A Survey on Collaborative SLAM with 3D Gaussian Splatting

TL;DR

This survey analyzes multi-robot SLAM that uses explicit 3D Gaussian Splatting ($3DGS$) to deliver real-time, photorealistic scene representations. It develops a taxonomy of CoGS-SLAM architectures (centralized vs distributed) and dissects core components including multi-agent consistency, communication-efficient optimization, Gaussian representations, fusion, and semantic distillation, supported by datasets and metrics. The authors synthesize design patterns such as shared Gaussian representations, asynchronous fusion, and compression-based data sharing, and discuss open challenges like lifelong mapping, robust inter-agent loop closure, and Sim2Real generalization. The work argues that CoGS-SLAM offers a powerful path toward scalable, high-fidelity digital twins and collaborative perception for robotics, AR/VR, and disaster-response applications, while highlighting the need for robust, bandwidth-aware, large-scale solutions.

Abstract

This survey comprehensively reviews the evolving field of multi-robot collaborative Simultaneous Localization and Mapping (SLAM) using 3D Gaussian Splatting (3DGS). As an explicit scene representation, 3DGS has enabled unprecedented real-time, high-fidelity rendering, ideal for robotics. However, its use in multi-robot systems introduces significant challenges in maintaining global consistency, managing communication, and fusing data from heterogeneous sources. We systematically categorize approaches by their architecture -- centralized, distributed -- and analyze core components like multi-agent consistency and alignment, communication-efficient, Gaussian representation, semantic distillation, fusion and pose optimization, and real-time scalability. In addition, a summary of critical datasets and evaluation metrics is provided to contextualize performance. Finally, we identify key open challenges and chart future research directions, including lifelong mapping, semantic association and mapping, multi-model for robustness, and bridging the Sim2Real gap.

Figures (5)

• Figure 1: Illustration of the relationship between the Collaborative Gaussian Splatting SLAM and related fields.
• Figure 2: Timeline of some noteworthy collaborative SLAM methods.
• Figure 3: Core Components of CoGS-SLAM: Centralized (left), Distributed (right).
• Figure 4: Compare with and without collaboration in MAC-Ego3D using Replica dataset (straub2019replica).
• Figure 5: Illustration of collaborative consistency with and without its application.