Our ultimate goal is introducing energy-efficient walking patterns to actual humanoid robots. It is known that Limit Cycle based walking methods, e.g. Passive Dynamic Walking, have such a desired property. Unfortunately, the methods have been limited to simple planar biped models. In this paper, we propose a way of extending Limit Cycle based walking pattern generation to apply it to a 7DOF 3D biped with ankles, knees, an upper body and with flat feet. Our approach decouples roll and pitch motions in the frontal/sagittal planes, respectively. We derive the limit-cycle based walking pattern with the help of an extended planar model including ankles, knees and the upper body. Robustness of the motion is ensured via a feedback control method based on mechanical energy. The performance of the walking pattern generator in 3D and the controller is confirmed via numerical simulations.