A real-time implementation of a neural-network controller for industrial robotics

This thesis outlines the development and implementation of a real-time neural network in a commercial robot controller. The research is based on the modular MOVE neural network architecture (Graham & D'Eleuterio 1990), and the underlying Cerebellar Model Arithmetic Controller (Albus 1971). The neural network has been implemented on the CRS Robotics C500 Controller for operation of a 6-degree-of-freedom CRS A465 robot. Rather than replacing the standard independent-joint, fixed-gain PID control algorithm, the neural network augments and enhances the existing equipment. All six joints are controlled by the neural network.; It has been demonstrated that significant improvements in tracking errors can be made by the additional neural network. Tests have shown that RMS joint position errors, as well as peak joint position errors, can be reduced by 30%--90% compared to the baseline controller. Although velocity RMS errors are generally reduced by up to 30%, high frequency oscillations around the commanded signals sometimes occur, leading to a larger overall velocity RMS error.; Before details of the implementation are given, a neural network foundation is outlined. This review introduces both the concept of computations in neural networks, as well as their applicability to real-time control. The neural-network controller architectures are introduced.; All steps leading to the implementation on the C500 Controller are detailed. These are: the development of the neural-network controller architecture on a simulation model of a planar two-degree-of-freedom manipulator, the development of the parallel distributed processing model of the neural-network controller, and its implementation on the Radius testbed, and finally, the implementation on the commercial robot controller.; It is hoped that this thesis increases industry awareness of the benefits that a neural network enhanced controller can bring to existing applications. Neural networks are ready for commercialization in the real-time robot control environment of industrial automation.