TrafficDiff: diffusion model based adversarial traffic scenario controllable generation for autonomous driving robust evaluation

As foundation models (FMs) are increasingly applied in safety-critical domains such as autonomous driving, their ability to handle rare, ambiguous, or adversarial conditions becomes essential for ensuring cognitive robustness. Generative AI offers a promising path for testing such capabilities by synthesizing diverse and realistic traffic scenarios. As a prominent class of generative models, diffusion models are known for their strong diversity, yet controllable generation remains a key challenge. To address this, we propose a controllable scenario generation framework based on diffusion models. First, a dynamic spatiotemporal fusion encoding mechanism integrates contextual factors (e.g., road layout, vehicle types) to enhance realism. To enhance diversity, we introduce a global–local optimizer that guides scenario generation while preserving physical and statistical consistency. To generate safety-critical long-tail scenarios, we design an adversarial induction method that enhances scenario criticality, while a system dynamics model improves long-tail scenario generation. Finally, a mechanism-based scenario filter ensures the safety and compliance of generated scenarios by eliminating unrealistic samples. We validate our method on benchmark datasets and real-vehicle tests. Compared to existing SOTA methods, traffic scenario diversity is enhanced by 6–8 times on average. In real-vehicle evaluations, TrafficDiff increase the collision rate by 25.5% and leads much mission failure, effectively challenging system robustness. This approach provides a scalable solution for virtual scenario validation, driving advancements in autonomous driving safety assessment.Our code of TrafficDiff is available at https://github.com/Moresweet/TrafficDiff.