grow-ai-like-a-child/mechanical-reasoning

Mechanical Reasoning

Mechanical Reasoning is a multimodal benchmark designed to evaluate mechanical and physical reasoning in large vision-language models (VLMs). This dataset is introduced in the paper:

Probing Mechanical Reasoning in Large Vision Language Models
arXiv:2410.00318

It is part of the Grow AI Like a Child benchmark initiative, which explores structured, curriculum-aligned evaluation for embodied intelligence.

---

🧠 Dataset Overview

This dataset focuses on mechanical reasoning grounded in either static images or short video clips. Each item presents a physical scenario and poses a question requiring spatial, causal, or commonsense understanding.

The dataset contains:

• 148 samples
• Two modalities: image or video
• Two question formats: multiple-choice (MC) or true/false (TF)

---

🧩 Reasoning Types

The questions span a variety of physical reasoning challenges:

• Gear and pulley systems
• Fluid and buoyancy dynamics
• Torque and balance
• Conveyor motion
• Mechanical advantage
• Temporal prediction and causality

---

📁 Dataset Format

Each sample includes:

Field	Description
index	Unique ID (e.g., d0001)
media_type	Either image or video
media	Local file name (e.g., d0001.png, d0131.mov)
question_type	Either MC or TF
sec_label	Integer from 1 to 6 (see section mapping below)
question	Natural language question with choices embedded (if MC)
correct_answer	The ground-truth answer (e.g., A, B, Yes, No)

🔢 sec_label Categories

Label	Category
1	System Ability
2	Pulley Systems
3	Gears & Rotations
4	Leverage Principle
5	Inertia and Motion
6	Fluid Mechanics

---

📂 Folder Structure

data/
├── data.csv
├── images/
│   ├── *.png
│   └── metadata.jsonl
├── videos/
│   ├── *.mov / *.mp4
│   └── metadata.jsonl

• The metadata.jsonl files store structured sample entries.
• For video samples, file_name in metadata is mapped to .png format to avoid embedding issues on Dataset Viewer.

---

💡 Example

{
  "file_name": "d0001.png",
  "media_type": "image",
  "question_type": "MC",
  "sec_label": 6,
  "question": "In which direction is the leaking cylinder moving in the water in the picture? A. up; B. down",
  "correct_answer": "B"
}

---

📚 Citation

If you use this dataset, please cite:

@misc{sun2025probingmechanicalreasoninglarge,
      title={Probing Mechanical Reasoning in Large Vision Language Models}, 
      author={Haoran Sun and Qingying Gao and Haiyun Lyu and Dezhi Luo and Yijiang Li and Hokin Deng},
      year={2025},
      eprint={2410.00318},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2410.00318}, 
}

---

🤝 Acknowledgments

This dataset is developed by the Grow AI Like a Child community.