Hierarchical Gaussian Mapping with Controlled Growth for Compact Robotic Navigation

Existing 3D Gaussian Splatting (3DGS) methods in robotic navigation suffer from multi-scale information coupling and the continual accumulation of redundant Gaussians, which makes map size and storage overhead difficult to control and limits deployment on resource-constrained platforms. To address this issue, a hierarchical Gaussian mapping framework is proposed to restructure 3DGS. The framework organizes the scene into three layers, L0, L1, and L2, corresponding to global structure, mid-scale geometry, and fine details. This design turns coupled multi-scale information into a decomposable map representation. Based on this layered representation, a controlled growth mechanism uses visibility, scale, gradient response, and reconstruction error to govern selective refinement across layers. A progressive Level-Of-Detail (LOD) strategy is further introduced, enabling layer-wise access and on-demand composition of map representations under different task requirements. Experimental results show that the method significantly reduces map size and storage overhead while preserving reconstruction quality.