Abstract
With the increasing integration of intelligent driving functions intoserial-produced vehicles, ensuring their functionality and robustness posesgreater challenges. Compared to traditional road testing, scenario-basedvirtual testing offers significant advantages in terms of time and costefficiency, reproducibility, and exploration of edge cases. We propose aTransformer-enhanced Conditional Variational Autoencoder (CVAE-T) model forgenerating multi-agent traffic scenarios in roundabouts, which arecharacterized by high vehicle dynamics and complex layouts, yet remainrelatively underexplored in current research. The results show that theproposed model can accurately reconstruct original scenarios and generaterealistic, diverse synthetic scenarios. Besides, two Key-Performance-Indicators(KPIs) are employed to evaluate the interactive behavior in the generatedscenarios. Analysis of the latent space reveals partial disentanglement, withseveral latent dimensions exhibiting distinct and interpretable effects onscenario attributes such as vehicle entry timing, exit timing, and velocityprofiles. The results demonstrate the model's capability to generate scenariosfor the validation of intelligent driving functions involving multi-agentinteractions, as well as to augment data for their development and iterativeimprovement.