Print View - Return to Normal View
No Drivers Wanted
By Anne Burke
Published Jan 1, 2006 12:00 AM
Copyright ©2006 Amanda Friedman
Photos, charts, video, audio, background stories.
Golem 2, a Dodge Ram pickup truck and UCLA's entry in the Pentagon's robotic-vehicle race, is tearing through the Mojave Desert like a thing possessed. Bam! She mows down a yucca tree. Whoa! She bounds into the air. The front grille grabs a creosote bush. Rocks ping against her sides. The taillights spring off and the tailgate, hanging by a cable, flails wildly.
Ed Robinson, a seasoned off-road racer from Sacramento, approaches from behind in a chase vehicle. Riding shotgun, his thumb and index finger lightly grasping an emergency-stop switch, is Brian Hearing, who works for the Defense Advanced Research Projects Agency (DARPA), the race’s sponsor. Bouncing in the back seat, manning the radios, is Greg Duckworth, who also works at DARPA. Even with his many years off-road driving, Robinson is a little weirded out. “Have you ever followed a vehicle with nobody driving it?” he asks.
HIGH STAKES IN THE HIGH DESERT
Richard Mason, Jim Radford, Eagle Jones, Jason Meltzer and a dozen others who have spent the last three years preparing for this day are about 10 miles west in Primm, the Nevada gambling town that straddles Interstate 15 just over the state line. The DARPA Grand Challenge started here at 6:40 a.m. on Oct. 8, when the first of 23 “bots” lumbered across the start line in pursuit of $2 million in prize money and the glory of winning a race that had captured the world’s imagination.
Golem 2 was seventh out of the gate at 7:10 a.m., right behind “Dad, Are We There Yet?,” a modified Toyota Tundra designed by Morgan Hill, Calif., scientists, and ahead of “Spider,” a military, light-strike vehicle from Cornell University. It’s now 8:20 a.m. and most of the Golem Group, as these UCLA-based roboticists call themselves, are sitting around a table in the spectator’s tent, watching with gathering confidence a big scoreboard that shows Golem 2’s progress as a horizontal bar creeping rightward. There is no camera feed, so no one knows the first thing about the truck’s wild romp. As far as they know, this four-wheeled robot that they have loved, cursed and sweated over has been behaving exactly as they hoped.
Late the previous night, Mason and Radford, the team’s leaders, had decided to pursue a somewhat risky “throttle up” strategy. The Grand Challenge was not a race in the pedal-to-the-metal, NASCAR sense. At a qualifying event at Fontana Speedway the previous week, 35 mph was considered double fast. Nevertheless, Mason and Radford had come to believe that the race would go to the swift.
The exact route of the Grand Challenge was kept secret until 4:10 a.m. on race day, when DARPA officials handed each team a CD containing global positioning system waypoints. These latitude and longitude markers act like bread crumbs to guide the bots along the 131.6-mile course. With the exception of the treacherous Beer Bottle Pass through the mountains, the course was mostly a bunch of straight lines between Primm and Jean, about 12 miles north on I-15. Huddled over a laptop computer in the aging RV that was command central for the Golem Group/UCLA, Mason and Radford, unshaven and fueled by Coca Cola, made a critical decision: Once Golem 2 is out of the chute and gets her bearings, she would be under orders to go 50 mph.
DIVINE INSPIRATION FOR A ROBOT RACE
The name Golem comes from Jewish legend. In 16th-century Germany, God commanded a rabbi to fashion out of clay an automaton, or golem, that would protect Israel against her enemies. As it happened, this golem could see and understand its environment in a robot- like way, and so performed good deeds for the Jews. In 2000, the U.S. Congress, looking to save American lives in war zones, passed a law ordering that one-third of all ground combat vehicles be unmanned by 2015. As any insomniac who has watched the TV infomercial for the Roomba robotic vacuum cleaner can attest, robots left the purview of science fiction a long time ago. Robots, some partly autonomous and others remote-controlled, are disabling bombs in Iraq and Afghanistan, exploring the surface of Mars, probing the seabed around Japan, and sucking up cat hair from your neighbor’s carpet.
But building a totally self-operating vehicle that could traverse difficult terrain in wartime had confounded even the nation’s top defense contractors. The problem, explains Stefano Soatto, a computer science professor who runs the UCLA Vision Lab and led the UCLA branch of the Golem Group, is far more difficult than the public is given to understand. (If it were easy, we’d all be reading the paper while stuck on the 405.) An autonomous vehicle must be capable of deciding future action based on sensory data drawn from an unfamiliar and changing environment.
