sEMG, as a kind of bioelectrical signal reflecting muscle motion state, generally applies to motion recognition and human interface. Healthy subjects are selected in most studies, while for hemiplegic patients, especially patients with severe hemiplegia, high accuracy motion recognition is difficult to acquire due to the non-ideal sEMG signal from dysfunction muscles. Therefore, this paper presents an upper limb exercise therapy, based on 5 defined motions and 6 Muscle-Units, for patients with severe hemiplegia. Through the sampling and analysis of sEMG signals from 8 subjects, including 4 healthy and 4 hemiplegic patients, we draw a conclusion of the relevance between specific motions and Muscle-Units, which can be used as a reference for paralyzed arm training. According to this relevance, six Muscle-Units can be classified into two categories: major Muscle-Units and minor Muscle-Units. In order to improve the interest and positivity of patients, a PC based virtual interactive platform is established. The sEMG signal from major Muscle-Units is processed with a moving average algorithm, and the result is used as the control signal for training interaction.