⚡ electrical engineering

Towards Networked One Search Agent Systems: Multilateration of WiFi Fine Time Measurement Responders Using GNSS References

This paper presents a proof-of-concept system that enables an uncrewed aerial vehicle to localize ground-based WiFi access points in real-time by combining IEEE 802.11mc Fine Time Measurement ranging with GNSS-referenced multilateration, achieving sub-5-meter horizontal accuracy for line-of-sight targets in challenging mountainous terrain.

Original authors: Juan Bravo-Arrabal, Javier Serón-Barba, Carlos Simón Álvarez-Merino, J. J. Fernández-Lozano, Alfonso García-Cerezo, Anders Lyhne Christensen

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Juan Bravo-Arrabal, Javier Serón-Barba, Carlos Simón Álvarez-Merino, J. J. Fernández-Lozano, Alfonso García-Cerezo, Anders Lyhne Christensen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a search and rescue drone flying over a rugged, mountainous landscape. Its mission is to find a lost hiker. But the hiker isn't holding a high-tech beacon; they are just carrying a standard smartphone or a common WiFi router. The drone's job is to figure out exactly where that device is, even if it's hidden behind bushes or rocks.

This paper describes a "proof-of-concept" system that turns a regular drone into a super-smart detective using nothing but WiFi and GPS.

Here is how the system works, broken down into simple concepts:

1. The "Ping" Game (The Core Idea)

Usually, to find something with WiFi, you just look at how strong the signal is (like how loud a radio station is). But signal strength is tricky; a wall can make a signal look weak even if you are close.

Instead, this system uses a special feature built into modern WiFi called Fine Time Measurement (FTM). Think of this as a game of "ping-pong" with time.

The drone (the "initiator") sends a signal to the ground device (the "responder").
The ground device bounces it back immediately.
The drone measures exactly how long that round trip took.
Since light travels at a known speed, the drone can calculate the exact distance: Time = Distance.

2. The Moving Anchor (The GPS Twist)

Here is the clever part. Usually, to triangulate a position, you need three fixed towers. But this drone is moving!

The drone is equipped with a high-precision GPS receiver.
Every time the drone "pings" the ground device, it knows exactly where it was at that precise moment.
Imagine the drone flying in a circle around the hidden device. Every time it pings, it leaves a "virtual anchor" in the air at its current GPS location.
After dozens of pings from different angles, the system has a cloud of virtual anchors. It then uses math to find the one spot on the ground where all those distance lines intersect. This is called Multilateration.

3. The "Smart Filter" (Cleaning Up the Mess)

In the real world, things aren't perfect. The signal might bounce off a tree (multipath), or the GPS might wobble a bit. The system uses a "smart filter" (a mathematical brain) to clean up the data:

The Bouncer: It throws out obvious bad data points (outliers) that don't make sense, like a measurement that says the device is 100 meters away when it was just 10 meters away a second ago.
The Weighting System: It knows that a signal coming from far away or through a wall is less reliable, so it gives those measurements less "weight" in the final calculation.
The Bias Tracker: It notices if the drone's clock is slightly off and corrects for it on the fly.

4. The Results: What Worked and What Didn't

The researchers tested this on a university campus with three different "victims" (WiFi routers):

The Open Field (Line-of-Sight): When the router was out in the open, the system worked beautifully. After flying around for a while, the drone pinpointed the router's location with an error of about 4.4 meters (roughly the length of a car). If the drone flew very close at the end, the error dropped to just 1.1 meters.
The Bushes (Moderate Obstruction): When the router was hidden under thick vegetation, the system struggled. The signal was too weak and the geometry (the angles the drone flew) wasn't quite right. The system couldn't lock on in real-time, though a computer could find a rough location (about 7 meters off) by analyzing the data later.
The Concrete Block (Severe Obstruction): When the router was buried under concrete and bushes, the signal was completely blocked. The system found nothing.

5. Why This Matters

The paper argues that this is a building block for a future where multiple drones work together.

No Special Gear Needed: The "victims" don't need to carry special equipment; a standard WiFi router or phone is enough.
One Agent, Many Jobs: The drone acts as the sensor, the GPS receiver, the computer, and the internet gateway all in one.
Networked Search: The ultimate goal is to have a team of these drones talking to each other (via a system called ROS 2) to share what they see, creating a networked search team.

In summary: This paper shows that a drone with a smartphone and a GPS can act like a high-tech radar, using standard WiFi signals to find hidden devices on the ground. It works great in open areas, struggles a bit in thick bushes, and fails completely if the signal is totally blocked, but it proves that we don't need expensive, custom hardware to do this kind of search.

Technical Summary: Towards Networked One Search Agent Systems

Problem Statement
Accurate localization of stationary wireless emitters is critical for robotic applications such as search and rescue (SAR) and infrastructure mapping. While Global Navigation Satellite Systems (GNSS) are standard, they degrade in complex environments. Existing solutions often rely on signal-strength fingerprinting (requiring prior calibration) or dedicated ranging technologies like Ultra-Wideband (UWB), which demand purpose-built infrastructure. The IEEE 802.11mc Fine Time Measurement (FTM) protocol offers a middle ground, providing meter-level range estimates using commodity WiFi hardware without requiring clock synchronization. However, leveraging FTM for outdoor localization of ground-based access points (APs) from an Uncrewed Aerial Vehicle (UAV) presents challenges: the system must perform real-time multilateration using a moving baseline (the UAV) while handling measurement noise, non-line-of-sight (NLoS) conditions, and limited computational resources on the edge device.

