Clustering-Based Semantic SLAM for Autonomous Underwater Vehicles

Reliable data association remains one of the main challenges in landmark-based Simultaneous Localization and Mapping (SLAM), especially in perceptually ambiguous environments. Many approaches rely on explicit correspondences and modality-specific assumptions, which can limit robustness and generality across different sensing configurations. This paper proposes a clustering-based semantic SLAM framework in which data associations are inferred through an iterative clustering-optimization process rather than prescribed by explicit observation-to-landmark matching. Observations are grouped into landmark hypotheses through density-based clustering in an augmented metric-semantic space, allowing landmark identities to emerge implicitly from data consistency rather than enforced correspondences. The resulting associations are used to build and optimize a factor graph, jointly refining vehicle poses and landmark positions. The proposed framework is modality-agnostic and does not rely on appearance-based matching, making it applicable to both visual and non-visual sensing scenarios. Robustness to noise and spurious detections is further enhanced through a life-points policy that stabilizes clustering across iterations. The approach is validated in simulation using a model of the FeelHippo Autonomous Underwater Vehicle equipped only with sonar-based perception, demonstrating the effectiveness of the proposed method for non-visual SLAM scenarios.