Robust model-based neural network control

A particular type of robust model-based control, the robust model-based neural-network controller (RMBNNC), is proposed and experimentally evaluated. It combines feedforward payload adaptation based on multilayer perceptron artificial neural networks with a robust feedback controller design based on pseudocontinuous-time analog quantitative feedback theory (QFT). The neural network payload estimator adapts the controller to payload variations while the QFT design process implicitly accounts for the uncertainty in manipulator dynamics due to unmodeled drive system effects. The artificial neural networks are trained through repetitive training on multijoint trajectory tracking error data to provide an estimate of payload. QFT feedback is implemented by a series of simple backwards difference equations. The result is a computationally efficient direct form of robust adaptive control. Tracking performance was experimentally validated on the first three links of a PUMA-560 robot over a standard set of test conditions. The performance improvement potential and limitations of the RMBNNC approach are illustrated