Self Driven Cars in Our Future
http://www.sciencemag.org/cgi/content/full/318/5853/1060
Robotic Cars Tackle Crosstown Traffic--and Not One Another
Adrian Cho
In DARPA's Urban Challenge, cars that drive themselves face off in a strange, soulless rush hour. Are
human drivers about to go the way of the buggy whip?
VICTORVILLE, CALIFORNIA--The Land Rover bristles with sensors like a mechanical porcupine. John
Leonard, an engineer at the Massachusetts Institute of Technology (MIT) in Cambridge, ticks off the
robot's features. On the roof spins a conical laser range finder called a lidar that sees in three
dimensions. A dozen lidars that see in one direction, 15 radars, and six digital cameras look out every
which way. Computers fill the back of the truck, and a generator supplies the 3.5 kilowatts of power they
need. It's impressive. But all this so the truck can turn left across traffic by itself?
The robot is one of nearly three dozen vying in the Urban Challenge, a competition sponsored by the U.
S. Defense Advanced Research Projects Agency (DARPA). It's the third and most demanding in a
series that aims to spur the development of autonomous vehicles, which the U.S. military hopes to press
into service by 2015. In 2004 and 2005, robots raced one at a time across open terrain. This time, they
must navigate the streets of an abandoned air base here in the Mojave Desert without colliding with one
another or with human-guided vehicles.
The competition showcases some of the world's best talent in robotics. "We were drawn to the Urban
Challenge because it requires real-time decision-making in a dynamic environment and in the presence
of uncertainty," Leonard says. It also serves the higher purpose of trying to save lives, as worldwide, 1.2
million people die each year in traffic accidents that robotic cars might help avoid.
And yet the Urban Challenge is at least slightly absurd. It looks a bit like a real race. Engineers wear
bright shirts emblazoned with the logos of sponsors--GM, Ford, Intel, Google. Teams have hauled in
tractor trailers full of equipment and plastered their robots with decals. Besides the $2 million first prize,
the appeal of the challenge is obvious. It's hard, and by pitting idea against idea and technology against
technology, "it determines what technical DNA moves to the next generation" in the evolution of
autonomous vehicles, says roboticist William "Red" Whittaker of Carnegie Mellon University in
Pittsburgh, Pennsylvania.
Still, the task the robots will attempt seems so ordinary. They must obey the California traffic laws
(although if two collide, they won't have to exchange insurance information as human drivers are
required to do). We're all here to watch traffic. But we won't see with our own eyes. Instead, we'll have to
watch it on television. It's not even clear what DARPA gets out of this well-crafted media circus. The
competition is meant to stimulate the development of cars that drive themselves--and it has--but DARPA
does not require winners to reveal to the agency the details of their technologies.
Comeback kid. After a delayed start, Carnegie Mellon's Boss cruises briskly to a victory.
CREDIT: A. CHO/SCIENCE
Go ahead, bend the rules
At the decaying fighter base, across the road from the new federal prison here in Victorville, 11 teams
have made the final competition. Three years ago, not one robot traversed more than a dozen
kilometers of the 230-kilometer off-road course. A year later, four completed a similar course. And this
year's robots are far more capable than last year's crop, possessing better sensors, more powerful
computers, and, most important, more sophisticated programming. "Driving is a software problem, not a
hardware problem," says engineer Michael Montemerlo of Stanford University in Palo Alto, California. "At
Stanford, we can't build a better car, but we can make a smarter car."
Computationally, this year's challenge is much more difficult than the first two, researchers say. In the
desert races, the robots had only to identify obstacles in a static landscape and plot a safe path around
them. This time, the vehicles will have to avoid other cars, including other robots, while at the same time
obeying the relatively arbitrary traffic laws. To do that, each robot's computer must calculate the likely
trajectories of all the objects around it and plan to miss them. Of course, a robot cannot know exactly
where another car will go, so the machines generally employ layers of probabilistic algorithms to decide
their next moves.
MIT has decked its robot, Talos, with the most sensors. The radars see distant objects, the lidars see at
an intermediate range, and the cameras spot things close by, explains David Barrett, a team member
from the Franklin W. Olin College of Engineering in Needham, Massachusetts. Talos depends mainly on
its sensors to navigate, Barrett says. That's because the team assumed, incorrectly it turns out, that
DARPA would not let robots use signals from the satellite-based Global Positioning System (GPS) all
over the course.
