Adaptive neural network for identification and tracking control of a robotic manipulator

Effective control strategies for robotic manipulators require on-line computation of the robot dynamic model in real-time. However, the complexity of robot dynamic model makes this difficult to achieve in practice. Neural networks are an attractive alternative for identification and control of robotic manipulators, because of their ability to learn and approximate functions. This paper presents the development of an adaptive Multilayer Neural Network (MNN) as a feedforward controller for a robotic manipulator. The MNN is trained to identify the unknown nonlinear plant (inverse dynamics of a robotic manipulator) using a modified back-propagation technique. A PD controller is used in the feedback loop to guarantee global asymptotic stability. Also, the output of the PD controller is used as a learning signal for the on-line learning to adjust the weights of the MNN to capture any parameters variation and/or disturbances. The controller architecture developed has been simulated and its effect on the trajectory tracking performance of a manipulator has been evaluated and compared to the conventional adaptive controller.