This paper presents a control scheme using human arm dynamics to simulate the subject's volitional movement. To achieve this goal, first, it has to obtain the torques purely generated from the subject from force/torque sensors. With this result, the controller can implement the therapeutic exercises, consisting of the active range of motion (AROM), the active-assistive range of motion (A-AROM), and the resistive range of motion (RROM) modes on our self-built upper limb rehabilitation robot arm by feeding back the torques purely generated from the subject with different gains to the associated arm controller. Besides, by modifying the gravity term in human arm dynamics, this controller can achieve different levels of gravity compensation of the upper limb. To validate our design, some realistic experiments are conducted and the performance of our design is demonstrated.