Model-based compensation and comparison of neural network controllers for uncertainties of robotic arms

The paper proposes a decentralised compensation scheme for unstructured uncertainties and modelling errors of robotic manipulators. The scheme employs a central decoupler and independent joint neural network controllers. Recursive Newton Euler formulae are used to decouple robot dynamics to obtain a set of equations in terms of the input and output of each joint. To identify and suppress the effects of uncertainties associated with the model, each joint is controlled separately by neural network controllers. Gaussian radial basis neural networks, using the direct adaptive technique for weight updates, and multilayered perceptrons, using the backpropagation learning algorithm, are used as the adaptive elements in the control scheme. The effectiveness of the proposed scheme is demonstrated by controlling the trajectories of the three primary joints of a PUMA 560. Simulation results show that this control scheme can achieve fast and precise robot motion control under substantial model inaccuracies. Properties of both types of compensators are compared with conventional adaptive control, and suitability for real-time control is discussed