Abstract <p>The increasing deployment of robot navigation and mapping systems in autonomous driving, intelligent manufacturing, and indoor services highlights the growing importance of visual simultaneous localization and mapping (SLAM) technologies. However, traditional SLAM methods that rely on geometric feature matching often suffer from poor performance in dynamic indoor environments due to unstable feature correspondences caused by moving objects. These limitations lead to degraded localization accuracy and, in extreme cases, system failure. To address these challenges, this paper presents a novel adaptive RGB-D SLAM algorithm that integrates multi-modal generator-based semantic segmentation. Our approach combines RGB and depth data using a multi-modal prompt generator (MPG) and a multi-modal feature adapter (MFA) to achieve robust, high-precision segmentation. The segmentation results are further refined using a motion-level initialization and cross-frame propagation mechanism, which effectively filters out dynamic disturbances. By incorporating weighted static constraints in the pose optimization process, our method enhances pose estimation accuracy and overall system robustness. Extensive experiments conducted on public datasets demonstrate that our approach significantly outperforms traditional and state-of-the-art SLAM systems, offering a promising solution for SLAM in dynamic environments.</p> Graphic abstract <p></p>