User: DarrenO

Neural Network based road navigation

On: Ask Jordan Pollack About AI - Or Anything Else

Hi. I am involved with the Autonomous Robotic Vehicle Project at the University of Alberta. It is a student group that sets out to build autonomous ground robots for competition at the International Ground Vehicle Competition held in Orlando this year. The competition is for purposes of motivation of future engineers/scientists in the area of autonomous driving. The competition course follows a typical road layout of two white lines on a contrasting background with construction barrels, etc. placed throughout the driving area to test the AI of the robot's controller. I am involved heavily with the AI side of the project at U of A, and am presently in the process of developing a neural network based controller for our robot that takes a downsampled video image of the road in front of the robot and decides on a direction for the vehicle to move in. We are presently using a downsampled 8-bit grayscale image of 8x32 pixels that is fed into a back propagation trained network with all 256 pixels as inputs. There has been discussion of using a neural network with less inputs to get quicker convergence and have less required training points. These inputs would be extracted features from the picture like the membership of the particular input picture to a predetermined set of ~20 approximate vector directions. On this note, what do you suggest would create the most robust of neural network controllers: A neural network based on raw input data or a neural network based on feature extractions from input data? This is a subject that seems to be of much debate with neural network experts, but it would be nice to have some clarity on the issue.

Lastly, with the present state of AI technologies, what is your opinion on the likelyhood of autonomous cars becoming reliable and robust enough to become commonplace in the upcoming years? It would be neat to buy a autonomous car in my lifetime...

Thanks for your time.

Regards,