Researchers from Stanford, who won the 2005 competition, say that they focus on the algorithms
programmed into their robot. Merely encoding the traffic laws can leave the robot stymied, says
Stanford computer scientist Sebastian Thrun. For example, when two robots arrive at a four-way stop
simultaneously, each may try to yield to the other endlessly. To avoid such deadlock, the team lets its
robot skirt the laws. "Our car has a hierarchy it follows," Thrun says. "At the top, it obeys strict rules.
And if it gets stuck, it ignores more and more rules." Fair enough. Why expect more from a robot than a
human?
Team AnnieWAY, one of two German squads in the final, has taken a minimalist approach to guiding its
Volkswagen Passat. The team relies almost entirely on the $75,000 three-dimensional (3D) lidar, which
Bruce Hall and colleagues at Velodyne Acoustics Inc. in Morgan Hill, California, developed to compete in
the 2005 DARPA challenge. The sensor may be all you need, says Sören Kammel of the Karlsruhe
Institute of Technology in Germany. "I think some teams have a lot of sensors because they have a lot
of sponsors, and everybody wants their sensor on the car," he says.
Even that one sensor is beyond the means of Donald Harper and his six teammates from the University
of Central Florida in Orlando. They've outfitted Knight Rider, a 1996 Subaru Outback that belonged to
Harper's wife and has 99,257 miles (159,705 km) on it, with just enough gizmos to get around the
course--they hope. Instead of the spinning 3D lidar, they use two lidars that see in one direction and
rock them back and forth. "If just one wire falls off, something essential is not going to work," Harper
says. Still, the team made the final having invested only $130,000 in the project.
Robots, start your engines!
Race day usually brings the intoxicating smell of high-octane fuel and the electrifying scream of
engines. But not here. At 8:00 a.m., the robots leave the starting area, one by one, like rental cars
leaving a lot. There's a glitch. Interference from a jumbo TV screen knocks out the GPS receiver of first
qualifier, Boss, Carnegie Mellon's Chevy Tahoe. The team replaces the unit and has to wait 30 minutes
to regain the signal. Meanwhile, Odin, a Ford Escape from Virginia Polytechnic Institute and State
University in Blacksburg; Junior, Stanford's Volkswagen Passat; and the others head out, hesitating and
swerving as if driven by octogenarians. After a half-hour, all 11 robots--plus their chase cars and 37
other cars--are on the road.
There's only one curve from which to glimpse the robots, so DARPA has hired a helicopter and is
televising the event on three huge screens in a vast tent. Jamie Hyneman and Grant Imahara of the
geeky cable-television reality show Mythbusters provide commentary. It's like watching a hybrid of a
NASCAR race and the infamous O. J. Simpson lowspeed police chase.
After you. Stanford's Junior and Virginia Tech's Odin negotiate an intersection.
CREDIT: DARPA
Each robot has to complete three "missions" comprising six or seven "submissions," such as parking in
exactly the right space in a lot, traversing an off-road passage, or navigating between two places. After
each mission, the robots return to the start area to download the specifications for the next, and each
machine must travel 60 miles (97 kilometers) in less than 6 hours.
At first, the action comes fast and heavy. An hour into the race, TerraMax, the hulking vehicle entered
by military contractor Oshkosh Truck Corp. in Wisconsin, turns toward a pillar and gets stuck staring at
it. Forty-five minutes later, Central Florida runs straight toward a house. Caroline, the robot from Team
CarOLO, the other German squad, collides with MIT's Talos and loses sensors. By 11:00 a.m., five
robots have either failed or been disqualified.
Then things settle down. The remaining robots' "personalities" emerge. Carnegie Mellon's Boss zooms
confidently away from stops, a hard charger like team leader Whittaker. Stanford's Junior glides around
smoothly, so much so you hardly notice it. MIT's Talos is aggressive in traffic--it also clips Cornell's
Chevy Tahoe, Skynet--but skittish off-road, stopping and starting like a cat creeping down a steep slope.
