WiFi-based Global Localization in Large-Scale Environments Leveraging Structural Priors from osmAG

arXiv:2508.10144 · cs.RO · Submitted 2025-08-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "WiFi-based Global Localization in Large-Scale Environments Leveraging Structural Priors from osmAG".

Title and authors: Dev: Let’s look at who wrote this, Rosa and Dev are leading on this work. They’ve put together a framework that uses WiFi APs to build a global map of an indoor area by combining signal physics with the osmAG structure.

Rosa: It really boils down to moving away from having to manually survey every single Access Point in a building. Instead, the authors use their framework to localize those APs automatically during an offline phase, which is a huge time saver for anyone trying to map a large facility.

Taro: So it’s not just about mapping the physical space; it’s about using that structure—the graph of nodes and ways in osmAG—to guide how we find the WiFi signals themselves.

Dev: That's right, Taro. They create a unified map format where you can store both the robot's past positions and the WiFi data, which is what they call fingerprint data, directly onto the osmAG structure itself to make it very space efficient.

Rosa: And that efficiency is key because it means you don't need to store massive amounts of raw fingerprint data anymore; typically, you only need about one order of magnitude less information than a traditional method requires.

Taro: So instead of storing huge datasets for every corner, the system stores the structural insight from the map itself and uses that to pinpoint those APs reliably.

Dev: Precisely. The goal here is to create a scalable solution for large indoor environments where exhaustive manual surveying would just be impossible or take way too long.

The paper's summary: Rosa: So, what the paper actually does is lay out a two-phase process: first, an offline phase where they use an iterative optimization algorithm to figure out exactly where every WiFi Access Point is located by modeling wall attenuation.

Taro: That sounds like the hard part—getting those AP locations right when you’re only guessing based on signal strength measurements and some initial map data.

Dev: They tackle that by first estimating the line-of-sight parameters, which are things like RSSI zero and a path loss exponent, then they use non-line-of-sight measurements to estimate a single wall attenuation factor for all walls.

Rosa: Once they have those environmental parameters—the RSSI zero, the path loss exponent, and that wall attenuation factor—they can then iteratively refine the position of each Access Point based on how far away the robot is estimated to be.

Taro: And in that offline phase, they store these localized AP positions back into the augmented osmAG map permanently so it’s ready for the next step.

Dev: The online phase is where things get real-time for a robot. The robot queries this pre-calibrated map to find the known AP locations and then performs a robust, real-time global localization using that information.

Rosa: And the results they got were pretty solid: in areas with good fingerprint data, they saw an error of three point twelve meters, which is an eight point seven seven percent improvement over traditional KNN fingerprinting methods.

The paper's improvements: Taro: What stands out to me is how much better the performance becomes when the robot moves into areas where it hasn't seen any specific training data before.

Dev: That’s where they show a massive jump in performance, achieving a mean error of three point eighty-three meters in non-fingerprinted areas, which is an eighty-one point zero five percent improvement over KNN fingerprinting in those zones.

Rosa: And the paper highlights that this method is much more space efficient than the fingerprint-based approaches because you don't need to store that huge amount of data.

Taro: So it’s not just about accuracy; it’s about making the system practical for deployment where you can't afford massive storage or constant retraining on new locations.

Dev: The authors also claim this approach eliminates the need for manual Access Point surveying because they fuse that physics-based signal modeling directly with the geometric and topological priors from osmAG.

Rosa: That fusion is what makes it robust; it uses the map structure to compensate for signal distortions caused by walls, which is something traditional trilateration just can't handle well on its own.

Conclusion: Dev: So to wrap this up, the authors introduce a novel physics-informed global WiFi localization method that leverages structural priors from osmAG to enable robust and efficient robot localization in large-scale environments.

Rosa: Essentially, they show that by first localizing the Access Points using signal modeling against the map structure, you get a superior real-time system that performs better than traditional fingerprinting methods across both fingerprinted and non-fingerprinted areas.

Taro: The main implication for us is that this approach creates a foundational data layer—the pre-calibrated AP map stored in osmAG—which can actually serve as a prior for future lifelong indoor LiDAR localization systems.

Dev: That’s right. It shifts the problem from constantly collecting massive amounts of data to efficiently using structural information and physics models to achieve high accuracy, especially when training data is scarce.

Rosa: The paper also releases the code and the dataset, which is important because it lets other researchers test this out in their own large-scale environments immediately.

Taro: It’s a practical step toward making indoor localization a more scalable and reliable thing for real robots operating in complex settings.

