CoVLM-Bench: A Real-World Benchmark for Cooperative Driving Question Answering and Planning
Author1, Author2, Author3, Author4, Author5
Abstract
The paper presents CoVLM-Bench, a benchmark that enhances cooperative driving question answering and planning by integrating infrastructure-side observations into vision-language models.
Reality Card
CoVLM-Bench enables the joint evaluation of cooperative driving question answering and trajectory planning on real-world vehicle-infrastructure scenes, improving accuracy and reducing planning errors.
CoVLM-Drive achieved a lower final destination error (FDE) compared to existing V2X planners.
The reliance on model-assisted drafting and human verification may limit reproducibility and generalizability of the results.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.