Methodology
The authors propose a proof-of-concept system where a UAV acts as an FTM initiator and GNSS receiver, localizing ground-based WiFi APs acting as FTM responders. The core architecture inverts previous models: instead of the victim carrying a smartphone, the search agent (UAV) carries the computation, while the target carries a simple, passive AP.

Hardware Platform: The system utilizes a DJI Matrice 350 RTK UAV carrying a Google Pixel 7 Pro smartphone and a u-blox ZED-F9P multi-band GNSS receiver. The smartphone runs a custom Android application that interfaces with the WiFi RTT API and the GNSS receiver.
Localization Pipeline: The system treats the flight trajectory as a temporal multilateration problem. The pipeline includes:
1. Data Association: Synchronizing FTM ranging timestamps with GNSS positions using a monotonic clock, with a 1.2-second association window.
2. Outlier Gating: A sliding-window Median Absolute Deviation (MAD) filter rejects gross outliers while preserving valid Line-of-Sight (LoS) measurements.
3. Initialization: A two-stage Gauss-Newton bootstrap computes an initial 3D position estimate once sufficient geometric spread is achieved.
4. Sequential Filtering: The system employs an Iterated Extended Kalman Filter (IEKF) with bias tracking or an Unscented Kalman Filter (UKF). The state vector includes the AP position $(x, y, z)$ and a residual range bias.
5. Noise Modeling: Six distinct measurement-noise configurations were evaluated, ranging from a constant baseline model to adaptive models that adjust variance based on distance, Received Signal Strength Indicator (RSSI), and NLoS detection (using range-window skewness and RSSI consistency).
Constraints: The system operates autonomously without network connectivity, logging data to JSONL files, though it supports ROS 2 for multi-agent fusion when a link is available.

Key Contributions

Real-Time Estimation Pipeline: A complete pipeline running on a commodity smartphone, integrating ranging, positioning, and estimation (IEKF/UKF) with bias tracking and outlier gating.
Noise Configuration Analysis: A comparative evaluation of six measurement-noise models, demonstrating how distance- and RSSI-dependent weighting preserves measurement diversity and enables late-stage corrections during close flybys.
Field Validation: Proof-of-concept testing in unstructured, mountainous terrain with three APs under varying concealment conditions: LoS, moderate NLoS (vegetation), and severe NLoS (vegetation + debris).
Networked Architecture Integration: Preliminary demonstration of software-level interoperability, showing heterogeneous agents (UAV and ground) can share FTM observations via ROS 2, serving as a building block for "Networked One Search Agent" (OSA) systems.

Results
Experiments were conducted at the University of Málaga with three APs.

LoS Performance (AP_1F): For an AP with 455 ranging measurements:
- The online real-time pipeline (BASELINE mode) achieved a final horizontal error of 4.4 m and a 3D error of 5.5 m.
- Offline replay using the EMPIRICAL-R noise model (distance and RSSI dependent) achieved a time-weighted mean (TWM) horizontal error of 4.7 m.
- Under the best noise model, a close flyby allowed the system to refine the estimate to a final horizontal error of 1.1 m.
- The ADAPTIVE-R model achieved the smallest internal uncertainty ( $\sigma_H = 1.0$ m) but a final error of 4.3 m.
NLoS Performance:
- Moderate NLoS (AP_AB): The system failed to converge in real-time due to sparse measurements (31 total), signal attenuation, and near-collinear geometry (condition number $\kappa = 28.5$ ). However, an offline robust least-squares solver recovered a coarse solution with approximately 7 m horizontal error.
- Severe NLoS (AP_AA): The AP, blocked by vegetation and concrete debris, produced zero FTM responses.
Filter Confidence: The authors note that the filter's internal covariance ( $\sigma_H$ ) significantly underestimates the actual absolute error (e.g., $\sigma_H \approx 1.1$ m vs. actual error $\approx 4.4$ m). This is attributed to unmodeled correlated errors in the GNSS-referenced initiator poses, as the system did not include a separate GNSS covariance term in the comparison models.

Significance and Claims
The paper positions this work as an architectural advancement for Networked OSA systems rather than a claim of superior accuracy over previous datasets. By inverting the localization architecture—placing the computational burden on the search agent and requiring only a simple FTM-capable responder on the target—the system eliminates the need for software on the victim side.

The authors emphasize that the system is a building block for future multi-agent networks. While the single-agent pipeline demonstrates meter-level LoS performance, the paper modestly acknowledges that NLoS operation remains a significant challenge due to geometry and signal blockage. The 1.1 m result is characterized as a "favourable close-pass outcome" dependent on specific flight geometry, not a guaranteed system bound. Future work is directed toward network-level fusion, improved flight trajectory planning (e.g., helical or grid patterns) to optimize multilateration conditioning, and the development of damped robust solvers for degenerate geometries.