We present a three level architecture for humanoid robot navigation. The perception layer creates a 2D environment model from range data and foot step information and classifies areas into one of six different types. The control layer is organized into behavior modules each representing different motion capabilities of the robot. Based on the environment classification, a suitable sequence of actions is selected for navigating the robot to a goal location. On the planning layer, collision-free paths are searched on the environment map and provided to the control layer. The system is modular as behaviors can be added and removed. We verify our approach on Sony's QRIO robot and present experimental results in an environment containing stairs, a table, and obstacles of different height