Say you’re driving on Interstate 5 and a tumbleweed blows into your lane. Most of us would correctly conclude that the tumbleweed poses little danger. But say the tumbleweed is in the path of a robotic vehicle equipped with a laser beam that sweeps a 90-degree arc. If the beam bounced against a branch, it would alert the robot to trouble ahead. But if the beam passed through spaces in the tumbleweed, it might send the all-clear signal. Even for humans, Soatto notes, processing sensory information is so difficult that “half the brain is devoted to this task.”
ASSEMBLING THE TEAM
With research into autonomous war vehicles moving at a sluggish pace, DARPA cast a wider net. In January 2003, the agency put out a call to roboticists, university researchers, artificial intelligence developers, video-game makers, garage tinkerers and pipe dreamers to enter the DARPA Grand Challenge in March 2004. The bot that most quickly completed a 142-mile course within 10 hours, from Barstow, Calif., to Las Vegas, would win $1 million.
Mason, an ’03 Caltech Ph.D. in mechanical engineering, heard about the Grand Challenge at his job at the RAND Corporation, where he researches unmanned ground and aerial vehicles. In his spare time, Mason is a quiz bowler, sometime actor, and blogger who enjoys pondering brain teasers like, “If I get in a helicopter and go straight up and just hover there for 12 hours, will Icome down on the other side of the world?” In 2002, Mason won $52,000 on Jeopardy! While his wife, Maribeth, had in mind a down payment on a house, she did not begrudge her husband instead spending the money on a half-ton Ford 150 pickup truck for the Grand Challenge. To help roboticize the truck Mason turned to Radford, his classmate at Caltech and also a Ph.D. in mechanical engineering. Working at the time for a Santa Ana company called Indigita Corporation, Radford was a crackerjack at developing and analyzing the computer algorithms that control and coordinate robotic motion.
In order for a robot to drive on its own, it needs to see. To give their truck eyes, Mason and Radford, now calling themselves the Golem Group, thought of Jason Meltzer, whom they had known at Caltech. Meltzer, from Long Island, N.Y., had done his undergraduate work at Caltech but for his doctoral studies, he came to UCLA to work under Soatto in the Vision Lab, which was quickly establishing itself as one of the best in the country.
The Italian-born Soatto has Caltech credentials himself, having done his Ph.D. there. In 2000, after teaching at Washington University and Harvard, Soatto came to UCLA and launched the Vision Lab. Aided by about 13 students and four postdocs, Soatto investigates a field called dynamic vision, which involves a computer’s ability to take visual data about a changing environment and apply it to the performance of tasks. Soatto sees dynamic vision as one of technology’s most exciting frontiers, with the potential to improve nearly every aspect of our lives. Already, dynamic vision is used in Japanese automobiles to detect a pedestrian stepping off a sidewalk, and by freightliner trucks to sound an alarm when a driver starts to nod off. “Think of everything you do with your eyes and then think of a machine being able to do that,” Soatto says. “The potential impact is humongous!”
Meltzer mentioned the Golem Group to Eagle Jones, a fellow Caltech grad who had also come to UCLA to pursue a Ph.D. under Soatto. Raised in Oregon and California, Jones, a rangy 6-foot-5, early on developed a gargantuan appetite for life. While still in high school, he programmed a computer game that enjoyed modest commercial success. At Caltech, he was an NCAA All-American in water polo, set two swimming records that stand today, and earned a pilot’s license. After coming to UCLA, he completed his first Ironman event.
In the way that one might put together a team for a bank heist, Mason and Radford rounded up another dozen or so friends and friends of friends. Maribeth Mason recruited Jim Swenson, Brent Morgan and Jerry Fuller from The Aerospace Corporation, where she works. Robb Walters and Deepak Kumar joined from Caltech, where they are pursuing Ph.D.s. Josh Arensberg signed up from the film industry. Also on board were Radford’s boss at Indigita, Bill Caldwell, and Caldwell’s son, David. Having reached critical mass, what the Golem Group needed was backing from a big university.
Jones approached Soatto, whose answer was an unequivocal no. The Grand Challenge, Soatto warned, was more like Mission Impossible. The course was too long and the time limit too restrictive. “Nobody was going to be able to make it under the conditions DARPA set out,” he said.
Soatto called it right. None of the 15 bots completed even 10 miles in the first race in 2004. The top finisher was Carnegie Mellon University, whose Humvee veered off course at 7.4 miles, despite the best efforts by CMU’s venerated roboticist, William ” Whittaker. Other bots crashed, flipped or became disoriented and wandered off course. Though the next day’s headlines rang of mockery, the media found an inspirational, little-guy story in the Golem Group’s truck. With no big sponsors and operating only on GPS, Golem 1 completed 5.2 miles for a fourth-place finish. Even the team was shocked. “We were, by all accounts, the lowest budget team there and sort of the only true garage team,” Jones recalled. “There were just no expectations for us.”
