CAEZ: CSI Acquisition at ETH Zurich

Table of Contents

• Reference Positioning Systems
• CAEZ-5G Datasets
• CAEZ-WIFI Datasets
• Download and Usage
• Related Real-World CSI Datasets
• Dataset License
• Acknowledgements
• References

Reference Positioning Systems

Most CAEZ datasets include ground-truth UE position labels from one of two reference positioning systems used at ETH Zurich. The summaries below explain the Position tracking entries in the following dataset descriptions.

WorldViz PPT

The external page WorldViz Precision Position Tracking (PPT) system provides external ground-truth UE position and orientation estimates. Infrared cameras placed around the measurement area track active IR markers mounted on the UE-carrying platform, yielding millimeter-level accuracy at approximately 240Hz. Tracking requires uninterrupted optical line-of-sight between markers and cameras and is sensitive to infrared interference (e.g., from sunlight or reflections).

Pose estimates are logged and timestamped for alignment with CSI captures. For the CAEZ-5G-INDOOR and CAEZ-5G-OUTDOOR datasets, position logs share the same PTP (Precision Time Protocol) time reference as the 5G NR testbed. For the CAEZ-WIFI-INDOOR-LSHAPE dataset, position timestamps use the same NTP (Network Time Protocol) source as the Wi-Fi sniffers. In post-processing, ground-truth positions are interpolated to the CSI sample timestamps.

Used in: CAEZ-5G-INDOOR, CAEZ-5G-OUTDOOR, CAEZ-WIFI-INDOOR-LSHAPE

Custom Robot SLAM

The custom robot platform "Chopper" carries the UE on a mount attached to its robot arm and navigates autonomously within a designated area. It runs simultaneous localization and mapping (SLAM) using LiDAR and wheel odometry and provides position estimates in its local map coordinate system with centimeter-level accuracy at approximately 10–20Hz. This onboard reference is used when external WorldViz tracking is impractical, e.g., in large indoor spaces with non-line-of-sight conditions or under infrared interference.

Chopper's position-logging script accounts for the UE mount relative to the robot arm and directly records the UE position in the local map coordinate system. The position logs use NTP timestamps. For CAEZ-WIFI-OFFICE-2WEEKS, the CSI and position logs use the same NTP time reference and are aligned by timestamp in post-processing. For CAEZ-5G-HALLWAY, our post-processing script provided external page here aligns the NTP/UTC position timestamps with the PTP/TAI time reference used by the 5G NR testbed by applying the 37-second TAI–UTC offset. In both cases, the ground-truth positions are interpolated to the CSI sample timestamps.

Used in: CAEZ-5G-HALLWAY, CAEZ-WIFI-OFFICE-2WEEKS

CAEZ-5G

We publish four real-world wideband multi-antenna multi-O-RU (Open RAN Radio Units) CSI datasets from the 5G NR uplink channel, specifically from the Physical Uplink Shared Channel (PUSCH): an indoor lab/office room dataset, an outdoor campus courtyard dataset, an indoor hallway dataset spanning one floor, and a device classification dataset with six commercial-off-the-shelf (COTS) user equipments (UEs). These datasets enable high-accuracy CSI-based sensing tasks including external page neural positioning, external page channel charting in real-world coordinates, and external page closed-set device classification. For further details, please refer to [1].

ETH Zurich 5G NR Testbed

The CAEZ-5G datasets were collected using a 5G NR testbed at ETH Zurich. The testbed builds upon the external page NVIDIA Aerial Testbed (ATB) software platform and is a full-stack software-defined 5G NR system with COTS UEs and four COTS O-RAN Radio Units (O-RUs), where one O-RU is used for 5G NR communication, while the other three operate as passive listeners. The system operates in the full Swiss private 5G band, i.e., 100MHz of the 5G NR N78 band centered at 3.45GHz. All components (except for the UEs) are connected via a fiber optical switch. A Supermicro NVIDIA MGX GH200 server runs the full-stack 5G NR system, comprising the NVIDIA Aerial L1, OAI L2, and OAI core network. A PTP (Precision Time Protocol) grand master clock with GNSS (Global Navigation Satellite System) time reference synchronizes the fiber-optical network.

O-RU Specifications

We use the following O-RU models and antenna configurations.

The transceiver chains of each O-RU are internally phase-synchronized. The O-RUs and 5G NR base station (gNB) are time-synchronized over the network by a PTP grand master clock using GNSS time reference.

Measurement Setup Details

Our standard O-RU–to–Cell ID association is shown below. For more information on the NVIDIA Aerial multi-cell setup, see external page this link.

In the CAEZ-5G CSI samples, the O-RU dimension uses the following index order.

Uplink Power Control

The gNB applies closed-loop uplink power control and sends transmit-power-control commands to the COTS UEs to keep the received PUSCH SNR close to a configured target, which is typically set to 28 dB for the CAEZ-5G measurements.

