Sounds simple enough. Go 170 miles, collect $2 million. Here’s the catch: the traveler is a completely autonomous vehicle developed by Cornell students and the route, to be announced just hours before the race, is off-road through the rugged terrain of the desert Southwest.
By Melanie Bush
IF all goes well in the semifinal round, Cornell engineers will advance to the final round of the Defense Advanced Research Projects Agency (DARPA) Grand Challenge in October. Cornell’s autonomous vehicle will compete against 19 other robotic cars in an attempt to negotiate 170 miles of punishing desert terrain in the Southwest somewhere between Los Angeles and Las Vegas. The stakes are high: winner takes home
$2 million and bragging rights for accomplishing something that’s never been done. In the first DARPA Grand Challenge, held in March 2004, the most successful team, Carnegie Mellon, completed exactly 7.4 miles before being incapacitated by a boulder.
The event is part of a congressionally mandated program that authorizes DARPA to conduct tests and award prizes for advances in vital technologies, in this case, the 2001 Defense Authorization Act’s goal that one third of all combat vehicles be unmanned by 2015. DARPA is the central research and development agency of the Department of Defense and has pioneered major technological breakthroughs such as the Internet, Stealth aircraft, and smart bombs.
Ephrahim Garcia, associate professor of mechanical and aerospace engineering, is the main faculty adviser of the Cornell team, although he swiftly explains that this is a completely student-run project. Garcia feels confident that Cornell will reign in the desert. “We aim to finish the Grand Challenge. Go 170 miles, collect two million dollars. We aim to end this thing once and for all.”
And what would Team Cornell do with the $2 million grand prize? Their goal is to establish an endowment for future student projects that would allow Cornell to make a larger impact at competitions like DARPA. But actually, Garcia says, “The kids don’t talk about the prize money. It’s not the culture of the team, which is to execute well technically.”
Garcia has himself blazed a rigorous path on his journey to Cornell. He grew up in New York City as the child of Cuban immigrants and earned his Ph.D. at the State University of New York in Buffalo in 1990, concentrating on smart controls in microelectromechanical systems. Since then, he’s focused on mechatronics, using entomological models to create a new generation of bio-inspired robots. Though he doesn’t consider himself a roboticist, he’s excited by the challenge of building machines that can travel efficiently and change their shapes to adapt to the environment around them.
Garcia has conducted research for NASA, the Air Force, the Office of Naval Research (ONR), and the Army Research Office; run his own small company; taught at Vanderbilt University; and, most recently, worked in the Defense Sciences Office at DARPA. He was named a Presidential Faculty Fellow by Bill Clinton and an ONR Young Investigator. Now, as part of what he calls his “strange but fun path,” he has come to Cornell.
Arriving on campus in 2002, Garcia admitted at the time that a typical career trajectory would lead to a high-profile job in industry after a stint at DARPA. But, he says, “I wanted to get back to the trenches of academia, to go someplace where there were really good students and really good colleagues.”
And if Cornell’s Grand Challenge team has a secret weapon, Garcia is convinced it’s those really good students, who this year put together not one but two potential entries for the event. “The energy of college students is not to be underestimated, and Cornell kids are a zealous group, working incredibly hard and achieving things that no one would ever expect from undergrads. I’ve been very impressed with the students’ ability to equip two teams.”
Garcia believes Cornell students are especially well suited to large, ambitious projects like the Grand Challenge in part because of classes like MAE 225 Mechanical Design and Synthesis. “Synthesizing so many elements in class motivates them to create real projects, to realize that products are not just something you buy. It makes them want to do everything themselves. I think my mechatronics class—combining electromechanical systems—has also helped. It has further demented their thinking, in a good way.”
Daniel Huttenlocher, co-faculty adviser and the Neafsey Professor of Computing, Information Science and Business, is also impressed with the student team. “The most amazing thing about this whole project is that last summer they had nothing. Now they have two vehicles that run autonomously. And not only do they work hard; they work smart. They’ve done a thorough analysis of what went wrong last year, and they know where to get their hands on things and when to go look for experts.”
Cornell’s involvement with the Grand Challenge dates from the summer of 2004, when then-sophomore Noah Zych ’06 ME used the proceeds of his undergraduate research grant from the College of Engineering to do a feasibility study on how much it would cost to field a vehicle. “It all started with the students coming to me, and me saying, ‘Wait a minute—let’s think about this,’” Garcia says.
