ALVINN

Carnegie Mellon's 1988 driving neural network, which maps camera images of the road directly to a steering command.

ALVINN is work by Dean Pomerleau at Carnegie Mellon University, published at NIPS 1988. Self-driving at the time relied on hand-designed image-processing rules that broke easily when lighting or road conditions changed. ALVINN used a backpropagation network with just one hidden layer of 29 units instead: it took in a 30×32 camera image and an 8×32 laser range-finder image, and produced 45 output units representing the steering curvature to follow, with no hand-written road-detection rules in between. The network was first trained on 1,200 synthetically generated road images, then installed on CMU's NAVLAB test vehicle, which it drove autonomously at 0.5 m/s along a 400-meter wooded path on campus. It's often regarded as one of the earliest examples of behavior cloning and end-to-end driving, and NVIDIA's 2016 end-to-end driving system, DAVE-2, cites it as an inspiration in its paper.

ExampleFor each frame of road image it reads in, the network's 45 steering output units light up most strongly at the position that determines direction: the center unit means drive straight, and units further toward either side mean turn more sharply left or right.

Also called

Autonomous Land Vehicle In a Neural Network

Related

End-to-End · Behavior Cloning · Autonomous Driving · Imitation Learning · Multilayer Perceptron · Tesla FSD v12

Sources

ALVINN: An Autonomous Land Vehicle in a Neural Network (NIPS 1988 proceedings)
End to End Learning for Self-Driving Cars (NVIDIA, arXiv 1604.07316)

As of

2016-04