The application of artificial neural networks to tactile robot visualization

Abstract

Artificial Neural Network (ANN) computation is relatively new as an applications technology and is based on the paradigm of massively parallel distributed processing where the computation is performed over a large number of interconnected simple processing elements. The aim of this thesis is to apply ANN techniques to tactile robot visualisation which is the tactile sensing and perception of objects in a robot's workspace. The thesis is interdisciplinary in nature, combining topics from robotics, pattern recognition, ANNs and aspects of industrial engineering. The main focus is on robot trajectory planning, control of the tactile data acquisition process, and shape recognition, with an emphasis on providing a framework for the utilisation of inexpensive tactile sensors by industrial robots and the application of ANN techniques. Since every industrial process will have a different set of specifications, the development of the trajectory planning and shape recognition systems are addressed from a generalised point of view which may then be tailored for specific operational systems using the methodologies developed. A new angle-length perimeter (ALP) structural shape representation for tactile shape recognition was developed, implemented and tested within the prescribed framework. The ALP representation was shown to offer a consistent representation of two-dimensional polygonal shapes from highly variable tactile data. This enabled the modular ANN recognition systems that were developed to achieve good results. The general guidelines Iaid down in this thesis form a good basis for some general heuristic rules in the use of ANNs for solving complex problems, as well as a foundation for hardware implementation.