Most of the biped robots are controlled using pre-computed trajectory methods or methods based on multi-body dynamics models. The pre-computed trajectory-based methods are simple; however, a system becomes highly vulnerable to the external disturbances. In contrast, dynamic methods make a system act faster, yet extensive knowledge is required about the kinematics and dynamics of the system. This fact gave rise to the main purpose of this study, i.e., developing a controller for a biped robot to take advantage of the simplicity and computational efficiency of trajectory-based methods and the robustness of the dynamic-based approach. To do so, this paper presents a two-layer hierarchical control framework for an under-actuated, planar, five-link biped robot model. The upper layer contains a centralized dynamic-based controller and uses all system sensory data to generate stable walking. The lower layer in this structure is a decentralized trajectory-based controller network, which learns how to control the system based on the upper layer controller output. When the lower controller fails to control the system, the upper layer controller takes action and makes the system stable. Then, when the lower layer controller gets ready, the control of the system will be handed to this layer. © 2018 Sharif University of Technology. All rights reserved.