🤖 machine learning

OmniLoc: A Geometry-Aware Foundation Model for Anchor-Free UE Localization Across Diverse Indoor Environments

OmniLoc is a pioneering geometry-aware foundation model that achieves robust, anchor-free indoor user equipment localization across diverse environments by unifying heterogeneous wireless measurements through a specialized tokenization module, a geometry-aware Transformer for feature extraction, and a geometrically consistent regression mechanism.

Original authors: Lei Chu, Yuning Zhang, Omer Gokalp Serbetci, Anushka Katiyar, Bassel Abou Ali Modad, Andreas F. Molisch

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lei Chu, Yuning Zhang, Omer Gokalp Serbetci, Anushka Katiyar, Bassel Abou Ali Modad, Andreas F. Molisch

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 you are trying to find your way inside a massive, confusing shopping mall. You don't have a map, and the store signs (which act like Wi-Fi signals) keep changing depending on which floor you are on or which building you are in. Some floors have dozens of stores; others have only a few. Some signals are loud and clear; others are faint and garbled by walls and crowds. This is the daily struggle of indoor localization for your phone or device.

The paper introduces OmniLoc, a new "smart guide" designed to solve this problem without needing a pre-drawn map of every single building.

Here is how OmniLoc works, explained through simple analogies:

1. The Problem: The "Chaos" of Indoor Signals

Current methods are like trying to navigate that mall using a list of store names that changes every time you turn a corner.

Different Buildings, Different Rules: A signal that means "You are near the coffee shop" in one building might mean "You are in the parking lot" in another because the walls and layouts are different.
Missing Pieces: Sometimes your phone sees 20 Wi-Fi routers; sometimes it only sees 3. Old systems get confused when the number of signals changes.
Noisy Data: The strength of a signal (how loud the "store sign" is) can be misleading. A loud signal might just mean a wall is thin, not that you are close to the router.

2. The Solution: OmniLoc (The "Universal Translator")

The authors built OmniLoc as a Foundation Model. Think of this not as a specific map for one building, but as a super-smart student who has studied thousands of different malls, airports, and office parks. Instead of memorizing a specific floor plan, it learns the language of wireless signals.

OmniLoc uses three main "superpowers" to navigate:

A. The Universal Translator (Unified Input Tokenization)

Imagine you have a list of ingredients: some are measured in cups, some in grams, some in ounces. A normal chef gets confused. OmniLoc acts like a translator that converts all these different measurements into a single, standard "recipe language."

How it works: It takes messy, different types of data (signal strength, signal quality, and which routers are visible) and turns them into a neat, standardized sentence that the computer can easily read, regardless of how many routers are actually present.

B. The Smart Detective (Geometry-Aware Transformer)

Once the data is translated, OmniLoc uses a "detective" (a Transformer model) to figure out where you are.

The "Star Witness": Usually, one strong Wi-Fi signal is the most important clue. OmniLoc knows to pay extra attention to this "star witness."
The Supporting Cast: But it doesn't ignore the others. It gathers "testimony" from the weaker signals to confirm the story.
The Analogy: If you are in a crowded room, you don't just listen to the loudest voice; you listen to the whole conversation to understand the context. OmniLoc does this with Wi-Fi signals, understanding that the pattern of signals matters more than just one loud signal.

C. The Contextual Map (Geometry-Aware Location Estimation)

This is the final step where the detective makes a guess.

The "Address" Check: Before guessing your exact coordinates (X and Y), the model first asks: "Which building is this? Which floor?"
Why it matters: Knowing you are in "Building A, Floor 2" changes how the model interprets the signals. It's like knowing you are in a library versus a gym; the same sound means something different in each place. By forcing the model to identify the building and floor first, it ensures the final location guess makes geometric sense.

3. The Results: Why It Matters

The researchers tested OmniLoc on a huge dataset covering 16 different buildings with thousands of Wi-Fi routers.

Better Accuracy: It found locations much more accurately than previous methods, especially in tricky situations where signals were weak or the building layout was complex.
The "Plug-and-Play" Effect: The authors showed that if you take other existing localization tools and just add OmniLoc's "translator" and "detective" parts to them, those tools also get significantly better. It's like giving a regular car a turbocharger; the car was already good, but now it's a race car.
Learning from Few Examples: Even if they showed the model a brand new building it had never seen before, it could adapt very quickly with just a tiny bit of new data, unlike older systems that would fail completely.

Summary

OmniLoc is a new AI system that learns to "speak" Wi-Fi signals fluently. Instead of trying to memorize every building, it learns the underlying rules of how signals behave in different shapes and sizes. By translating messy data into a clear language, listening to the most important signals while ignoring the noise, and always checking the "address" (building and floor) before making a guess, it can find your location anywhere indoors with high precision, even without a pre-installed map.

