Evaluating Spatial Grounding and ۳D Understanding in Vision-Language Models for Autonomous Driving

We study how spatial grounding and three-dimensional (۳D) understanding should be evaluated in vision-language models for autonomous driving. Our goal is to distinguish plausible descriptions of road scenes from responses that correctly identify objects, represent their spatial relationships, and remain consistent with available geometric and temporal evidence. We propose a theoretical evaluation framework that connects language queries to structured scene representations and assesses object grounding, metric geometry, spatial relations, viewpoint consistency, temporal understanding, and uncertainty handling. The framework accepts driving observations, natural-language queries, sensor metadata, and independently established reference annotations. It produces a structured response record and a profile of spatial capabilities rather than a single language-based score. Our conceptual analysis establishes that object identification, geometric accuracy, and relational correctness are logically distinct: success in one does not guarantee success in the others. We also show through constructed counterexamples that internally consistent answers can remain physically incorrect and that agreement across viewpoints is meaningful only when it is accompanied by correctness. These analytical results motivate a joint correctness criterion, explicit treatment of unobservable information, and separate reporting of response accuracy and answer coverage. Accordingly, the contribution is an evaluation specification and a set of analytical findings, without claims of measured performance improvements. The proposed framework provides a basis for examining whether a model’s language responses are supported by the spatial evidence required for their intended use within a driving system.

کلیدواژه ها:

Autonomous Driving ، VLM ، Spatial Grounding ، Three-Dimensional Understanding.