What does a pipe-inspection robot actually see? A TUBERS dataset now available on Zenodo offers a direct answer: endoscopic images collected by a soft robot inside an experimental pipe network designed to reproduce realistic inspection challenges.

Inspection data under varied conditions

The dataset was produced by Bendabl and the University of Essex. Images were captured in a network containing polyethylene pipe sections, transparent acrylic pipes and steel fittings, including connectors, elbows, junctions and reducer couplings. The robot was tested in horizontal, vertical and inclined configurations, in both dry and water-filled conditions and with different lighting, viewing angles and inspection distances.

The collection includes intact sections as well as introduced defects, including small holes and curved cracks. Two endoscopic camera systems were used: the camera integrated into the soft robot produced 1280 x 720 images, while a second camera produced 1920 x 1080 images.

Ready for machine-learning research

The images are supplied in JPG format and organised into training, validation and test sets. Each split contains crack and no-crack folders, while a labels.csv file provides a binary label for every image. This structure makes the dataset easier to reuse for model development, comparison and validation.

The data can support research in robotic inspection, computer vision, infrastructure monitoring and automated anomaly detection. Just as importantly, the range of water, lighting and camera conditions reflects a central TUBERS lesson: inspection tools must cope with variation, not only with ideal laboratory images.

Opening the view to others

By publishing the 55.3 MB collection under a Creative Commons Attribution 4.0 licence, TUBERS enables researchers and technology developers to reuse the images with attribution and extend the work toward more robust inspection systems.

View the record and download Visual Data from Water Pipe Inspection on Zenodo. Access the dataset on Zenodo