Download the datasets on IEEE Dataport
We provide 2 datasets: iV2V (industrial Vehicle-to-Vehicle) and iV2I+ (industrial Vehicular-to-Infrastructure + sensor). Both datasets provide information from several sources in different granularity. For ease of use, parquet files containing direct translations of the raw data are provided in respective sources folders.
In the following, an overview of the data is provided. For a detailed description of the measurement campaigns, please refer to the paper.
We strongly recommend to work on Python with the following libraries:
Furthermore, we suggest some additional libraries to process and analyze the data, such as:
- numpy for common mathematical tools
- jupyter to run the interactive examples
- matplotlib for plotting
- scikit-learn for ML analysis
- bagpy for ROS bag files
The following data is provided both for iV2V and iV2I+:
- A combined dataframe with selected features for direct usage, in parquet format.
- A compressed file
*-sources.zip, with the different data sources transformed to individual parquet files. - Several compressed files with the unedited raw data sources in their original file formats, as
*-sources/raw/*.zip.- The bag files
iV2Ip-sources/raw/sensors/*.bagconstitute an exception, since they are already compressed and a single zip file with all of them would have a size of >250 GB.
- The bag files
- Metadata from all features in the combined dataframes, as
*_info.csv
Head to the iV2V.parquet file, and load it in pandas to inspect the columns. Some general information on each column can be found in iV2V_info.csv
The dataframe contains information from the sidelink as extracted from RUDE and Crude, including:
- time of arrival and signal strength measurements
- the location of AGV1 within the test track
- labels to identify source and destination AGVs.
Sidelink and location data have been matched on the epoch timestamps with a small error tolerance. The wall scenarios "A" and "B" are also provided as labels as noted down during the measurements.
Sources:
- RUDE & CRUDE for sidelink communication
- Localization data provided from the AGV's sensors.
The packets were transmitted roughly every 20 ms. This reference value, together with the provided timestamps, can serve as a basis to estimate packet error rate.
The sidelink data extracted from the incoming messages for any given AGV are provided as separate parquet files among the iV2V sources (e.g., sidelinkX_df.parquet with X the id of the AGV). For a detailed insight of the sidelink signal parameters, check the dataset publication or RUDE's documentation.
The AGV1 localization data is provided in .txt format as tab-separated-values containing the following fields:
- Sidelink Epoch Time [sec] - As unix epoch timestamps
- X-coordinate [m]
- Y-coordinate [m]
The update period for the localization data is approximately 50 ms.
Head to the iV2Ip.parquet file, and load it in pandas to inspect the columns. Some general information on each column can be found in iV2Ip_info.csv
The dataframe contains:
- radio data (RSRP, RSRQ, SINR, RSSI)
- basic sensor data (x and y location, speed)
- throughput and delay measurements
- additional calculated features like the Line of Sight (LoS) or the cell load
Each source has a particular update period, so they all have been resampled to 1 second while merging. As a result, information loss can be expected.
For detailed information about the columns and information in higher resolution, read below.
Sources:
Except for the Sensor Data, all measurement software is Open Source and free of use.
All information captured by MobileInsight from available LTE channels. The available information also depends on the modem of the measurement device.
The "rs_intra_all" file contains RSRP and RSRQ information in a resolution of 40 ms for the measurement campaign.
Alternatively, RSRP and RSRQ values were logged together with SNR and RSSI directly into the AGV-mounted mini PC every 200 ms as cell_info_yyyymmdd-HHMMSS.log. These logs are merged and available in the sources as "cell_df.parquet"
TCP Dump is a packet analyzer that allows tracking transmitted packets and their properties (e.g. payload, size of the packet).
During the measurement campaign, TCP Dump was run on both server and the mini-PC which was attached to the AGV. This allows reconstructing e.g. the packet delay (
The parquet files contain already information from both server and client side and each packet is listed with respective fields if it is either an ICMP, TCP or UDP packet. For the UDP packets also the delay between sending and receiving entity is included, as well as if a packet has arrived. The parquet files are split between Uplink and Downlink and also between the different days.
