An online walking pattern generation is important for a biped humanoid robot to move in dynamic environments. This chapter describes a novel online walking pattern generation using Fast Fourier Transform (FFT). Most previous studies about online walking pattern generation use an inverted pendulum model, which requires other methods to compensate for errors between the model and a real robot. Because our method uses a multi-body robot model, a biped robot can dynamically change its motions online by utilizing only the proposed method. If a robot has external sensors such as a stereo vision system to recognize dynamic environments, the robot can follow a moving target. Verification of the proposed method is conducted through experiments with the human-sized humanoid robots WABIAN-2R and KOBIAN.