As sophisticated as machines are today, they still cannot navigate an automobile through crowded city streets as well as experienced human drivers. But the Defense Advanced Research Projects Agency (DARPA) is working to change that with an eye toward sending automated robotic ground vehicles into battle to evacuate wounded soldiers, collect reconnaissance and carry out other dangerous missions.

This weekend DARPA, the U.S. Department of Defense's central research and development arm, will move closer to that goal when it hosts the final round of its 2007 DARPA Urban Challenge, a competition testing the driving prowess of experimental unmanned autos. The agency has whittled the field down from 89 to 11 teams of gearheads, scientists and students who will test their autonomous creations at the former George Air Force Base in Victorville, Calif. The winner will drive away with $2 million and those snagging second and third places will get $1 million and $500,000, respectively.

Unlike the DARPA Grand Challenges held in 2004 and 2005 in the Mojave Desert, this year's competition tests whether a vehicle—which must run entirely on its own using a system of sensors, global positioning systems and computers—can handle the types of driving conditions that city dwellers face every day, such as changing lanes, merging onto roadways with fast-moving traffic and traversing busy intersections. (Fortunately, they will not have to contend with iPod-wearing pedestrians, cell phone–gabbing drivers or unpredictable cabbies.)

Congress six years ago passed a measure mandating: "It shall be a goal of the Armed Forces to achieve the fielding of unmanned, remotely controlled technology such that…by 2015, one third of the operational ground combat vehicles are unmanned" with an eye toward one day handing over "dull, dirty or dangerous" tasks to machines.

The first Grand Challenge was held in March 2004 and featured a 142-mile (228.5-kilometer) desert course. Fifteen autonomous ground vehicles entered but nary a one made it to the finish line. The following year four autonomous vehicles successfully completed a 132-mile (212-kilometer) desert route within the required 10-hour limit. That year, "Stanley," a modified Stanford University Volkswagen Touareg R5 featuring six wireless laptop computers powered by Intel Pentium M processors running custom artificial intelligence software, claimed the top $2-million prize.

Despite the initial dearth of finishers (not to mention the absence of a competition last year as DARPA retooled the event), the agency is satisfied that the military is on the right track, DARPA public information officer Jan Walker says. But an autonomous vehicle must be able to traverse all types of terrain—from the city to the desert to the jungle to the mountains—to meet military demands. "The biggest difference with this year's competition," Walker says, "is that they have to be able to respond to a dynamic, unpredictable environment."

Despite the dramatic exploits of fictionalized robots in books and movies (who could forget the Terminator's skill behind the wheel?), artificial intelligence may not be up to the multitasking required to analyze the constantly changing conditions and other demands of driving. "The really interesting challenge in technology is building machines that do the types of things that people do so you can put [these machines] deep under the ocean or in outer space," says Richard Murray, a professor of control and dynamical systems at the California Institute of Technology's Division of Engineering and Applied Science. Murray's Caltech team entered the DARPA competition with a 2004 Ford E-350 van named Alice, whose artificial intelligence was created by the joint effort of 80 undergraduate and graduate students, postdoctoral candidates and faculty as well as engineers from NASA's Jet Propulsion Laboratory in Pasadena, Calif.

Alas, despite some early success, Alice failed to advance to this weekend's final round. Equipped with eight cameras, two radar devices and eight laser range finders that measure distance by emitting infrared pulses, Alice nonetheless had trouble stopping and going at intersections and merging into traffic to DARPA's satisfaction. In one instance, Alice stopped at an intersection, advanced when the road was clear of traffic, but slowed down again because she mistakenly perceived a concrete guardrail to be an obstacle. "We would stop and not go and get honked at," Murray says. Alice at times also failed to identify the line in the road signaling her to stop at intersections and either stopped too late or not at all. (Although that may be good enough to get her a gig as a taxi in New York City.)

Alice has since been shipped back to Pasadena so that Murray and his team can study her computer's data logs and better understand her behavior. The van contains about a dozen computers running the open-source Linux operating system and a number of open-source software applications designed to process information gathered by the cameras, radar and laser range finders. Students plan to continue using Alice to test their research into the development of autonomous vehicles.

"Driving is a great example of something that humans [can] do really well that requires the need to think about a number of problems simultaneously, like working the pedals, steering the wheel and watching the road, all while thinking about where you're going," Murray says. "The DARPA challenges have really pushed the state of the art in that area." DARPA wants artificial intelligence that can measure a situation and make wise decisions based on the information it collects. "Right now we have people doing that," he says, "because we don't know how to make machines do it."