Unmanned surface vehicle (USV) is a kind of robot that can control and execute water surface tasks by satellite positioning and self sensing without human intervention. However, the high costs of labor and the environment-sensitivity property, such as the sea storm and weather condition, contributes to a numerous and complicated testing process of USV control system. Except the aboved, this task also faces the risk of equipment hardware deterioration. In total, the test on USV control system now suffer from various kinds of dilemma, such as high cost on human and material resources, high requirements for conditions, low efficiency, etc. In order to solve the aboved problems, the most urgent matter is to put forward an USV virtual simulation module to support the needs of testing the control system. With the assistant of imitation learning algorithm, we propose a virtual simulation module from a large number of the historical data to construct the virtual simulation environment of USV. Specifically, we combine the behavioral cloning algorithm and the generative adversarial imitation learning algorithm, and get further improvement in several practical problems. Experimental results show that the improved generative adversary imitation learning algorithm converges faster than the baseline generative adversary imitation learning algorithm, and effectively reduce the compound error when comparing with behavior cloning.