<p>While 3D Gaussian Splatting (3DGS) shows immense potential in dense SLAM, moving objects in dynamic environments severely degrade camera tracking and introduce destructive artifacts like “rendering ghosts” into maps. To address this, we propose DynGS-SLAM, a robust, dynamic-aware 3DGS SLAM framework based on SGS-SLAM. For the tracking front-end, we design a joint “semantic-geometric” dual-layer dynamic filtering strategy. For the tracking front-end, we propose a joint “semantic-geometric” dual-layer dynamic filtering architecture that establishes a rigorous deterministic boundary for pose estimation. In the semantic dimension, a bi-directional masking strategy achieves strict consensus between real-world observations and rendered views to bilaterally exclude dynamic interference. In the geometric dimension, dynamic sample-aware MAD filtering is introduced to accurately reject depth anomalies. For the mapping back-end, we propose a dynamic-aware pure mapping mechanism spanning the Gaussian lifecycle: during the densification phase, it blocks the initialization of dynamic targets at the source via a semantic exclusion mechanism; during the optimization phase, it relies exclusively on a unidirectional real-world observation mask to generate strong gradients, effectively eliminating historical rendering ghosts; and during the maintenance phase, it cooperates with a multi-dimensional joint semantic pruning strategy to eradicate residual dynamic entities and geometric distortions. Experimental results demonstrate that the proposed DynGS-SLAM algorithm outperforms SGS-SLAM across all metrics on dynamic sequences of the TUM dataset and other datasets. Specifically, the average RMSE of absolute trajectory error is reduced by 82.7% over five dynamic sequences from the TUM dataset.</p>