Progress in neural network-based vision for autonomous robot driving

This paper describes recent improvements to the ALVINN system (Autonomous Land Vehicle In a Neural Network) for neural network based autonomous driving. The authors perviously (1991, 1992) reported a technique which allows an an artificial neural network to quickly learn to steer by watching a person drive. But the faster the network is trained, the less exposure it receives to novel or infrequent scenarios. For instance, during a typical four minute training run, the network sees few if any examples of passing cars. When a rare situation like this occurs during testing, its lack of coverage in the training set canresult in erratic driving. By modeling the appearance of infrequent scenarios and then using the model to augment the training set, one can teach the network to generalize to situations not explicitly represented in the live training data. Using this technique, a network trained over a two mile stretch of highway was able to drive autonomously for 21.2 miles at speeds of up to 55 miles/hour.>