FZI-AURA

A multimodal autonomous-driving dataset featuring the largest LiDAR sensor suite of any public autonomous-driving dataset.

Download | Python SDK | Data format | License

2,473 scenes | 13.70 hours | 8 cameras | up to 12 LiDARs | 4.11 million 3D-box annotations | 30.01 billion segmented LiDAR points

Public preview release: 1,081 out of the 2,473 FZI-AURA scenes are currently available. The remaining scenes will be added soon. Pin a Hugging Face commit revision for reproducible experiments.

Overview

FZI-AURA is a multimodal autonomous-driving dataset recorded by FZI Research Center for Information Technology across southern Germany. The data was recorded using CoCar NextGen, FZI's research platform for automated and connected driving. The dataset contains 2,473 scenes captured with eight surround-view cameras, up to twelve LiDAR sensors, and up to three radar sensors. It also includes 3D bounding boxes, semantic LiDAR labels, ego and vehicle data, as well as road and weather information. To the best of our knowledge, FZI-AURA provides the densest point-cloud coverage and the largest number of semantically labeled LiDAR points of any publicly available autonomous-driving dataset. Most scenes were recorded in urban traffic, but the dataset also covers overland roads and motorways, with different weather and lighting conditions. It can be used for work on perception, sensor fusion, tracking, motion compensation, and end-to-end driving.

Why FZI-AURA?

• The most segmented LiDAR points of any public non-synthetic autonomous-driving dataset: FZI-AURA contains more than 30 billion ( $3\times {10}^{10}3\backslash times10^\{10\}$ ) human-annotated semantic LiDAR points across over 450,000 annotated LiDAR scans.
• The largest public autonomous-driving LiDAR suite: FZI-AURA provides up to 12 synchronized LiDAR sensors, comprising four Ouster OS1-64, two Ouster OS2-128, and six Aeva Aeries II FMCW sensors. This configuration combines rotating and FMCW LiDAR technologies in a single public dataset.
• Large-scale 3D annotations: The dataset includes more than 4.1 million 3D-box annotations across nearly 100,000 annotated keyframes.
• Rich multimodal sensing: Eight synchronized surround-view cameras, up to three radar sensors, IMU/GNNS and vehicle signals complement the recorded data.
• Context beyond the sensors: Samples are associated with OpenStreetMap-derived road information and weather metadata from OpenWeather.
• Raw and motion-compensated point clouds: LiDAR data is available both as raw sensor clouds and as motion-compensated clouds, enabling research on sensor motion, temporal alignment, and downstream perception.
• Multiple perception and driving tasks: FZI-AURA supports research in 3D object detection, LiDAR semantic segmentation, sensor fusion, tracking, motion compensation, multimodal scene understanding, and end-to-end driving.
• Framework-neutral access: The files can be inspected directly or accessed through the companion FZI-AURA Python SDK with minimal dependencies.
• Open for research and commercial use: FZI-AURA is released under CC BY-SA 4.0, allowing the dataset to be used, shared, and adapted for both academic and commercial purposes, provided appropriate attribution is given and adaptations are shared under the same license. Unlike many widely used autonomous-driving datasets, FZI-AURA does not impose a non-commercial-use restriction.

Sensor Setup

The FZI-AURA dataset was recorded using FZI's research vehicle CoCar NextGen. It includes 12 lidar sensors: 6 ouster 360° sensors (4x OS1, 2x OS2), and 6 AEVA Aeries II FMCW LiDAR Sensors, providing 360° coverage with both the AEVA and Ouster sensor sets. The setup additionally includes 360° camera coverage using 8 Cameras with global shutter. 3 ARS 548 RDI Radar Sensors are also included in some scenes. While some scenes do not contain the full sensor setup, most scenes include at minimum 6 Ouster LiDAR sensors. The below image provides an overview of sensor positions and FOVs, with more precise, scene-specific values provided via the SDK.

