Flexible and robust walking is the most basic guarantee for humanoid robots to complete various tasks, which requires complex mechanical control systems, environmental perception, and action planning capabilities. However, robots are inevitably disturbed in actual walking environments, which can affect the stability of walking and even disrupt the robot’s periodic motion, causing it to fall. This paper is concerned with bipedal walking based on nonlinear disturbance observer. By constructing dynamic equations and utilizing Lyapunov stability theory to design observers, real-time estimation and compensation of composite disturbances, such as sudden changes in ground reaction forces, joint friction, can be achieved. We adjust the gain rate by designing the gain matrix of the observer. Finally, the effectiveness of the mentioned method is illustrated by simulations.