“DARPA sees the future of the military in machine-augmented human forces,” he continues, “robotic vehicles that will follow soldiers on foot into battle carrying firepower and ferrying wounded soldiers back to hospitals.” This year DARPA has made the course not only longer and tougher but more akin to a real battlefield, laying down obstacles like tank traps, for example, along the route. “Those will stop a vehicle that can’t sense them,” says Garcia. “DARPA wants to get people really thinking about this; accomplishing their own research objectives is the purpose of the contest.”
For the 2005 event, 190 teams from both academia and private industry applied, but only 20 would ultimately be selected to compete on the basis of numerous safety inspections and qualifying rounds. The Cornell team already aced its site visit from DARPA on March 1. In an e-mailed report to colleagues, Garcia wrote: “The DARPA Grand Challenge team did extremely well yesterday in front of two DARPA Program Managers (PMs).... The first vehicle’s performance was a flawless three runs of 200 meters, following a path set by GPS waypoints. The DARPA PMs put trashcans at blind spots ... but the artificial intelligence algorithms were robust enough to avoid them in all cases.
“When the second vehicle was about to begin its trial, a computer power supply fried, smoking spectacularly…. The team adapted brilliantly, cannibalizing parts from a desktop [computer] that was being used for diagnostics. The second vehicle also performed its requisite runs with about ten minutes to spare. The PMs were very impressed with Team Cornell’s ability to adapt under a very adverse situation.”
The university teams seem to have several significant advantages over industry professionals. “Industry teams made up of paid workers tend to be smaller than school teams,” says Isaac Miller, a Ph.D candidate in mechanical engineering. He’s one of the few grad students involved in the project; about 90 percent of the team members are undergrads. “We not only have more manpower but more time. Kids are working 10-hour days, seven days a week. If you get a group of crazy undergrads together, they can do anything.”
“Technically, the team that completes the race the fastest will win,” says Matthew Grimm ’06 ME, head of the business team. “Realistically, whoever finishes will win.”
Grimm calls himself “the overall problem solver, the logistics man.” His focus has been on helping the team navigate the university bureaucracy and cultivating the kinds of large-scale sponsors a project like this needs. And large-scale sponsors have been found. AMD has provided substantial cash and equipment. Singapore Technologies (ST) donated the entire second vehicle, Titan. When the vehicle had engine trouble, ST sent an engineer to Cornell to assess the problem, shipped the truck to Detroit for repairs, then returned it to the team race-ready.
“It serves both of us,” says Grimm of the ST sponsorship. “DARPA is one of the huge contractors in the military; ST wants to break into the military market. This is a way for them to get their name and products into the forefront.” (Additionally, Cornell has the second-largest Singaporean college population in the U.S., after Stanford.)
The other vehicle, Code Red, was assembled out of different sponsors’ parts. “We never ask for gifts, just for at-cost or discounted donations, which are incredibly valuable. Companies have been exceptionally generous; in turn, we give them something to brag about.” Grimm estimates the team has raised approximately $90,000 in cash and $250,000 in in-kind donations, all in less than one year.
The three major components involved in mastering the Grand Challenge are the artificial intelligence system, the data-gathering sensors, and the vehicles themselves. “This whole project is a system of systems, as well as one of getting all the systems to talk to each other appropriately,” explains Garcia. “There are servomechanisms that control things like the steering and the brakes; I think of this level as a vehicle’s musculoskeletal system. At the next, middle level there’s a GPS (Global Positioning System) for basic path planning. With that come the LIDAR (Light Detection and Ranging) system and the stereoscopic vision that the car will use to process information collected from the world. This is a higher-end thinking in which the vehicle will see the landscape and the obstacles and fuse this information into its view of the world. Then the artificial intelligence (AI) will make decisions about where to go in that world by massive computation—algorithms based on cost: ‘In which direction will I be least likely to get stuck?’ A vehicle with an accurate view of its world will have a huge advantage.
“Realistically, the battle will be won or lost by how good the AI is,” says Garcia. “It needs the ability, for example, to navigate complexities like the difference between a dense shrub and a rock. The LIDAR, along with other sensors, must answer the question does this object flex? Is it a bush or a rock? There’s a decision tree constantly being generated: ‘What if…what if…what if?’”
rian Schimpf ’06 ORE is the AI team leader. “A lot of teams had trouble last year because their systems couldn’t handle sudden changes; they were designed to avoid what was right in front of them,” says Schimpf. “The path-planning systems we’re implementing are much more intelligent—they’re not just dealing with immediate problems. If the vehicle has a small rock in front of it and a corridor ahead that swerves to the right, our AI will synthesize this information completely. Also, our AI will learn as it’s going, adapting simulations in real time in response to the terrain. Being able to adapt is a huge advantage.”
