Neural Positioning Without External Reference

TL;DR

This work tackles the challenge of CSI-based UE positioning without relying on external high-precision ground-truth labels. It introduces a neural positioning framework that learns a mapping from CSI features to positions using only off-device CSI data and relative displacement commands from a COTS robot, augmented by a novel triangle displacement loss and a small anchor loss. Real-world validation across Wi‑Fi and 5G NR testbeds in LoS and NLoS environments shows centimeter-level accuracy approaching that of externally supervised methods, with robust performance in diverse indoor scenarios. The approach enables scalable training and retraining of neural positioning functions over large areas using inexpensive hardware, offering practical impact for indoor localization and network management tasks.

Abstract

Channel state information (CSI)-based user equipment (UE) positioning with neural networks -- referred to as neural positioning -- is a promising approach for accurate off-device UE localization. Most existing methods train their neural networks with ground-truth position labels obtained from external reference positioning systems, which requires costly hardware and renders label acquisition difficult in large areas. In this work, we propose a novel neural positioning pipeline that avoids the need for any external reference positioning system. Our approach trains the positioning network only using CSI acquired off-device and relative displacement commands executed on commercial off-the-shelf (COTS) robot platforms, such as robotic vacuum cleaners -- such an approach enables inexpensive training of accurate neural positioning functions over large areas. We evaluate our method in three real-world scenarios, ranging from small line-of-sight (LoS) areas to larger non-line-of-sight (NLoS) environments, using CSI measurements acquired in IEEE 802.11 Wi-Fi and 5G New Radio (NR) systems. Our experiments demonstrate that the proposed neural positioning pipeline achieves UE localization accuracies close to state-of-the-art methods that require externally acquired high-precision ground-truth position labels for training.

Figures (7)

• Figure 1: Illustration of the triangle displacement loss formed from a geometric triangle. Each vertex is associated with a CSI feature; arrows indicate the displacement vectors between endpoints used for training. The gray curve represents the true movement in space; the smaller green arrows indicate noisy displacement measurements.
• Figure 2: Overview of the UE positioning pipeline: (a) Measurement setup: A UE moves from time stamp $n$ to $n+1$ and records the displacement measurement $\tilde{\bideltad\xspace}_n$ (in our case, corresponding to the executed displacement command). Distributed receive APs $b,b' \in \{1,\dots,B\}$ collect the CSI at each time stamp, which is preprocessed on a CPU and then used for model training on a GPU. (b) Neural positioning function learning: CSI is transformed into CSI features and relative displacement data is recorded. The CSI features and displacement measurements are grouped into sets of triangle vertices to form the collection of CSI features $\mathcal{F}\xspace$ and large relative displacements $\mathcal{D}\xspace$. Finally, the neural positioning function parameters are trained using the triangle displacement loss.
• Figure 3: Wi-Fi and 5G testbeds in three scenarios: (a) and (d) show the small meeting room Wi-Fi scenario, (b) and (e) the large meeting room Wi-Fi scenario with partial NLoS, and (c) and (f) show the 5G office scenario with human activity. The top row shows photos of each scenario; the bottom row shows the respective floorplans. The UE platform's dock position is denoted by a cross with an arrow indicating the initial orientation.
• Figure 4: Test set ground-truth (GT) positions (top row) and estimator outputs (bottom row) across the three measured scenarios. (a) and (d) correspond to the Wi-Fi small meeting room scenario, (b) and (e) to the the Wi-Fi large meeting room scenario, and (c) and (f) to the 5G office scenario with human activity. In all plots, the anchor position and its orientation are indicated by a black cross with an arrow. The arrow denotes the initial orientation of the UE platform.
• Figure 5: CDF of positioning errors for different methods, including Baseline 1 using external ground-truth position labels, Baseline 2, Baseline 3, and the proposed approach. For the Wi-Fi datasets, we used disp.-dep. features. Results are shown for (a) the Wi-Fi small meeting room dataset, (b) the Wi-Fi large meeting room dataset, and (c) the 5G office dataset.

Theorems & Definitions (12)

• Remark 1
• Remark 2
• Remark 3
• Remark 4
• Remark 5
• Remark 6
• Remark 7
• Remark 8
• Remark 9
• Remark 10