Abstract

Reliable traffic scenario initialization is essential for autonomous driving simulation, safety evaluation, and scenario completion. Most learning-based initializers rely on high-definition map priors to constrain agent placement and attribute generation. However, in practical deployment, map priors may be unavailable, incomplete, outdated, or inconsistent with temporary road changes. This study investigates traffic scenario initialization under strict map-prior removal, where both external HD-map inputs and agent-internal road-related cues are masked to reduce implicit structural leakage. We propose an interaction-only initialization framework, InterInit, which retains a fixed decoding-slot space and conditions initialization on observed agent states and multi-agent interactions. Experiments on the Waymo Open Motion Dataset first provide a vertical prior-removal analysis from full-map-prior initialization to weak-map initialization and then to strict map-prior removal. The results reveal attribute-asymmetric degradation: position and heading remain relatively stable, whereas speed becomes the dominant failure dimension. Further diagnostic analysis shows that the speed degradation is mainly caused by speed-magnitude collapse, characterized by nearly zero predicted variance rather than a large mean-speed bias. Based on this diagnosis, InterInit-SpeedFix is introduced as a localized repair strategy that models speed probabilistically in log-speed space without changing the overall architecture or inference pipeline. It reduces global Speed MMD from 0.2461 to 0.1210, corresponding to a 50.9% improvement, while maintaining stable position and heading performance. Density-stratified Speed MMD and absolute speed-magnitude error statistics further verify that the repair is consistent across different scene-density regimes and improves both typical and high-error cases. These results provide an application-oriented diagnosis-and-repair workflow for traffic scenario initialization when map priors are unavailable, and clarify the practical boundary of interaction-only initialization under strict map-prior removal.

IPC Classification

Keywords