This paper presents an original method in the use of neural networks and backpropagation algorithm to learn control of robotics systems. The originality consists to express the control objective as a criterion of which the gradient is backpropagating through the network instead of the classical quadratic error used in standard backpropagation. This technic allows on-line learning that is impossible to do with standard backpropagation. Experimental validation is realised by the position and the orientation control of a faster industrial mobile robot. Results show the feasability of the method, and particularly establish that on-line learning scheme permit to refine the weights of the network in front of the kinematics constraints of the robot.