Dataset Statistics

Quantity	Current value
Scenes	2,473
Split sizes	1,979 train / 247 validation / 247 test
Total scene duration	13.70 h (13 h 42 min)
Samples	493,754 at 10 Hz
Cameras per scene	8
LiDAR sensors per scene	4–12
Radar sensors per scene	0–3
LiDAR frames indexed	4,239,982
Radar frames indexed	2,707,107
Semantic LiDAR labels	30,008,440,687 points in 450,762 sensor-frame arrays
3D-box annotated keyframes	98,920
3D-box rows	4,106,333

The shared class metadata for both semantic labels and 3D boxes is stored in each scene's labels/semantic/classes.json. This file is present independently of whether the scene contains semantic label payloads. The SDK preserves all exported semantic IDs; FZI-AURA does not define a universal ignored-class set. Training pipelines should explicitly decide whether to retain, ignore, or remap classes such as unlabeled and noise.

Class Distribution

Download

We suggest to use the FZI-AURA SDK downloader. The SDK will handle the downloading and extraction for you and allows you to select only the data you need. After accepting the dataset terms and authenticating with hf auth login:

git clone https://github.com/fzi-forschungszentrum-informatik/fzi-aura-sdk.git
cd fzi-aura-sdk
python -m pip install ".[download]"

# Inspect the selected blocks and compressed size without downloading.
fzi-aura-download /data/fzi-aura --dry-run

# Download and extract camera and raw-LiDAR keyframes from the train split.
fzi-aura-download /data/fzi-aura \
  --splits train \
  --layers camera_keyframes,lidar_motion_compensated_keyframes \
  --jobs 8 --verify

The release is packaged in synchronized blocks of up to 20 scenes. base_keyframes is added automatically because it contains the indexes, calibration, ego data, and annotations required by the SDK. Camera, raw LiDAR, motion-compensated LiDAR, and radar are available at keyframes. Camera, raw LiDAR, and radar also provide non-keyframe layers for dense temporal context.

LiDAR sizes reported by --dry-run are compressed. Extracted PCD files and the uncompressed TAR files used for mounting require approximately 40–60% more space for those layers. During normal extraction, each archive is removed as soon as it has been extracted successfully.

Re-running the downloader discovers newly published complete blocks. See the SDK's download documentation for scene selection, verification, archive retention, and mounting.

FZI-AURA SDK

The companion Python SDK provides lazy scene/frame access, camera and point-cloud loading, calibration transforms, ego poses, 3D boxes, semantic labels, and projection helpers. See the FZI-AURA SDK repository for installation instructions, example notebooks and helper functions.

from fzi_aura import FZIAURADataset

dataset = FZIAURADataset("/path/to/fzi-aura", split="train")
scene = dataset[0]

# Requirements respect frame references and the selected release layers.
sensor_frames = scene.frames(
    sample_filter="sensor_available",
    require_cameras=["front_medium"],
    require_lidars=["top_left"],
)
if len(sensor_frames):
    frame = sensor_frames[0]
    image = frame.load_camera("front_medium")
    cloud = frame.load_lidar("top_left", stage="motion_compensated")

box_frames = scene.frames(sample_filter="boxes_3d")
if len(box_frames):
    boxes = box_frames[0].load_boxes(frame="base_link")

semantic_frames = scene.frames(
    sample_filter="semantic_lidar",
    require_lidars=["top_left"],
    require_semantics_for=["top_left"],
)
if len(semantic_frames):
    cloud, labels = semantic_frames[0].load_lidar_semantic_pair(
        "top_left", stage="motion_compensated"
    )

Data organization

After extraction or mounting, the downloaded layers form the following directory:

fzi-aura/
  dataset.json
  available_data.json
  README.md
  LICENSE
  THIRD_PARTY_NOTICES.MD
  splits/
    v1.0/
      train.txt
      val.txt
      test.txt
  scenes/
    <scene_directory>/
      scene.json
      samples.jsonl
      samples.parquet
      calibration.json
      camera/<camera_id>/*.jpg
      lidar/raw/<sensor_id>/*.pcd
      lidar/motion_compensated/<sensor_id>/*.pcd
      radar/<sensor_id>/*.pcd                             # when available
      labels/
        boxes_3d.jsonl                                    # all 3D BBoxes in base_link
        boxes_3d_sensor_frame/<sensor_id>/boxes_3d.jsonl  # in sensor frame and trimmed to boxes with at least one point
        semantic/<sensor_id>/*.label                      # when available
        semantic/classes.json
      ego/poses.parquet
      ego/vehicle_signals.parquet

dataset.json is the dataset-level scene index. scene.json describes the scene's sensors, modalities, counts, and relative paths. samples.jsonl and samples.parquet provide frame-level references; paths in those indexes are relative to the scene directory. More information is available in the SDK documentation.

