A Robust Wall-Following Robot That Learns by Example

We designed and implemented a wall-following robot that learned appropriate actions from a small set of noisy training examples. We then assessed the effectiveness of this strategy by physically modifying the robot prior to training. Our system used nearest-neighbour classification to make motor command predictions, but we also tested two regression techniques for continuously varying action selection: linear regression and support vector regression (SVR). We found both classification and regression to be satisfactory prediction strategies for this task, although the advantage of SVR for predicting a continuum of actions became clear when the number of possible class labels was reduced.