In a closed loop control system, a six-degree-of-freedom robot with a CCD-camera mounted on its end-effector, is operated. An object moves freely in 3D space(translation + rotation). The aim is to servo the camera on the object, so that the image of the object is always close to the reference image", defined by a given reference position of the object with respect to the camera. Classical kinematics equations are first studied in order to determine the significant parameters of the problem. Two neural approaches are then proposed: in the first solution, a Multi-Layer Perceptron (MLP) is fed with the image coordinates of feature points and with previous robot commands. In the second solution, different neural input parameters are used, that are based on affine transformations between succeeding image coordinates. Results and comparisons with the classical approach (uniform and non-uniform translations of the object) are presented, and show the interest of the approach vs. classical methods.