Watching the race on the Internet from his Pasadena home, Soatto was “ecstatic.” When the team announced they would return for the 2005 Grand Challenge, Soatto was in, as was Assistant Professor Emilio Frazzoli, a roboticist who specializes in unmanned aircraft and would write the navigational software for Golem 2. The two professors contributed the bulk of their discretionary funds to the effort. Combined with money from the Henry Samueli School of Engineering and Applied Science and Vice Chancellor for Research Roberto Peccei, the Golem Group/UCLA, as the team was officially known, had a war chest of $260,000 cash. The team invited all comers to help out. Kerry Connor ’03 and Brian Fulkerson, another Ph.D. student from the Vision Lab, signed on.
Golem 2 would rely on a far more sophisticated array of visual sensors than its predecessor. Attached to a metal frame over the front grille were five so-called LADAR sensors from the German manufacturer SICK. Each scanned a field of 90 or 180 degrees using a pulsed laser beam that would reflect back to the sensor if it collided with an object. A small camera from the Israeli company MobilEye would find the edges of the road from its mount on the hood. While some of the 23 entries in the 2005 Grand Challenge carried computer systems the size of refrigerators, Golem 2 relied on a single laptop computer strapped with bungee cords to the front seat of the Dodge Ram.
The race started behind Buffalo Bill’s casino, just as daylight broke over the desert. The grandstands were packed. The favorites were Carnegie Mellon, back this year with two entries, a customized Hummer called H1ghlander and a red Humvee called Sandstorm, and Stanford’s Stanley, a Volkswagen Touareg, which enjoyed generous financial and technical help from Volkswagen of America.
Soatto and Frazzoli watched Golem 2 roll out of the gate from a berm behind the start line. She faltered momentarily, steering toward a concrete barrier. Soatto’s jaw tightened, then relaxed, as the truck found her bearings and rounded a curve, picking up speed along the way. Farther into the race, Golem 2 was driving like an aggressive teen-ager - hard on the pedal. On a straight shot along a dry lakebed, she clocked 47 mph, kicking up a cloud of dust so thick that Robinson had to let the chase vehicle drop back.
Then, at 22 miles, still moving at a fast clip, the truck did something that no one could have anticipated. She veered off-course and headed northwest into open desert. A wide-eyed Robinson followed about 25 feet behind, off to the side. Hearing, seated beside him, frantically flipped a “pause” switch that should have temporarily stopped Golem 2, but it had no effect.
After about a quarter mile, just as Hearing was about to flip the emergency-stop switch that would have put her out of the race, Golem 2 came to a dead halt. Robinson parked the chase vehicle and he, Hearing and Duckworth got out, eyeing the motionless robot warily. “I walked over and touched the tailpipe and there was no exhaust coming out, so we knew the engine had died,” Robinson said.
About 6 p.m., Golem 2, wearing a moustache of brush on her front grille, was towed back to the parking lot behind Buffalo Bill’s. Mason, Radford, Jones and the rest of the team circled round her, touching their fingers to her sides. Opening the hood, they could tell why she had come to a sudden stop: The battery case dislodged and knocked a fuse box loose. Jones unfastened the laptop from the front seat and carried it to Radford’s hotel room at the casino, where he combed through log files.
The problem was easy to spot. At mile 22, the computer ran out of memory and crashed. Controls remained fixed at “steering angle 2 degrees left” and “accelerate hard.” Because the software wasn’t running, the pause switch that Hearing was flipping didn’t work. The memory problem, which amounted to a tiny misallocation, “would have come out in testing if we had done a long, 200-mile run,” Mason remarked, but as it was, the team didn’t have time.
As fast as she was moving, Mason thinks Golem 2 was on course to beat Stanley, which won the race in 6 hours, 53 minutes and 8 seconds. Golem 2 finished 14th, ahead of Caltech’s Alice, which went berserk at 8 miles and crashed into a row of concrete barriers.
Soatto says the team has nothing to feel bad about. Stanford, CMU and some of the other DARPA competitors “have very, very well established robotics programs. We had to essentially create a robotics lab from scratch,” he says.
Jones had hoped for a finish in the top 5 but, all in all, he thinks the team should be proud. He’s working on his thesis now but expects he’ll change the topic to something more closely related to what he learned working on the Grand Challenge. “More than anything,” he says, “this has made me more focused.”
And after he gets his doctorate from UCLA? Jones thinks he’ll start a robotics company.