Interaction-aware scene-level multi-vehicle trajectory synthesis with generative latent representations

AI-ready brief

Traffic world models and scenario-based safety testing benefit from large libraries of interaction-rich multi-agentscenes,yetcollectingsufficientlydiversenaturalistictrajectoriesremainsexpensive. Existing multi-vehicle generative approaches often under-represent explicit relational structure or rely on fixed role assumptions, which limits generalization to heterogeneous role-mixture traffic scenes.

Author abstract

Traffic world models and scenario-based safety testing benefit from large libraries of interaction-rich multi-agentscenes,yetcollectingsufficientlydiversenaturalistictrajectoriesremainsexpensive.Existing multi-vehicle generative approaches often under-represent explicit relational structure or rely on fixed role assumptions, which limits generalization to heterogeneous role-mixture traffic scenes. This paper proposes an interaction-aware framework for conditional scene-level multi-vehicle trajectory resynthesis (stochastic replay), where vehicle-to-vehicle relations are represented as temporal interaction graphs and encoded into structured latent variables. A graph-structured variational encoder learns a tokenized latent scenerepresentation,andabelief-conditionedDiffusionTransformerperformslatentdiffusiontosample multiple plausible, interaction-consistent latent codes. A latent-conditioned decoder then reconstructs coherent multi-vehicle trajectories while maintaining kinematic plausibility. We validate the framework on a role-fixed cut-in benchmark (highD) and a role-agnostic mixed-role benchmark (CitySim), and evaluate fidelity, distributional similarity, kinematic feasibility, and safety-relevant interaction metrics (e.g., Time To Collision (TTC) and Time Headway (THW), etc.). Results indicate that the proposed method achieves strong conditional fidelity and distributional alignment while maintaining near-zero feasibility violations, and it reproduces safety-relevant post cut-in headway patterns consistent with real observations. Overall, the proposed conditional resynthesis approach provides a practical mechanism for stochastic replay and augmentation of interaction-rich multi-agent scenarios for traffic world modeling.

AI retrieval note

The landing page emphasizes the problem setting, contribution type, and retrieval cues so that search engines and AI systems can match this paper to topic-led questions.

Questions this page answers

Retrieval cues

Citation-ready BibTeX

@unpublished{noh2026interactionawarescenelev,
  title   = {Interaction-aware scene-level multi-vehicle trajectory synthesis with generative latent representations},
  author  = {S. Hong and B. Noh},
  year    = {2026},
  journal = {Transportation Research Part C: Emerging Technologies},
  note    = {Under review}
}

Source links

No external article link is currently attached to this archive record.