Off-line programming of robots requires the determination of the joint coordinate angles. However, inverse kinematic solutions for robot control and positioning are intended for the "ideal" robot and, therefore, do not take into account the robot's inherent deficiencies. These deficiencies include inaccuracies in the arm-element dimensions, internal play of joints, internal non-linearities of gearing, deflection of arm elements, and servo-positioning errors. Also, traditional approaches to solve such problems are computationally intensive and require frequent calibration to maintain positional accuracy. In this paper we report a technique that uses a neural network to learn the idiosyncrasies of a robot. In particular, a multi-layer feed-forward network is trained by the back propagation algorithm to learn the inverse kinematic solution that is unique to the Rhino robot used here.