We highlight recent research results on AI-driven methods for visual simultaneous localization and mapping (VSLAM) in underwater environments.

We developed deep-learning-based algorithms that incorporate geometric constraints to enhance robustness and efficiency under low visibility and dynamic conditions. Building on these foundations, we extended the work toward real-time, camera-based navigation for autonomous underwater vehicles. Results from experimental evaluations on field datasets demonstrate how the integration of deep learning and geometric reasoning can advance the resilience, precision, and reliability of underwater robotic localization beyond traditional methods. The work was performed within a Marie-Curie Action REMARO (Horizon2020) and DeepODO project (Innovation Fund Denmark).