The pit crews behind DARPA's robot race

People in downtown Ithaca, N.Y., got a glimpse this spring of the vehicular equivalent of a headless horseman--a Chevy Tahoe gutted and modified with computers, wire controls and sensors so that it can drive city streets by itself.

Isaac Miller, a doctoral student in mechanical engineering at Cornell University, said bystanders had one of several responses when they saw the Chevy shuttled through town on the way to a remote testing site: fear, disgust or appreciation. Like many other teams, Cornell is preparing for the Defense Advanced Research Projects Agency's Urban Challenge, a $2 million military-sponsored race of autonomous vehicles on city roads that's set for November 3.

Then there's the response Miller calls "the joke," for "when the onlooker laughs, thinking we're crackpots who just threw a bunch of crap on our car for no reason."

"Not surprisingly, we get very few looks of appreciation and quite a good number of fearful glances. The latter are especially commonplace in Ithaca and the surrounding area, where there are a good number of closet hippies and environmentalists unwilling to believe the Vietnam War is over," Miller wrote on the team's blog.

Pedestrians' incredulity aside, there are 53 teams working on autonomous vehicles that see the Urban Challenge as anything but a joke. The month of June is crucial for contestants because DARPA is making the rounds for so-called site visits, prequalification meetings in which the military's research and development arm will evaluate each team's viability to compete.

That means that to proceed in the challenge, each team's robot must prove basic navigation skills by driving on a prescribed course, and demonstrate traffic skills by negotiating a four-way intersection with two human-driven cars and another robotic vehicle. The feat, as well as the final race, will require the robots to sense its surroundings, discern moving objects from static ones, read their positions and predict behaviors, among other abilities. As one team leader put it: "It's a very unpredictable scenario."

The third of DARPA's robot races since 2004, the Urban Challenge will easily garner more attention than past races. That's largely because when the Stanford University Racing Team took first prize in the 2005 Grand Challenge, a 132-mile Nevada desert race of autonomous vehicles, more people began to take notice and envision a future of robotic vehicles--both in and outside the military's domain.

So when DARPA raised the stakes with an urban race, more universities, including the Massachusetts Institute of Technology, joined the fray. Computer scientists and engineers seem to agree that building artificially intelligent cars to navigate city streets is a much harder challenge than what was previously on the table. And as in any competition, every team has a slightly different story. Here are three teams to watch:

The heavy hitter
After passing over the two previous DARPA competitions, MIT, an academic stalwart in computer science and robotics, will try to play catch-up to race veterans like Stanford this year, according to John Leonard, one of the team's leaders. MIT joins the challenge with support from the university, Quanta Computer, Draper Laboratory and Ford Motor, which donated a Land Rover LR3 to the cause.

"We sort of like to pose ourselves challenge problems, and we want to push the field...and try to envision 10 years from now."

--John Leonard, team lead, MIT team

Leonard, associate professor in mechanical engineering at MIT, said the school's group is different than others because instead of employing one visionary leader--such as robotics professor Red Whittaker, who's led the Carnegie Mellon University team throughout the previous two DARPA challenges--it draws on the expertise of various faculty for "distributed decision making."

For example, Dave Barrett, a professor at Olin College, is building the vehicle with a team of students. Seth Teller, a professor in the electrical-engineering group at MIT, is leading perception development for the robot. And Jonathan How, in MIT's Department of Aeronautics and Astronautics, is working on the car's planning and control system.

Leonard said the robot is still a work in progress but that the team is making obvious safe bets in some cases and taking risks in other areas. For example, it is using an Electronic Mobility Controls drive-by-wire system, which Team Gray proved reliable in the 2005 Grand Challenge effort. But he said the team is being aggressive with a novel perception strategy. He said the robot will deploy tens of sensors to help perceive surroundings and infer the state of the world, but he did not go further into the technology.

Rather, Leonard gave a higher-level perspective on the contest.

"We sort of like to pose ourselves challenge problems, and we want to push the field...and try to envision 10 years from now. In that context, we thought urban driving was it," Leonard said. "We're drawn to it precisely for how hard we think it is."

Leonard specifically wants to build a robot that can operate forever autonomously, capable of dealing with change in the world. From an algorithm perspective, he said, that means building sophisticated maps of a fluid world.

"We're tremendously advanced in mapping static environments, but there's been very little progress in mapping dynamic environments. To reason about them and make intelligent decisions about things that are moving in the world--that's the challenge."

The dropouts
New to the DARPA races this year, the Arizona team "A Bunch of Dropouts," is likely the only outfit running out of a remote 75-acre workshop powered by solar panels. It's also an easy bet that the Dropouts--named because the team is led by grown-up high-school dropouts--will have the one and only 1941 Dodge Power Wagon painted like the American flag in the race.

Kevin Jackson, leader of the team, said he tried to finish his robot for the 2005 Grand Challenge but couldn't get it ready in time. In contrast to teams such as Stanford's, which has up to 50 people working on the robot, Jackson said his team largely consists of "two crazy guys," including himself and J.R. Johnson, who handles the mechanical aspects of the vehicle. But the team's size doesn't dampen Jackson's optimism.

"Sometimes, small enterprising groups can do things that bigger organizations with more money won't think of," he said. "We're constrained by our size to do things very creatively."

The team is funded by Jackson, who runs Jackson Digital Imaging, which designs and manufactures digital-photography systems for theme parks. He got inspired to build a robotic vehicle after reading about how some of the best minds in the artificial-intelligence industry failed in the 2004 Grand Challenge.

"If Red (Whittaker) flopped in 7 miles, then I could do it," Jackson said after learning that the DARPA event isn't limited to big industry and big academia.

"Sometimes, small enterprising groups can do things that bigger organizations with more money won't think of."

--Kevin Jackson, team lead, A Bunch of Dropouts

So after the 2004 Challenge, he bought the 65-year-old wagon, stripped it down to the frame and outfitted it with a Chrysler V8 engine, custom off-road racing suspension and a striking paint job.

Jackson said the team is taking a different approach in terms of computer hardware and software too. For example, some other teams are using $10,000 precise Global Positioning System devices to track location, which strikes Jackson as "kind of cheating." His team is using two $120 GPS devices from Garmin, units that can be found in passenger cars. "We don't need superprecision if we do a good job with the rest of our sensors," he said.

The team is also using 6 Gigabit Ethernet progressive-scan digital cameras to do the car's visual sensing, or what Jackson equates to a "third guy riding shotgun watching the rear."

Jackson, who is doing all the computer programming, said the typical approach to artificial intelligence is writing many algorithms that solve a lot of little problems. That can lead to glitches evident in programs like Windows and Linux, he said. So his plan is to develop goal-based algorithms.

"It continuously adjusts its behaviors to its changing environment in order to achieve those goals," he said. For example, to execute a turn when the vehicle loses traction, the robot would detect that motion from its inertial measurement unit, among other sensors, so that it could correct itself and finish the task.

Jackson said he's also focused on the social psychology of driving just as much as the physics of it. "One of the things we're trying to do is to look at other vehicles driving, evaluate if they appear to be following the driving rules--for example, if that vehicle is weaving, it's an impaired driver. In the robot world, we can assign a threat level to that driver or object."

The one indulgence Jackson said he's given to the vehicle is a "super reliable safety-critical" $200,000 operating system from the team's lone sponsor.

Whatever the outcome for Jackson, the project seems like a labor of love.

"My normal job is cutting into my robotics project," he said.

The undergrads
Cornell's Miller is the only doctoral student on a team largely of undergraduates in computer science and mechanical engineering. Cornell competed in the 2005 Grand Challenge, but because of a blip in its GPS system, the team's robot ran into a bridge after 9 miles. Back for their second race, Cornell's team is rare because most of the original 15-member team is intact. One team member left his job to return for the 2007 race, and others have extended their schooling, Miller said.

"There's a lot of camaraderie," he said. "Each of us knows what he needs to complete."

The team's primary financial backer is DARPA, which has doled out $500,000 to the team, along with many of the other teams slated to compete. Cornell is also backed by TechKinetics, a Singapore-based defense company.

The team's bias, or where it's overcompensating, is in its probability algorithms surrounding the GPS system, thanks to its failure in the 2005 race. Miller said that from the team's experience, sensors can easily be wrong, and it now treats every input as a little bit of noise.

"With every input, we do a cost basis analysis to find what's the most efficient route and fastest route. Nothing's for certain, the way I perceive it," he said.

Instead of using a GPS device to tell the computer where the robot is, the team is using a GPS receiver from a European company to get raw measurements of distance by satellite, and then combining it with data from other sensors. "That's less prone to fail," he said.

For Miller, too, the challenge is about something bigger, given that he spends 12 to 16 hours a day working on the project. He's team leader on "sensor fusion."

"I really hate driving, and I wouldn't mind having a robot do it for me," he said. "More philosophically speaking, making cars safer and participating in that process is worthwhile."

That makes it worth all the strange looks.