Around 1:30 p.m., three teams have nearly completed their missions, and spectators swarm back to the
grandstands. At 1:42, Stanford cruises across the finish line, followed a minute and a half later by
Carnegie Mellon. Upstart Virginia Tech cruises home third--even without the 3D lidar. "We knew we were
good," says Virginia Tech's Alfred Wicks. "We'd done our homework." The University of Pennsylvania's
Toyota Prius, Little Ben, straggles in an hour later. Sometime past 3:30 p.m., MIT slips in just before
Cornell.
The outcome seems obvious. Carnegie Mellon spotted Stanford and Virginia Tech a 20- minute head
start and made up almost all of it. It seems the victory should be theirs. DARPA officials will make the
final call, however. And, some participants grumble, DARPA never fully explains its judgments.
Make it out to …
But the next morning brings no surprises. Carnegie Mellon walks off with the win. Stanford takes second
and $1,000,000, Virginia Tech takes third and $500,000. "There's tremendous satisfaction in what the
whole field accomplished," Whittaker says. "That was a day that stunned the world." DARPA Director
Anthony Tether also gushes. "Quite frankly, I watched these things and I forgot after a while that there
wasn't anybody in there," he says. "It's a historic day--'bot on 'bot for the first time!"
Maybe there's something to the grandiose rhetoric. Now only a Luddite could doubt that soon cars will
guide themselves, at least in a pinch to avoid collisions. In fact, the technology already seems ripe for
low-risk applications, such as automating farm equipment, and the leading teams are pushing to
commercialize their software. "I think it's going to come in bits and pieces," says Charles Reinholtz,
leader of the Virginia Tech team and an engineer at Embry-Riddle Aeronautical University in Daytona
Beach, Florida.
Ironically, the success of the Urban Challenge could reduce the chances that DARPA will stage another
competition. "DARPA never finishes anything," Tether says. "All we do is show that it can be done" in
the hope that industry takes over and pushes further development. Clearly, when it comes to making
robotic cars, the Urban Challenge has shown that it is possible.
Still, many engineers are eager for another competition. Their robots aren't nearly ready for the open
road, they say, and many already know what they would like to see in the next challenge: a contest for
autonomous cars that must communicate and work together. Suddenly, that doesn't seem quite so
absurd.
Next
http://www.sciencemag.org/cgi/content/full/318/5853/1060
Robotic Cars Tackle Crosstown Traffic--and Not One Another
Adrian Cho
In DARPA's Urban Challenge, cars that drive themselves face off in a strange, soulless rush hour. Are
human drivers about to go the way of the buggy whip?
VICTORVILLE, CALIFORNIA--The Land Rover bristles with sensors like a mechanical porcupine. John
Leonard, an engineer at the Massachusetts Institute of Technology (MIT) in Cambridge, ticks off the
robot's features. On the roof spins a conical laser range finder called a lidar that sees in three
dimensions. A dozen lidars that see in one direction, 15 radars, and six digital cameras look out every
which way. Computers fill the back of the truck, and a generator supplies the 3.5 kilowatts of power they
need. It's impressive. But all this so the truck can turn left across traffic by itself?
The robot is one of nearly three dozen vying in the Urban Challenge, a competition sponsored by the U.
S. Defense Advanced Research Projects Agency (DARPA). It's the third and most demanding in a
series that aims to spur the development of autonomous vehicles, which the U.S. military hopes to press
into service by 2015. In 2004 and 2005, robots raced one at a time across open terrain. This time, they
must navigate the streets of an abandoned air base here in the Mojave Desert without colliding with one
another or with human-guided vehicles.
The competition showcases some of the world's best talent in robotics. "We were drawn to the Urban
Challenge because it requires real-time decision-making in a dynamic environment and in the presence
of uncertainty," Leonard says. It also serves the higher purpose of trying to save lives, as worldwide, 1.2
million people die each year in traffic accidents that robotic cars might help avoid.
And yet the Urban Challenge is at least slightly absurd. It looks a bit like a real race. Engineers wear
bright shirts emblazoned with the logos of sponsors--GM, Ford, Intel, Google. Teams have hauled in
tractor trailers full of equipment and plastered their robots with decals. Besides the $2 million first prize,
the appeal of the challenge is obvious. It's hard, and by pitting idea against idea and technology against
technology, "it determines what technical DNA moves to the next generation" in the evolution of
autonomous vehicles, says roboticist William "Red" Whittaker of Carnegie Mellon University in
Pittsburgh, Pennsylvania.
