New York and Bath, England – – September 30, 2026 -- A joint research study by Dynamic Infrastructure and the University of Bath has found that leading frontier vision-language models do not yet deliver the robustness needed for real-world infrastructure inspection, even as more specialised AI approaches perform substantially better on the same task.
Researchers test how AI models generalise across unseen infrastructure sites
The collaboration, carried out with Thomas Kjeldsen, Professor of Hydrology and Water Engineering at Bath's Department of Architecture and Civil Engineering, forms part of Arline Osorio Moreno's MSc research. The study examines whether computer vision models can reliably detect blockage and obstruction in drainage assets from inspection images, and whether that performance holds when applied to sites the models have never seen. Construction methods, camera equipment and inspection practices vary widely between regions, creating a generalisation gap that determines whether agencies can trust an AI system for deployment.
Four AI approaches are benchmarked on the same drainage-inspection task
The research compares vision-language models, vision foundation models, supervised deep learning and classical machine-learning methods, each carrying different trade-offs in accuracy, data requirements and cross-site transferability. Early results show significant performance differences between both methods and sites.
Frontier vision-language models underperform on specialised infrastructure tasks
Despite their broad capabilities, current frontier vision-language models do not yet provide the robustness required for reliable real-world implementation on this specialised task, according to the study's early findings. More targeted approaches performed substantially better.
"The biggest, most general models are not always the strongest choice here, and more specialised approaches can be much more consistent," said Arik Voronov, AI R&D Lead at Dynamic Infrastructure. He added that model choice combined with robustness across real-world variation is where the critical differences emerge.
Professor Kjeldsen said the partnership allows academic research expertise to be tested against practical infrastructure challenges, ensuring new AI tools are scientifically rigorous while delivering measurable benefits for infrastructure operators.
Partners aim to publish findings in a peer-reviewed journal
Dynamic Infrastructure, which manages thousands of structures across 15 states and countries through its Engineering AI platform, and the University of Bath intend to continue the collaboration with the goal of producing a peer-reviewed publication.