Person2Drive Dataset Details
This document provides additional details about the organization, content, and release status of the Person2Drive dataset.
Dataset Overview
Person2Drive is a closed-loop personalized end-to-end autonomous driving dataset collected in the CARLA simulator. The dataset is designed to support research on human driving style modeling, driver-level personalization, and closed-loop autonomous driving evaluation.
The dataset contains human driving records collected from anonymized drivers under controlled CARLA driving environments. Each driver is represented by an anonymized identifier, such as driver01, driver02, and so on.
Dataset Scale
The full Person2Drive dataset contains:
- 50 anonymized human drivers
- 4 driving routes
- multiple repeated drives for each driver-route setting
- driver-level metadata and route-level driving records
The dataset is being progressively uploaded to this Hugging Face repository. Please refer to RELEASE_MANIFEST.md for the current upload status.
Data Organization
The released dataset is organized by driver. Each driver folder contains compressed route-level driving record archives and a corresponding metadata file.
The current repository structure is:
Person2Drive/
├── drivers/
│ ├── driver01/
│ │ ├── Town04_drive_1.tar.zst
│ │ ├── Town04_drive_2.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
Each driverXX/ folder corresponds to one anonymized human driver.
Route-Level Archives
The .tar.zst files are compressed route-level driving record packages. Each archive contains data collected from a specific driver under a specific route or driving setting in CARLA.
Examples include:
Town04_drive_1.tar.zst
Town04_drive_2.tar.zst
Town05_drive_1.tar.zst
Town05_drive_2.tar.zst
The file names indicate the CARLA town and drive index used in the released data organization.
Data Content
Depending on the released archive, the dataset may include:
- 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 exact internal structure of each archive will be further documented as the release is completed.
Metadata Files
Each driver folder contains a metadata file:
b2d_infos_train.pkl
This file stores driver-specific metadata used for data loading and organization.
The root directory contains:
b2d_map_infos.pkl
This file stores map-level metadata used by the dataset and benchmark pipeline.
Anonymization
Driver identities are anonymized. Public driver folders use identifiers such as driver01, driver02, and so on. No real driver names are used in the released dataset structure.
Release Status
The dataset is currently being uploaded and organized. Files already available in this Hugging Face repository are part of the public release.
The release status of uploaded and in-progress driver folders is tracked in:
RELEASE_MANIFEST.md
Notes
This document focuses on the dataset organization and released data files. Benchmark protocols, evaluation metrics, and code-level usage instructions will be documented separately.