Still, the task the robots will attempt seems so ordinary. They must obey the California traffic laws
(although if two collide, they won't have to exchange insurance information as human drivers are
required to do). We're all here to watch traffic. But we won't see with our own eyes. Instead, we'll have to
watch it on television. It's not even clear what DARPA gets out of this well-crafted media circus. The
competition is meant to stimulate the development of cars that drive themselves--and it has--but DARPA
does not require winners to reveal to the agency the details of their technologies.
Comeback kid. After a delayed start, Carnegie Mellon's Boss cruises briskly to a victory.
CREDIT: A. CHO/SCIENCE
Go ahead, bend the rules
At the decaying fighter base, across the road from the new federal prison here in Victorville, 11 teams
have made the final competition. Three years ago, not one robot traversed more than a dozen
kilometers of the 230-kilometer off-road course. A year later, four completed a similar course. And this
year's robots are far more capable than last year's crop, possessing better sensors, more powerful
computers, and, most important, more sophisticated programming. "Driving is a software problem, not a
hardware problem," says engineer Michael Montemerlo of Stanford University in Palo Alto, California. "At
Stanford, we can't build a better car, but we can make a smarter car."
Computationally, this year's challenge is much more difficult than the first two, researchers say. In the
desert races, the robots had only to identify obstacles in a static landscape and plot a safe path around
them. This time, the vehicles will have to avoid other cars, including other robots, while at the same time
obeying the relatively arbitrary traffic laws. To do that, each robot's computer must calculate the likely
trajectories of all the objects around it and plan to miss them. Of course, a robot cannot know exactly
where another car will go, so the machines generally employ layers of probabilistic algorithms to decide
their next moves.
MIT has decked its robot, Talos, with the most sensors. The radars see distant objects, the lidars see at
an intermediate range, and the cameras spot things close by, explains David Barrett, a team member
from the Franklin W. Olin College of Engineering in Needham, Massachusetts. Talos depends mainly on
its sensors to navigate, Barrett says. That's because the team assumed, incorrectly it turns out, that
DARPA would not let robots use signals from the satellite-based Global Positioning System (GPS) all
over the course.
Researchers from Stanford, who won the 2005 competition, say that they focus on the algorithms
programmed into their robot. Merely encoding the traffic laws can leave the robot stymied, says
Stanford computer scientist Sebastian Thrun. For example, when two robots arrive at a four-way stop
simultaneously, each may try to yield to the other endlessly. To avoid such deadlock, the team lets its
robot skirt the laws. "Our car has a hierarchy it follows," Thrun says. "At the top, it obeys strict rules.
And if it gets stuck, it ignores more and more rules." Fair enough. Why expect more from a robot than a
human?
Team AnnieWAY, one of two German squads in the final, has taken a minimalist approach to guiding its
Volkswagen Passat. The team relies almost entirely on the $75,000 three-dimensional (3D) lidar, which
Bruce Hall and colleagues at Velodyne Acoustics Inc. in Morgan Hill, California, developed to compete in
the 2005 DARPA challenge. The sensor may be all you need, says Sören Kammel of the Karlsruhe
Institute of Technology in Germany. "I think some teams have a lot of sensors because they have a lot
of sponsors, and everybody wants their sensor on the car," he says.
Even that one sensor is beyond the means of Donald Harper and his six teammates from the University
of Central Florida in Orlando. They've outfitted Knight Rider, a 1996 Subaru Outback that belonged to
Harper's wife and has 99,257 miles (159,705 km) on it, with just enough gizmos to get around the
course--they hope. Instead of the spinning 3D lidar, they use two lidars that see in one direction and
rock them back and forth. "If just one wire falls off, something essential is not going to work," Harper
says. Still, the team made the final having invested only $130,000 in the project.
Robots, start your engines!
