Drive-P2D: A Progressive Perception-to-Decision Benchmark for VLMs in Autonomous Driving

Paper (arXiv): https://arxiv.org/abs/2601.14702
Hugging Face Dataset: https://huggingface.co/datasets/ColamentosZJU/Drive-P2D

Drive-P2D is a progressive perception-to-decision benchmark for evaluating Vision-Language Models (VLMs) across object-level perception, scene understanding, and driving decision-making in autonomous driving. It contains 6,650 questions across three levels: Object, Scene, and Decision. The benchmark uses objective choice-answer scoring while supporting reasoning-based error-mode analysis with a lightweight analyzer model.

This release provides annotations only.
Please obtain the original images from the official sources (nuScenes / KITTI / BDD100K) and align them locally if you want to run image-based evaluation.

---

Overview

Key Facts

• Protocol: progressive perception-to-decision evaluation — Object → Scene → Decision
• Tasks: 6 tasks, with two tasks at each level
• Scale: 6,650 QA items built from 1,295 front-facing driving images
• Risk-aware evaluation: each item includes a 5-level risk label danger_score ∈ {1,2,3,4,5}
• Reasoning analysis: separated answer scoring and reasoning-based error-mode analysis with an analyzer model

Contributions

• We introduce Drive-P2D, a progressive perception-to-decision benchmark with a three-level protocol, six tasks, 6,650 questions, and risk-aware splits.
• We evaluate mainstream VLMs with objective choice-answer scoring, covering perception-decision dependencies, similar-scene robustness, and model scaling behavior.
• We analyze model-generated reasoning to reveal error modes, and introduce a lightweight analyzer model for automated error-mode tagging.

---

Benchmark Structure

Drive-P2D follows a progressive perception-to-decision evaluation protocol:

• Object Level: object-level perception of key objects and object states
• Scene Level: scene-level understanding of weather, illumination, and special context factors
• Decision Level: decision-level driving action selection and action-risk assessment

Comparison with Existing Benchmarks

Benchmark	Progressive Evaluation	P-D Analysis	Risk-aware Split	Similar-scene Pairs	Reasoning Analysis	Error Modes	Auto Tagger
nuScenes-QA	✗	✗	✗	✗	✗	✗	✗
SURDS	✗	✗	✗	✗	✗	✗	✗
DriveLM	✓	✗	✗	✗	✓	✗	✗
DriveVLM	✓	✗	✗	✗	✓	✗	✗
DriveBench	✗	✗	✗	✗	✓	✗	✗
OmniDrive	✓	✓	✗	✗	✓	✗	✗
Drive-P2D	✓	✓	✓	✓	✓	✓	✓

---

Task List

The dataset contains six tasks, each provided as a JSON file:

Object Level (single-choice)

• Object-1 (Object-1.json): Identify the most important object influencing the driving decision.
• Object-2 (Object-2.json): Determine the state of a designated key object, such as a traffic light.

Scene Level (multiple-choice)

• Scene-1 (Scene-1.json): Recognize weather and illumination, such as daytime, nighttime, rain, snow, or heavy fog.
• Scene-2 (Scene-2.json): Identify special scene factors that may affect driving decisions, such as accidents or construction zones.

Decision Level (single-choice)

• Decision-1 (Decision-1.json): Select the optimal driving action for the ego vehicle.
• Decision-2 (Decision-2.json): Evaluate the risk level of a specified, potentially suboptimal action.

---

Data Format

Each file is a JSON array. Each element contains:

• image_name (string): image identifier/path
  • In this release, we provide annotations only; image_name is intended to be mapped to local image storage.
• taskX_q (string): question text for task X
• taskX_o (string): option list as a single string, such as "A....; B....; C...."
• taskX_a (string): answer letters
  • single-choice tasks: one letter, such as "C"
  • multiple-choice tasks: comma-separated letters, such as "A,C"
• danger_score (int or string): scenario risk label on a 5-level scale, where 1 is minimal and 5 is severe

Example

{
  "image_name": "images/xxxx.jpg",
  "task1_q": "...",
  "task1_o": "A....; B....; C....",
  "task1_a": "C",
  "danger_score": "2"
}

---

How to Use

1) Download Annotations

Download the six JSON files from the Hugging Face dataset page:

• https://huggingface.co/datasets/ColamentosZJU/Drive-P2D

2) Load Annotations in Python

import json

with open("Object-1.json", "r", encoding="utf-8") as f:
    data = json.load(f)

print(len(data), list(data[0].keys()))

3) Local Image Alignment

To evaluate with images, you must:

1. Download the source datasets from the official providers:
   • nuScenes
   • KITTI
   • BDD100K
2. Prepare a local image folder, such as ./images/.
3. Map each image_name in JSON to an existing local file path.

---

Citation

If you use Drive-P2D in your research, please cite:

@misc{tang2026drivep2dprogressiveperceptiontodecisionbenchmark,
      title={Drive-P2D: A Progressive Perception-to-Decision Benchmark for VLMs in Autonomous Driving},
      author={Zecong Tang and Zixu Wang and Yifei Wang and Weitong Lian and Tianjian Gao and Haoran Li and Tengju Ru and Lingyi Meng and Zhejun Cui and Yichen Zhu and Qi Kang and Kaixuan Wang and Yu Zhang},
      year={2026},
      eprint={2601.14702},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2601.14702},
}

---

License

This project is released under the Apache License 2.0. Some components or third-party implementations may be distributed under different licenses.

---

Acknowledgments

We thank the open-source community and dataset providers (nuScenes, KITTI, BDD100K) that make this benchmark possible.