This is a study of an application of neural technics to the learning of control laws within the framework of the evolutionary design of robotics systems. The present paper proposes the replacement of the evolutionary synthesis of the individual's control law by its learning. The learning of neural controller is carried out on-line when the robot undergoes evaluation tests. Thus, a robot that is a priori inadequate to solve a task can, thanks to the training it goes through, improve its performance. It participates then to the global improvement of the population while it would have been eliminated without learning. A mobile robot that could be equipped with up to 4 independent driving wheels and that must attain a given configuration will be taken as an example. The whole unit uses a simulation of the robot and its environment in which all dynamic effects are taken into account. Results show the accuracy and strength of the method since even the structures which would have been in fact eliminated to carry out this kind of task, are controlled with reasonable efficiency.