DARPA Grand Challenge update #3: Interview with Team Cornell

Improving software code

The software walks the vehicle through all the available paths and takes the best one that minimizes steering wheel turn. Grimm says that there are two ways of avoiding obstacles and the vehicle must balance between the two approaches. In some conditions, the vehicle will avoid obstacles at all cost, which results in repeated stop and go motion commonly seen in last year's competition. Grimm says, "Going forward and turning on a dime at 20 mph not going to happen." In other cases, smooth glancing turns are needed and the software walks itself through the paths that require the least turning. "We run numerical simulations using simple Newton's formulas which compute if the vehicle will slip, spin or roll over," says Miller. If any of these occur, the path is thrown out.

Miller told us that the sensor package and path finding software is so effective that the vehicle can see up to four seconds and 80 meters in advance. According to Miller, the vehicle could theoretically run with just the center LIDAR, but if the team had unlimited money, he would install two LIDAR in front of the wheels and one in the back. Unfortunately, the team doesn't have unlimited funds and has spent $145,000 so far. "We raise all of our own money and we can't just go up to daddy and ask for $40,000 or $50,000," said Grimm.

Spider keeps all of the terrain data in memory, which has lead to some early unexpected behavior. On a practice run on campus, the team told us, someone walked in front of the vehicle and the collision avoidance code stopped the vehicle. On subsequent runs, the vehicle would stop in the exact same place, even though the person wasn't there, because the obstacle information was never erased from memory. To solve the problem, Miller wrote a new algorithm to allow recent data to override old data. He says, "Old data will decrease in confidence exponentially, so the computer will be 10 percent less confident about data seen two seconds ago. It has some idea about the old terrain, but is very willing to replace it."

Team members told us the vehicle can back up even though it doesn't have any rear-mounted sensors. Since the terrain data is saved in memory, the software simply flips the dynamic information. Aaron Nathan, a Cornell team member said, "We hold the entire race in memory, if you had to, you can go all the way to the beginning," said Miller.

According to Grimm, STK has provided substantial support to the team. While the outside of the vehicle may be indestructible, the transmission and engine have been on the verge of death. Grimm told us that it takes 3500 rpm to go anywhere and that the vehicle needs to be redlined to back it up into the garage. "The transmission fluid is pasty grey-black with visible metal shavings," said Grimm. STK was informed about the problem on Thursday and in that same evening an engineer was being flown out from Singapore with a new transmission as checked baggage. Cornell plans on putting in the new transmission after completing their second round qualifications. Earlier in the year, STK sent the same engineer to help replace a bad engine. "STK told him to not go home until the engine worked. He stayed here 7 weeks," Grimm said.