Target-Guided Trajectory Generation for Controllable Traffic Scenarios - A novel conditioning method for diffusion-based trajectory generation enabling controllable traffic scenario synthesis

Abstract

Autonomous vehicles (AVs) must be evaluated under rare and hazardous driving conditions to ensure safety and reliability. However, creating such safety-critical scenarios is made difficult for several reasons. They occur infrequently in real-world data and are costly to reproduce through physical testing, while existing simulation methods often yield unrealistic behaviors. This thesis explores generative modeling as a tool for producing realistic and controllable scenarios for closed-loop evaluation of AV systems. We introduce a novel diffusion-based method for generating adversarial trajectories, with a focus on Classifier-Free Guidance (CFG) to steer agents toward defined targets. The approach incorporates target information during training, uses data augmentation to improve robustness, and applies trajectory optimization to enhance accuracy. Building on the Versatile Behavior Diffusion (VBD) framework, our method strengthens controllability while preserving realistic motion patterns. The experimental results show that CFG improves guidance performance without any additional computational cost during inference, which has been a major limitation of prior approaches, while still matching the accuracy of classifier-based guidance. When combined with classifier-based guidance, CFG yields substantial improvements in target accuracy and reduces the number of required guidance iterations. Furthermore, direct trajectory optimization is shown to further refine target accuracy, although it introduces trade-offs with respect to adherence to traffic regulations. Collectively, these findings establish an efficient and versatile framework for the generation of safety-critical driving scenarios, thereby advancing the methodological foundation for rigorous evaluation of autonomous vehicle systems.