Unstructured scene benchmark (USB): Which VLM performs better in autonomous driving

Vision-language models (VLMs) offer the potential for unified perception and language-guided decision-making in autonomous driving. However, existing benchmarks predominantly focus on structured road environments and high-quality imagery, leaving limited evidence on model performance (e.g., reasoning, explanation, and decision traceability) under unstructured scenes or degraded sensing conditions. This study develops a trustworthy test and evaluation framework to systematically assess VLM performance in these challenging contexts. Impromptu vision-language-action (VLA) samples are reorganized into six synchronized camera views augmented with vehicle state information, and twenty realistic input perturbations—covering illumination, weather, sensor reliability, and occlusion—are introduced for each scene. Moreover, six original non-open-ended questions are reformulated into traffic decision templates that combine structured choice sets with template-constrained free-text responses. Two types of evaluation methods are formulated. Multiple-choice questions (MCQs) are based on exact answer matching, aggregated with importance weighting according to expert rankings to prioritize planning tasks, and subjective questions (SQs) are graded using tailored, multidimensional large language model (LLM)-based scoring prompts. Experiments are conducted to assess the performance of seven open-source VLMs, including MCQ accuracy, reasoning coherence, visual fidelity of their responses to SQs, and their robustness to input perturbations. Overall, the proposed framework addresses a critical evaluation gap for supporting the deployment of VLMs in autonomous driving applications.