High-value multi-vehicle scenario generation for autonomous driving testing using a fine-tuned generative transformer

Abstract

Rare but high-risk multi-vehicle interactions are essential for evaluating the safety boundaries of autonomous driving systems, yet are difficult to cover efficiently through naturalistic replay, single-agent perturbation or manual parameter combinations. This study develops a high-value multi-vehicle scenario generation framework based on SMART target-domain adaptation. Two representative highway scenarios, constrained cut-in and zipper merge, are mined from highD and exiD using event-triggered rules and converted into a unified multi-agent scene representation. A SMART-based generative Transformer discretizes vehicle motions and road structures into motion and road tokens, is pretrained on WOMD, and is fine-tuned on the mined target-domain samples. During inference, Top-p sampling generates multimodal motion-token sequences, while physical-consistency filtering after continuous trajectory reconstruction rejects or resamples candidates with abnormal acceleration and jerk. Generated scenarios are evaluated across trajectory, scenario and system-response levels through open-loop quality assessment, target-scenario construction and closed-loop validation with an IDM + Pure Pursuit SUT. Compared with CS-LSTM, the proposed method reduces minADE and minFDE by 12.7% and 12.8%, respectively, and improves map compliance to 99.10%. Construction success reaches 42.0% for constrained cut-in and 35.6% for zipper merge. Selected critical scenarios reduce the SUT pass rate from 86.7% to 20.0% and expose collision and motion-paralysis failures. These results demonstrate map-compliant, replay-compatible and testing-oriented high-value interaction scenarios for operational-boundary evaluation and safety validation of autonomous driving systems within the validated highway domains.