Person2Drive

Person2Drive is a dataset and benchmark for closed-loop personalized end-to-end autonomous driving. This repository hosts its dataset release.

Driving Like Yourself: A Benchmark for Closed-Loop Personalized End-to-End Autonomous Driving

ECCV 2026 · Paper on arXiv · PDF

Updates & Feedback

Actively maintained · Last updated: 2026-10-14

We actively maintain the Person2Drive dataset and its documentation. For current file availability and release history, see RELEASE_MANIFEST.md.

Questions and reports of problems are welcome. If you encounter missing files, download or extraction issues, or problems loading the data, please open a discussion in the Hugging Face Community. Include the affected driver/file path, the command you ran, and any error messages so we can investigate. Documentation questions are welcome too.

Overview

Person2Drive is a CARLA-based benchmark for studying personalized end-to-end autonomous driving. It contains human driving records from anonymized drivers and supports research on driving style modeling, driver-level personalization, route generalization, and closed-loop evaluation.

The full benchmark covers 50 drivers and four routes. Routes 103 and 108 provide driving records and selected sensor data. Routes 100 and 101 are used for generalization evaluation on unseen routes; their published driving records do not include sensor data.

Dataset Contents

Data component	Routes	Description
Six-camera RGB and LiDAR archives in sensor_data/	103 (Town05), 108 (Town04)	50 complete drivers; 800 archives available
Driver-level b2d_infos_train.pkl	103, 108	50 files available
Original driving records in driving_logs/	103, 108	50 drivers; 800 records / 4000 files
Driving records in driving_logs_without_sensor_data/	100, 101 (Town06)	Original driving records for unseen-route evaluation, without sensor data; see the release manifest for driver coverage
Shared map metadata	Benchmark maps	b2d_map_infos.pkl available

Repository Structure

Person2Drive/
├── sensor_data/driverXX/
│   ├── Town04_drive_1.tar.zst ... Town04_drive_8.tar.zst
│   ├── Town05_drive_1.tar.zst ... Town05_drive_8.tar.zst
│   └── b2d_infos_train.pkl
├── driving_logs/
│   └── driverXX/
│       ├── Town04_drive_N/
│       └── Town05_drive_N/
├── driving_logs_without_sensor_data/
│   └── driverXX/
│       ├── Town06_route100_drive_N/
│       └── Town06_route101_drive_N/
├── b2d_map_infos.pkl
├── RELEASE_MANIFEST.md
├── DATASET_DETAILS.md
└── README.md

driverXX is an anonymized public identifier. Each complete sensor package contains eight Town04 drives, eight Town05 drives, and one driver-level metadata file.

Naming convention. TownXX identifies the CARLA map, while route IDs identify individual driving routes. In this dataset, Town05_drive_N corresponds to route 103, and Town04_drive_N corresponds to route 108. Routes 100 and 101 both use Town06, so their directory names explicitly include route100 or route101 to distinguish them. N is the drive index within each driver and route.

Sensor Archives and Metadata

Each .tar.zst archive contains one drive directory with six RGB camera views and LiDAR point clouds. The camera views are rgb_front, rgb_front_left, rgb_front_right, rgb_back, rgb_back_left, and rgb_back_right.

Frame counts vary by drive; retained sensor files keep their original names and frame numbers.

b2d_infos_train.pkl provides driver-specific training metadata and frame references. The shared b2d_map_infos.pkl provides map-level metadata. See DATASET_DETAILS.md for the internal structure and file descriptions.

Original Driving Records

driving_logs/ stores original driving records, route information, and run results for routes 103 and 108. Corresponding camera and LiDAR data remain in sensor_data/.

driving_logs/driver01/Town04_drive_1/
├── metric_info.json
├── route.json
├── result.json
├── live_results.txt
└── RouteScenario_108_rep0.log

driving_logs_without_sensor_data/ provides the same five file types for routes 100 and 101 in Town06, organized as driverXX/Town06_route100_drive_N/ or driverXX/Town06_route101_drive_N/. There are two recorded drives per route for each currently available driver. Camera, LiDAR, and annotation files are not included in this part of the release.

