ReMAV: Reward Modeling of Autonomous Vehicles for Finding Likely Failure Events

Autonomous vehicles are advanced driving systems that are well known for being vulnerable to various adversarial attacks, compromising the vehicle's safety, and posing danger to other road users. Rather than actively training complex adversaries by interacting with the environment, there is a need to first intelligently find and reduce the search space to only those states where autonomous vehicles are found less confident. In this paper, we propose a blackbox testing framework ReMAV using offline trajectories first to analyze the existing behavior of autonomous vehicles and determine appropriate thresholds for finding the probability of failure events. Our reward modeling technique helps in creating a behavior representation that allows us to highlight regions of likely uncertain behavior even when the baseline autonomous vehicle is performing well. This approach allows for more efficient testing without the need for computational and inefficient active adversarial learning techniques. We perform our experiments in a high-fidelity urban driving environment using three different driving scenarios containing single and multi-agent interactions. Our experiment shows 35%, 23%, 48%, and 50% increase in occurrences of vehicle collision, road objects collision, pedestrian collision, and offroad steering events respectively by the autonomous vehicle under test, demonstrating a significant increase in failure events. We also perform a comparative analysis with prior testing frameworks and show that they underperform in terms of training-testing efficiency, finding total infractions, and simulation steps to identify the first failure compared to our approach. The results show that the proposed framework can be used to understand existing weaknesses of the autonomous vehicles under test in order to only attack those regions, starting with the simplistic perturbation models.