PriorAug: Reliability-Aware Traffic Scenario Initialization Under Sensor-Perceived Structural and Interaction Priors

Traffic scenario initialization is a key component of autonomous driving simulation. The generated initial traffic states directly affect downstream motion prediction, planning evaluation, closed-loop simulation, and safety assessment. Existing traffic scenario generators usually rely on structural and interaction priors, but these priors are commonly treated as deterministic and fully reliable during both training and deployment. In practical autonomous driving systems, however, such priors are reconstructed from heterogeneous perception pipelines involving sensors, localization, tracking, online mapping, map alignment, and behavior estimation. Their reliability may therefore vary continuously because of sensing noise, localization drift, incomplete observations, occlusion, tracking instability, and imperfect interaction inference. To address this issue, this paper proposes PriorAug, a reliability-aware prior augmentation framework for robust traffic scenario initialization under sensor-perceived prior uncertainty. PriorAug formulates traffic scenario initialization as conditional generation over sensor-perceived priors and introduces a two-dimensional reliability space to characterize structural reliability and interaction reliability separately. Based on this space, reliability-conditioned prior operators are designed to construct diverse prior configurations during training, enabling the generator to learn from continuously varying perception conditions without modifying the underlying generation architecture or introducing additional learnable inference modules or iterative decoding stages. Experiments on the Waymo Open Motion Dataset show that PriorAug maintains competitive full-set distribution fidelity, improves robustness under shifted-prior conditions and traffic density variations, and improves the physical plausibility of generated initial states. Reliability landscape analysis further demonstrates that coverage-oriented prior augmentation yields smoother performance transitions across the reliability space. These results indicate that explicitly modeling sensor-perceived prior reliability provides an effective and practical route to robust traffic scenario initialization for autonomous driving simulation.