Hierarchical Feature-based Localization using Scene Graphs in Off-road Navigation

While numerous advancements have been made in autonomous navigation for structured indoor and outdoor environments, these solutions often do not generalize well to off-road settings. There are unique challenges in such settings such as unreliable GPS, limited computational and memory resources, and sparse environmental features, making navigation particularly difficult. In our work, we propose a novel data structure called Hierarchical Dynamic Scene Graphs (HDSG) to address these challenges. HDSG captures environmental information at different resolutions, integrating both geometric and semantic features. It enables various navigation tasks such as localization, loop closure, and human interaction through the visualization of environmental features for remote operators. We evaluated the performance of localizing a robot's position within the world frame by comparing compact spatial descriptors extracted from semi-consecutive scene graphs, derived from 3D LiDAR point clouds. Compared to directly applying traditional Iterative Closest Point (ICP) algorithms on point clouds, our approach demonstrates that localization on scene graphs is more efficient and accurate. In evaluations using the RELLIS-3D dataset, the HDSG is constructed in at most 5 seconds using only commodity hardware. The overall memory footprint of the HDSG is very compact accounting for only 450-500 MB. Moreover, using the HDSG for robot localization has demonstrated faster and more precise than using traditional ICP approaches directly on the input point cloud. These results highlight the potential of scene graph-based localization to deliver faster, more memory-efficient, and more accurate performance in unstructured off-road environments, showing a promising foundation for future enhancements and applications.