A number of studies have measured kinematics, dynamics and oxygen uptake while a person walks on a treadmill. In particular, during walking on a split-belt treadmill, in which the left and right belts have different speeds, remarkable differences in kinematics are observed between normal subjects and subjects with cerebellar disease. In order to construct a gait adaptation model of such human split-belt treadmill walking, we proposed a simple control model and developed a new two-dimensional biped robot walk on a split-belt treadmill. We combined the conventional limit-cycle-based control consisting of joint PD control, cyclic motion trajectory planning and a stepping reflex with a newly proposed adjustment of P-gain at the hip joint of the stance leg. The data obtained in experiments on the robot (normal subject model and cerebellum disease subject model) have highly similar ratios and patterns to data obtained in experiments on normal subjects and subjects with cerebellar disease carried out by Bastian et al. We also showed that the P-gain at the hip joint of the stance leg was the control parameter of adaptation for symmetric gaits in split-belt walking and that P-gain adjustment corresponded to muscle stiffness adjustment by the cerebellum. Consequently, we successfully proposed a gait adaptation model for human split-belt treadmill walking, and confirmed the validity of our hypotheses and the proposed model using the biped robot.