Race day usually brings the intoxicating smell of high-octane fuel and the electrifying scream of
engines. But not here. At 8:00 a.m., the robots leave the starting area, one by one, like rental cars
leaving a lot. There's a glitch. Interference from a jumbo TV screen knocks out the GPS receiver of first
qualifier, Boss, Carnegie Mellon's Chevy Tahoe. The team replaces the unit and has to wait 30 minutes
to regain the signal. Meanwhile, Odin, a Ford Escape from Virginia Polytechnic Institute and State
University in Blacksburg; Junior, Stanford's Volkswagen Passat; and the others head out, hesitating and
swerving as if driven by octogenarians. After a half-hour, all 11 robots--plus their chase cars and 37
other cars--are on the road.
There's only one curve from which to glimpse the robots, so DARPA has hired a helicopter and is
televising the event on three huge screens in a vast tent. Jamie Hyneman and Grant Imahara of the
geeky cable-television reality show Mythbusters provide commentary. It's like watching a hybrid of a
NASCAR race and the infamous O. J. Simpson lowspeed police chase.
After you. Stanford's Junior and Virginia Tech's Odin negotiate an intersection.
CREDIT: DARPA
Each robot has to complete three "missions" comprising six or seven "submissions," such as parking in
exactly the right space in a lot, traversing an off-road passage, or navigating between two places. After
each mission, the robots return to the start area to download the specifications for the next, and each
machine must travel 60 miles (97 kilometers) in less than 6 hours.
At first, the action comes fast and heavy. An hour into the race, TerraMax, the hulking vehicle entered
by military contractor Oshkosh Truck Corp. in Wisconsin, turns toward a pillar and gets stuck staring at
it. Forty-five minutes later, Central Florida runs straight toward a house. Caroline, the robot from Team
CarOLO, the other German squad, collides with MIT's Talos and loses sensors. By 11:00 a.m., five
robots have either failed or been disqualified.
Then things settle down. The remaining robots' "personalities" emerge. Carnegie Mellon's Boss zooms
confidently away from stops, a hard charger like team leader Whittaker. Stanford's Junior glides around
smoothly, so much so you hardly notice it. MIT's Talos is aggressive in traffic--it also clips Cornell's
Chevy Tahoe, Skynet--but skittish off-road, stopping and starting like a cat creeping down a steep slope.
Around 1:30 p.m., three teams have nearly completed their missions, and spectators swarm back to the
grandstands. At 1:42, Stanford cruises across the finish line, followed a minute and a half later by
Carnegie Mellon. Upstart Virginia Tech cruises home third--even without the 3D lidar. "We knew we were
good," says Virginia Tech's Alfred Wicks. "We'd done our homework." The University of Pennsylvania's
Toyota Prius, Little Ben, straggles in an hour later. Sometime past 3:30 p.m., MIT slips in just before
Cornell.
The outcome seems obvious. Carnegie Mellon spotted Stanford and Virginia Tech a 20- minute head
start and made up almost all of it. It seems the victory should be theirs. DARPA officials will make the
final call, however. And, some participants grumble, DARPA never fully explains its judgments.
Make it out to …
But the next morning brings no surprises. Carnegie Mellon walks off with the win. Stanford takes second
and $1,000,000, Virginia Tech takes third and $500,000. "There's tremendous satisfaction in what the
whole field accomplished," Whittaker says. "That was a day that stunned the world." DARPA Director
Anthony Tether also gushes. "Quite frankly, I watched these things and I forgot after a while that there
wasn't anybody in there," he says. "It's a historic day--'bot on 'bot for the first time!"
Maybe there's something to the grandiose rhetoric. Now only a Luddite could doubt that soon cars will
guide themselves, at least in a pinch to avoid collisions. In fact, the technology already seems ripe for
low-risk applications, such as automating farm equipment, and the leading teams are pushing to
commercialize their software. "I think it's going to come in bits and pieces," says Charles Reinholtz,
leader of the Virginia Tech team and an engineer at Embry-Riddle Aeronautical University in Daytona
Beach, Florida.
Ironically, the success of the Urban Challenge could reduce the chances that DARPA will stage another
competition. "DARPA never finishes anything," Tether says. "All we do is show that it can be done" in
the hope that industry takes over and pushes further development. Clearly, when it comes to making
robotic cars, the Urban Challenge has shown that it is possible.
Still, many engineers are eager for another competition. Their robots aren't nearly ready for the open
road, they say, and many already know what they would like to see in the next challenge: a contest for
autonomous cars that must communicate and work together. Suddenly, that doesn't seem quite so
absurd.
Next
 | |  | |  |