Our research group recently submitted a paper on on-board, evolved path following to the journal Evolutionary Intelligence. Very briefly a human drives the robot for a few minutes on a training track, during which time the robot (we used the Rover 5 chassis with an Android smartphone) collects data on the color pattern of the track and information on what it was seeing when the drive chose to turn or go forward (image-action pairs). After a few minutes of driving the robot paused and ran an on-board evolutionary algorithm to train a NN network to learn how to imitate the human driver’s actions.
About 5 minutes of evolution was sufficient for the robot to navigate fairly well. We tested it on novel tracks with much sharper turns than it saw during training. It was ~95% successful on the training track and ~50% successful on the novel test track, with most of the failures occurring on the sharp corners it hadn’t seen before. I.e. just a few minutes of training and evolution is sufficient to train a robot to follow a path fairly well.