A dynamically sized Radial Basis Function neural network for joint control of a PUMA 500 manipulator

We present the design and analysis of a neural control structure for joint control of a PUMA 500 robot manipulator. We lay out the design considerations and steps to build an experimental electronic control system to control the shoulder joint of the manipulator. We review the use of neural networks for on-line learning closed-loop control applications. The ´curse of dimensionality´, a problem encountered when using Radial Basis Function (RBF) neural networks, is addressed and a neuron-node resource-allocating algorithm is investigated to overcome this problem. An on-line learning neural-control structure, employing this resource-allocating algorithm, is proposed, implemented and successfully tested to improve the position accuracy of the robot manipulator. All the implementations are executed on a 16-bit microcontroller in real-time, developed using integer arithmetic in the programming language C. The program listings are available upon email request.