The model-based control of humanoid robots requires the availability of accurate mechanical models that can be hard to obtain in practice. One approach to this problem consists in calling upon machine learning methods. In this paper, using a standard control approach based on visual servoing, we compare the accuracy of two supervised learning methods, namely LWPR and XCSF, to extract the forward velocity kinematics of the upper body of the iCub robot. Experiments are performed in simulation, using one arm and the head for reaching tasks. We show that both methods provide accurate models of the robot, with a slight advantage to XCSF over LWPR.