Three years ago, news stories all but laughed aloud at the driverless vehicles in their now-famous race across the Mojave Desert, as they gamely tried to cross a 142-mile desert course on their own. In “comic” fashion, the stories said, the robotic vehicles drove in dizzy circles, stalled atop rocks, and slowly halted themselves, apparently confused by the breadth of their electronic chores. As if on cue, a robotic motorcycle added to the already-mirthful tone of the day by traveling just two feet before toppling over. A San Francisco Chronicle headline summed it up: “Robot race suffers quick, ignoble end.”
How times change. On Oct. 8, 2005, just 19 months after that “quick, ignoble end,” five driverless vehicles stunned the American public by crossing the finish line in the second DARPA Grand Challenge. Stanford University’s Volkswagen Tourag – dubbed “Stanley” by its designers – crossed the rock-strewn, 132-mile course in an astounding six hours and 51 minutes. Three others completed the off-road course in less than eight hours; a final vehicle finished in 12 hours, 51 minutes.
The unspoken lesson was obvious: Robotic technology was advancing faster than anyone had imagined. Moreover, during the ensuing months since that second government-sponsored race, many have come to believe that engineers might yet prevail in an effort to put autonomous vehicles on our roads in the next quarter-century.
“The biggest thing ‘Stanley’ taught us is that autonomous cars are really possible,” says David Stavans, a Ph.D. student at Stanford and co-creator of Stanley, which won a $2 million purse on that day. “Stanley drove flawlessly in the Grand Challenge, and in the national qualifying event, and for many hundreds of miles before the race.”
To be sure, it wasn’t easy. Most of the robotic driving was on an off-road course strewn with obstacles, natural and man-made. Sensors, microprocessors, and software struggled to recognize roads, let alone rocks, ruts, tunnels, bridges, tumbleweed, fence posts, barbed wire, competing vehicles, and mechanical traps that the Defense Advanced Research Projects Agency (DARPA) had laid out for them. Still, Grand Challenge engineering teams were supposed to make it happen with off-the-shelf hardware.
“It was like teaching a baby to walk,” says William “Red” Whittaker, Fredkin research professor of robotics at Carnegie Mellon University, and team leader for the school’s two race entries. “At this point in history, computers simply aren’t as competent at driving as humans are.”
Urban Challenge
Still, some robotics experts believe that robotic vehicles could one day be up to the task of controlling everyday transportation, given their extraordinary pace of advancement. If that’s so, robots will one day enable drivers to snooze or read while their vehicles drive them to the office.
The greatest challenge in making that happen, however, may arrive in the earliest years, when robotic vehicles will undoubtedly share the road with aggressive and unpredictable humans.
“Being able to drive in traffic is a much harder problem than the vehicles faced at the Grand Challenge,” Stavens says. “The biggest problem is other drivers. The robot has to use reasoning to predict what other drivers will do.”
Moreover, many experts warn that social and legal obstacles could prevent robotic cars from reaching the road for many, many years. Most drivers are likely to distrust machines, at least in the beginning. Moreover, our increasingly litigious society is likely to make automakers balk at the thought of taking responsibility for the operation of a robot.
Still, many high-level engineers believe the technology will be available and reliable long before society expects it, which is why the Defense Advanced Research Projects Agency (DARPA) is sponsoring a 2007 yet another robot race – this one across urban terrain. Known as the DARPA Urban Challenge, it will force the robotic vehicles to interact with city traffic. Participating vehicles will have to change lanes to pass slower-moving vehicles, stop at traffic lights and stop signs, and do virtually every task that city drivers do on a routine basis.
Stanford’s 50-member engineering team for the race believes that machine-learning may be one of their biggest keys to success, as well as in eventual efforts to put robotic vehicles on our roadways. The team used the same strategy in the development of Stanley, which incorporated more than 100,000 lines of computer code to examine old driving data, which enabled Stanley to learn whether it should stop, swerve, or just keep moving when it “saw” a potential obstacle. As a result, Stanley didn’t pause to “hallucinate” as some other vehicles did when they thought they saw obstacles in their paths.
