Resolve open challenges in 3DGS-based SLAM: sparse views, scalable memory, and semantic consistency

Investigate and develop Simultaneous Localization and Mapping (SLAM) methods based on 3D Gaussian Splatting that (i) operate robustly under sparse-view input, (ii) employ scalable memory mechanisms suitable for large environments, and (iii) maintain consistent semantic representations across large-scale scenes to enable reliable mapping and high-level reasoning.

Background

The survey categorizes 3DGS-based SLAM into geometry-focused approaches (RGB-D and RGB) and semantics-aware SLAM, noting that RGB SLAM must infer geometry via multi-view optimization or depth prediction and can be unstable in dynamic or outdoor settings. Semantics-aware SLAM integrates learned features or language-guided Gaussian representations to enhance high-level reasoning.

Despite these advances, the paper explicitly notes remaining challenges for SLAM within 3D Gaussian Splatting frameworks, specifically the need to handle sparse views, to design memory systems that scale to large scenes, and to ensure semantic consistency across extensive environments. Addressing these issues is important for robust real-world deployment and downstream tasks such as navigation and interaction.