This work proposes a new control approach for biped walking robots. Its purpose is to make human-like robots walk more smoothly and more efficiently with regard to energy. Thus, it is based on the decomposition of a step into two phases: a control phase which prepare a ballistic phase. As a first step towards more complex studies, the tools are simple and efficient: Lagrangian model, Newton's impact law, non-linear quadratic optimization problems used for trajectory planning and partial feedback linearization used for trajectory tracking. Although the final prototype will be the biped robot SHERPA, this control law has been implemented and tested on a simpler one: the cart-table. Numerous simulation results are presented with two concrete examples.