Image-goal navigation requires a global planner that can connect visually specified start and goal points while supplying observations that a local policy can follow. The paper introduces GS-PRM, a Gaussian Visual Roadmap, to address this task.
The method leverages 3D Gaussian Splatting (3DGS) to create map-registered RGB representations. These representations serve as the basis for a roadmap that links visual endpoints, giving the local policy the perceptual context it needs to navigate.
By using 3DGS as the underlying map representation, the approach aims to make global planning more effective in image-goal navigation scenarios. The abstract does not provide experimental details or comparisons, so the reported contribution is the planner design itself.