Generating Test Cases for Autonomous Vehicles With Controllable Levels of Difficulty

Abstract

Autonomous vehicle (AV) safety validation increasingly relies on scenario-based testing. However, existing approaches to test scenario generation do not provide mechanisms to systematically regulate scenario difficulty. To address this critical limitation, this paper introduces a novel game-theoretic framework for adversarial safety validation. The interaction between the AV-under-test and a strategically obstructing rear vehicle is modelled as a Stackelberg game. The level of adversarial intensity, which reflects the level of difficulty, of a scenario can be controlled by a single tunable parameter called aggressiveness at both the action level and the interaction level. The efficacy of this approach is studied through the highway lane-changing operational design domain. Simulation experiments demonstrate that increasing aggressiveness reduces the success rate of lane-changing and prolongs maneuver duration for the successful attempts. These results confirm that this parameter can effectively and systematically control the difficulty level of test scenarios, providing a valuable tool for rigorous and reproducible AV safety validation.