Xu Ma, Jiajie Zhang, Fujing Xie, Soren Schwertfeger

ShanghaiTech University

cs.RO

Submitted: 2025-08-13

Updated: 2026-11-13

Comments: IEEE ROBIO 2026

Code: https://github.com/XuMa369/osmag-wifi-localization

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 75/100

The gist: Global localization is essential for autonomous robotics, especially in indoor environments where the GPS signal is deniedThis capability is particularly vital for addressing the robot kidnapping…

Key concepts

WiFi AP Localization
This is the offline process where the system identifies the physical locations of all WiFi Access Points within a building. It uses signal propagation models (like wall attenuation) combined with geometric constraints from osmAG to calculate where each AP is situated. This replaces time-consuming manual surveying.
osmAG Map Representation
The map structure combines physical geometry (nodes representing corners and ways representing corridors) with WiFi data. It uses XML tags within this structure to store both robot movement history (fingerprints) and the calculated locations of APs, creating a rich, multi-purpose spatial model.
Physics-Informed Framework
This approach integrates mathematical models of radio wave propagation (log-distance path loss) with the map's geometry. During localization, the system doesn't just use raw signal strength; it uses these physical rules to accurately compensate for signal weakening caused by walls, leading to more precise distance and position estimates.
Two-Phase System
The framework operates in two distinct stages. The offline phase builds a foundational layer by precisely locating all APs and storing this knowledge in the osmAG map. The online phase then uses this pre-calibrated map to perform robust, real-time robot tracking without needing continuous re-surveying.

Terminology

Summary

Global localization is essential for autonomous robotics, especially in indoor environments where the GPS signal is deniedThis capability is particularly vital for addressing the robot kidnapping problem in GPS-denied environments, such as indoor spaces, underground facilities, and urban canyonsThe authors propose a novel WiFi-based localization framework that leverages ubiquitous wireless infrastructure and the OpenStreetMap Area Graph (osmAG) for large-scale indoor environmentsOur approach integrates signal propagation modeling with osmAG’s geometric and topological priorsIn the offline phase, an iterative optimization algorithm localizes WiFi Access Points (APs) by modeling wall attenuation, achieving a mean localization error of 3. p. 079 m (35.3% improvement over trilateration)In the online phase, real-time robot localization uses the augmented osmAG map, yielding a mean error of 3. p. 012 m in fingerprinted areas (8. p. 077% improvement over KNN fingerprinting) and 3. p. 083 m in non-fingerprinted areas (81. p. 005% improvement)This framework offers a scalable, cost-effective solution for indoor robotic localization, solving the kidnapped robot problemThe primary contributions of this work are: A WiFi AP localization framework that fuses physicsbased signal modeling with osmAG’s geometric and topological priors, eliminating the need for manual AP surveyingAn efficient real-time robot localization system exploiting the pre-calibrated AP map and osmAG’s structural insights for superior accuracyComprehensive validation across a 11,025 m2 multifloor environment, outperforming traditional fingerprinting methods in localization precisionThe public release code of our complete localization framework and dataset, alongside a detailed specification for an osmAG-compatible representation of WiFi data

Framework Architecture

The proposed large-scale indoor localization framework is architecturally divided into two primary phases: an offline phase for Access Points (APs) localization and an online phase for real-time robot localizationIn the offline phase, we populate the osmAG map with the building’s geometry and a WiFi dataset comprising robot poses (w.r.t. the map frame) and synchronized WiFi RSSI measurementsAn optimization algorithm then leverages this integrated data to compute the locations of all WiFi APs, explicitly modeling wall attenuation, and permanently stores these locations back into the mapIn the subsequent online phase, the robot queries this augmented map to retrieve the precise AP locations and performs robust, real-time global localizationThe system comprises an offline phase, where an iterative optimization algorithm is employed to accurately localize all WiFi Access Points (APs)In the online phase, the robot queries the augmented osmAG map to retrieve AP positions and performs robust, real-time global localization

Map Representation

The osmAG map is meticulously structured to encapsulate both the physical layout of the indoor environment and the specialized data required for WiFi-based localizationIt uses XML format to store data obtained from CAD, 2D grid maps, or 3D point cloud mapsThis representation is composed of two primary components: a geometric-topological component and a WiFi-associated componentNodes are discrete points defined by geographical coordinates (latitude and longitude), functioning as the vertices in the map’s graph structure, representing key locations such as corners of rooms, ends of corridors, or doorwaysWays are ordered sequences of nodes that define polygonal areasTogether, nodes and ways form a comprehensive graph that describes not only the geometry (the shape and size of rooms) but also the topology (the connectivity between different spaces, such as how corridors link various rooms)Our osmAG format can be used to store both: fingerprint data, which is also used for AP localization, and AP dataFingerprint data contains the time-synchronized robot position, acquired from an accurate Simultaneous Localization and Mapping (SLAM) algorithm, e.g. [14] and the corresponding WiFi Received Signal Strength Indicator (RSSI) measurementsFor this we utilize of osmAG tags, which are key-value pairs associated with nodes, as illustrated in Figure 2By integrating these two components into a unified osmAG representation, we create a rich, multi-purpose map that simultaneously provides a detailed model of the physical world and the essential data for robust indoor localization