Technical Summary of OmniLoc: A Geometry-Aware Foundation Model for Anchor-Free UE Localization

Problem Statement
Indoor user equipment (UE) localization remains a significant challenge in large-scale deployments due to substantial variations in building geometry, the number of detectable access points (APs), and the heterogeneity of received signals. Existing learning-based methods often struggle to generalize across diverse environments, particularly in anchor-free settings where localization relies directly on wireless measurements without explicit infrastructure calibration. Key difficulties include:

Geometric Diversity: Differences in floor plans, materials, and room density create distinct blockage and multipath patterns, causing signal observations to vary significantly between buildings and even within a single building.
Variable AP Visibility: The number of observable APs fluctuates across locations, leading to variable-length input sequences and irregular input spaces.
Signal Heterogeneity: Received signal strength (RSSI) and Channel State Information (CSI) exhibit high variability due to path loss, shadowing, and device-environment interactions, making fixed power signatures unreliable for robust localization.

Methodology: OmniLoc
The paper proposes OmniLoc, an environment-interactive foundation model designed to perform anchor-free localization directly on raw wireless measurements. Unlike previous approaches that rely heavily on synthetic data or single-environment training, OmniLoc is built on three core architectural designs:

Unified Input Tokenization (UIT):
To handle heterogeneous measurements (CSI magnitudes, RSSI, SINR, and AP presence indicators), the model converts these signals into a "sentence-like" representation.
- Input Fusion: CSI, RSSI, and SINR are concatenated per AP to form a unified feature vector.
- Soft Gating: AP presence flags modulate these vectors via element-wise multiplication, preserving gradient flow for absent APs rather than hard-zeroing them.
- Sequence Construction: A global AP signal vector is encoded and pooled to create a "GLOBAL" token, which is concatenated with the gated AP tokens. This allows the model to capture both local AP observations and the broader database context without assuming a fixed AP ordering.
Geometry-Aware Transformer:
The backbone utilizes a customized Transformer encoder with a Key-Padding Mask (KPM) to exclude padded (non-observed) AP slots from attention calculations.
- Multi-Probe AP Aggregation (MPAA): Instead of simple mean pooling, the model employs a multi-head attention pooling mechanism. This allows different attention heads to specialize: some may focus on dominant, nearby APs, while others aggregate complementary evidence from supporting APs. This addresses the issue where a single strong AP might be obscured by averaging in traditional pooling methods.
Geometry-Aware Location Estimation (GLO):
The output head implements a hierarchical localization process (Building $\to$ Floor $\to$ Coordinates) to enforce geometric consistency.
- The model first predicts building and floor indices.
- These discrete predictions are projected into continuous geometric embeddings.
- These embeddings are fused with the aggregated feature representation to condition the coordinate regression. This ensures that the predicted $(x, y)$ coordinates are spatially consistent with the identified building and floor structure.

Key Contributions

First Foundation Model for Raw Measurements: OmniLoc is presented as the first foundation-model-based approach built directly on wireless measurements for anchor-free indoor localization, explicitly addressing geometric diversity and variable AP counts.
Novel Architectural Components: The introduction of the Unified Input Tokenization, the Multi-Probe AP Aggregation mechanism, and the Geometry-Aware Location Estimation module.
Extensive Evaluation: The framework is validated on a large-scale in-house dataset (16 buildings, 180 corridors, 3,293 APs) and a public benchmark (WILD dataset).

Results

Superior Accuracy: In same-environment evaluations, OmniLoc significantly outperforms state-of-the-art baselines (CNN, BERT, and LWM). It achieves a median error of 0.553 m and a Root Mean Squared Error (RMSE) of 0.990 m, with particularly strong improvements in tail-error metrics (P90).
Component Transferability: Integrating OmniLoc's design components (tokenization, MPAA, GLO) into existing backbones (e.g., CNN+, BERT+) consistently improves their performance, demonstrating the generalizability of the proposed principles.
Cross-Environment Generalization: In "leave-one-floor-out" and "leave-one-building-out" scenarios, OmniLoc demonstrates robust generalization. When combined with parameter-efficient fine-tuning (PEFT) strategies like LoRA (Low-Rank Adaptation), the model achieves high accuracy with very limited labeled data (1%–10%), significantly outperforming models trained from scratch in low-data regimes.
Efficiency: Despite its performance gains, OmniLoc maintains a parameter count and computational complexity comparable to existing large models (e.g., LWM), with inference latency around 6.49 ms.

Significance and Claims
The authors claim that OmniLoc represents a shift toward unified, accurate, and extensible solutions for wireless indoor localization. By treating localization as a foundation model task, the work demonstrates that:

Robust anchor-free localization is achievable across diverse indoor geometries without relying on pre-surveyed infrastructure.
The specific challenges of wireless signal heterogeneity and variable AP visibility can be effectively managed through unified tokenization and geometry-aware attention mechanisms.
Pretraining on large, diverse datasets enables strong transferability to unseen environments, reducing the reliance on extensive environment-specific data collection.

The paper concludes that while current limitations exist (such as the exclusion of CSI phase information and the use of coarse geometric priors), the framework establishes a promising direction for achieving millimeter-level precision by potentially integrating fine-grained environmental context in future work.