PDD: Personalized Driving Dataset

Dataset Description

PDD (Personalized Driving Dataset) is a multi-driver, multi-scenario driving dataset collected in CARLA 0.9.15. It captures real human driving behavior from 30 individual drivers, each performing 21 challenging driving scenarios. The dataset is designed for research on personalized autonomous driving, where models learn to mimic individual driving styles.

Each driver has a detailed profile capturing demographics, driving experience, habits, and self-reported driving style. The driving data includes front-camera RGB images, 3D bounding boxes for surrounding objects, and per-frame vehicle telemetry (speed, acceleration, steering, throttle, brake, etc.).

Dataset Statistics

Metric	Value
Drivers	30
Scenarios per driver	21
Total scenario instances	630
Total image frames	70,087
Total bounding box files	70,087
Dataset size	~13 GB
Simulator	CARLA 0.9.15
Frame rate (saved)	4 FPS

Dataset Structure

PDD/
├── driver_01/
│   └── data/
│       ├── Accident/
│       │   ├── images/          # Front-camera RGB images (JPEG)
│       │   │   ├── 0.jpg
│       │   │   ├── 1.jpg
│       │   │   └── ...
│       │   ├── boxes/           # 3D bounding boxes (compressed JSON)
│       │   │   ├── 0.json.gz
│       │   │   ├── 1.json.gz
│       │   │   └── ...
│       │   └── metric/
│       │       ├── metrics.json       # Per-step control inputs
│       │       └── metric_info.json   # Per-frame telemetry
│       ├── BlockedIntersection/
│       │   └── ...
│       └── ... (21 scenarios)
├── driver_02/
│   └── ...
├── ... (30 drivers)
└── user_profiles/
    ├── driver_01.json
    ├── driver_02.json
    └── ... (30 profiles)

Data Fields

Images (images/*.jpg)

Front-forward RGB camera images captured at 4 FPS during driving.

Bounding Boxes (boxes/*.json.gz)

Gzip-compressed JSON files, one per frame. Each contains a list of detected objects:

• class: Object type (ego_car, car, walker, static)
• position: [x, y, z] relative to ego vehicle
• extent: [length, width, height] of bounding box
• yaw: Heading angle
• speed: Object speed
• id: Unique object identifier
• distance: Distance from ego vehicle

Telemetry (metric/metric_info.json)

Per-frame driving telemetry indexed by frame number:

• location: [x, y, z] world coordinates
• rotation: [pitch, roll, yaw]
• speed: Current speed (m/s)
• speed_limit: Road speed limit (m/s)
• acceleration: [x, y, z] acceleration vector
• velocity: [x, y, z] velocity vector
• angular_velocity: [x, y, z]
• distance_to_front_vehicle: Distance to lead vehicle (m)
• lane_change_count: Number of lane changes
• lane_info: Current lane information
• target_point, target_point_next: Navigation waypoints
• expert_target_speed: Expert reference speed
• expert_control_steer/throttle/brake: Expert reference controls
• other_vehicles: Nearby vehicle information
• walkers: Nearby pedestrian information

Control Inputs (metric/metrics.json)

Sequential list of control commands applied at each simulation step:

• steer: Steering angle [-1, 1]
• throttle: Throttle input [0, 1]
• brake: Brake input [0, 1]
• gear, hand_brake, reverse: Additional vehicle state

Driver Profiles (user_profiles/driver_XX.json)

• basic_information: Age, gender, occupation
• driving_experience: Years of experience
• driving_frequency_per_week: Typical weekly driving hours
• driving_purposes: Common driving use cases
• driving_habits_preferences: Self-reported driving habits
• health_and_driving_records: Health conditions, accident history
• driving_style: Self-classified style (Aggressive / Assertive / Balanced / Calm / Cautious)
• international_driving_experience: Driving experience in other regions

Usage

from huggingface_hub import snapshot_download

# Download the full dataset
snapshot_download(repo_id="tasl-lab/PDD", repo_type="dataset", local_dir="./PDD")

# Download a specific driver only
snapshot_download(repo_id="tasl-lab/PDD", repo_type="dataset", local_dir="./PDD",
                  allow_patterns=["driver_01/**", "user_profiles/**"])

Or use the provided loading script (load_pdd.py) for a structured PyTorch-compatible loader:

# Copy load_pdd.py to your project, then:
from datasets import load_dataset

dataset = load_dataset("./load_pdd.py", name="driver_01", trust_remote_code=True)
sample = dataset["train"][0]
print(sample["driver_id"])        # "driver_01"
print(sample["scenario"])         # "Accident"
print(sample["speed"])            # 0.001
print(sample["image"])            # PIL Image
print(sample["driver_profile"])   # {...}

Citation

If you use this dataset in your research, please cite:

@misc{wang2026drivewaypreferencealignment,
      title={Drive My Way: Preference Alignment of Vision-Language-Action Model for Personalized Driving},
      author={Zehao Wang and Huaide Jiang and Shuaiwu Dong and Yuping Wang and Hang Qiu and Jiachen Li},
      year={2026},
      eprint={2603.25740},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2603.25740},
}