Abstract
Lifelong indoor localization is a fundamental requirement for autonomous mobile robots. Traditional map representations, such as occupancy grids or point clouds, are often "brittle" because they capture transient environmental features like furniture, which change over time. This paper introduces AGLoc, a robust localization system that utilizes the Area Graph—a hierarchical, topometric semantic map representation. By matching 3D LiDAR point clouds against permanent architectural features (walls and doors) defined by polygons, AGLoc achieves stable, lifelong localization in cluttered environments. Our experiments demonstrate that AGLoc provides high-accuracy global localization and efficient pose tracking, outperforming standard methods in dynamic indoor spaces.
Introduction
Most autonomous robots rely on 2D occupancy grids or 3D point clouds for navigation. While precise, these maps require frequent updates as the environment changes—for instance, when furniture is moved or partitions are rearranged. To achieve true lifelong autonomy, robots need maps that are compact, stable, and easy to acquire from existing architectural data.
AGLoc addresses this by shifting the focus from transient clutter to the building's permanent "skeleton." By using an Area Graph, the system represents rooms, corridors, and buildings as a series of polygons. This representation is not only orders of magnitude smaller than traditional grids but also remains valid for years without requiring updates.
The AGLoc Methodology
The AGLoc system processes 3D LiDAR data through a specialized pipeline designed to filter out noise and focus on architectural features. The process begins with a broad initial guess from WiFi and barometric data to determine the correct floor and general vicinity.
Key Technical Innovations
- Clutter Removal Subsampling: To handle cluttered environments, the system projects 3D points into 2D and retains only the furthest points in each angular column. This heuristic assumes that the furthest returns are most likely to be permanent walls rather than foreground clutter like chairs or people.
- Global Localization via Sampling: AGLoc solves the "kidnapped robot" problem by sampling thousands of potential poses within a WiFi-defined radius and scoring them against the Area Graph polygons using a custom error function.
- Corridorness Downsampling: In long, featureless corridors, standard ICP (Iterative Closest Point) algorithms often "slide" due to a lack of longitudinal features. AGLoc calculates a "corridorness" score based on wall orientation histograms and downsamples points on dominant parallel walls to emphasize rare features like door indentations.
- Weighted Point-to-Line ICP: For active tracking, the system uses a specialized weight function that ignores points likely to be clutter (inside rooms) or reflections (from glass), ensuring the robot stays locked onto the structural map.
Experimental Results
The authors evaluated AGLoc using datasets captured in heavily cluttered laboratories and office corridors. The system achieved a 98-99% success rate in global localization, finding the robot's position within 1.0 meter starting from a broad 6-meter search radius.
In terms of efficiency, pose tracking is remarkably fast, requiring only 12.5ms per frame on modest hardware. Head-to-head comparisons showed that AGLoc significantly outperformed standard AMCL (Adaptive Monte Carlo Localization) in corridor environments, particularly when maps were sparse.
Conclusion and Future Work
AGLoc demonstrates that architectural features are a superior foundation for lifelong robot localization compared to traditional occupancy grids. By leveraging the Area Graph, robots can navigate dynamic environments without the overhead of constant map maintenance. Future work will focus on optimizing global localization speed, integrating topological connectivity data for enhanced robustness, and developing more sophisticated semantic state estimation for dynamic elements like doors.
Summary updated