| Currently deployed robotic manipulators are handicapped by the use of rigid structural members or links. The mass required to make each link rigid limits the acceleration and, consequently, the speed at which a manipulator can perform a given task. Rigid-link requirements set upper limits on the size and workspace of current manipulators. Replacing massive, rigid links with light, flexible links offers the potential to improve both the performance and application of robotic manipulators.; The application of artificial neural network (ANN) methods to the control of these 'flexible-link' manipulators has been limited because they used 'rigid-link' solutions. Although the first study of ANN learning by McCulloch and Pitts occurred over 50 years ago, the use of ANNs did not become popular until more recently. Researchers have applied a variety of ANN technologies to the control of mechanical structures, and survey papers specifically related to ANN control of rigid mechanical systems are found in the literature. The use of ANN learning control for nonminimum phase systems has also been studied, although less extensively. In general, the nonminimum phase studies related to robotics have focused on a particular class of systems, flexible-link manipulators.; The purpose of this investigation was to explore the feasibility of using ANNs to control a flexible-link manipulator and cause it to follow a desired trajectory. Two ANN-based control architectures were investigated: an adaptive-type architecture, where learning was performed on-line after each sample, and a learning-type architecture, where learning was performed off-line. Tracking performance for both architectures was demonstrated and characterized using both simulated and experimental manipulators. |
