---
license: apache-2.0
viewer: false
tags:
- autonomous-driving
- carla
- benchmark
- end-to-end-driving
- personalized-driving
---


# Person2Drive

**Person2Drive** is the dataset and benchmark repository for the ECCV 2026 paper:

**Driving like yourself: A Benchmark for Closed-Loop Personalized End-to-End Autonomous Driving**

## Overview

Person2Drive is a CARLA-based benchmark for studying personalized end-to-end autonomous driving. It contains human driving records collected in closed-loop simulation environments and is organized at the driver level.

The goal of the benchmark is to support research on human driving style modeling, driver-level personalization, route-level generalization, and closed-loop evaluation of end-to-end driving models. Unlike conventional driving datasets that mainly capture generic driving behavior, Person2Drive provides multiple driving records from anonymized human drivers under shared or comparable simulation settings.

## Release Status

This Hugging Face repository is the stable public access page for Person2Drive.

The dataset is being uploaded and organized progressively. Files already available in this repository are part of the public release. The current upload status is summarized in [`RELEASE_MANIFEST.md`](RELEASE_MANIFEST.md).

Additional metadata, benchmark documentation, and usage instructions will be added to this repository as the release is finalized for the ECCV 2026 camera-ready version.

## Dataset Statistics

The full Person2Drive release contains:

* 50 anonymized human drivers;
* 4 driving routes in closed-loop CARLA environments;
* driver-level folders under `drivers/`;
* route-level compressed archives for each driver;
* metadata files for driver-level and map-level organization.

Each `driverXX/` folder corresponds to one anonymized human driver. Route-level data are stored as `.tar.zst` archives.

## Dataset Contents

The released data include the information required for personalized end-to-end driving research, including:

* ego-vehicle states;
* control signals such as steering, throttle, and brake;
* ego trajectories;
* route and navigation information;
* selected sensor data used by end-to-end driving models;
* metadata for driver-level and route-level organization.

The dataset can be used for both open-loop behavior analysis and closed-loop personalized driving evaluation.

## Repository Structure

The repository is organized as follows:

```text
Person2Drive/
├── drivers/
│   ├── driver01/
│   │   ├── Town04_drive_1.tar.zst
│   │   ├── ...
│   │   ├── Town04_drive_8.tar.zst
│   │   ├── Town05_drive_1.tar.zst
│   │   ├── ...
│   │   ├── Town05_drive_8.tar.zst
│   │   └── b2d_infos_train.pkl
│   ├── driver02/
│   │   ├── Town04_drive_1.tar.zst
│   │   ├── ...
│   │   ├── Town05_drive_8.tar.zst
│   │   └── b2d_infos_train.pkl
│   └── ...
├── b2d_map_infos.pkl
├── RELEASE_MANIFEST.md
├── DATASET_DETAILS.md
└── README.md
```

The file `b2d_infos_train.pkl` stores metadata associated with the corresponding driver data. The file `b2d_map_infos.pkl` contains map-level metadata used by the benchmark and evaluation pipeline.

Detailed data organization and field descriptions are provided in [`DATASET_DETAILS.md`](DATASET_DETAILS.md).

## Download and Extraction

The released data are stored as `.tar.zst` archives. Each archive corresponds to a route-level driving record package for one anonymized driver.

To clone the repository with Git LFS:

```bash
git lfs install
git clone https://huggingface.co/datasets/dongxr7/Person2Drive
```

To extract an archive on Linux:

```bash
tar -I zstd -xvf Town04_drive_1.tar.zst
```

Alternatively, extraction can be performed in two steps:

```bash
zstd -d Town04_drive_1.tar.zst
tar -xvf Town04_drive_1.tar
```

On Windows, `.tar.zst` archives can be extracted with archive tools that support Zstandard compression, such as 7-Zip, PeaZip, or Bandizip.

## Benchmark Usage

Person2Drive is intended for evaluating whether an end-to-end driving model can adapt to individual human driving styles while maintaining safe closed-loop performance.

Typical evaluation settings include:

* driver-level personalization;
* route-level generalization;
* trajectory prediction quality;
* driving style similarity;
* closed-loop driving performance.

Benchmark scripts, evaluation protocols, and additional usage instructions will be released in this repository.

## Privacy and Anonymization

Driver identities are anonymized as `driverXX`. The dataset does not use real driver names in the public release.

## Citation

Citation information will be added after the ECCV 2026 camera-ready version is finalized.

## Contact

For questions about the dataset, please contact Xiaoru Dong at [xrdong@cs.hku.hk](mailto:xrdong@cs.hku.hk).