Transforming Decoder-Only Transformers for Accurate WiFi-Telemetry Based Indoor Localization

Nayan Sanjay Bhatia, Katia Obraczka · 2025

Read Transforming Decoder-Only Transformers for Accurate WiFi-Telemetry Based Indoor Localization by Nayan Sanjay Bhatia, Katia Obraczka (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-15835.txt.

Introduction

Nayan Bhatia and Katia Obraczka repurposed chat-style transformer models to locate devices indoors in a 2025 paper. WiFi signals can pinpoint a phone inside a building for navigation and asset tracking, but fading, reflections, and interference corrupt the signal, and each hardware vendor reports telemetry in its own format. No single model handled all that variety well.

Their system, WiFiGPT, treats telemetry as a language the model can learn, capturing faint spatial patterns in noisy readings. In experiments it acted as a reliable regressor of position, matching leading specialized methods while accepting varied inputs. A unified model could spare each site the cost of hand-building its own locator for every device mix.

Catalog record

CREATORNayan Sanjay Bhatia, Katia Obraczka
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-15835

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Transforming Decoder-Only Transformers for Accurate WiFi-Telemetry Based Indoor Localization Nayan Sanjay Bhatia, Katia Obraczka Abstract Wireless Fidelity (WiFi) based indoor positioning is a widely researched area for determining the position of devices within a wireless network. Accurate indoor location has numerous applications, such as asset tracking and indoor navigation. Despite advances in WiFi localization techniques- in particular approaches that leverage WiFi telemetry- their adoption in practice remains limited due to several factors including environmental changes that cause signal fading, multipath effects, interference, which, in turn, impact positioning accuracy. In addition, telemetry data differs depending on the WiFi device vendor, offering distinct features and formats; use case requirements can also vary widely. Currently, there is no unified model to handle all these variations effectively. In this paper, we present WiFiGPT, a Generative Pretrained Transformer (GPT) based system that is able to handle these variations while achieving high localization accuracy. Our experiments with WiFiGPT demonstrate that GPTs, in particular Large Language Models (LLMs), can effectively capture subtle spatial patterns in noisy wireless telemetry, making them reliable regressors. Compared to existing state-of-the-art methods, our method matches and often surpasses conventional approaches for multiple types of telemetry. Achieving sub-meter accuracy for RSSI and FTM and centimeter-level precision for CSI demonstrates the potential of LLM-based localisation to outperform specialized techniques, all without handcrafted signal processing or calibration. I Introduction Indoor positioning technologies have various applications, ranging from indoor navigation in environments such as airports, hospitals, and shopping malls, accessibility, health- and elderly care to improving smart building systems [ 1 , 2 , 3 ] . However, its adoption in the real world has been slow due to practical issues. Many localization systems are trained and tested on static environments in a controlled manner. However, the real world is dynamic, and even slight environmental variations can cause catastrophic loss in accuracy [ 4 ] . In order to make real impact, localization approaches must be able to adapt to different dynamic scenarios while maintaining high accuracy [ 5 ] . More recently, localization using WiFi telemetry has been receiving considerable attention from researchers and practitioners due to WiFi’s ubiquity, cost-effectiveness and scalability [ 6 ] . WiFi telemetry offers valuable information about the Radio Frequency (RF) environment through metrics such as the Received Signal Strength Indicator (RSSI), Channel State Information (CSI) and Fine Timing Measurement (FTM) [ 7 , 8 ] . This telemetry, especially the CSI, characterizes how a signal propagates from a transmitter to a receiver, encompassing effects like scattering, fading, and power decay [ 9 ] . As AI and ML tools become increasingly prevalent, they have been used for indoor localization systems to improve accuracy, adaptability, and real-time performance in a variety of environments [ 10 ] . Indoor localization has traditionally relied on rule-based algorithms that need telemetry features like RSSI, CSI, or FTM to be manually customized. While effective in custom settings, these models often fail to generalize across environments and require significant engineering effort to adapt to different target scenarios. Common