Tethys Robotics

Once submerged, a robot has no GPS. Underwater platforms therefore rely mostly on acoustic navigation — USBL, Doppler velocity logs, sonar — which works, but is costly, depends on the environment and offers limited spatial resolution.

Cameras are compact, cheap and already on board most inspection robots. The catch is the imagery: turbidity, scattering, blur and a strong blue-green cast reduce contrast and texture, which is exactly what visual tracking depends on.

The question of the thesis: is monocular visual SLAM reliable in underwater scenarios, and how can its robustness be improved under real operating constraints?

Bachelor's thesis at the Mobile Robotics Lab, ETH Zürich (Prof. Dr. Stefan Leutenegger), in collaboration with Tethys Robotics, building on the SWARM focus project.

I reviewed visual SLAM methods for underwater suitability, built an underwater-adapted monocular pipeline on ORB-SLAM3, integrated a learned feature front end, and evaluated it against ORB-SLAM3, SuperPoint and DROID-SLAM on underwater sequences from Tethys Robotics.

Raw monocular frames are converted to grayscale, enhanced with CLAHE (contrast-limited adaptive histogram equalisation) and passed to a SuperPoint-style feature extractor running UFEN weights. UFEN — Underwater Feature Extraction Network — distils SuperPoint into a model trained for underwater degradation; here it is used as a drop-in replacement for the SuperPoint weights.

The network runs as a TorchScript model on the GPU. Its 256-dimensional descriptors are binarised and packed into 32 bytes so the unchanged ORB-SLAM3 back end — Hamming matching, tracking, local mapping, loop closing — can consume them. Non-maximum suppression keeps a capped set of strong, well-spread keypoints.