RoDyn-SLAM is a novel dynamic dense RGB-D SLAM method that uses neural radiance fields (NeRF) to achieve robust performance in dynamic environments. The method addresses the limitations of traditional approaches that assume static environments by introducing a motion mask generation technique to filter out invalid rays from dynamic regions. This technique fuses optical flow and semantic masks to enhance motion mask precision. A divide-and-conquer pose optimization algorithm is designed to improve pose estimation accuracy, distinguishing between keyframes and non-keyframes. An edge warp loss is used to enhance geometry constraints between adjacent frames. The method is evaluated on two challenging datasets, demonstrating state-of-the-art performance in both accuracy and robustness. The implementation of RoDyn-SLAM will be open-sourced to benefit the research community. The paper introduces a dynamic SLAM framework with neural radiance fields, focusing on robust pose estimation and dense reconstruction in dynamic scenes. The method uses a motion mask generation algorithm to filter out rays in dynamic regions, combining optical flow and semantic masks for improved accuracy. A divide-and-conquer pose optimization algorithm is designed to enhance pose estimation, and an edge warp loss is used to improve geometry consistency. The method is evaluated on two dynamic datasets, showing superior performance compared to existing NeRF-based RGB-D SLAM methods. The paper also discusses related work, including conventional visual SLAM with dynamic object filtering, RGB-D SLAM with neural implicit representation, and dynamic object decomposition in NeRFs. The proposed method improves upon these by introducing a motion mask generation strategy that combines optical flow and semantic masks, and a novel pose optimization algorithm that enhances geometry consistency. The method is evaluated on two dynamic datasets, demonstrating state-of-the-art performance in both accuracy and robustness. The method is implemented with an implicit map representation using multi-resolution hash encoding and MLP decoders for color and depth rendering. A motion mask generation algorithm is used to filter out rays in dynamic regions, and a divide-and-conquer pose optimization algorithm is designed to improve pose estimation. The method is evaluated on two dynamic datasets, showing superior performance compared to existing NeRF-based RGB-D SLAM methods. The method is also evaluated on the TUM RGB-D dataset, demonstrating improved tracking performance in dynamic scenes. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes. The method is also evaluated on the BONN RGB-D dataset, demonstrating superior performance in complex and challenging scenarios. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes. The method is also evaluated on the TUM RGB-D dataset, demonstrating improved tracking performance in dynamic scenes. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes. The method is also evaluated on the BONN RGB-D dataset, demonstrating superior performance in complex and challenging scenarios. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes.RoDyn-SLAM is a novel dynamic dense RGB-D SLAM method that uses neural radiance fields (NeRF) to achieve robust performance in dynamic environments. The method addresses the limitations of traditional approaches that assume static environments by introducing a motion mask generation technique to filter out invalid rays from dynamic regions. This technique fuses optical flow and semantic masks to enhance motion mask precision. A divide-and-conquer pose optimization algorithm is designed to improve pose estimation accuracy, distinguishing between keyframes and non-keyframes. An edge warp loss is used to enhance geometry constraints between adjacent frames. The method is evaluated on two challenging datasets, demonstrating state-of-the-art performance in both accuracy and robustness. The implementation of RoDyn-SLAM will be open-sourced to benefit the research community. The paper introduces a dynamic SLAM framework with neural radiance fields, focusing on robust pose estimation and dense reconstruction in dynamic scenes. The method uses a motion mask generation algorithm to filter out rays in dynamic regions, combining optical flow and semantic masks for improved accuracy. A divide-and-conquer pose optimization algorithm is designed to enhance pose estimation, and an edge warp loss is used to improve geometry consistency. The method is evaluated on two dynamic datasets, showing superior performance compared to existing NeRF-based RGB-D SLAM methods. The paper also discusses related work, including conventional visual SLAM with dynamic object filtering, RGB-D SLAM with neural implicit representation, and dynamic object decomposition in NeRFs. The proposed method improves upon these by introducing a motion mask generation strategy that combines optical flow and semantic masks, and a novel pose optimization algorithm that enhances geometry consistency. The method is evaluated on two dynamic datasets, demonstrating state-of-the-art performance in both accuracy and robustness. The method is implemented with an implicit map representation using multi-resolution hash encoding and MLP decoders for color and depth rendering. A motion mask generation algorithm is used to filter out rays in dynamic regions, and a divide-and-conquer pose optimization algorithm is designed to improve pose estimation. The method is evaluated on two dynamic datasets, showing superior performance compared to existing NeRF-based RGB-D SLAM methods. The method is also evaluated on the TUM RGB-D dataset, demonstrating improved tracking performance in dynamic scenes. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes. The method is also evaluated on the BONN RGB-D dataset, demonstrating superior performance in complex and challenging scenarios. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes. The method is also evaluated on the TUM RGB-D dataset, demonstrating improved tracking performance in dynamic scenes. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes. The method is also evaluated on the BONN RGB-D dataset, demonstrating superior performance in complex and challenging scenarios. The method is compared with other SLAM methods, showing competitive results in both dynamic and static scenes.