PROJECT TITLE :

Adversarial Evaluation of Autonomous Vehicles in Lane-Change Scenarios

ABSTRACT:

Before being allowed to operate in cities or on highways, autonomous vehicles need to be thoroughly tested and evaluated. However, the majority of the existing evaluation methods for autonomous vehicles are fixed and lack adaptability; as a result, they are typically ineffective when it comes to generating challenging scenarios for vehicles that are being tested. In this paper, we propose an adaptive evaluation framework as a method for evaluating the performance of autonomous vehicles in hostile environments that are produced by deep reinforcement learning. We use ensemble models to represent different local optimums for diversity because of the multimodal nature of risky scenarios. This is done in order to maximize the amount of information obtained. After that, we cluster the adversarial policies by using a Bayesian approach that isn't parameterized. The proposed method is validated by applying it to a standard lane-change scenario. This scenario involves frequent interactions between the ego vehicle and the vehicles that are around it. The results indicate that the challenging conditions that are produced by our methodology significantly hinder the effectiveness of the vehicles that were put through their paces. We also provide examples of the various patterns of generated hostile environments. These examples can be used to infer the vulnerabilities of the vehicles that have been tested.