Neural network near-optimal motion planning for a mobile robot on binary and varied terrains

Abstract

Presents an efficient approach to plan a near-optimal collision-free path for a mobile robot on binary or varied terrains. Motion planning is formulated as a classification problem in which class labels are uniquely mapped onto the set of maneuverable robot motions. The neural network motion planner is an implementation of the popular adaptive error backpropagation model. The motion planner learns to plan 'good', if not optimal, collision-free path from supervision in the form of training samples. A multi-scale representational scheme, as a consequence of a vision-based terrain sampling strategy, maps physical problem domains onto an arbitrarily chosen fixed size input layer of an error back propagation network. The mapping does not only reduce the size of the computation domain, but also ensures applicability of a trained network over a wide range of problem sizes.<>