THIRTY-FIVE TEAMS started the DARPA Urban Challenge, a competition to create a driverless car to successfully navigate an urban landscape.

Cornell was one of only six to complete the trial. Mark Campbell, associate professor of mechanical and aerospace engineering and the team's faculty advisor, thinks Cornell was the only team for which most of the technical work was done by students – about 15 of them, led by PhD student Isaac Miller.

DARPA provided $1 million in funding. Some teams had as many as 40 or 50 people and considerable industry involvement. The biggest challenge, he says, wasn't automating the vehicle. Most of those problems were solved in DARPA's 2005 Grand Challenge, several veterans of which joined this effort. (Cornell's 2005 vehicle got its nose in a guard rail and couldn't reverse.)

Cornell's SUV included 17 rack-mounted computers, home-built out of off-the-shelf components including mobile processors. "We sacrificed a little bit of performance, but they didn't generate a lot of heat, so we could use the car's air conditioning. "

The computers ran on a streamlined version of Windows XP. "We had the car running for a year plus, and they never crashed," Campbell said. The two hardest problems they had to solve were first, tracking vehicles, and second, locating their own car in the lanes. The first included differentiating cars from bushes and rocks, identifying their lane and behaviours, and telling whether they were stopped. The second was difficult because of limitations in GPS. "Even if it's high-precision it might be off by half a meter."

The lanes' sizes varied and parked cars at the sides could start moving. The car's GPS was cued by two cameras, one mounted on the roof looking down and the other inside the radiator grille. The first of these looked for white lines: right, left, and stop. "Those were surveyed to the centimeter. Once you came up to a stop line, you knew almost exactly where you were on the map. It reinitialised every time it came up to one of those so we knew we could rely on it. Much more accurate than GPS."

The team tried using cameras to help with recognising and tracking other cars, but decided the sensors weren't good enough. With the Challenge completed, "We have a wonderful vehicle testbed for research that would have been difficult to build otherwise. It's a good catalyst for us." Over the next couple of years, Campbell's research group hopes to tackle problems they set aside while working on the competition, such as how best to use vision, and how to make decisions.

"It wasn't probablistic," he said of the car. "It didn't evaluate options and pick the one that made the most sense. It was hard-coded for different situations." One thorny problem, he said, surfaced two weeks before the competition. "We had a throttle problem we still haven't quite figured out." This problem only surfaced after more than 20 miles of autonomous driving: the car would start to slow down.

The upshot was that in the second half of the final race the car slowed from 25 miles an hour to five to ten. That problem cost Cornell the chance to win the final race, which unexpectedly became a time trial. Up until then, the cars had been graded by 100 officials around the course watching for mistakes at intersections, in parking, lane behaviour, and so on.

There were, Campbell says, no major incidents in the final, so the final standings were all done by time. "One of the things we disliked about time was that it took away from what the cars are really doing," he said.

"If you added up how many times we merged into traffic over seven hours of finals it was at leat 100 just for our car. It merged by itself, so it recognised what cars were there, tracked them, and pulled out appropriately at intersections. That's amazing. By just focusing on time they de-emphasised that. " µ