Shoestring budget won't stop Princeton’s Urban Challenge team
Princeton (NJ) – Princeton University students are hoping that a donated car, an eBay-bought GPS and a lot of hard work are enough to win the upcoming DARPA Urban Challenge race. Despite being short on cash, the all-undergraduate team has developed some impressive software to power their “Prowler” car through the city course. In a phone interview, team spokesman Gordon Franken explains how Princeton is able to compete with other, much more well-funded, teams.
Princeton's "Prowler" Ford Escape Hybrid
Franken, a senior majoring in Mechanical Engineering, said the team’s vehicle choice was actually quite simple. “We picked it because it was free,” said Franken. Ford donated a 2005 Escape Hybrid after the team talked with several Princeton alumni, who also worked at Ford.
The donated car was a huge help for the cash-strapped team and it saved Princeton around $25,000. Franken estimates that this was actually more than the team has spent on sensors. While other teams will rely on a multitude of sensors to guide their robotic cars through the course, Prowler only uses three stereo video cameras to see and avoid obstacles. Franken says three more cameras may be added.
“Other teams have a triple play combination of spinning lasers, high-precision GPS and electronic mobility control,” said Franken. But all this firepower can cost more than $150,000, according to Franken, something that his team just cannot afford.
The sensor data feeds into five custom-built computers mounted in a shock-isolated rack in Prowler’s trunk. All the computers are identical which allows the team to easily swap failed components. Interestingly enough, these computers are using hard-drives which could be prone to head crashes from all the jostling around, but Franken said this was much less of a concern than in the desert-bound Grand Challenge race.
“We feel in the urban environment that the overall speed is substantially less than 30 mph, if your vehicle has a collision, a hard drive failure is the last of your concerns,” Franken told us.
For navigating the waypoints, Princeton is counting on a single GPS unit that was bought on eBay for just $150. Because the team isn’t using any high-precision inertial navigation or differential GPS units, the students have had to focus on making the stereo vision system as accurate as possible.
“From the beginning we didn’t really want to rely on GPS sensors.”
Princeton loves the stereo “Bumblebee” cameras from Point Grey because they provide a lot of quality data for low cost. Franken said the cameras can see out to 60 meters and the team can write just one obstacle detection algorithm and then apply it to all the cameras.
The team also likes the cameras because they are a passive sensor and don’t rely on beaming out laser or radar waves. In addition, the cameras shouldn’t be affected by stray laser or radar beams coming from a competing vehicle. Bill Kehaly from Axion Racing, one of Princeton's competitors, told us in an interview that the extra radiation being flung around the course could cause cars to act erratically.
Despite the advantages, there are some limitations with using stereo cameras. Franken said the maximum range is approximately 60 meters, but there are instances where the car will need to see farther. Stereo cameras work on the disparity between images on the right and left cameras, but at a far enough distance both cameras get the same picture and the disparity becomes zero.
There are two ways of getting around this problem. First the resolution can be increased which will provide more pixel distance before the left and right images merge. The downside to this, according to Franken, is that the extra pixels require much more computer power to process the results.
Franken said the second way of seeing farther is to actually change the optics of the camera and focus farther down the field. This is similar to how human drivers constantly changing their focus from cars up close to other cars down the road or freeway. The problem here is that you narrow your field of vision and could end up crashing into a close object.
Stereo cameras also require a lot of texture in the scene to differentiate objects. Trees are easy because they have braches and leaves, but cars and roads can sometimes cause trouble. Franken explained that long stretches of road can have uniform color and texture and cars often have flat shiny surfaces. Fortunately, detecting other competitors’ cars won’t be a problem.
“Everyone has sponsor stickers plastered on their cars… that’s going to make it very easy for us to see,” said Franken.
Princeton has completely rewritten their path-finding and obstacle avoidance software to deal with the city-based obstacles of the upcoming race. The robotic cars must also dodge moving cars. “In the Grand Challenge you could make the really solid assumption that everything was stationary like bushes, rocks and parked cars, but in this race we assume that everything is moving,” Franken told us.
When the computer and sensors detect an obstacle, the Prowler will wait about five seconds for it to move. If the object is still there, then the car will treat it as a stationary object and drive around it. The Princeton team can fiddle with the waiting time, but Franken says they want their robot to be more cautious than aggressive.
“People don’t think this, but driving down Main Street is much more difficult than the desert.”
Since the Prowler doesn’t use expensive inertial navigation, the team has written some ingenious algorithms to continuous correct the car’s course. Regular GPS units, like the one Princeton is using, will eventually drift off the programmed course due to clock errors and interference from trees and buildings. Initially the difference will be a few inches, but could eventually grow to dozens of feet if the problem isn’t corrected. Princeton plans on stopping those errors by resetting the GPS readings at many of the course intersections.
All of the robotic vehicles must stop at so-called “stop lines” marked by GPS waypoints. These spots correspond to stop sign controlled intersections you see on everyday streets. These intersections are also marked by large white lines in the road, lines that the Prowler’s cameras can detect.
As the Prowler pulls up to the stop line, computers will reset the car’s GPS coordinates to the waypoint coordinates given by DARPA. In effect the team is using the intersections as a very inexpensive differential GPS system.
Despite the good software, Princeton almost didn’t make it to the semifinals because their car was only able to complete 50% of the runs during DARPA’s site visit back in July. Franken blames the poor showing on the intense heat and small software and sensor bugs that crept up during the day. The team hammered out all the problems over two weeks and filmed a successful run from several different camera angles. Princeton then hand delivered the video to DARPA headquarters.
Franken told us that the team has survived on the low budget by getting a fair amount of their technology for low prices or even free. But of course, Princeton could always use extra equipment or money. After all, it’s going to be very expensive to send all the students and the car over to Victorville from Princeton.
“We’re probably going to spend more on logistics for the race than on the technology and the vehicle,” Franken said.