This paper presents an underactuated control system for a two-wheeled mobile robot (TWMR). TWMR is expected to use as an autonomous mobile robot (AMR) that works in human-collaborative spaces. TWMR has a smaller footprint and higher center of gravity (CoG), than typical AMRs. Furthermore, in this study, TWMR has an arm. Therefore, TWMR is superior in handling loads in human-presence spaces. However, TWMR is the underactuated system in the body’s translational position and tilting angle. In addition, TWMR is affected by disturbances caused by loads. As a robust underactuated control system, the cascade PI/PD and synthesized pitch angle disturbance observer (SPADO) are considered. The cascade PI/PD results in significant translational position error and oscillatory responses. To cope with this problem, a proposed method minimizes both errors of the translational position and tilting angle within torque constraints. The proposed method utilizes the predictive functional control (PFC) for the underactuated control considering constraints. SPADO improves the accuracy of an internal model of PFC. The validity of the proposed method was confirmed compared with the cascade PI/PD through the simulation.