In order to solve the problems of rehabilitation robots based on brain-computer interface (BCI) at present that long time of multi-target classification and recognition accuracy. This study was designed an upper limb rehabilitation robot controlled by EEG signals to classify the steady-state visual evoked potential (SSVEP) in the EEG signals, then judged the intention of the subject and output the corresponding action instruction. Based on MATLAB PsychToolbox a flicker interface with five stimulus rectangles was designed as a visual stimulator to stimulate the brain to generate SSVEP signals, which corresponding to the five control instructions of the upper limb rehabilitation robot. Multivariate synchronization index (MSI) algorithm was used to classify the collected signals and output control instructions, the robot performed specific action after receiving instruction. Through the experiment, the optimal accuracy of robot action and average information transmission rate are 98.33% and 23.11 bit/min, respectively. The results show that the upper limb rehabilitation robot controlled by SSVEP signal has a good application prospect in aided therapy, can effectively improve the rehabilitation effect of limb hemiplegia patients.