The RouteScenario_*.log name varies with the route. These are CARLA binary recorder files, rather than plain-text console logs. JSON and text files are uploaded directly, without compression. Original file contents, timestamps, and indices within the files are preserved; motion record counts can differ from camera frame counts.

Download and Extraction

Install the Hugging Face CLI:

pip install -U huggingface_hub

Try one driving record first

Download the sensor archive and the five original driving-record files for driver01/Town04_drive_1. The download is approximately 325 MiB, excluding training and map metadata. All sizes below refer to downloaded files; allow additional disk space for extraction.

hf download dongxr7/Person2Drive --repo-type dataset \
  --include "sensor_data/driver01/Town04_drive_1.tar.zst" \
            "driving_logs/driver01/Town04_drive_1/*" \
  --local-dir Person2Drive

On Linux, install tar and zstd, then extract the archive:

tar -I zstd -xf \
  Person2Drive/sensor_data/driver01/Town04_drive_1.tar.zst \
  -C Person2Drive/sensor_data/driver01

This extracts the six RGB camera directories and the LiDAR directory under Person2Drive/sensor_data/driver01/Town04_drive_1/. The five driving-record files remain under Person2Drive/driving_logs/driver01/Town04_drive_1/.

On Windows, use an archive tool that supports Zstandard and TAR, preserving the directory structure.

Download training metadata when needed

To inspect the training metadata for the same driver, download it separately:

hf download dongxr7/Person2Drive \
  sensor_data/driver01/b2d_infos_train.pkl \
  --repo-type dataset --local-dir Person2Drive

The training metadata covers multiple drives. Downloading one archive is sufficient for inspecting that drive, but using all metadata entries requires the corresponding sensor archives. Extract those archives into the same driver directory so the drive folders and b2d_infos_train.pkl are located together.

Download a complete driver package

Download all sensor archives, training metadata, and route 103/108 driving records for driver01. This example downloads approximately 11.8 GiB, excluding route 100/101 records and shared map metadata.

hf download dongxr7/Person2Drive --repo-type dataset \
  --include "sensor_data/driver01/*" "driving_logs/driver01/*" \
  --local-dir Person2Drive

Replace driver01 in the paths to select another driver. Download sizes vary by driver.

Download route 100/101 driving records

Download the available route 100/101 driving records for the same driver. See RELEASE_MANIFEST.md for driver coverage.

hf download dongxr7/Person2Drive --repo-type dataset \
  --include "driving_logs_without_sensor_data/driver01/*" \
  --local-dir Person2Drive

Download shared map metadata

The root-level b2d_map_infos.pkl is shared across drivers and can be downloaded separately:

hf download dongxr7/Person2Drive b2d_map_infos.pkl \
  --repo-type dataset --local-dir Person2Drive

For reproducible experiments, record the Hugging Face commit revision used for your download. The CLI supports --revision <commit_hash>.

Privacy and Anonymization

Public driver identifiers use driverXX. The private name-to-ID mapping is not published. Original driving records retain run metadata, including recording timestamps.

Citation

If you use Person2Drive in your research, please cite our paper:

@inproceedings{dong2026person2drive,
  title={Driving Like Yourself: A Benchmark for Closed-Loop Personalized End-to-End Autonomous Driving},
  author={Dong, Xiaoru and Li, Ruiqin and Han, Xiao and Wu, Zhenxuan and Wang, Jiamin and Chen, Jian and Jiang, Qi and Yiu, Siu Ming and Zhu, Xinge and Ma, Yuexin},
  booktitle={Computer Vision -- ECCV 2026},
  series={Lecture Notes in Computer Science},
  volume={17039},
  pages={326--344},
  year={2026},
  publisher={Springer},
  address={Cham},
  doi={10.1007/978-3-032-37435-6_18},
  url={https://doi.org/10.1007/978-3-032-37435-6_18}
}

Contact

For direct inquiries, please contact Xiaoru Dong at xrdong@cs.hku.hk.