Animal walking is one of the most robust and adaptive locomotion mechanisms in the nature, involves sophisticated interactions between neural and biomechanical levels. It has been suggested that the coordination of this process is done in a hierarchy of levels. The lower layer contains autonomous interactions between muscles and spinal cord and the higher layer (e.g. the brain cortex) interferes when needed. Inspiringly, in this study we present a hierarchical control architecture with a state of the art intrinsic online learning mechanism for a dynamically walking 5-link biped robot with compliant knee joints. As the biological counterpart, the system is controlled by independent control units for each joint at the lower layer. In order to stabilize the system, these units are driven by a sensory feedback from the posture of the robot. A central stabilizing controller at the upper layer arises in case of failing the units to stabilize the system. Consequently, the units adapt themselves by including online learning mechanism. We show that using this architecture, a highly unstable system can be stabilized with identical simple controller units even though they do not have any feedback from all other units of the robot. Moreover, this architecture may help to better understand the complex motor tasks in human. © 2018, Institute of Control, Robotics and Systems and The Korean Institute of Electrical Engineers and Springer-Verlag GmbH Germany, part of Springer Nature.