Csi-based Positioning and Device Classification Achieves 95% Accuracy with Real-World 5G NR Data

Channel-state information (CSI) holds immense promise for future cellular networks, yet a significant barrier to progress has been the lack of real-world data from functioning 5G systems. Reinhard Wiesmayr, Frederik Zumegen, and Sueda Taner, all from ETH Zurich, alongside Chris Dick and Christoph Studer from NVIDIA, address this challenge by releasing three comprehensive CSI datasets captured from a live 5G new radio system.

The team deployed a software-defined 5G testbed at ETH Zurich and recorded data in both indoor and outdoor environments, alongside a dataset designed for identifying different devices. This achievement unlocks new possibilities for research into neural user positioning, accurate channel charting in real-world coordinates, and robust device classification, with the team demonstrating impressive results including positioning accuracy down to 0. 6cm and device classification exceeding 95% accuracy. These publicly available datasets and accompanying tools represent a crucial step towards realising the full potential of CSI-based sensing in next-generation wireless networks. Summary of the CAEZ Datasets and Research. This document details the creation and characteristics of the CAEZ (Channel Awareness for Efficient Zero-effort) datasets and associated research, focusing on three main tasks: neural UE positioning, channel charting, and device classification. The goal is to provide publicly available datasets and tools to advance research in these areas, particularly leveraging machine learning techniques for wireless systems. Key contributions include publicly available datasets for indoor and outdoor environments, collected using a distributed massive MIMO system and various robots for mobility, encompassing Channel State Information (CSI), robot trajectories, and ground truth positioning data. The research achieves centimeter-level accuracy (0. 7cm outdoors) in neural UE positioning using neural networks to estimate the location of a User Equipment (UE), creates a map of the wireless environment with channel charting achieving a mean absolute error of 73cm, and identifies devices based on their radio frequency fingerprints with accuracies of 99% (same-day) and 95% (next-day). Datasets and simulation code are publicly available at [https://caez. ch]. Detailed Breakdown of Each Task: Neural UE Positioning utilizes neural networks to learn the relationship between CSI and UE location, achieving centimeter-level accuracy in both indoor and outdoor environments using distributed massive MIMO and machine learning. Channel Charting learns a map of the wireless environment based on channel characteristics, allowing for prediction of channel conditions at different locations and achieving a mean absolute error of 73cm in predicting channel characteristics, utilizing triplet-based learning and distributed massive MIMO. Device Classification identifies devices based on their unique radio frequency fingerprints, achieving high accuracy (99% same-day, 95% next-day) using radio frequency fingerprinting and machine learning. Technical Details and Tools include a distributed massive MIMO system, robots (iRobot Create), and standard-compliant 5G NR equipment, alongside simulation code and tools for data processing and machine learning. Future work will focus on expanding datasets to include more diverse scenarios (mixed LOS/NLOS, larger areas, 3D trajectories) and validating model-based and NN-based receivers in real-world deployments. In essence, the CAEZ project aims to provide a comprehensive platform for researchers to explore and develop machine learning-based solutions for wireless localization, channel modeling, and device identification. Real 5G Channel Data for Sensing Scientists have published three real-world datasets of channel-state information (CSI) collected from a 5G New Radio (NR) system, addressing a significant gap in resources for developing and validating advanced sensing algorithms. The research team deployed a software-defined 5G NR testbed, leveraging commercial-off-the-shelf hardware, to capture uplink CSI from real network traffic, providing a uniquely representative dataset for future 6G wireless systems. Experiments demonstrate the utility of these datasets for three distinct CSI-based sensing tasks. Neural UE positioning, utilizing the collected CSI, attains a mean absolute error of 0. 7cm outdoors, representing a substantial improvement in positioning precision. Channel charting, a method for creating real-world maps from channel information, achieves a mean absolute error of 73cm outdoors, demonstrating the potential for detailed environmental mapping. Furthermore, the team achieved 99% accuracy in device classification using data collected on the same day, and 95% accuracy when classifying devices on the following day, highlighting the robustness of the system. The datasets, including ground-truth UE position labels, CSI features, and simulation code, are publicly available, enabling researchers to develop and validate new algorithms without the limitations of synthetic data or custom testbeds. This work delivers a valuable resource for the wireless research community, facilitating advancements in areas such as off-device neural positioning, device classification, and real-world channel mapping, all crucial components of next-generation wireless systems. Real 5G Channel Data for Positioning and Charting This work presents the first publicly available datasets of real-world 5G New Radio (NR) channel state information, accompanied by positioning data, and enables advancements in several key areas of wireless communication. Researchers have collected and released three datasets, one indoor, one outdoor, and one focused on device classification, using a software-defined 5G NR testbed. These datasets facilitate the development and validation of algorithms for neural user equipment positioning, channel charting in real-world coordinates, and accurate device classification. Experimental results demonstrate high performance across all three tasks. Neural network-based positioning achieves mean absolute errors of just 0. 7cm outdoors, while channel charting in real-world coordinates achieves 73cm accuracy. Device classification, using location-independent features, reaches 99% accuracy on the same day and remains high at 95% accuracy when tested on a subsequent day. The availability of these datasets and accompanying simulation code allows other researchers to build upon these findings and explore new possibilities in 5G and future wireless systems. The authors acknowledge that the datasets currently focus on specific scenarios and plan to expand their collection to include more diverse environments, such as mixed line-of-sight and non-line-of-sight conditions, larger measurement areas, and three-dimensional user trajectories. Future work will also investigate the validation of model-based and neural network-based receivers in real-world conditions, further enhancing the impact of this research. 👉 More information 🗞 CSI-Based User Positioning, Channel Charting, and Device Classification with an NVIDIA 5G Testbed 🧠 ArXiv: https://arxiv.org/abs/2512.10809 Tags: