Now comes the hard part: a race on mock city streets that will raise the bar for artificial intelligence in the 21st century.
A team of officials from the Defense Advanced Research Projects Agency (DARPA) visited a parking lot here next to Google headquarters to test Stanford University's autonomous passenger car,, in what was its first big qualifying test for the upcoming Urban Challenge, DARPA's third Grand Challenge competition for driverless vehicles.
DARPA will make so-called "site visits" this summer to evaluate all 53 prospective Urban Challenge contestants, homing in on whether the robots can perform basic driving skills, including navigating a four-way stop with live traffic, passing a stationary car and executing a U-turn.
"It's a steep ladder to get up to the Urban Challenge. What you saw today was the first rung of the ladder," Norm Whitaker, program manager for DARPA, said to a crowd of people following a two-and-a-half-hour test of Stanford Racing Team's Junior.
Driving just 15 miles per hour, Junior passed with flying colors in three of four "missions" on a parking-lot course, including a three-point U-turn and successful navigation of a four-way stop with human-driver cars. But it got stuck on a course passing a car.
"Stanford had a little bit of a hiccup, then recovered," Whitaker added.
said: "We have reason to be optimistic for the next round, and that we'll be in the top 30 (finalists)."
Stanford Racing Team is notable because it, a $2 million robot race across 132 miles of Nevada desert. At the time, Stanford was a new entrant to the DARPA Grand Challenges--which first began in 2004--and kept a lower profile than teams organized by institutions such as Carnegie Mellon University, a pioneer in robotics that entered two cars in the race. But Stanford's robot, Stanley, finished the course in the fastest time--under 10 hours--making it the first to accomplish DARPA's mission for autonomous cars and set a new standard in artificial intelligence.
Now the Stanford team has drawn widespread attention and sponsors such as Google, Red Bull and Intel. A Google representative at the test site said that Google co-founder Larry Page called him up last year and said that "he wanted to sponsor the best teams."
The DARPA Urban Challenge, scheduled for November 3, will be highly competitive, however. In the 2005 race, the robots' primarily needed to detect static objects like rocks and then maneuver around them while navigating the terrain. But in the Urban Challenge--designed to replicate the challenges of navigating an urban environment--the robots will need to discern between stationary and moving objects, and then predict their behaviors. Mike Montemerlo, senior research engineer at Stanford University's Artificial Intelligence Lab and a team leader, said that the robot also must factor in how different objects, such as a car or a bicycle, behave. That's on top of obeying simple traffic laws.
"To drive well, the robot has to have a more sophisticated understanding of the world," Montemerlo said during Thursday's site visit. "That's really the holy grail."
Knowing the rules, and when to break them
"Driving is about rules, but also knowing when you can break the rules," he added.
Although the site visit tested the robot's basic understanding of traffic and navigation, the race itself will prove much harder, according to DARPA officials. Competitors' vehicles will have to drive on streets populated by multiple cars, including other robots and human-driven vehicles. They'll have to merge with traffic, pass moving vehicles and navigate a parking lot, among other feats. Some engineers following the race expect that it could be so hard that it might produce results to those of DARPA's first Grand Challenge, in which all the contestants flamed out.
But DARPA is investing heavily in this challenge to produce "synergistic effects" among many universities, corporations and individuals that are ultimately working toward the same goal, Whitaker said. To foster development, and in acknowledgment of the complexity of the challenge, DARPA has doled out $1 million grants to 11 "track A" teams, including Stanford, Montemerlo said. DARPA's ultimate goal in the challenges is to develop artificial intelligence for use in vehicles on the battlefield.
But the Stanford team is highly focused on advancing AI for consumer cars. More than 40,000 people die in vehicular accidents every year, and Thrun's team believes that if it can build effective autonomous controls for cars, it could significantly reduce the number of accidents. Montemerlo speculated that we could see real autonomous vehicles on the road in 25 years.
Junior's sponsors, including Volkswagen of America and NXP Semiconductors, are also keenly interested in new safety applications that emerge from race development. "There's a fertile ground to break. But spinning off applications (like adaptive cruise control) is more of an incremental approach than a quantum leap," said Mark Hoffberg, senior director of alliances at NXP, which conducts research for the auto industry.
For this race, Junior, a modified 2006 VW Passat, has more-advanced sensors that can "see" the world in a 360-degree view and process that data in as close to real time as possible. Unlike its predecessor, Junior has a new, high-definition "lidar" (light detection and ranging) system, which measures the range to a target in much the same way radar does.
The system, which was developed by high-definition lidar specialist Velodyne, spins around to give the robot an omnidirectional view of its surroundings. The vehicle also has a Ladybug 2 video system, from Point Grey Research, with six video cameras to capture near-high-definition video in all directions. In all, Junior has about eight external sensors, including two white GPS saucers on top.
Junior's software is also brainier than Stanley's. Junior has software components that deal with perception and decision-making, for example. One algorithm Stanford's AI lab has developed is for object tracking, which helps the robot understand when it sees a bike, car, curb, road markings or other moving objects. The algorithm will classify objects--such as a car moving 10 mph--and run that through a planning tool that can match the data to rules of the road in order to make a decision about how to proceed.
Junior's software runs on two quad-core servers from Intel, which is roughly four times the computing power of Stanley. With 2GB of memory each, one server is dedicated to processing data from the range finder and detecting obstacles. The other is devoted to the mechanics and navigation of the car, including following instructions for driving the course. (DARPA gave Stanford a USB drive with the robot's mission instructions for its qualification round, which will be similar to how the Urban Challenge race is handled.)
Since Stanford won the 2005 Grand Challenge, DARPA-sponsored races have become much higher-profile, and this year's competitors are stronger as a result. New contestants include a handful of defense contractors and universities such as MIT, known for its AI department. AI stalwarts like University of California at Berkeley, Georgia Institute of Technology and Carnegie Mellon will also be among the contenders.
"The common goal of the challenge has synergistic effects even among all the universities," Whitaker said.