Offline Phase Details

The primary objective of the offline phase is to localize all Access Points (APs) from previously collected fingerprint data and embed their positions into the osmAG map, thereby establishing the foundational data layer required for the subsequent online robot localizationIn this stage we utilize the ground truth positions of a few APs (estimated by hand within the map), together with the map and some fingerprints, to estimate certain model parametersWe adopt the log-distance path loss model, a standard for describing indoor radio wave propagationThe goal is to estimate the parameters RSSI0, n, and an average wall attenuation factor WStage 1: LOS Parameter Estimation. We first isolate a set of Line-of-Sight (LOS) measurementsUsing this LOS dataset (SLOS), we solve for the baseline parameters (RSSI∗0, n∗) via least-squares regressionStage 2: NLOS Wall Attenuation Estimation. p. 2 With the baseline parameters fixed, we then use NLOS measurements to estimate a single, effective wall attenuation factor W, that will be used for all wallsWith the environmental parameters (RSSI∗0, n∗, W) established, we proceed to estimate the unknown position of each APInitialization: For a target AP, we first convert all its associated RSSI measurements from the WiFi dataset into initial distance estimates ˆdi between AP and robot, using the LOS modelWe then compute an initial 3D position estimate p (0) AP by solving a least-squares minimization problem:p(0)AP = arg min p X N i=1∥p − pi∥ − ˆdi 2Iterative Refinement: At each iteration t, we perform wall intersection tests between the current AP position estimate p(t−1)AP and all robot positions piFor LOS measurements, the distance estimate remains uncompensatedWe solve the least-squares problem again with these refined distance estimates to obtain the updated AP position p(t)AP

Online Phase Execution

In the online phase, the robot uses the augmented osmAG map from the offline phase containing the now-known AP locations and environmental parameters to determine its own position in real timeThe robot continuously scans for WiFi signals, collecting RSSI measurements over a brief temporal window (e.g., 3-5 seconds) and averaging them to ensure signal stability for the subsequent localization processAn initial 3D position estimate p(0) robot is computed via trilateration using the detected APs and their known locations from the osmAG mapAt each iteration t, the system performs wall intersection checks between the current robot position estimate p(t−1)robot and all visible APsIt then calculates compensated distance estimates and solves a leastsquares problem to find the updated position p(t)robotThe process terminates when the change in position between iterations falls below a predefined threshold (∥p(t)robot −p(t−1)robot ∥ < ϵ) or a maximum number of iterations is reached

Experimental Validation

Our experiments were conducted in a multi-floor indoor environment at ShanghaiTech University, covering 11,025 m2 across two floors and two buildingsWe used an Agilex Hunter SE mobile platform equipped with an EDUP EPAC1681 WiFi receiver and a Hesai 64-channel LiDAR for data acquisitionThe environment, comprising classrooms, labs, and offices with various wall types (concrete, drywall, glass), provided realistic signal propagation conditionsWe selected the first floor of the SIST building as our training areaThe optimization yielded the following values: Reference Signal Strength (RSSI∗0): -28.79, Path Loss Exponent (n∗): 2. p. 25, Wall Attenuation Factor (W): -10. p. 277Applying our iterative optimization algorithm, which leverages geometric data from the osmAG map to compensate for wall attenuation, yielded a significant improvementOur physics-informed framework reduced the mean localization error to 3. p. 479 m with a standard deviation of 2. p. 452 m, a 35.3% improvement over the baselineIn this scenario, our method achieves a slightly lower mean localization error, demonstrating its effectiveness compared to the fingerprinting algorithmIn this second, more challenging scenario was designed to assess the system’s generalization capabilityIn contrast, as detailed in Table I, our method of robot localization via localized AP positions produced a significantly lower localization error, demonstrating its robustness and practical applicability for large-scale deployment where exhaustive fingerprinting is infeasible

Conclusion

In this paper, we introduced a novel physics-informed global WiFi localization method that leverages structural priors from osmAG to enable robust and efficient robot localization in large-scale environmentsOur experiments show that our AP localization works well, that storing AP information in osmAG is much more space efficient than fingerprints and that the robot localization of our approach is superior to fingerprint approaches, especially in areas for which no fingerprint data is availableThis WiFi localization can act as the prior for our approach to robust lifelong indoor LiDAR localization with osmAG [16]

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AP localization accuracy by fusing physics-based signal modeling with osmAG’s geometric and topological priors, eliminating manual AP surveying. This allows for a scalable solution that is much more space efficient than fingerprint data, reducing storage requirements by over one order of magnitude.

  2. Enhanced robot localization robustness in non-fingerprinted areas, achieving an 81.05% improvement over KNN fingerprinting by leveraging the pre-calibrated AP map and physics-informed model. This enables reliable operation in areas where traditional methods fail due to lack of training data.

  3. Created a two-phase system that first focuses on high-fidelity AP localization offline, storing AP locations as semantic tags within the osmAG map for long-term support. This provides a foundational data layer that supports robust lifelong indoor LiDAR localization with osmAG.

Sources

Related papers