Recent studies have suggested that virtual reality (VR) has considerable potential for motor rehabilitation. Wearable devices using electromyography (EMG) signals have been rapidly developed for medical use. In this study, a new training system was designed for motor rehabilitation. The system is based on the fusion of VR and EMG signal processing. In this system, after EMG signals are acquired and denoised, and their features are extracted, a mathematical approach using a support vector machine (SVM) is applied to classify gestures. Subsequently, the gestures are sent to a VR environment for rehabilitation training tasks. Stroke patients can use this system. Experimental results indicate the effectiveness and convenience of the system. The approach also shows several advantages. First, no camera is necessary; thus, the space occupied by the system is reduced. Second, gesture control provides a convenient and motivational approach to use the system in rehabilitation tasks. Future research directions include adding gestures; improving recognition accuracy; and integrating various sensors, including those for recording inertia and ECG signals.