This paper introduce an original approach to the Neural Network learning architecture for the control of robotics systems. The basic idea is to use multi-layer-network and the backpropagation algorithm without desired outputs, but with a quadratic criterion which spezify the control objective. Use a criterion avoid the computing of the inverse model of the robot which is required in the case of the desired outputs and difficult to obtain for complex systems. The method allows on-line learning that is impossible with classic backpropagation. We argued that on-line learning scheme made feasible adaptive control. To illustrate the idea of a control criterion on backpropagation algorithm, we first considered a cartesian control problem of a simplified simulated planar manipulator. Next, we presents the results of a more difficult problem where the goal is to control the dynamic stability of a simulated planar biped. In this study, the results of the on-line learning scheme validate the adaptive control hypothesis. Finally, we outline a more real problem, that is to control the dynamic equilibrium of a quadruped robot. We are specially studying reflex behavior to control unexpected situations.