell said, and he expects Skynet will be used for further research. The purpose of the DARPA challenge, he noted, was to encourage the development of technology that can be applied to future military vehicles, and based on their submitted proposals, DARPA provided up to $1 million in startup funds to each of 11 teams, including Cornell. The car was built to be a research platform, he said, more than just as an entry to win a competition. "We have to try to transfer some of this [technology] back to the government," Campbell said.

Media Contact

Media Relations Office