TRACE - TRaffic Autoregressive sCene gEneration

Validating Connected and Autonomous Vehicle (CAV) algorithms requires diverse, realistic traffic scenarios that traditional simulators cannot provide. For this reason, in this work, we present TRACE (TRaffic Autoregressive sCene gEneration), a text-conditioned framework for multi-agent traffic scene generation that addresses this limitation through two contributions. First, we propose an automated annotation pipeline that processes driving scenarios from the Waymo Open Motion Dataset (WOMD) using the Large Language Model Meta AI (LLaMA) 3.2 3B, a compact open-source language model, to generate natural language descriptions without reliance on proprietary APIs. The pipeline, which is fully reproducible on consumer hardware, produces a large-scale, text-annotated dataset. Secondly, we introduce an autoregressive transformer that populates map scenes with heterogeneous traffic participants conditioned on natural language textual prompts. Evaluated on the full Waymo validation set (40,075 scenes), TRACE achieves a 99.6% cardinality exact-match rate, 89.4% entity-type accuracy, and a maximum per-sector spatial deviation of 4 percentage points. Failure modes are traced to distributional bias in the training data, with targeted augmentation identified as a direction for future work.