The Looming Robot Menace, part I

Continuing with my earlier post about robotic research, I was reminded of the Hummer that was driven around campus when I was in school. CMU's robotics department was busy teaching a computer how to drive. A friend of mine who was a doctoral candidate in the Computer Science Department told me all about the use of neural networks and the theory behind this approach to artificial intelligence.

From what I remember, the Hummer had all sorts of input devices, mostly video cameras I believes, that recorded reams of data while a human drove the vehicle. The computer would look all of the data that was recorded, for example, when the human stopped the vehicle. At first, there would be no way for the computer to determine which data was significant, and which was not. The next time the vehicle stopped, however, that data could be compared to the data from the first stop. And so on and so on. Eventually patterns could be detected (like every time the video camera recorded an image of a red octagon with the white letter "STOP" on, it the human stopped the vehicle, therefore the computer could "learn" that a stop sign was consistent with the human stopping the vehicle)

The robotics department was pretty good at this, and last fall won a $2 million prize from DARPA in what to date has been the most prestigious robot race, the Urban Challenge.

The grounds of the former George Air Force Base in Victorville, CA, served as a mock city that the robots had to navigate. The course consisted of 60 miles of roads and parking lots and took about six hours to complete. The whole time, the robotic cars needed to obey traffic laws and avoid both cars driven by professional stunt drivers and the other robots on the course.

While DARPA is interested in autonomous vehicles for military purposes, obviously, the technology will likely be used to eventually add safety features to consumer cars.