Robots shift car tech into high gear
The wizardry behind vehicles powered with artificial intelligence is heading toward popular cars.
If you ask the masterminds behind the robots, the answer is "yes, it's just a matter of time."
Vehicles powered with artificial-intelligence software and sporting the ability to "see" the road with external sensors will be a staple in the U.S. military within 10 years, under a mandate from Congress that spurred the desert robot rally. The underlying technology also will find its way into popular cars with features like collision and lane-departure warnings and adaptive cruise controls. The technology is also relevant, experts say, for the disabled and for automating machines.
"It's not just about a bright idea. There's a lot of work to do. That business of development and productization and building an enterprise is a lot harder than creating a technology," said William "Red" Whittaker, a professor of robotics at Carnegie Mellon University. So much robotics research has been done at Carnegie Mellon that CMU's hometown of Pittsburgh is jokingly called "Roboburgh" in some science circles.
For Stanford University, the winner of the DARPA Grand Challenge robot desert race and its $2 million prize, the goal has long been to make vehicles safer for the road.
Stanford set out with the particular interest of developing technology that would help carmakers include aids that could cut down on the number of traffic deaths caused by inattention or intoxication. Now Stanford has $2 million to invest in improving its technology and artificial intelligence research, under the direction of robotics professor Sebastian Thrun. Volkswagen, which sponsored Stanford's vehicle, Stanley, and donated a Touareg V5 for the race, is also developing this technology for its line of cars.
The next frontier will be to develop technologies that can help vehicles improve city driving, as opposed to motoring off-road or on highways, where there are no stoplights or pedestrians. The race gave the robots a structure for driving the course. But on the highway or in cities, intelligent vehicles won't have that direction.
"We've been working on the war on cancer, but with this technology we're a lot closer to saving more lives--young lives--through accidents, by giving attentional aids," said Gary Bradski, a machine-learning expert at Intel who worked on Stanley. "The question is how to alert people without causing an accident."
Four autonomous vehicles--Stanford's Stanley, Carnegie Mellon's H1ghlander and Sandstorm, and Gray Insurance's Kat 5--drove a tricky 132-mile course in the Mojave Desert. They were the first unmanned cars ever to complete the race designed by DARPA, or the Defense Advanced Research Projects Agency, the research and development arm of the U.S. Department of Defense.
And after only two years of hosting the DARPA Grand Challenge, the U.S. military says it has accomplished its goal of fostering innovation in artificially intelligent designed vehicles. Within 10 years, such vehicles are supposed to make up a third of the U.S. army's transportation.
The technology has already made its way into contractor machinery, and some of the cars showed surprising resilience.
H1ghlander had engine problems the day of the race, which caused it to finish 40 minutes late and ultimately lose the $2 million to Stanford University. Despite its trouble, the car still finished--a testimony to the system software's sturdiness, Carnegie Mellon's Whittaker said. "On its worst day it can run anything," he said.
The same sturdiness is exhibited in robots used by two companies spun out of Whittaker's robotics research. One is RedZone Robotics, which uses robotic machines to make maps of sewer systems. Whittaker also founded Workhorse Technologies, unmanned robots to explore and make maps of mines.
The technology is also being deployed in earth-moving and construction machines from Caterpillar, which was one of the major sponsors of CMU's Red Team.
For its future combat system, the government plans to build a family of 18 vehicle types that will be faster and lighter on the road. The group will include so-called drive-by-wire vehicles, as well as some with autonomous and semi-autonomous capabilities. Drive by wire is a car that can be driven without a steering wheel from an internal protected seat or from a remote location, but it is still human operated. The car, perhaps a supply vehicle, could be lightly armored if driven remotely. A semi-autonomous vehicle, on the other hand, can be programmed to travel from point A to point B, or to follow another car.
The autonomous vehicle would be heavily armored and could take GPS (Global Positioning System) coordinates of the road and create a map of obstacles and pass that data back to the semi-autonomous truck. All this could help keep soldiers out of the line of fire.
"There will be programs in the next four or five years, such as
tests of semi-autonomous or convoy vehicles in the military," said Bill Klarquist, vice president of engineering at PercepTek, a robotics company that has contracts with the government, Ford and others.PercepTek creates software for perception planning and control. That's what the car observes about its environment, and given that information, how it will travel. The company develops technology that helps vehicles manage their speed, follow the road and avoid obstacles.
Carmakers are headed toward total drive-by-wire systems, the route airlines took roughly a decade ago. That means they take the physical actions of the driver, such as pressing on the gas pedal, and turn them into digital messages for the car's central control system. Sensors measure how far the driver turns the steering wheel, for example, and translate that to a message to turn the wheels the appropriate amount. More sophisticated controls can be added for things like emergency braking and traction control.
Many modern cruise control systems already use drive-by-wire throttle functions. With the addition of radar and laser sensors, a car can measure the distance between it and one ahead of it. That way, carmakers can add "adaptive cruise controls" that will regulate the speed of the car to maintain a safe distance between it and other vehicles.
Pricey models from Jaguar, Mercedes and Lexus are already offering that feature.
"Lane keeping" is another benefit of radar and laser sensors. The technology maintains a path down the center of the road and alerts the driver when the car begins to drift into another lane. The feature is already used in the trucking industry, but General Motors has said it plans to offer the feature in cars by 2007.
"The issue is when you introduce new technology there's also the liability. You normally see cars like this introduced in Europe and Japan first, and as they're embraced there, the bugs and characteristics are worked out," Klarquist said.
Adaptive cruise showed up in Germany long before it did in the United States, he said.
In Japan, carmakers have already been testing systems that warn drivers if they're drifting too far out of a lane or if they are about to hit something.
"I think technologically, we're within five to 10 years of having good systems for this," said Intel's Bradski.
PercepTek, which backed a robot in the Grand Challenge called Intelligent Design Systems, said that what it will gain from the race is the knowledge of how to use multiple sensors together for road and obstacle detection and avoidance. Commercially, that knowledge will inform what's called precrash applications.
With a combination of laser and radar sensors, a car system could "see" an oncoming collision, if an object ahead was stopping at a faster speed than expected, for example. By detecting how fast the car is traveling in relation to another on the road, the car's system could prepare airbags or cinch seat belts tighter. It could even regulate how the airbags inflate in relation to expected harm from the impact.
John Davidson, whose investment firm Mohr Davidow Ventures sponsored Stanley, predicts that within five years, sophisticated technology for collision avoidance will be in cars.
"Brake systems are already smart, and we'll slowly walk up this curve," Davidson said. "I'm skeptical about ever sitting in the backseat and pressing a button. But the technology has applicability in lots of commercial and industrial applications where autonomy is important."
Machine-learning technology has already touched industries like drug discovery, e-mail processing and financial forecasting. But technology is still a long way from allowing an autonomous machine to handle every unpredicted situation that pops up. "The robot must be able to learn from a situation and think it's way through the problem," Davidson said. "But there are other problems they have to solve, like dealing with contingencies. What if a computer dies?"