ALVINN drives a van with a neural network

ALVINN, the Autonomous Land Vehicle in a Neural Network, was a 30-by-32-pixel camera image fed into a network with one hidden layer of a few units and thirty outputs, each corresponding to a steering angle. Dean Pomerleau trained it by driving Carnegie Mellon's Navlab, a converted Chevrolet van full of computers paid for by DARPA's Strategic Computing programme, and recording what a person did with the wheel. After a few minutes of examples the network could keep the van on the road by itself, at first at walking pace and by the mid-1990s at motorway speeds across most of the width of the United States.

It was imitation learning before the term existed: rather than program the rules of driving, show the machine a driver. Pomerleau also had to invent tricks that are still used, generating extra training examples by shifting and rotating the camera images so the network would learn to recover from positions the careful human driver never got into.

The self-driving cars of the 2010s and 20s use the same idea at scale, cameras into a network trained on human driving, with three decades of compute and data behind it. ALVINN's network had fewer parameters than a modern model has in one attention head.