| This thesis investigates neural-adaptive control of a two-link robotic manipulator, based on estimating the forward dynamics of the system. Modeling the forward dynamics is achieved on-line using the Cerebellar Model Articulation Controller (CMAC) associative memory neural network. The CMAC takes advantage of any previous knowledge of the dynamics of the system while compensating for external disturbances. This method contrasts previous results in the literature that use an approximation of the inverse dynamics, and to the best of the author's knowledge, is a new scheme that uses forward dynamic approach to compensate for both the structured and unstructured uncertainties on-line. The weight updates take advantage of the neural network property that a nonlinear function can be uniformly estimated using different sets of weights, and guide the weights toward their real values using previous knowledge of the system.; In this thesis, three different approaches of forward-dynamic CMAC control are compared: state error-based weight update, supervised learning, and combined state error and supervisory teaming. All the approaches are shown to be Lyapunov-stable (Ultimately Uniformly Bounded). Simulation experiments are conducted with a two-link rigid planar manipulator arm that tries to follow a desired trajectory and adapt to payload changes. The conclusion is that training based on Lyapunov-stable weight updates combined with supervised teaming using a model estimate of the inertia matrix is sufficient to achieve neural-adaptive control using a forward dynamics approach. |