Iperf is a speed test application that enables measuring the bandwidth and jitter of a UDP or TCP connection.
In the measurement campaign, Iperf was run on both a mini PC and server to receive throughput measurements with a granularity of 1s. For experiments that require high accuracy, it is recommended to use the TCP dump based information since the information was collected one time per second, but not at the beginning of each second.
Collected from the console command ping.
Sensor data was stored using ROS (Robot Operating System) as .bag files, which can be read e.g. with the bagpy library in Python. An excerpt of the available information can be seen below.
| Topic | ROS message type | Update period | Description |
|---|---|---|---|
| Map static elevation | nav_msgs/OccupancyGrid | - | Single precomputed map of the whole area |
| Far map obstacles | nav_msgs/OccupancyGrid | 50 ms | 400 |
| Near map obstacles | nav_msgs/OccupancyGrid | 20 ms | 36 |
| Odometry | nav_msgs/Odometry | 10 ms | Sensor-fused position, orientation and speed of the AGV |
| Inertial Measurement Unit | sensor_msgs/Imu | 10 ms | Conventional IMU data |
| LIDAR | sensor_msgs/PointCloud2 | 100 ms | 3D point cloud with obstacles |
The static map and the odometry data are provided in the sources as "static_map" and "ros_df", respectively, while the remaining information can be extracted from the original bag messages.
It is important to note that "Odometry" does not refer to pure wheel odometry but sensor-fused dead reckoning using other sensor sources, e.g., IMU. Within "ros_df", the odometry data has been downsampled to 40 ms (considering the AGV's low speed and update rate of the communication data) and extended with:
- distance_to_bs: The distance to the base station, whose position was fixed to (9,9).
- obstacles_sum: A summation of the obstacles for Line-of-Sight (LoS) estimation. For this, the elevation values above a small threshold lying within a Fresnel ellipse between AGV and base station were added together. The threshold here serves to neglect the grid values that account for ground.
- line_of_sight: A boolean estimate of LoS, obtained as the condition
obstacles_sum < 1000(The threshold value "1000" has been selected a posteriori).
These added fields are computed within odom_parser.py
For a complete code example to explore the datasets, check the Jupyter notebooks iV2Ip-visualize.ipynb and iV2V-visualize.ipynb.
AI4Mobile is a research project funded by the Federal Ministry for Education and Research (BMBF), from the announcement Artificial Intelligence in Communication Networks within the scope of the High-Tech Strategy of the German Federal Government.
The scope of the project is the study of AI-aided wireless systems for mobility in industry and traffic. More information at ai4mobile.org.
If you use the dataset, please cite it as:
@article{hernangomez2024aienabled,
title = {Toward an {{AI-Enabled Connected Industry}}: {{AGV Communication}} and {{Sensor Measurement Datasets}}},
shorttitle = {Toward an {{AI-Enabled Connected Industry}}},
author = {Hernang{\'o}mez, Rodrigo and Palaios, Alexandros and Watermann, Cara and Sch{\"a}ufele, Daniel and Geuer, Philipp and Ismayilov, Rafail and Parvini, Mohammad and Krause, Anton and Kasparick, Martin and Neugebauer, Thomas and {Ramos-Cantor}, Oscar D. and Tchouankem, Hugues and Calvo, Jose Leon and Chen, Bo and Fettweis, Gerhard and Sta{\'n}czak, S{\l}awomir},
year = {2024},
month = apr,
journal = {IEEE Communications Magazine},
volume = {62},
number = {4},
eprint = {2301.03364},
primaryclass = {cs},
pages = {90--95},
issn = {1558-1896},
doi = {10.1109/MCOM.001.2300494},
copyright = {All rights reserved},
keywords = {Artificial intelligence,Computer Science - Artificial Intelligence,Computer Science - Machine Learning,Computer Science - Networking and Internet Architecture,Fingerprint recognition,Line-of-sight propagation,Quality of service,Robot sensing systems,Service robots,Vehicular ad hoc networks,Wireless communication,Wireless sensor networks}
}