In a passenger-carrying walking humanoid robot, it is preferable to use the walking pattern generation and balance control method that can accommodate variable passenger weights. This is because the range of possible payloads for passenger weight is relatively wide, from zero to approximately one hundred kgf. If the variable weight is not considered, the walking performance and walking stability of the robot decrease when the passenger's weight was much heavier or lighter than the predefined passenger weight. Therefore, in this paper, the walking pattern generation and ZMP control methods are developed to adaptively cope with variable passenger weights. The walking pattern generation method using the convolution sum can effectively calculate the walking pattern for variable passenger weights since it contains the analytic function of variable mass. The ZMP controller that performs the real-time sensor feedback control is designed as state space forms where the gains can be easily changed for variable passenger weights. The experiments show that the walking performance of the proposed method is well maintained and superior to those of the non-adaptive controllers for three different payload levels.