John Chamberlain
Developer Diary
 Developer Diary · You Heard It Here First · 2 December 2003
Can AI Write Algorithms?
The DARPA Grand Challenge off-road rally scheduled for next March will be an interesting test of current AI and sensor technology. The race will be up to 300 miles long and to qualify for the $1 million prize the winner must complete the route in 10 hours or less. To accomplish this the vehicle must go at least an average of 30 miles per hour--a stiff pace for off-road travel. There will be a lot of difficulties because the route will include minor obstacles like ditches and narrow defiles which must be transited. Will any one of the 19 teams be able to build a vehicle capable of doing this?...

It will be difficult to do this for several reasons. First of all notice that the average vechicle speed will actually be greater than 30 mph because slowing down to deal with obstacles will take time. For example, if you take 10 minutes to deal with each of 20 obstacles that will soak up an additional 3 hours. Including the pit stop the vehicles might have to travel over 40 miles per hour to have a chance at qualifying.

If your vehicle hits a 2-foot deep ditch with steep sides going 40 miles per hour it will not be a pretty picture. For this reason the AI on the vehicle must be able to recognize rocks and ditches in the path ahead. Right now this is not a solved problem. The human ability to look ahead and tell the difference between a ditch and a shadow or between a rock and a tumbleweed is something computers are not capable of doing, especially in real-time mode where the decision must be made in milliseconds. Remember that a video image is streaming in at 24 frames a second giving the computer a 50 millisecond window to analyze each image.

It will be interesting to see how the teams will deal with these problems and if they will come up with any innovations.

Developer Diary · · bio · Revised 15 January 2018 · Pure Content