---
license: mit
task_categories:
- robotics
tags:
- CARLA
- Robotics
- autonomous-driving
- end-to-end-driving
- multi-modal
- 123D
pretty_name: LEAD 123D
size_categories:
- 1K<n<10K
---

<p align="center">
  <picture>
    <source media="(prefers-color-scheme: dark)" srcset="https://kesai.eu/py123d/_static/123D_logo_transparent_white.svg">
    <source media="(prefers-color-scheme: light)" srcset="https://kesai.eu/py123d/_static/123D_logo_transparent_black.svg">
    <img alt="123D" src="https://kesai.eu/py123d/_static/123D_logo_transparent_black.svg" width="320">
  </picture>
</p>

<h2 align="center">LEAD 123D</h2>

<p align="center">
  Five hours of CARLA expert driving logs in the unified <a href="https://github.com/kesai-labs/py123d">123D</a> Apache Arrow format.
</p>

<p align="center">
  <a href="https://ln2697.github.io/lead">Project</a> · <a href="https://github.com/kesai-labs/lead">Code</a> · <a href="https://github.com/kesai-labs/py123d">py123d</a>
</p>

## Overview

**LEAD 123D** is a CARLA Leaderboard 2.0 driving dataset collected with LEAD's rule-based privileged expert and stored in the [123D](https://github.com/kesai-labs/py123d) unified driving-data format. Each route is a self-contained directory of Apache Arrow IPC files, one per modality.

## Data structure

| File | Content |
| :--- | :--- |
| `ego_state_se3.arrow` | Ego vehicle pose, velocity, acceleration |
| `camera.pcam_{f,b,l,r}{0,1}.arrow` | RGB camera streams (front, back, left, right) |
| `lidar.lidar_top.arrow` | Top LiDAR point clouds |
| `box_detections_se3.arrow` | 3D bounding boxes for all dynamic actors |
| `traffic_light_detections.arrow` | Per-frame traffic light states |
| `sync.arrow` | Cross-modality synchronization timestamps |
| `maps/carla/*.arrow` | HD map: lanes, intersections, crosswalks, road edges, road lines (WKB geometry) |

- **Coordinate conventions:** ISO 8855 (vehicle / body), OpenCV (cameras)

## Usage

Install [py123d](https://github.com/kesai-labs/py123d) and pull the dataset:

```bash
pip install py123d

git lfs install
git clone https://huggingface.co/datasets/ln2697/lead123d
```

See the [py123d documentation](https://kesai.eu/py123d/) for loading, visualization, and Scene/Map API usage.

## License

Released under the [MIT License](https://github.com/kesai-labs/lead/blob/main/LICENSE).
CARLA assets and OpenDRIVE maps remain under their respective upstream licenses.

## Citation

If you use LEAD 123D, please cite the papers:

```bibtex
@inproceedings{Nguyen2026CVPR,
  author    = {Long Nguyen and Micha Fauth and Bernhard Jaeger and Daniel Dauner and
               Maximilian Igl and Andreas Geiger and Kashyap Chitta},
  title     = {LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving},
  booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
}

@software{Contributors123D,
  title   = {123D: A Unified Library for Multi-Modal Autonomous Driving Data},
  author  = {123D Contributors},
  year    = {2026},
  url     = {https://github.com/kesai-labs/py123d},
  license = {Apache-2.0}
}
```