Generation and Adversarial Testing for Assessing Autonomous Vehicles in Complex Environments

Autonomous cars are the new future of the world so they must be reliable and trusted before handing them over to the general public. In order to ensure their reliability, it is important to realize that when naturalistic data is used for training, the idea of safety does not change. This is because, in natural settings, the likelihood of an automobile being involved in an accident is relatively low, making training for such situations challenging. This work proposes an efficient technique by which an adversarial environment is created to train the autonomous vehicle for difficult scenarios. This technique facilitates the evaluation of autonomous vehicles in a multi-agent setting, where adversarial car policies are taught and compared. The experiments are carried out in a simulated urban driving environment that relies on the difficulty of scenarios and provides a high level of accuracy. The findings and training are used to enhance the effectiveness of deep reinforcement learning-based autonomous driving rules.