Ultimately, the birth of autonomous vehicles may also benefit from the parallel emergence of smart highways, which are already being discussed by automakers, electronics companies, and government agencies. Using a concept known as a Dedicated Short Range Communication (DSRC) system, the new smart highway will employ a 5.9-GHz transmission frequency to enable roadside transceivers to talk to vehicle-based transceivers and GPS units. By knowing the number, location and speed of nearby participating vehicles, the system could create a local area network in which vehicles and traffic lights will share data.
“You could put one of these radio devices at a traffic light, stop sign or intersection, and it could broadcast information to approaching traffic,” says Bob Lange, executive director of structure and safety integration for General Motors Corp. “In the event of a collision threat, it could alert drivers. In extreme cases, it might even be aggressive enough to intervene and apply the brakes or steering.”
To be sure, DSRC isn’t being targeted at autonomous vehicles. But some engineers see it as a piece of the puzzle that might one day speed the development of robot cars.
“That’s a very important part of making autonomous vehicles a reality,” Stavens says. He adds that it will serve as a good way to reliably detect what’s happening at traffic lights and on roadways across the country.
“The goal is to create a system that works well and is safe,” Stavens concludes. “Certainly within 20 years, and maybe even in ten, we will have autonomous cars capable of driving reliably on any kinds of roads.”
In 2005, five autonomous vehicles successfully traversed DARPA's 132-mile desert course, setting the stage for robotic vehicle technology in the future. (Photo courtesy of Standford University)
Machine learning software makes the difference. Vehicle: Volkswagen Touareg R5 Sensors:Roof rack holds five Sick laser range finders. Also on board: color camera; radar; GPS system; inertial navigation. Computing: Trunk-mounted rack carries six Pentium M blade computers and Gigabit Ethernet switch. Actuation: Touareg R5 natively offers throttle- and brake-by-wire. Stanford team added DC motor to steering column to provide steer-by-wire. Finish time: 6 hours, 53 minutes.
When Stanford University’s “Stanley” crossed the Grand Challenge finish line first on Oct. 8, many observers were shocked. Two Carnegie Mellon University vehicles, after all, had dominated the National Qualifying Events, emerging as top qualifiers in every event.
Still, Stanford’s vehicle won the race, and for good reason. The University’s 50-member team spent countless hours developing 100,000 lines of software code, much of which was devoted to “teaching” its vehicle how to drive before the race ever started.
“Stanley’s key advantage was the machine learning technology we put in there,” notes David Stavens, a Stanford Ph.D. student who worked on the vehicle. “We gave Stanley a framework to approach the problems he had to tackle, particularly finding the road and avoiding obstacles.”
Indeed, Stanley’s software framework enabled it to learn from pre-race data, thus improving the algorithms that determined whether it would stop, swerve, or just keep moving straight ahead when it “saw” a potential obstacle. As a result, Stanley didn’t stop to “hallucinate,” as some other competing vehicles did when they thought they saw obstacles in their paths.
To accomplish that, Stanford engineers wrote special machine learning code that enabled Stanley to examine old data, particularly sensor data, which served as the basis for its decisions during the October 8th race.
“Using machine learning, we reduced Stanley’s error rate from about 12 percent in the beginning to about one in 50,000 by race time,” Stavens says.
Like many who developed race technology, Stavens believes that the advancement of such computing systems, and particularly improvements in software, will one day enable vehicles to drive in traffic.
“The project is just beginning,” Stavens says. “We feel that, ultimately, you’ll be able to go to the dealer and buy a car that drives itself.”
Carnegie Mellon two robotic vehicles finished second and third, narrowly missing the $2 million prize. Vehicle: Sandstorm (1986 AM General Hummer) and H1ghlander (1999 AM General Hummer) Sensors:Long-range LIDAR (three-axis gimbal-controlled), short-range LIDAR, 360° radar, GPS, inertial navigation system. Computing: Four Pentium III PC104 stacks; seven Pentium M CompactPCI-based computers, along with numerous device-based DSPs, FPGAs, and ASICs. Actuation: Carnegie Mellon (CMU) teams added throttle-, brake-, and steer-by-wire. Finish time: 7 hours, 4 minutes (Sandstorm); 7 hours, 14 minutes (H1ghlander)
Prior to the race, Carnegie Mellon’s much-publicized Sandstorm and H1ghlander vehicles logged more than 4,000 miles of autonomous driving. Sandstorm, in particular, proved itself during high-speed navigation on a 200-mile endurance run during which it hit peak speeds of 54 mph. What’s more, both vehicles emerged as top qualifiers in the National Qualifying Event before the race.
