Abstract:

The main purpose of the thesis is to design a learning-based controller for robotic manipulators. This controller can estimate the nonlinear system dynamics to minimize tracking errors in motion and eventually achieve a zero-tracking-error performance. Among learning-based control techniques, the adaptive neural network control has an on-line learning ability, and the cerebellar model articulation controller (CMAC) has the properties of rapid convergence, lower computational complexity, and local generalization, which are advantageous to allow the microcontroller to execute the control algorithm in real-time. In order to prevent from getting stuck in local minima and have faster learning convergence, a new CMAC controller is proposed. The controller consists of two main approaches: a grey learning rate and a modified tracking error. The grey learning rate, which is based on a grey relational analysis, is utilized to adjust the learning rate on-line. The modified tracking error, which is defined according to synchronization control, can achieve asymptotic convergence of both tracking errors and synchronization errors simultaneously. To demonstrate the performance of the proposed controller, ADMAS and MATLAB/Simulink are used for simulation. The NI sbRIO-9642 is employed to realize the control algorithm on the NTU arm, which is developed by our laboratory. In comparison with conventional controllers, the proposed learning-based controller can provide better tracking performance.