Splits

Version 1.0 provides scene-level train, validation, and test splits under splits/v1.0/: 1,979 train scenes, 247 validation scenes, and 247 test scenes. The split is stratified by labels, weather, and time, with no geographic overlap between the three partitions. Use these files rather than deriving a split from scene order.

Sampling and annotation conventions

• Scenes contain 10 Hz samples. 3D boxes and semantic LiDAR labels are provided at 2 Hz annotation keyframes.

• labels/boxes_3d.jsonl contains canonical boxes in base_link. A box with boxes_3d_count: 0 is a valid annotated negative frame, not a missing annotation.

• Per-sensor box files are expressed in the corresponding sensor frame and are filtered to boxes containing points.

• Box object_id values are stable within a scene, but are not globally unique across scenes.

• Semantic label arrays are point-index aligned with both the raw and motion-compensated clouds and use uint32 values:

  semantic_id = value & 0xFFFF
  instance_id = value >> 16
  

• Raw and motion-compensated point clouds use sensor coordinate frames from calibration.json.

• Raw LiDAR clouds retain the original sweep geometry, and their timestamp marks the start of the sweep. Motion-compensated clouds express all points at the end-of-sweep reference time, and their timestamp marks the end of the sweep. samples.jsonl records the association between samples and sensor files, and the SDK resolves these associations directly.

• Camera projection matrices and sensor-to-base_link transforms are stored in calibration.json. Ego poses describe the transform from base_link to the exported odometry frame.

• base_link follows the standard ROS convention: X forward, Y left, and Z up. Positive box yaw is counter-clockwise when viewed from +Z. Boxes use a center origin, full size_lwh dimensions along their local X/Y/Z axes, and a full rotation_xyzw quaternion.

Always use the class metadata shipped with the exact revision being evaluated.

Context metadata

FZI-AURA augments the recorded sensor data with contextual road and historical weather information associated with the recording time and location. Coverage is scene-dependent, and individual values can be missing.

details

Metadata group	Source	Representative fields
Road context	OpenStreetMap	osm_highway_tag, osm_highway_distance, road_type
Weather context	OpenWeather One Call API 3.0	weather condition, temperature, humidity, visibility, cloud cover, wind, rain, and snow

Road context

Recorded positions are matched to road geometries derived from OpenStreetMap. The metadata contains:

• the matched OpenStreetMap highway tag
• the distance to the matched road geometry
• a derived coarse road-context classification

The available road_type values are:

• motorway
• overland
• urban
• unknown

The classification uses the matched road category and, where available, additional OpenStreetMap traffic-context and administrative-boundary information.

Weather context

Historical weather data was retrieved for the recording time and location using the OpenWeather One Call API 3.0 Time Machine endpoint. Requests were grouped into 10-minute time bins and associated with the corresponding dataset samples.

Available fields can include:

• weather condition
• temperature and feels-like temperature
• pressure, humidity, and dew point
• cloud cover and visibility
• wind speed, gust, and direction
• sunrise and sunset
• rain and snow

Exact field names and availability are recorded in the scene metadata shipped with the selected dataset revision. The SDK documents the supported access patterns; version 1.0 does not publish a separate JSON Schema.

Metadata Attribution

Contains information from OpenStreetMap, made available under the Open Data Commons Open Database License 1.0.

Attribution: OpenStreetMap contributors.

Weather data provided by OpenWeather.

Detailed source, transformation, attribution, and licensing information is provided in THIRD_PARTY_NOTICES.MD.

Intended use

FZI-AURA is intended for research and education in autonomous-driving perception, 3D object detection, LiDAR semantic segmentation, sensor fusion, calibration, tracking, motion compensation, end-to-end driving, and multimodal scene understanding. The version 1.0 release is suitable for data inspection, model training, and evaluation. Pin the dataset revision and split version when reporting results.