neural networks or non-linear regressor based approaches often require data preprocessing steps such as filtering and data scaling. These methods rely on fixed schemas and struggle with missing or incomplete data, limiting flexibility and scalability across heterogeneous environments [ 11 ] . As Transformers [ 12 ] become more ubiquitous and accessible, we have explored their use to address the challenges posed by accurate WiFi-based indoor localization. Transformers [ 12 ] were originally built for language translation, where the goal is to understand textual sequences and generalize across different contexts. Their generalization capability makes Transformers a good candidate for modeling WiFi telemetry, which is often dynamic, noisy, sparse, and affected by signal propagation impairments. In environments that are heavily affected by the effect of multipath, like corridors or indoor spaces, even small changes in position can lead to significant signal variations, which traditional methods struggle to capture. Transformers, and more broadly LLMs, provide a flexible way to learn patterns from spatio-temporal data without needing hand-crafted features. Their generalization ability, which has already been shown in NLP, computer vision, robotics, and even biomedical signals [ 13 , 14 , 15 , 16 ] , can be applied here to make sense of complex patterns in WiFi environments. In other words, instead of building different models for every setup, we can train a single model that learns to work across environments. Motivated by the cross-generalization capabilities of GPTs, in particular LLMs, we investigate their potential to address indoor localization tasks using WiFi signals to accurately estimate positioning in dynamic environments subject to noise and multipath effects. In this paper, we introduce WiFiGPT, a decoder-only large language model adapted and re-tooled to model real-world WiFi telemetry for accurate indoor localization. LLMs have been typically trained to model textual sequences in natural language. In WiFiGPT, we leverage decoder-only LLMs’ learning capability to model the distribution of telemetry signals, treating each signal feature as an input and mapping the final answer to the device position. We demonstrate that LLMs are not limited to text-based systems and can operate in a noisy environment utilizing wireless telemetry data for indoor positioning. The system is designed with portability, agility, and adaptability in mind. It can quickly adjust to different environments while maintaining high accuracy. This allows fast learning, even from just a few samples, and enables strong cross-generalization. Because of this, the resulting model can be rapidly trained and deployed in real-world scenarios with minimal computational overhead where LLMs already exist, such as smart home devices, smartphones. TABLE I: Comparison of Wireless Localization Techniques Feature CSI RSSI Time-Based AoA Accuracy High Low to Moderate Moderate to High High Resolution Fine-grained Coarse-grained Fine-grained Fine-grained H/W Requirements Specialized (e.g., Intel 5300 NIC) Standard WiFi Timing hardware Multiple antennas Computational Complexity Moderate to High Low Moderate High Multipath Resilience Good (can leverage multi-path) Poor Moderate Good Ease of Implementation Moderate to Complex Simple Moderate Complex In this paper, we make the following contributions: • To the best of our knowledge, WiFiGPT is the first system repurposing decoder-only LLMs for indoor localization. • We present a generalized framework with deterministic outputs that integrates diverse telemetry sources and adapts across radio environments. Extending LLMs beyond language, we apply a simple training scheme to map telemetry sequences to distance predictions for sensor-based regression. • We show that LLMs can handle WiFi telemetry directly without manual feature engineering. WiFiGPT performs well on real-world indoor datasets and adapts quickly, even with limited training data, getting sub-meter accuracy. • We analyze how the accuracy of different LLaMA model sizes varies under changing environmental conditions. The rest of the paper is organized as follow: In Section II, we provide an overview of the background and related work setting the stage for our contributions. Section III describes in detail our large WiFi-telemetry model for indoor localization, while in Section IV, we present its implementation details. Section V describes how we evaluated the performance of our system and Section VI presents and discusses our results. Finally, Section VII concludes the paper and discusses directions for future work. II BACKGROUND AND RELATED WORK In this section, we start with a brief overview of the background on WiFi telemetry

Read the complete text