Adversarial Driving Behavior Generation via Fuzzy Reward Reinforcement Learning Incorporating Human Risk Cognition

Exposing failures in autonomous vehicles (AVs) through adversarial environments is crucial for evaluating their reliability before deployment. This paper presents an innovative fuzzy reward reinforcement learning (RL) framework that seamlessly incorporates influential human factor to generate adversarial driving behavior from background adversarial vehicles (ADVs). The primary aim is to effectively and rationally expose risky events for AV evaluation. Specifically, a three-layer fuzzy logic for driving reward is designed to represent human reasoning regarding satisfaction in response to varying driving conditions. Building upon this, inspired by the human risk cognition factor in potential risky scenarios, the Cumulative Prospect Theory (CPT) is employed to model cognitive biases associated with human risk perception within the proposed fuzzy reward. It is then integrated into a CPT-RL framework, facilitating the controllable generation of a spectrum of driving behaviors, ranging from adventurous to conservative. The paper further provides assessment of the benefits derived from integrating fuzzy reward. Finally, comparative case studies of AVs evaluation are conducted within a Hardware-in-the-Loop (HiL) platform and specific metrics are designed to evaluate the effectiveness of the generated adversarial behaviors.