X

Caltech's spin on DARPA's robot race

In a race of driverless vehicles--and especially in collisions between them--size matters. Photos: 'Alice' in DARPA-land

Stefanie Olsen Staff writer, CNET News
Stefanie Olsen covers technology and science.
Stefanie Olsen
3 min read
PASADENA, Calif.--Even in a collision between robot cars, size might make all the difference.

That's the thinking of Team Caltech, a robotics group at the California Institute of Technology that is fine-tuning a large self-driving van, an 8,000-pound Ford E-350 reinforced with armor plating, for next month's semifinals of the Urban Grand Challenge. The government research arm known as the Defense Advanced Research Projects Agency invited 35 teams to compete in its race of driverless vehicles, with a $2 million first-place price.

According to one of the road rules, if a team's robot hits another vehicle--which seems highly possible for a newbie machine driver--that team is automatically disqualified from the competition. But if a car is the victim of a collision, that car's team has 30 minutes to fix the robot and continue in the race. If the damage is too great, it's out.

Photos: Alice in DARPA-land

"We're one of the biggest robots in the race. We think we'd be a physical victor in any crash," said Joel Burdick, professor of mechanical engineering at the California Institute of Technology, and Team Caltech's co-leader.

Of course, physical hardware is just one consideration in a competition that's ultimately about testing artificial intelligence software on the road. Teams must program software that can synthesize and understand data coming in from multiple external sensors on the car--including radar, ladar, cameras and GPS. Then the software must plot a drivable path on city roads, avoid obstacles in real time, and obey traffic laws. That's a mission even people have a hard time with.

As a result, Caltech is among nearly three dozen teams furiously testing and tweaking their robots' software before the race semifinals on October 26, to be held in Victorville, Calif., at the site of the former George Air Force Base. After the qualifying event, DARPA will hold the finals at the same site on November 3.

Caltech has competed in all three challenges, but this year the largely student-run team has much more money and support. In 2005, Caltech spent roughly $200,000 building its robot, "Alice," including travel expenses. And despite qualifying for the finals, Alice had a hardware failure early in the race and nearly ran into the media viewing section off course. For the Urban Challenge, DARPA, in acknowledgement of the complexity of the race, offered 11 track "A" teams $1 million in grant money as a measure of support, and Caltech was one of them.

The team repurposed Alice for the Urban Challenge. But according to Burdick, only about 15 percent of the original software remains; the team has written mostly new code for this year's race. The software has three parts: a mission planner that takes in data from DARPA to sequence the course of the race; a traffic planner that gets data from sensing technology on factors like the speed of other cars on the road; and a trajectory planner to execute on the above data. Burdick said that in the coming weeks, the team is largely trying to iron out the bugs in its software, such as any small errors in programming or logical flaws, while taking Alice on test drives.

The Caltech team is comprised of about seven full-time graduate students, one post-doctoral student and two faculty members. Team members come from several departments, including the computer science and mechanical engineering departments. And they have received help on the computer vision system from Caltech's Jet Propulsion Lab, NASA's research center. Richard Murray, the team's three-time leader, is a professor in Caltech's control and dynamical systems department, which specializes in control theory.

Burdick said he believes the Caltech team will excel because of its expertise in control theory, a branch of engineering and mathematics that studies the behavior of dynamic systems. "This is a large-scale computing problem. We're going to have some unique, more structured approaches to planning algorithms" that involve a hybrid control theory, he said.