Autonomous movement inside a pipe requires a robot to interpret an environment that is dark, repetitive and difficult to access. To help researchers work on that challenge, TUBERS partner DEMCON has released the project’s synthetic pipeline dataset through Zenodo.

More than a camera image

The dataset contains images generated at 31,974 locations within a three-dimensional model of the KWR drinking-water test network. Four aligned image types are provided at each location: a rendered view of what the robot’s head camera would see, a surface-normal map and two ground-truth masks identifying ridges and pipe branches. This allows models to learn from the camera view while being checked against known reference information.

Connecting simulation with real images

The release also includes camera images captured at 1,161 locations in a clean pipe-section test setup. These are accompanied by manually identified branch pixels. All images have a resolution of 512 x 512 pixels, and the complete record contains approximately 4.5 GB of data. Synthetic data make controlled training possible across many positions and viewpoints; the real test images help connect that work with the variability of physical environments.

Supporting reproducible navigation research

Within TUBERS, the dataset supported work on recognising pipe-network features needed for autonomous navigation. Its publication now gives other robotics and computer-vision teams a reusable benchmark for exploring feature detection, perception and navigation in confined infrastructure.

Explore DS5_DMC_synthetic_pipeline_data and download the open files from Zenodo. Access the dataset on Zenodo