Shifting from Adversarial Behavior to Sensor-Level Generation for Testing Autonomous Vehicles

We contribute an abstract paper, where we pinpoint the necessity of shifting from adversarial agent interactions to sensor-level generation, enabling the safety evaluation of end-to-end autonomous vehicles. To achieve this, a unified taxonomy is proposed to form the key technical steps. Representative approaches to instantiate the taxonomy are presented and compared. We found that by combining 3D Gaussian Splatting with generative models is a practical and promising way to address the shifting, enabling spatial and temporal consistent sensor observations and a scalable testing framework.