This paper developed an surface electromyography (sEMG) -based controller for an upper limb rehabilitation robot system, in order to provide variable resistance training to the impaired arm of stroke patients. A four-muscles based musculoskeletal system was implemented to predict the muscle torque and force of upper limb. Thus, the value of resistance provided by the rehabilitation robot system can be adjusted according to the patent's participant effort instead of a fixed value. The sEMG-based torque and force prediction model was tested by preliminary experiments and the results showed that the prediction model can reflect the muscle torque and force in real-time with acceptable accuracy. Therefore, three healthy subjects took part in the preliminary experiments. The sEMG-based rehabilitation robot system and the experiments showed promising results for the sEMG-based controller in an upper limb rehabilitation protocol.