SurfSLAM

SLAM (simultaneous localisation and mapping) is how a robot builds a map of its surroundings while working out where it is within that map. SurfSLAM is by Onur Bagoren, Seth Isaacson and colleagues with Ram Vasudevan and Katherine A. Skinner at the University of Michigan.¹ Its stereo network was trained first on simulated underwater scenes with haze, caustics, lighting and particle effects, then fine-tuned on real footage without labels.^1,4 It produces metric depth maps.¹

A pose graph, the running estimate of the robot's path, combines those depth maps with readings from an inertial measurement unit, a pressure sensor and a Doppler velocity log.^1,4 A Doppler velocity log is a sonar that measures the robot's speed over the seabed. Global registration against the depth maps limits drift.¹

It was developed on shipwreck surveys by an underwater robot in Thunder Bay National Marine Sanctuary.¹ The paper compares it with five other SLAM systems, including ORB-SLAM3, DROID-SLAM and MASt3R-SLAM, on scenes such as Boiler and Engine.¹ The accompanying dataset, SUDS, holds over 24,000 stereo pairs with dense photogrammetry models and reference trajectories, hosted on the university's Deep Blue repository.^1,2

The SLAM system and the stereo training and inference code are in two GitHub repositories.^2,3 Both are packaged to run in Docker and pull in upstream stereo models (FoundationStereo, DEFOM-Stereo, IGEV++, Underwater_Stereo) as patched submodules.^2,3 The project page is licensed CC BY-NC-SA 4.0; the repositories have no licence file of their own.^1,2,3 Each repository has a single commit and no releases.^2,3 The preprint, arXiv 2601.10814, was first posted on 2026-01-15 and revised on 2026-09-14.⁴