MANG-SLAM: Multi-Agent Neural Submap and Gaussian Representation for Dense Mapping

Abstract

Simultaneous Localization and Mapping (SLAM) is a core component in robotic navigation and augmented/virtual reality technologies. <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Multi-agent collaborative SLAM</b> improves scalability and consistency by distributing mapping tasks and sharing observations, making it suitable for large-scale environments. However, current 3DGS-SLAM systems have yet to fully utilize this architecture. This is primarily limited by the lack of capabilities in current <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3D Gaussian Splatting (3DGS)</b>-based SLAM methods for addressing global drift, loop closure detection, and scene gap filling in multi-agent collaborative scenarios. This paper proposes <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MANG-SLAM</b>, a progressive dense SLAM system that integrates multi-agent collaboration with neural submaps. By leveraging <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Neural Submaps</b> to guide 3DGS rendering, MANG-SLAM achieves efficient optimization from local to global levels and supports parallel mapping by multi-agents. Each agent independently generates submaps, which are fused on the server side for submap integration, loop closure detection, and global optimization. This approach effectively corrects accumulated drift, improves geometric consistency, and enables gap filling. Experimental results demonstrate that MANG-SLAM enhances both global consistency and local detail recovery on the Replica and ScanNet datasets, achieving efficient and accurate reconstruction and localization.