The brain-computer interface (BCI) plays an important role in assisting the disabled in life support and entertainment. In this study, a wireless retrieval robot system for steady-state visual evoked potential (SSVEP) based BCI system was proposed. In order to improve the recognition accuracy and system stability of BCI system for subjects' instructions, in this study, offline and online experiments are used to study and solve the performance parameters, obstacle avoidance and retrieval performance of BCI respectively. In the offline experiment, 5 types of SSVEP data of 4 subjects were used. Based on CCA algorithm, the multi-channel fusion recognition accuracy, time window length (TWL), command activation margin and other aspects are studied. The optimal experimental parameters are determined. Online experiments are based on the optimal test parameters of offline experiments, 8 subjects were used to study the performance of obstacle avoidance and retrieval. The online experiments achieved an average accuracy of 87.89 ± 6.62% with an ITR of 23.6 ± 4.76 bit/min. Furthermore, the rate of success in grasping three kinds of different difficulty preset targets was 61.38%.