Avatar: Adversarial Vehicle Trajectory Attack Targeting Autonomous Driving Planner

With the increasing adoption of autonomous vehicles, ensuring the planner safety is critical, as it directly impacts trajectory and safety. This study introduces AVATAR, a novel black-box attack framework exploiting planner vulnerabilities by subtly adjusting trajectories. The attack requires no prior knowledge of the planner internals. AVATAR uses reinforcement learning to optimize attack strategies and polynomial trajectory generation to maintain realism. Experiments across multiple traffic scenarios show significant degradation in safety and efficiency for both rulebased and learning-based planners. The impact is especially significant in high-traffic and collision-avoidance scenarios. Cross-environment and cross-planner evaluations confirm AVATAR's robustness, sometimes outperforming scenariospecific attack methods. This study shows that autonomous driving planners are susceptible to adversarial trajectory attacks, potentially resulting in unsafe or inefficient behaviors.