CAEZ-5G-INDOOR

The CAEZ-5G-INDOOR dataset contains CSI measurements from an indoor lab/office environment. The measurement area is a 3.5m x 3.5m square area between lab desks. The O-RUs were placed at the corners of the measurement area. Four WorldViz PPT (Precision Position Tracking) cameras were placed around the measurement area and two cameras were placed above the measurement area to provide ground-truth UE position tracking.

Measurement Details:

• Duration: 1h 47min
• Number of samples: 338,981
• UE type: Quectel RMU500EK
• Vehicle: iRobot Create 3 robot platform with Raspberry Pi and Quectel 5G modem
• Position Tracking: Yes, from WorldViz PPT
• PUSCH transmission: Every 20ms

The robot was controlled using random waypoint navigation. Four WorldViz PPT markers were mounted on the robot to enable tracking of position and rotation. One measurement operator was present in the lab/office during CSI collection and sometimes walked through the measurement area.

This dataset enables high-accuracy neural UE positioning, achieving 0.6cm mean absolute error on the test set. The simulation code is available in the external page neural positioning repository. For further details, please refer to [1].

CAEZ-5G-OUTDOOR

The CAEZ-5G-OUTDOOR dataset contains CSI measurements from an outdoor campus courtyard environment. The measurement area is a 10m x 10m square area in the ETH Zurich electrical engineering campus courtyard, surrounded by multiple buildings, trees, and other obstacles. The O-RUs were placed at the corners of the measurement area and six WorldViz PPT cameras were placed around the measurement area to provide ground-truth UE position tracking.

Measurement Details:

• Duration: 1h 38min
• Number of samples: 303,189
• UE type: Samsung Galaxy S23
• Vehicle: Custom robot platform with robot arm
• Position tracking: Yes, from WorldViz PPT
• PUSCH transmission: Every 20ms

A Samsung Galaxy S23 was mounted on Chopper's robot arm. Chopper was controlled manually. Four WorldViz PPT markers were mounted on the robot arm (which remained fixed) to enable tracking of the mounted UE's position and orientation. Two measurement operators were present near the measurement area during CSI collection.

This dataset enables high-accuracy neural UE positioning (5.7cm mean absolute error) and channel charting in real-world coordinates (73cm mean absolute error). The simulation code is available in the external page neural positioning repository and external page channel charting repository. For further details, please refer to [1].

CAEZ-5G-HALLWAY

The CAEZ-5G-HALLWAY dataset contains CSI measurements from the hallways on one floor of the ETH Zurich electrical engineering building. The measurement area spans approximately 50m × 30m, with the O-RUs distributed across the floor as shown in the floorplan below.

Measurement Details:

• Duration: 1h 35min
• Number of samples: 288,180
• UE type: iPhone 16e
• Vehicle: Custom robot platform with robot arm
• Position tracking: Yes, from custom robot's SLAM
• PUSCH transmission: Every 20ms

An Apple iPhone 16e was mounted on Chopper's robot arm. Chopper autonomously moved to randomly sampled waypoints throughout the hallways using LiDAR- and wheel-odometry-based SLAM. Its position-logging script accounts for the UE mount on the robot arm and records the ground-truth UE positions in the local map coordinate system. To obtain the O-RU positions in the same coordinate system, we drove Chopper to each O-RU, moved the robot arm to the O-RU position, and recorded the position reported by the same logging script. Two measurement operators were present within the measurement area during CSI collection.

The research codebase available in the external page neural-positioning-transformer repository demonstrates neural UE positioning and evaluates its generalization to unseen spatial regions and UE trajectories. For further details, please refer to [4].

CAEZ-5G-DEV-CLASS

The CAEZ-5G-DEV-CLASS dataset contains CSI measurements from six different COTS UEs for device classification tasks. The measurement setup is similar to that of CAEZ-5G-INDOOR, but the measurements were carried out on different days and no WorldViz PPT cameras were used. The measurement area is a joint lab/office space of about 4m x 4m between the lab desks. The O-RUs were placed at the corners of the measurement area.

Measurement Details:

• Duration: 6 x (2min + 30s) per UE
• Number of samples: 83,619 (1st day) + 21,805 (2nd day)
• UE types: Six COTS UEs (see figure above)
• Vehicle: Rotation table + human operator
• Position tracking: No
• PUSCH transmission: Every 10ms

The dataset consists of six consecutive measurements, each taken separately using one of the six UEs shown in the figure above. The measurement protocol for each UE comprises the following four steps: (i) rotation on a predefined fixed location with a rotation table, (ii) random human walk, (iii) additional rotations on the same location, and (iv) another random human walk on the next day.

On the first measurement day, each measurement started when the UE was mounted on the rotation table and taken out of flight mode. As soon as the UE connected to the 5G network, it was left slowly rotating for 30s. Then, the operator removed the UE from the phone mount and carried it randomly through the lab space for 60s. Finally, the UE was put back into the phone mount on the rotation table and left spinning for another 30s.

On the second measurement day (the following day), each measurement consisted only of 30s random human walk through the lab space. This next-day dataset is used for testing. Note that one of the O-RU's power supplies stopped working overnight, and the lab environment has slightly changed (e.g., chairs and equipment were moved). Therefore, this next-day evaluation dataset is not only recorded from different UE locations, but also in a slightly modified environment.

