Team MIT Squeaks Into Robot Car Finals
It wasn’t a shoo-in, according to MIT professor of aeronautics and astronautics Jonathan How. But Team MIT learned this afternoon that it has won a qualifying berth in the DARPA Urban Challenge finals on Saturday, when observers will learn which competing institution’s autonomous vehicle is best at navigating a complex mock-city environment replete with moving traffic, busy four-way intersections, and numerous obstacles.
A surprising number of prominent teams, including Team Caltech, Team Princeton, and Axion Racing, were definitively eliminated after qualifying events held at George Air Force Base in Victorville, CA, from October 26 through 31. And while MIT wasn’t one of the teams eliminated, it also wasn’t among a group of six teams named by DARPA as qualifiers on Wednesday, creating considerable suspense as to whether the MIT “supercomputer on wheels”—a Land Rover loaded up with sensors, servers, and software by a multidisciplinary team of faculty and students from MIT and Olin College—would be allowed to compete for the $2 million grand prize on Saturday.
But officials from the Defense Advanced Research Projects Agency, sponsors of this year’s challenge as well as two previous robot-car races in 2004 and 2005, released a list of five more qualifying teams this afternoon, including MIT. In the end, 24 of the 35 teams who traveled with their robot cars to Victorville were eliminated from the competition. MIT was the only team from New England to make the finals (or even the semi-finals).
How, who spoke with me from his office at MIT before boarding a plane for California, says he listened in on the announcement remotely as a colleague in Victorville held up a cell phone. “It’s nice to get this far,” How says. “There are good friends of mine who did not, and it would have been tough to have done this amount of work and not get through. I really feel bad for those guys. But I’m just hoping we can go out there and put on a good show so that MIT can be proud.”
How says the main point of the qualifying events was “to make sure that every vehicle in the race is sufficiently safe that they can just leave the robot driving around.” That’s probably a good thing, given that the environment on Saturday will be more complex than anything any of the teams have dealt with before. All 11 robot cars will be on the course simultaneously, along with as many as 80 other vehicles, including chase cars and spectator vehicles. “It’s going to be a lot of fun to see what happens,” How says.
According to How, Team MIT didn’t make it into the first group of qualifiers because the Land Rover had trouble with one particular event, a loop designed to test how well the robot cars could merge into moving traffic. “You’re facing two lanes of oncoming traffic,” explains How. “You turn left across that traffic, merge with moving traffic, drive around a loop, cut across traffic again, join moving traffic again, and so on. The question was how many laps could you do in 20 minutes.” Tartan Racing—one of the first six qualifiers—completed 16 loops in the given time. Team MIT completed six or seven. (Stanford’s team, the defending champions from the 2005 Grand Challenge, did well in the qualifing events and was on the list of six teams announced Wednesday.)
Every team had two chances at the merging-traffic event, and some were able to fine-tune their planning and navigation software sufficiently to improve on the second go-round. Others weren’t. “The traffic was only moving at 10 miles per hour, so a human would probably have been okay with it, but it is a hard problem” for robots, How remarks.
Past DARPA challenges simply involved getting a robot car from the start line outside Los Angeles to a finish line near Las Vegas. Saturday’s finals, where all team members will be forced to look on from spectator stands at the edge of a suburb-like test course, “will be quite a bit different,” How says. “We’ll get to see [the cars] over an extended time. And once the robots start meeting up, then it really becomes stochastic”—a big word for completely chaotic.
How’s own specialty, as I wrote after visiting MIT’s Urban Challenge team in early September, is developing path-planning algorithms like those used by unmanned aerial vehicles and deep space probes to cross unknown territory safely. And that, as it turns out, is the exact problem the robot cars will face on Saturday. “Say there’s a robot blocking the road. Do you sit there and wait? What if he is waiting for you? Do you move back? You need a very general strategy that’s able to handle many different scenarios,” How says. How well MIT’s car handles such unpredictable situations will be evident in a couple of days—when the event will be webcast live at http://www.darpa.mil/grandchallenge/.
UPDATE 11/2/07 2:40 p.m.: DARPA has released the list of all 11 teams that have advanced to Saturday’s finals.