Red Team leader William “Red” Whittaker, Fredkin Research Professor of Robotics at Carnegie Mellon University, says that two of the teams’ distinguishing technologies were their use of long-range LIDAR and pre-planning software.
The long-range LIDAR, Whittaker says, was the key to achieving higher speeds. “Because we could see the obstacles earlier, we could drive at higher speed thresholds,” Whittaker says. “Performance is very much a function of how well and how fast you do the computing.”
Moreover, the million lines of code in CMU’s vehicles incorporated so-called “pre-planning software” for the purpose of creating performance consistency by governing elapsed times and moderating the vehicles’ peak speeds. Moreover, the software is able to simulate the negotiation of extreme terrain, and thus help the autonomous vehicles choose optimal paths.
“With our pre-planning software, we could vastly improve the elapsed times of any other vehicles in the race,” Whittaker says.
Carnegie Mellon engineers say their vehicles had done successful test runs over far rougher topography than the desert terrain of race day. But they missed their target time of six hours and 19 minutes, they say, because of an as-yet unexplained engine failure that left their lead vehicle struggling toward the finish line.
“We can’t re-create the engine problem,” Whittaker says. “We’ve driven a thousand miles since, and we still have no idea why it happened.”
In less than six months, Team Gray put together a successful robotic vehicle that missed the $2 million grand prize by just 37 minutes. Vehicle: 2005 Ford Hybrid Escape Sensors: Front-mounted Sick LADAR systems; GPS and inertial navigation. Computing: Marine computer from Valhalla Enterprises, Inc. Actuation: Caterpillar diesel engine natively offers throttle-by-wire; electric-over-air valves enable brake-by-wire on air brakes at three axles; motor on steering shaft enabled steer-by-wire. Finish time:12 hours, 51 minutes.
Oshkosh Truck’s 27-ft-long, 30,000-lb defense truck might not be the image that comes to mind when the public thinks about racing across the desert, but it’s the ideal image in the minds of military officials searching for a way to haul supplies through battle zones. Indeed, prior to the race, Oshkosh had already built 7,000 of the vehicles for the United States Marine Corps.
None of those vehicles, however, offered the autonomous driving capabilities of the team’s Grand Challenge vehicle. Known as the Team Terramax vehicle, the truck incorporated a vehicle management system from Rockwell Collins and stereo-vision cameras from the University of Parma in Italy.
Oshkosh engineers say that the combination of sensor technologies enabled the truck to recognize different kinds of obstacles.
“The vision systems did an excellent job of picking up obstacles with edges, such as fence posts and barbed wire,” notes Gary Schmeidel, vice president of advanced products engineering for Oshkosh Truck. “And the multi-plane LIDAR did a great job because it reflected multiple beams off an object, so you could get a sense of size and distance.”
Indeed, Schmeidel says, the combination of technologies benefited Team Terramax when tumbleweed blew in front of its vehicle. “They were moving targets because of the wind, but the sensing system was able to recognize them,” Schmiedel recalls.
More important, Schmeidel says, the truck didn’t require a massive number of extra computers to get the job done. Rather, he says, the autonomous vehicle computers fit beneath the bench seat in the truck’s cab.
Says Schmeidel: “This demonstrated to our customers that this capability doesn’t need to destroy all the other electronic capabilities in the vehicle.”
Clippard Instrument Laboratory |
Product Brief VIEW NOW
SPONSORED CONTENT
ADVERTISEMENT
Design News Partner Zone
Mechatronics in action
Successful synergistic integration of controls, electronics, computers and mechanical systems is key to the 21st century design process. Unlock the secrets at the new Mechatronics Zone!
Webcast: Sensor Know-How Now
Join our moderator Randy Frank and John Keating from Cognex and explore Solving Industrial Inspection Problems. Read More
Design News RSS News Feeds
Webcast: Sensor Know-How Now