The authors propose a simple on-line method for generating a walking pattern for the biped humanoid robot KHR-3 (HUBO). The problem of realizing a walking action in humanoid robots involves two components: generation of the basic walking pattern and the compensation required to maintain the robot's balance. Dynamic walking can be realized by incorporating the real-time stabilizing control algorithm developed for KHR-1, KHR-2 and KHR-3. The walking pattern of KHR-3 has four modes: forward/backward, left/right, curved walking and turning around. In the previous pattern generation of the KHR series, the step time and stride of the robot were fixed, and the walking modes, step time and action of stride without stopping could not be changed. Hence, the flexibility of the walking pattern of the robot needed to be upgraded. The walking pattern in this paper allows variation in the walking mode, step time and stride for each step. The pattern uses a simple mathematical form of trajectory curves, specifically the sine, cosine, linear and third-order polynomial curves, and the superposition of these curves is used to minimize the complexity and burden of the computation. The authors used a third-order polynomial to generate the trajectory of the robot's pelvis. With the aid of a simplified zero-moment point (ZMP) equation, the pelvis trajectories have a direct relationship with the ZMP trajectories. An effective means of generating the trajectories is introduced, and the scheme is verified experimentally under various walking conditions that take into account the step time and stride. The experimental platform, which has human-like features and movement, is briefly introduced here. With a simple kinematical structure and distributed control hardware architecture, the platform was designed to consume relatively low levels of energy. Moreover, the scheme for generating the trajectory is realized for variations to flexible walking. (C) Koninklijke Brill NV, Leiden and The Robotics Society of Japan, 2008.