AutoDriDM: An Explainable Benchmark for Decision-Making of Vision-Language Models in Autonomous Driving

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

AutoDriDM is a decision-centric, progressive benchmark for evaluating the perception-to-decision capability boundary of Vision-Language Models (VLMs) in autonomous driving.

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: 3 progressive levels — Object → Scene → Decision
• Tasks: 6 tasks (two per level)
• Scale: 6,650 QA items built from 1,295 front-facing images
• Risk-aware evaluation: each item includes a 5-level risk label danger_score ∈ {1,2,3,4,5}
  • High-risk can be defined as average danger_score ≥ 4.0

---

🧩 Benchmark Structure

AutoDriDM follows a progressive evaluation protocol:

• Object Level: identify key objects and recognize their states
• Scene Level: understand global context (weather/illumination, special factors)
• Decision Level: choose driving actions and assess risk levels

---

📦 Task List (6 JSON Files)

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

Object Level (single-choice)

• Object-1 (Object-1.json): Identify the key object that most influences the driving decision.
• Object-2 (Object-2.json): Determine the state of a designated key object (e.g., traffic light state).

Scene Level (multiple-choice)

• Scene-1 (Scene-1.json): Recognize weather / illumination (e.g., daytime, nighttime, rain, snow, heavy fog).
• Scene-2 (Scene-2.json): Identify special scene factors that potentially affect driving decisions (e.g., accident scene, construction zone).

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 (JSON)

Each file is a JSON array. Each element is an object with the following fields:

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

Example (JSON)

{
  "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/AutoDriDM

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 (for image-based evaluation)

To evaluate with images, you must:

1. Download the source datasets from the official providers:
   • nuScenes
   • KITTI
   • BDD100K
2. Prepare a local folder (example):
   • ./images/
3. Map each image_name in JSON to an existing local file path in your environment.

---

📌 Citation

If you use AutoDriDM in your research, please cite:

@article{tang2026autodridm,
  title={AutoDriDM: An Explainable Benchmark for Decision-Making of Vision-Language Models in Autonomous Driving},
  author={Tang, Zecong and Wang, Zixu and Wang, Yifei and Lian, Weitong and Gao, Tianjian and Li, Haoran and Ru, Tengju and Meng, Lingyi and Cui, Zhejun and Zhu, Yichen and others},
  journal={arXiv preprint arXiv:2601.14702},
  year={2026}
}

---

⚖️ 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.