PhysicsMind Video Dataset

This is the video generation (VG) component of the PhysicsMind benchmark, a unified physics-grounded benchmark that evaluates both reasoning and prediction under the same physical laws in both simulated and real-world settings.

About PhysicsMind

PhysicsMind is designed to evaluate whether video generation models can produce physically plausible motion trajectories that obey center-of-mass, torque, and inertial constraints. The benchmark focuses on three canonical mechanics scenarios:

1. Center of Mass - Measuring center of gravity by suspension method
2. Lever Equilibrium - Lever balance conditions
3. Newton's First Law - Inertia verification experiments

Each scenario is realized with diverse real tabletop experiments and 2D simulations, enabling evaluation of whether models trained in simulation environments generalize well under real-world noise, dynamic differences, and visual variability.

Directory Structure

vieo_data/
├── Lever_Balance_Conditions/
│   ├── Boundary_Frames/
│   │   ├── Lever_Balance_Conditions_X_Initial_Frame.png
│   │   └── Lever_Balance_Conditions_X_Final_Frame.png
│   ├── video/
│   │   └── Lever_Balance_Conditions_X.mov
│   ├── data_lever_balance_en.json
│   ├── data_lever_balance_scene_en.json
│   └── data_lever_balance_scene_object_law_en.json
│
├── Verification_of_Newton's First_Law/
│   ├── Boundary_Frames/
│   │   ├── Newtons_First_Law_X_Initial_Frame.png
│   │   └── Newtons_First_Law_X_Final_Frame.png
│   ├── video/
│   │   └── Newtons_First_Law_X.mp4
│   ├── data_newton_first_law_en.json
│   ├── data_newton_first_law_scene_en.json
│   └── data_newton_first_law_scene_action_law_en.json
│
└── Measuring_Center_of_Gravity_by_Suspension_Method/
    ├── real/
    │   ├── Boundary_Frames/
    │   │   ├── Center_of_Gravity_Suspension_X_Initial_Frame.png
    │   │   └── Center_of_Gravity_Suspension_X_Final_Frame.png
    │   ├── video/
    │   │   └── Center_of_Gravity_Suspension_X.mp4
    │   ├── data.json
    │   ├── data_scene.json
    │   └── data_scene_action_law.json
    │
    └── sim/
        ├── Boundary_Frames/
        │   ├── Center_of_Gravity_Suspension_Sim_X_Initial_Frame.png
        │   └── Center_of_Gravity_Suspension_Sim_X_Final_Frame.png
        ├── video/
        │   └── Center_of_Gravity_Suspension_Sim_X.mp4
        ├── data.json
        ├── original_prompts.json
        ├── refined_prompts.json
        └── negative_prompts.json

Experiments

1. Lever Balance Conditions

Demonstrates lever equilibrium principles, showing how forces and distances from the fulcrum affect balance. Videos feature varying initial setups and object types.

2. Verification of Newton's First Law (Inertia)

Illustrates Newton's First Law of Motion through tablecloth-pull experiments with varying pulling speeds and object masses.

3. Measuring Center of Gravity by Suspension Method

Shows how to determine an object's center of gravity using the suspension method with varying object shapes and suspension points.

• real/: Real-world tabletop experiment recordings
• sim/: 2D simulated videos

File Descriptions

File	Description
Boundary_Frames/	Initial and final frame images for each video
video/	Video files (.mov or .mp4)
data*.json	Basic metadata with video name, path, category, and prompts
*_scene*.json	Scene-level descriptions
*_action_law*.json or *_object_law*.json	Physics law and action descriptions
original_prompts.json	Original text prompts for video generation
refined_prompts.json	Refined/improved prompts
negative_prompts.json	Negative prompts for video generation

JSON Fields

• Video Name: Filename of the video
• Video Path (Relative): Relative path to the video file
• Category: Experiment category
• Initial Frame: Path to the first frame image
• Final Frame: Path to the last frame image
• Prompt: Text description of the video content

Citation

If you use this dataset, please cite the PhysicsMind paper.