Team Cornell installed the same sensing configuration and AI algorithms in both vehicles. As Schimpf explains, “We put all our energy into designing one set of code.”
For data-gathering, the Cornell team is using a versatile combination of sensors. The vehicle will integrate a GPS, LIDAR, stereoscopic vision, and tactile feedback to provide the vehicle with information. These sensor systems act as eyes, providing the input needed to create a model of the world. Unlike the human eye, these sensors operate on a variety of wavelengths, acquiring much more information than even the most experienced driver can gather.
The racecourse itself is defined using GPS waypoints, so an accurate GPS receiver is essential. Most commercial units have 20-foot accuracy, and while that is suitable for most uses—hiking in the woods or giving directions in a car—Cornell’s receiver and antenna have an accuracy of 10 centimeters. The receiver also uses a signal from a subscription service called OmniSTAR that provides corrections to improve accuracy further. When the GPS service goes out or becomes less accurate, for instance in a valley, the vehicles will also use an inertial navigation unit to pinpoint position.
The stereoscopic image processing, which mimics a human driver’s eyes, will be used for intermediate and short-range obstacle detection, with original algorithms providing dramatically improved speed and object resolution. Motion parallax (a form of motion flow) will allow the cars to identify objects separate from the ground plane using sequential video frames from a single camera. These two complementary techniques will compensate for the shortcomings of each. The vehicles will also employ texture recognition to distinguish between penetrable and impenetrable obstacles.
The data from each individual sensor (LIDAR, stereo vision, GPS) are combined into one large map of the world, which the AI will use to track where the road is, where obstacles are, how big the obstacles are, where the vehicle is on the race course, and the current status of the vehicle (for instance, how hard it’s accelerating). A path-planning algorithm uses the data to decide on the best way to navigate to the next waypoint, and a separate function, called the system controller, translates the world model into actual driving commands: how much throttle is needed, for example, or which way the vehicle needs to steer.
The final crucial ingredient for success: a vehicle that can reliably travel the 170 desert miles that Cornell hopes will incapacitate their competition. Riding on 38-inch tires, Code Red can handle virtually any terrain that could be encountered. Its suspension geometry was originally designed for rock crawling, an extreme variant of off-road racing. Titan is based on a military light strike vehicle, designed to handle demanding combat situations, and is the perfect host for the sensing platform, lightweight and durable, allowing it to support sensors and computers without sacrificing performance.
“I do not think that these cars finishing this course is an impossible goal,” says Garcia. “Once a vehicle goes 10 miles—if its decision logic is that good—the next 10 miles will be easier. This is an incredibly difficult technical problem, if a team like Carnegie Mellon, with 20 years of robotics experience and 15 years of government contracts, can only make it seven miles. It’s true this year they have the edge of experience—they’ve seen things firsthand. But they also didn’t do very well, so maybe there are ruts in their thinking.”
“Carnegie Mellon represents the opposite end of the spectrum,” says Huttenlocher. “They have a huge robotics department, a huge budget, faculty research scientists, and professional engineers working on this. Cornell is taking the approach of general skill and spirit. And Ephrahim Garcia is a great role model, an energetic, positive guy who doesn’t think ‘no’ is ever the right answer.”
On June 6, DARPA announced the 40 teams selected to advance to the semifinals of the Grand Challenge competition. Those teams will compete head-to-head in the National Qualifying Event at the California Speedway in September, where half will advance to the finals.
Given only one entry slot for the semifinal round, Team Cornell had to make a choice between their vehicles. “We decided to enter the Titan since the design makes it easier to mount all the large components, such as computers and generators, that we need to include,” explains Schimpf. Although Code Red won’t go to the competition, it was nevertheless critical to the team’s success. “There was about six-week period in which Titan was out for repair,” Schimpf says. “Without Code Red, we would never have been able to finish the software and sensor systems in time for the site visit.”
The Cornell team has scheduled six full weeks of practice in the desert, testing their algorithms in the field. “I think the main mistake teams made last year is that they didn’t spend enough time practicing,” Garcia says. “But in terms of winning, I actually think we’ve already won. The educational goals these students have achieved cannot be underestimated; they’re enormous, much more than engineers in a corporation would gain in a year.”
