This paper presents a 5 DOF wearable rehabilitation robot which can implement single joint and multi-joint multiple motions for hemiplegic patients. The method of driving rehabilitation robot to assistant patients' impaired limb carry out rehabilitation exercises by healthy one of their own is present because hemiplegic patients' upper limb is usually unilaterally impaired. sEMG (surface electromyogram) signal is introduced into this method as the input of rehabilitation motion. Two algorithms-integral of absolute values (IAV) and Auto-regressive (AR) parameter model are adopted to compress data and extract feature of sEMG. Features worked out are sent into Levenberg-Marquardt (LM) based back propagation neural network (BPN) as the input, whose outputs are six upper limb rehabilitation exercise motions, to establish relationship of sEMG signal and motions. At the end of paper, for each motion 60 groups data is used to train and test network to get a good result. It laid the groundwork for study relationship of sEMG signal of patients' impaired upper limb and motions of which.