The abilities of massive computation parallelism and self-learning make the neural networks a promising candidate for intelligent robot control. In this investigation, a robot controller consisting of inverse kinematics and inverse dynamics algorithms, has been replaced by two neural networks. A different approach to the neural networks learning phase made this solution applicable for robots control from the point of view of learning time as well as accuracy. First, the learning algorithm has been changed from the traditional back-propagation method to the quasi-Newton method. Secondly, the initial neural network weights which usually are chosen arbitrarily, have been determined in a systematic way based on a geometric interpretation of the neuron function. Simulation results show the significant improvement of both learning time and accuracy, which practically enables the use of a natural network controller in robotic applications.