This dataset enables highly accurate device classification with location-independent radio frequency fingerprinting (RFFI) features, achieving 99% accuracy (same day) and 95% accuracy (next day). The simulation code is available in the external page device classification repository. For further details, please refer to [1].

Citation Key

If you use our CAEZ-5G datasets and/or our simulation code (or parts of it), then you must cite this reference:

CAEZ-WIFI

We publish two real-world IEEE 802.11a Wi-Fi CSI datasets recorded using multiple distributed, multi-antenna, software-defined Wi-Fi sniffers. The dataset enables high-accuracy CSI-based external page neural positioning, external page channel charting (CAEZ-5G-specific codebase), and analyses of spatial and temporal generalization. Further details can be found in [2] - [4].

ETH Zurich Custom Wi-Fi Sniffer Testbed

The CAEZ-WIFI datasets were collected using a software-defined Wi-Fi testbed with multiple Wi-Fi sniffers acting as receivers for CSI measurements. Each sniffer is equipped with a four-antenna software-defined radio (SDR) and runs a custom PHY-layer software stack capable of decoding Wi-Fi traffic. The Wi-Fi traffic itself is generated by UEs operating in an active 5GHz Wi-Fi network.

Using MAC addresses extracted from received Wi-Fi frames, each sniffer identifies whether a given CSI estimate corresponds to the channel between a targeted UE and its receive antenna array. The resulting CSI estimates are stored locally on the sniffer’s host PC and later processed to form a combined multi-sniffer CSI dataset. To enable temporal alignment during post-processing, all sniffer host PCs synchronize their clocks via a common Network Time Protocol (NTP) server.

Transmit Power and Receiver Gain

All SDR sniffers use the same fixed receiver gain. Because the sniffers are passive, they neither control nor record the UE transmit power; each COTS Wi-Fi UE selects its transmit power autonomously.

CAEZ-WIFI-INDOOR-LSHAPE

The CAEZ-WIFI-INDOOR-LSHAPE dataset contains CSI measurements from an indoor meeting room environment. The measurement area is L-shaped with outer side-lengths of 4m x 5m. The Wi-Fi sniffers were placed outside the measurement area. To create partial NLoS conditions for at least one sniffer at a time with respect to any position inside the measurement area, we placed a wall of RF absorbers on the inner side of the L-shape. Six WorldViz PPT cameras were placed around the measurement area to provide ground-truth UE position tracking.

Measurement Details:

• Duration: 4h 15min
• Number of samples: 33,411 (sniffer 1), 43,518 (sniffer 2), 40,339 (sniffer 3), 43,914 (sniffer 4)
• UE type: Wi-Fi USB Adapter Alfa AWUS036ACHM
• Wi-Fi bandwidth: 20MHz 
• Vehicle: iRobot Create 3 robot platform with Raspberry Pi and Wi-Fi adapter
• Position tracking: Yes, from WorldViz PPT

The robot was controlled using random waypoint navigation. Four WorldViz PPT markers were mounted on the robot to enable tracking of the UE’s position. Two measurement operators were present in the room during CSI collection and sometimes walked through the measurement area.

This dataset enables high-accuracy neural UE positioning, achieving 6.1cm mean absolute error on a randomly sub-sampled test set. The simulation code is available in the external page neural positioning repository. Further details can be found in [3], where the CAEZ-WIFI-INDOOR-LSHAPE dataset is referred to as “Wi-Fi Large Meeting Room” dataset. The implementation available in the external page neural positioning repository improves upon the supervised baseline presented in [3].

CAEZ-WIFI-OFFICE-2WEEKS

The CAEZ-WIFI-OFFICE-2WEEKS dataset contains CSI measurements from an indoor two-room office environment. The dataset comprises two measurement campaigns recorded one week apart (Week 1 and Week 2). The approximately 10m × 8m measurement area spans two rooms and a small part of the hallway in the upper-right corner of the floorplan below. The Wi-Fi sniffers were distributed across both rooms.

Measurement Details:

• Duration: 1h 51min and 1h 36min (Week 1 and Week 2)
• Number of samples: 115,716 and 98,170 (Week 1 and Week 2)
• UE type: iPhone 15 Pro
• Wi-Fi bandwidth: 40MHz
• Vehicle: Custom robot platform with robot arm
• Position tracking: Yes, from custom robot's SLAM

An Apple iPhone 15 Pro was mounted on Chopper's robot arm. Chopper autonomously moved to randomly sampled waypoints within the two office rooms using LiDAR- and wheel-odometry-based SLAM. Its position-logging script accounts for the UE mount on the robot arm and records the ground-truth UE positions in the local map coordinate system. Two measurement operators were present within the measurement area during CSI collection.

The research codebase available in the external page neural-positioning-transformer repository analyzes the spatial and temporal robustness of neural UE positioning, including generalization to unseen spatial regions and UE trajectories and consistency across measurement campaigns recorded one week apart. Further details are presented in [4].

Citation Key

If you use our CAEZ-WIFI datasets and/or our simulation code (or parts of it), then you must cite this reference: