Fontana (CA) - The university teams at the DARPA Grand Challenge qualifications at the Fontana Speedway are doing well, with many of them already having mastered the qualification round. The team from Cornell University in New York swiftly conquered the course with its "Spider" vehicle. Join us on a visit in Cornell's team garage for a detailed look on what makes Spider one of the best performing vehicles in the DARPA Grand Challenge so far.
Like other teams, Cornell selected a rugged vehicle built for desert driving. A military light strike vehicle was donated by Singapore Technologies Kinetics (ST Kinetics). Matt Grimm, team member of Cornell, told us that "it's virtually indestructible" and that it provides "a forgiving platform that allows making mistakes." While other teams try to avoid obstacles at all costs, the ruggedness of the LSV allows Cornell to drive over most rocks and small objects.
The Cornell team is composed of undergraduate and graduate students led by faculty advisor Ehpraim Garcia. Garcia told us that he likes taking a hands-off approach to team management and mainly handles the administrative and safety issues. He says, "My job is trying to maintain safety.-After all this is an automatic robot that can kill you."
The team runs their C++ coded software exclusively on Windows 2003 Server. The team told us that there were thoughts to use a 64-bit Windows version to get around the maximum 2 GByte memory limits, but the team "couldn't find available drivers for the hardware." Multiple AMD Opteron processors and computers, all donated are at the core of Spider's computer system.
Grimm said that "it was fairly easy to establish drive by wire" on the vehicle, since the LSV's throttle already is electronically controlled and the gas pedal basically is a potentiometer, which gives readings that vary with pressure. A flexible mount and rotor had to be made for the brakes, while the steering wheel required extensive tooling. Through all modifications, the steering wheel can be turned from the most left to the most right in about one second. "If someone stole our car, they'd get whiplash," Grimm joked.
Compared to other teams, Cornell's Spider vehicle uses very few sensors. Three LIDAR (light detection and ranging) units are mounted on the back end of the hood, facing forward. The left and the right LIDARS are fixed and respectively focus 15 and 20 meters ahead, giving the software increased confidence whether an obstacle really exists. The center LIDAR is mounted on a gimbal which can pitch and rotate to see around corners.
The center LIDAR is the secret to Cornell's success and helps reduce blind spots when the vehicle takes a turn. As the vehicle turns, the LIDAR will turn and scan the path. This is similar to how a human being will turn their head and adjust their eyes to focus farther. In addition, the sensor makes large up/down oscillations when traveling at slow speed. This helps sensing all obstacles in debris laden areas. As the vehicle speeds up, the amplitude will narrow and focus the LIDAR further down the path, simulating "high visual horizon" - a technique used by professional drivers.
During Cornell's successful first round qualification run, the vehicle managed an impressive 35 mph on the straight away while hitting just one cone. It even avoided a misplaced hay bale that the Carnegie Mellon team hit in the run directly before. The team's computer science and sensor fusion guru, Isaac Miller, was instrumental in Spider's success as he developed the path finding code. The code plots out a cost map of the upcoming terrain, by creating a grid of upcoming terrain. The grid is divided into 40 X 40 centimeter squares and the height is computed from the raw sensor data coming from the LIDAR units. The height gradient is recorded as two numbers, a mean height and a variance. "We have accurate estimates of the world, down to centimeters," says Miller.
