ITGAN: An Interactive Trajectories Generative Adversarial Network Model for Automated Driving Scenario Generation

Interactive scenarios are of great significance for the testing of high-level automated vehicles. Based on the structure of Generative Adversarial Network (GAN), the Interactive Trajectories GAN model (ITGAN) is proposed in this paper. This research comprehensively considers both of the dynamic interactions between agents and the responses of agents to the static road environment. ITGAN consists of a specially designed pair of generator and discriminator, which are trained iteratively in an adversarial manner. As a data-driven agent model, ITGAN has the ability to generate agent interactive trajectories just like those from the real world. In order to verify the utility of ITGAN, experiments are conducted with the co-simulation framework of CARLA and SUMO, where ITGAN is implemented as the external agent control model to help generate interactive driving scenarios on various types of urban roads. By the qualitative and quantitative analysis of the experiment results, it shows that ITGAN can produce more effective interactive scenarios compared with the traditional agent model and has great potential to supplement the deficiencies of existing testing methods for high-level automated vehicles.