ADAPTIVE NEURAL NETWORK CONTROL OF FLEXIBLE ROBOTS

Abstract

Flexibility of joints or links limits the performance of robots. Adaptive neural network controllers are presented in this thesis for two classes of flexible robots, namely, the flexible joint robots and the flexible link robots. The proposed controllers are developed by applying adaptive neural networks to well-established conventional flexible robot control schemes so as to improve their robustness to parameter uncertainties and unknown model changes. Suitable neural networks are first used to parametrize the control laws and direct adaptive techniques are used to update the weights of the neural networks on-line. The introduction of neural networks removed the need to find the linear-in-the-parameters (LIPS) model between the known nonlinearities and the unknown parameters as in conventional adaptive techniques; and the use of direct adaptive methods eradicates the time¬ consuming off-line training of neural networks and at the same time provide adaptive enhancements to dynamics changes. The first adaptive neural network control scheme for flexible joint robots is designed based on the feedback linearization approach. Adaptive neural networks are added to the conventional model-based or neural network based feedback linearization controller. The second scheme is based on model-order reduction using the singular perturbation approach. Taking advantage of the natural time-scale separation, the full model is decomposed into two reduced-order subsystems, namely, a slow subsystem and a fast subsystem, which are separately controlled using the composite control strategy. The adaptive neural network composite control is thus developed which consists of a slow adaptive neural network control and a fast control to damp the joint elasticities. This adaptive neural network composite control scheme is further extended to the case of flexible link robots where the slow joint variables are controlled by the slow adaptive neural network control and the tip deflections are effectively damped out by the fast control. For both approaches, a robust control is included for closed¬ loop stability in the presence of neural network reconstruction errors. Rigorous stability analyses are provided for the proposed adaptive neural network control schemes and it is shown that stable adaptation is assured and asymptotic tracking of the position reference signal is achieved. Intensive computer simulation results are included to verify the effectiveness of the proposed adaptive neural network controllers in handling unknown dynamics changes as compared to the respective conventional flexible robot control schemes.