Biped running can be conceptually reduced to a set of simple and quasi-independent tasks such as weight bearing, upper-body balancing, and energy injection through ankle push-off. We show in this paper that by appropriately designing multi-articular elastic actuators for biped robots in a manner inspired by human biomechanics, these tasks can be favorably expressed in a set of coordinates, in which the system is elastically decoupled. In these coordinates, the robot can be easily controlled by a set of simple and independent control laws. By exploiting the natural dynamics of the specially designed robot, the proposed controller requires only minimal model knowledge (mainly in terms of kinematic and static parameters) and is therefore robust to model uncertainties. It requires only state measurements and no measurement or model based computation of higher order state derivatives. Moreover, since the system is operated at a frequency dictated by the natural resonance, the running gait is energy efficient and resembles to a large extent natural human motion. Simulations validate the concept and demonstrate the independence of the approach from the knowledge of dynamics parameters.