Release notes

2026-09-09 - Version 1.0.0 official preview: Publication of the first 1081 scenes.

License and third-party data

Unless stated otherwise, FZI-AURA is released under the Creative Commons Attribution-ShareAlike 4.0 International license.

The contextual metadata contains information derived from third-party data sources:

• OpenStreetMap-derived metadata is subject to the Open Data Commons Open Database License 1.0 and the applicable attribution requirements.
• OpenWeather-derived metadata is subject to the applicable OpenWeather content and database licensing terms.

See THIRD_PARTY_NOTICES.MD for the exact attribution, provenance, transformations, and license notices associated with these metadata fields.

Maintainers and contact

• Rupert Polley - polley@fzi.de
• Marc Heinrich - heinrich@fzi.de
• Marc Uecker - uecker@fzi.de
• Sven Ochs - ochs@fzi.de
• Philip Schörner - schoerner@fzi.de

Contributors

Rupert Polley FZIKIT Principal Implementer · Lead Author · Data Platform Owner · SDK Owner · Anonymization · Label Integration & QA LinkedIn · G-Scholar

Marc Heinrich FZI Lead Author · Requirement Engineering · Labeling Guidelines · Project Management · Quality Control · SDK · Data Collection · Vehicle LinkedIn

Philip Schörner FZIKIT Lead Author · Research Direction · Project Management LinkedIn · G-Scholar

Marc Uecker FZIKIT Lead Author · Instance Segmentation · Calibration · Data Visualization · Quality Control · Odometry & Motion Correction LinkedIn · G-Scholar

Sven Ochs FZIKIT Lead Author · Localization · Odometry & Motion Correction · Data Collection LinkedIn · G-Scholar

Tobias Fleck FZIKIT Senior Author · Label Requirements · Research Direction LinkedIn · G-Scholar

Marc René Zofka FZI Senior Author · Research Direction LinkedIn · G-Scholar

J. Marius Zöllner KITFZI Senior Author · Research Direction G-Scholar

Acknowlegement of Further Contributors

The creation of the FZI-AURA Dataset includes additional contributions by the following colleagues, whom we thank for their assistance in data collection and discussions surrounding the dataset:
Melih Yazgan, Helen Gremmelmaier, Abhishek Vivekanandan, Dominik Heid, Tim Joseph, Christian Hubschneider, Nikolai Polley, Albert Schotschneider, and Yacin Boualili.

The manual labeling was conducted by Deepen AI. We are grateful for the great collaboration and results.

Citation

Please cite FZI-AURA as follows. We will replace this citation once the accompanying paper or technical report is available:

@misc{polley_fzi_aura_2026,
  author    = {Polley, Rupert and
               Heinrich, Marc and
               Sch{\"o}rner, Philip and
               Uecker, Marc and
               Ochs, Sven and
               Fleck, Tobias and
               Zofka, Marc Ren{\'e} and
               Z{\"o}llner, J. Marius},
  title     = {{FZI-AURA: A Multimodal Autonomous Driving Dataset}},
  year      = {2026},
  version   = {1.0.0},
  publisher = {FZI Forschungszentrum Informatik},
  url       = {https://huggingface.co/datasets/fzi-forschungszentrum-informatik/FZI-AURA}
}

@inproceedings{heinrich2024cocar,
  title={{CoCar NextGen: A Multi-purpose Platform for Connected Autonomous Driving Research}[,
  author={Heinrich, Marc and Zipfl, Maximilian and Uecker, Marc and Ochs, Sven and Gontscharow, Martin and Fleck, Tobias and Doll, Jens 
          and Sch{\"o}rner, Philip and Hubschneider, Christian and Zofka, Marc Ren{\'e} and Viehl, Alexander and Z{\"o}llner, J. Marius},
  booktitle={2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC)},
  pages={482--489},
  year={2024},
  organization={IEEE}
}

Funding

FZI-AURA was developed as part of the just better DATA (jbDATA) project and is partially included in the jbDATAset. jbData was supported by the German Federal Ministry for Economic Affairs and Energy (BMWE) and the European Union through NextGenerationEU. Visit the official jbDATA website for more information regarding the project and the jbDATAset: https://www.justbetterdata.de/en/.