DriveFusion-Data

DriveFusionQA

An Autonomous Driving Vision-Language Model for Scenario Understanding & Decision Reasoning.

---

DriveFusion-Data is a large-scale multimodal autonomous driving dataset collected in the CARLA simulator using a privileged rule-based expert policy (PDM-Lite). The dataset contains rich sensor data, vehicle measurements, and language annotations for training vision-language-action (VLA) models.

This dataset is part of the DriveFusion project.

---

Dataset Overview

DriveFusion-Data provides a comprehensive multimodal dataset for autonomous driving research, including:

• RGB camera images from 360° multi-camera coverage (front, front-left, front-right, back-left, back-right)
• LiDAR point clouds
• Semantic segmentation maps
• Depth maps
• Bounding boxes
• Vehicle and simulator measurements
• Natural language annotations (VQA, commentary, instruction following)

The dataset is generated using a CARLA-based data collection framework with multi-town, multi-scenario, and multi-sensor configurations.

---

Data Collection Framework

The data was collected using the DriveFusion CARLA Data Collection Framework, which provides:

• Rule-based expert driving using PDM-Lite
• Multi-camera 360° sensor recording and LiDAR
• Weather and lighting augmentation
• Scenario-based route execution
• Automated batch data generation on clusters (SLURM)
• Format conversion and dataset validation tools

Collection code repository:
https://github.com/DriveFusion/carla-data-collection

---

Dataset Sources and Attribution

DriveFusion-Data builds upon several open-source frameworks and datasets:

Core Simulation:

• CARLA Simulator
• CARLA Leaderboard 2.0
• Scenario Runner

Reference Methods:

• DriveLM (PDM-Lite autopilot and VQA generation)

Language Dataset Reference:

• SimLingo Dataset

Users must comply with the licenses of all referenced frameworks and datasets.

---

Dataset Format

Two main formats are provided:

Pre-DriveFusion Format

• Raw sensor data and measurements stored in compressed JSON and sensor files.

DriveFusion Format

• Standardized multimodal structure for end-to-end VLA training.
• Includes aligned sensor data and language annotations.

---

Intended Use

This dataset is designed for:

• Vision-Language-Action (VLA) model training
• Autonomous driving research and benchmarking
• Multimodal perception and planning research
• Language grounding in driving environments
• Embodied AI and robotics research

---

License and Attribution

This dataset is derived from simulation and public frameworks. Users must comply with:

• CARLA license
• CARLA Leaderboard and Scenario Runner licenses (MIT)
• DriveLM license
• SimLingo license

The DriveFusion framework code is released under Apache 2.0. Language annotations and third-party components may have additional license restrictions.

---

Citation

If you use DriveFusion-Data, please cite:

@misc{drivefusiondata2026,
  title={DriveFusion-Data: A Large-Scale Multimodal Dataset for Autonomous Driving},
  author={Samir, Omar and DriveFusion Team},
  year={2026},
  url={https://huggingface.co/datasets/DriveFusion/DriveFusion-Data}
}