SparseLoc: Sparse Open-Set Landmark-based Global Localization for Autonomous Navigation

SparseLoc is a novel framework for global localization (finding a robot’s position on a map without GPS) that addresses the high storage costs of traditional navigation methods. It does this by creating sparse, semantically-rich maps instead of dense point clouds. Most modern techniques rely on dense LiDAR maps, which are precise but become prohibitively large (often gigabytes) and computationally expensive to scale for city-wide navigation. Other sparse-map approaches often lack the robustness needed for reliable localization.

Our method uses vision-language foundation models (VLMs) in a zero-shot way to identify and map only the centroids of static, open-set landmarks (like ‘traffic pole’ or ‘hedge’). This creates a “semantic-topometric” map that is 1/500th the size of a traditional dense map. For localization, we use a robust Monte Carlo particle filter enhanced with a novel late optimization strategy to refine the pose estimate. SparseLoc achieves accuracy comparable to dense-map methods (below 5m and 2° error on KITTI) and dramatically outperforms other sparse-map techniques. It also demonstrates high robustness in challenging cross-dataset localization scenarios.

More information is available on the project page.