Towards End-to-End Underwater vSLAM Using Neural Radiance Fields

Underwater environments pose unique challenges for 3D reconstruction and pose estimation - critical tasks in robotic navigation, environmental mapping, and exploration. Traditional visual SLAM methods, which rely heavily on feature extraction and matching, often struggle in underwater conditions due to low visibility, non-uniform lighting, and optical distortions introduced by light absorption and scattering. While Neural Radiance Fields (NeRFs) have shown remarkable success for view synthesis and 3D reconstruction in terrestrial settings, their application to underwater scenarios remains underexplored, primarily due to the scarcity of labeled data required for training. In this work, we investigate the potential of applying NeRF-based methods underwater by leveraging FlowCam, a semi-supervised framework designed to reduce the dependence on large labeled datasets. We evaluate pre-trained NeRF models on underwater datasets from NOAA Okeanos Explorer Mission, Kattegatcentret pool and DeepOcean, to examine robustness and generalization in diverse, dynamic conditions. To contextualize its performance, we compare FlowCam against DROID-SLAM, a fully differentiable SLAM system that combines learned optical flow and recurrent modules with a classical bundle adjustment backend. This comparison allows us to explore the advantages and limitations of NeRFs in relation to more traditional learning-based and geometry-driven SLAM approaches. Our results highlight both the challenges and opportunities of deploying NeRFs in underwater domains, including improved reconstruction quality and pose estimation even under significant light-scattering effects. Our findings offer insights into the applicability of NeRF-based methods for underwater robotic perception and aim to motivate further research into generalizable, data-efficient solutions for challenging marine environments.