Yuhang253820/Onsite_OVPD

OVPD: OnSite Virtual-Physical Dataset

Paper

Dataset Summary

OVPD is a virtual-physical fusion testing dataset derived from the 2025 OnSite Autonomous Driving Challenge. It is designed to support autonomous driving research on planning, decision-making, replay-based analysis, and deployment-oriented evaluation in complex interactive scenarios. The dataset is organized at the clip level: each clip corresponds to one complete test run from one participating team. In total, OVPD contains 20 testing clips collected from 20 teams, covering a scenario chain composed of 15 atomic scenarios, with a total duration of nearly 3 hours. The original data were sampled and aligned using a unified global timestamp at 10 Hz.

OVPD is built upon a virtual-physical fusion testing paradigm. The ego vehicle executes on a real vehicle platform, while surrounding background traffic is generated and managed in a controllable testing environment. This allows the dataset to preserve real ego-vehicle dynamics feedback while maintaining replayability and scenario controllability.

At the current release stage, OVPD primarily provides:

• ego vehicle trajectories and motion states
• ego control commands
• background traffic participant trajectories and states
• traffic light states
• HD maps in OpenDRIVE format
• evaluation results under five dimensions:
  • safety
  • efficiency
  • comfort
  • rule compliance
  • traffic coordination

Note: The current release provides the full structured clip-level data and image data for all teams. Image data in this repository are downsampled from the original 10 Hz sampling rate to 2 Hz for easy access, and are distributed as WebDataset shards under image_wds/<team>/. The full-resolution 10 Hz image data can be accessed from Baidu Netdisk with extraction code ensc.

Dataset Motivation

Current public autonomous driving datasets are highly valuable for perception, prediction, and planning research, but many of them are still limited in terms of deployment-oriented evaluation. In particular, publicly available resources often lack real ego-vehicle dynamics constraints, controllable closed-loop interaction, and multi-dimensional diagnostic metrics for deployability analysis. OVPD is released to help bridge this gap by providing a replayable and diagnosable dataset collected in a virtual-physical fusion testing environment centered on real-vehicle-in-the-loop execution.

OVPD is especially intended to support:

• long-tail planning and decision-making validation
• replay-based scenario analysis
• failure diagnosis and iterative improvement
• capability-oriented benchmark construction
• research on rule-compliant and interaction-aware autonomous driving

Dataset Structure

The dataset is organized as follows:

OVPD/
├── README.md
├── LICENSE
├── SCHEMA.md
├── TIME_ALIGNMENT.md
├── data/
│   ├── map/
│   │   ├── opendrive_bj.xodr
│   │   └── opendrive_sh.xodr
│   └── clips/
│       ├── team_01/
│       ├── team_02/
│       └── ...
├── image_wds/
│   ├── team_01/
│   ├── team_02/
│   └── ...
└── meta/
    └── image_release_manifest.jsonl

The official semantic path for image data is data/clips/<team>/images/. In this Hugging Face release, images are published in WebDataset format for efficient storage and streaming. The uploaded image shards cover all teams and are sampled at 2 Hz to keep the release size manageable. They are stored under image_wds/<team>/.

Basic Unit

The basic unit of OVPD is a testing clip.

Each clip: • corresponds to one team • contains one complete test run • includes a scenario chain with 15 atomic scenarios • is synchronized by a unified global timestamp

Released Modalities

The current release includes the following modalities:

1. Ego vehicle trajectories and motion states

For the ego vehicle, OVPD releases motion-related data such as: • global trajectory • velocity • heading / yaw • angular acceleration

2. Ego control commands

OVPD releases low-level control-related signals generated during real-vehicle execution, including: • throttle • brake • steering wheel angle

3. Participants’ trajectories and states

For background traffic participants, OVPD releases structured state data for agents such as: • vehicles • pedestrians • cyclists / non-motorized traffic participants

Typical released attributes include: • category • global trajectory • velocity • heading / yaw • angular acceleration

4. Traffic light states

OVPD records traffic signal states for signal-controlled regions in the test map, supporting replay and rule-compliance analysis. 

5. HD maps

HD maps for the two test sites are provided in OpenDRIVE format. These maps describe static proving-ground elements such as: • lanes • intersections • traffic lights

6. Evaluation scores

OVPD releases score outputs associated with each test record. The evaluation protocol follows the competition setting and includes five dimensions: • safety • efficiency • comfort • rule compliance • traffic coordination

Scenario Taxonomy

OVPD contains 15 atomic scenarios grouped into three capability-oriented categories. 

A. Emergency Response • A1 Pedestrian pop-out crossing • A2 Suddenly exposed stalled vehicle • A3 Red-light-running emergency vehicle • A4 Lead vehicle hard braking • A5 Aggressive e-scooter cut-in • A6 Fast background-vehicle cut-in • A7 Work-zone obstacle avoidance

B. Traffic Efficiency • B1 Roundabout traversal • B2 Ramp merging • B3 Non-motorized spillover • B4 Unsignalized intersection negotiation

C. Rule Compliance • C1 Consecutive right turns with blockage • C2 Queue spillback at an intersection • C3 Complex signal-phase composition • C4 Autonomous parking under restrictions

These scenarios are designed to reflect three major deployment-critical capability dimensions: • emergency response • traffic efficiency • rule compliance 

Time Alignment

All modalities were originally sampled and aligned using a unified global timestamp at 10 Hz. Detailed assumptions and alignment rules are documented in TIME_ALIGNMENT. In this Hugging Face release, image data are downsampled to 2 Hz to reduce data volume; full 10 Hz image data are available from Baidu Netdisk with extraction code ensc.

Data Schema

Detailed field definitions, units, and file format descriptions are provided in SCHEMA.

Evaluation Protocol

OVPD adopts a multi-dimensional evaluation framework to better reflect deployability under realistic urban driving constraints. The five released dimensions are: • Safety • Efficiency • Comfort • Rule Compliance • Traffic Coordination 

These score outputs are released together with the dataset to support: • offline replay analysis • capability profiling • benchmark comparison • failure diagnosis

Intended Uses

OVPD is intended for: • academic research • autonomous driving planning and decision-making analysis • scenario replay and debugging • evaluation of deployability-oriented driving capabilities • capability diagnosis under complex interactive scenarios

Out-of-Scope Uses

OVPD is not intended for: • identifying individuals or institutions • direct public-road deployment claims without independent validation • safety certification without additional testing and verification • any use beyond the scope permitted by the dataset license

Limitations

•	OVPD is collected in closed-course proving grounds rather than open public roads.
•	Background traffic is generated within a virtual-physical testing pipeline and may not fully cover all real-world interaction diversity.
•	Competition-derived clips may reflect challenge-specific distributions of driving strategies and failures.

License

Please refer to LICENSE for the license terms of this dataset.

Citation

If you use OVPD in your research, please cite the corresponding paper:

@article{zhang2025ovpd,
  title={OVPD: A Virtual-Physical Fusion Testing Dataset of OnSite Autonomous Driving Challenge},
  author={Zhang, Yuhang and Zhang, Jiarui and Jian, Bowen and Zhou, Xin and Lv, Zhichao and Hang, Peng and Yu, Rongjie and Tian, Ye and Sun, Jian},
  year={2025}
}

Acknowledgement

OVPD is derived from the 2025 OnSite Autonomous Driving Challenge and built upon the virtual-physical fusion testing workflow described in the accompanying paper. We thank all participating teams and the platform developers for their support.