This paper presents a computational model of cortical organization, enhancing the fact that self-organization in connected cortical modules have to be joint. This architecture is validated on the learning of the control of a robotic arm device, through sensori-motor integration. In order to get rid of the use of an inverse model of the arm, the self-organization of the whole network is conditioned by a reward signal. The point of the model is that this reward signal is given at the motor level only, but its effects are spread to the whole architecture by a side effect of the coherence keeping function in the cortical modules. First experiments on this approach are discussed.