Dataset & Publication · ECCV 2026

NARRATE

A Multimodal Real-World Australian Driving Dataset for Human-Centred Explanations in Automated Driving

Ashkan Yousefi Zadeh · Zishuo Zhu · Xiaomeng Li · Andry Rakotonirainy · Sebastien Glaser · Ronald Schroeter · Patricia Delhomme · Zahra Mehraban

ECCV 2026 — DriveX: 6th Workshop on Foundation Models for Autonomous Driving · Archival Track

ECCV 2026 DriveX Workshop arXiv 2608.14767 Dataset · QUT RDF DOI 10.25912/RDF_1786669427992 Poster · DriveX License coming soon

GitHub Repository QUT Research Data Finder View Poster

Dataset Showcase

NARRATE dataset showcase — 5 annotated driving events with multimodal sensor data and natural language explanations.
Full HD · Stereo audio narration · 91 seconds

2,050

Annotated driving events

Drivers & driving instructors

Synchronised camera feeds

Brisbane

CBD, Queensland, Australia

About

NARRATE is a multimodal real-world driving dataset collected in Brisbane CBD, Queensland, Australia by Queensland University of Technology. The dataset was developed through the ARC Training Centre for Automated Vehicles in Rural and Remote Regions (AVR3).

NARRATE is designed to support research on human-centred explanations for automated driving. It captures driving events from experienced drivers and driving instructors on public roads, pairing synchronised vehicle sensor data with driver-produced explanations collected during the drive (in-vehicle) and/or during a post-drive video-cued interview.

Each of the 2,050 annotated events is grounded in multimodal sensor streams and includes annotations for driver action, driving scenario context, and Situational Awareness (SA) levels.

Sensor Modalities

📷

4 Camera Feeds Front, front-left, front-right, rear — synchronised event-level clips

📡

LiDAR Point Clouds 3D spatial data from rosbag recordings for each driving event

💬

Natural Language In-vehicle and post-drive driver explanations in natural language

🗂️

Rich Metadata JSON per event: speed, acceleration, SA levels, scenario context, action labels

Annotation Schema

Each event in NARRATE is annotated with a comprehensive schema covering:

Dataset Structure

The mediated-access release is organised as an anonymised publication package. All event files use non-identifying anonymised event identifiers.

NARRATE/
├── front_camera/          # Forward-facing centre camera clips
├── front_left_camera/     # Forward-left camera clips
├── front_right_camera/    # Forward-right camera clips
├── rear_camera/           # Rear-facing camera clips
├── metadata_json/         # Per-event JSON (speed, accel, SA, labels, text)
└── sensor_bags/           # LiDAR rosbag recordings

Dataset Access

The dataset files are not hosted directly in the GitHub repository. Access is provided through mediated access via QUT Research Data Finder to support responsible data sharing.

QUT Research Data Finder: https://doi.org/10.25912/RDF_1786669427992

Documentation on GitHub

Acknowledgements

This research was supported by Queensland University of Technology (QUT) and the Australian Research Council Discovery Project (DP220102598).

Citation

If you use NARRATE in your research, please cite:

@article{yousefizadeh2026narrate,
  title   = {NARRATE: A Multimodal Real-World Australian Driving Dataset
             for Human-Centred Explanations in Automated Driving},
  author  = {Yousefi Zadeh, Ashkan and Zhu, Zishuo and Li, Xiaomeng and
             Rakotonirainy, Andry and Glaser, Sebastien and Schroeter, Ronald
             and Delhomme, Patricia and Mehraban, Zahra},
  year    = {2026},
  note    = {ECCV 2026, DriveX Workshop}
}