PhysX-3D

# PhysXNet & PhysXNet-XL

<p align="left"><a href="https://arxiv.org/abs/2507.12465"><img src='https://img.shields.io/badge/arXiv-Paper-red?logo=arxiv&logoColor=white' alt='arXiv'></a>
<a href='https://huggingface.co/papers/2507.12465'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Paper-blue'></a>
<a href='https://physx-3d.github.io/'><img src='https://img.shields.io/badge/Project_Page-Website-green?logo=homepage&logoColor=white' alt='Project Page'></a>
<a href='https://youtu.be/M5V_c0Duuy4'><img src='https://img.shields.io/youtube/views/M5V_c0Duuy4'></a>

This dataset aims to bridge the critical gap in physics-annotated 3D datasets. It is the first physics-grounded 3D dataset systematically annotated across five foundational dimensions: **absolute scale**, **material**, **affordance**, **kinematics**, and **function description**.

## Dataset Details

🎉 Our paper has been accepted to **NeurIPS 2025 (Spotlight)**

🎉 We have released the code for converting our JSON files to URDF at: [urdf_gen.py](https://github.com/ziangcao0312/PhysX).

### Dataset Sources

- **Repository:** [PhysX-3D](https://github.com/ziangcao0312/PhysX)
- **Project page:** [PhysX-3D: Physical-Grounded 3D Asset Generation](https://physx-3d.github.io)
- **Demo video:** [Video](https://youtu.be/M5V_c0Duuy4)

## Dataset Structure

```
PhysX
--PhysXNet.zip
----finaljson
------103.json
------502.json
------...
----partseg
------103
--------img
----------0.png
----------1.png
----------...
--------objs
----------0.obj
----------1.obj
----------...
--PhysXNet-XL_bottle.zip
--PhysXNet-XL_knief.zip
...
```

The physical properties are included in the JSON file. It can be converted to URDF or XML files.

###### Example.json

```python
{
"object_name": "Folding Knife",
"category": "Tool",
"dimension": "20*3*2", # Physical scaling (cm)
"parts": [
{
"label": 0,
"name": "Blade",
"material": "Stainless Steel",
"density": "7.8 g/cm^3",
"priority_rank": 2, # Affordance rank
"Basic_description": "xxx",
"Functional_description": "xxx",
"Movement_description": "xxx",
"Young's Modulus (GPa)": xx,
"Poisson's Ratio": xx
},
{
"label": 1,
"name": "Handle",
"material": "Plastic",
"density": "1.2 g/cm^3",
"priority_rank": 1,
"Basic_description": "xxx",
"Functional_description": "xxx",
"Movement_description": "xxx",
"Young's Modulus (GPa)": xx,
"Poisson's Ratio": xx
}
],
"group_info": {
"0": [ # basement group index
1 # label of the part
],
"1": [ # child group index
[
0 # moveable parts in child group
],
"0", # parent group index
[
1, # rotation/movement direction x coordinate
0, # rotation/movement direction y coordinate
0, # rotation/movement direction z coordinate
0.0, # Revolute/Hinge location x coordinate
0.3, # Revolute/Hinge location y coordinate
-0.0, # Revolute/Hinge location z coordinate
0.0, # rotation/movement min range
1.0 # rotation/movement max range
],
"C" # Kinematic type (A,B,C,CB,D,E)
]
}
}
```

### Kinematic Details

**Rotation range:**

Rotation range = rotation angle / 180.

(Rotation range) [-1, 1] * 180° → (Rotation angle) [-180°, 180°].

**Movement range:**

Movement range = movement length in 3D coordinates.

(Movement range) [-1, 1] * Physical scaling → (Movement length) [-10cm, 10cm].

**Kinematic type:**

A. No movement constraints *(water in a bottle)*

B. Prismatic joints *(drawer)*

C. Revolute joints (*door*)

CB. Prismatic & Revolute joints (lid of the bottle)

D. Hinge joint (*a hose in a shower system*)

E. Rigid joint.

**Note:** For CB, there are more kinematic parameters.

```python
"group_info": {
"0": [ # basement group index
1 # label of the part
],
"1": [ # child group index
[
0 # moveable parts in child group
],
"0", # parent group index
[
1, # rotation direction x coordinate
0, # rotation direction y coordinate
0, # rotation direction z coordinate
0.0, # Revolute location x coordinate
0.3, # Revolute location y coordinate
-0.0, # Revolute location z coordinate
0.0, # rotation min range
1.0 # rotation max range
1, # movement direction x coordinate
0, # movement direction y coordinate
0, # movement direction z coordinate
0.0, #
0.3, #
-0.0, #
0.0, # movement min range
1.0 # movement max range
],
"CB" # Kinematic type (A,B,C,CB,D,E)
]
}
```

If you find our dataset useful for your work, please cite:

```
@article{cao2025physx,
title={PhysX: Physical-Grounded 3D Asset Generation},
author={Cao, Ziang and Chen, Zhaoxi and Pan, Liang and Liu, Ziwei},
journal={arXiv preprint arXiv:2507.12465},
year={2025}
}
```

### Acknowledgement

PhysXNet and PhysXNet-XL are based on [PartNet](https://huggingface.co/datasets/ShapeNet/PartNet-archive). We would like to express our sincere thanks to the contributors.

### License
If you use PhysXNet and PhysXNet-XL, you agree to abide by the [ShapeNet terms of use](https://shapenet.org/terms). You are only allowed to redistribute the data to your research associates and colleagues provided that they first agree to be bound by these terms and conditions.

# PhysXNet 与 PhysXNet-XL

<p align="left"><a href="https://arxiv.org/abs/2507.12465"><img src='https://img.shields.io/badge/arXiv-Paper-red?logo=arxiv&logoColor=white' alt='arXiv论文'></a>
<a href='https://huggingface.co/papers/2507.12465'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Paper-blue'></a>
<a href='https://physx-3d.github.io/'><img src='https://img.shields.io/badge/Project_Page-Website-green?logo=homepage&logoColor=white' alt='项目页面'></a>
<a href='https://youtu.be/M5V_c0Duuy4'><img src='https://img.shields.io/youtube/views/M5V_c0Duuy4' alt='YouTube观看量'></a>

本数据集旨在填补带物理标注的3D数据集领域的关键空白，是首个从**绝对尺度（absolute scale）**、**材质（material）**、**功能可供性（affordance）**、**运动学（kinematics）**与**功能描述（function description）**五个基础维度进行系统性物理锚定标注的3D数据集。

## 数据集详情

🎉 本文已被**NeurIPS 2025（Spotlight论文）**收录

🎉 我们已在以下地址开源了将JSON文件转换为URDF（统一机器人描述格式，Unified Robot Description Format）的代码：[urdf_gen.py](https://github.com/ziangcao0312/PhysX)。

### 数据集来源

- **代码仓库**：[PhysX-3D](https://github.com/ziangcao0312/PhysX)
- **项目主页**：[PhysX-3D：物理锚定3D资产生成](https://physx-3d.github.io)
- **演示视频**：[视频](https://youtu.be/M5V_c0Duuy4)

## 数据集结构

PhysX
--PhysXNet.zip
----finaljson
------103.json
------502.json
------...
----partseg
------103
--------img
----------0.png
----------1.png
----------...
--------objs
----------0.obj
----------1.obj
----------...
--PhysXNet-XL_bottle.zip
--PhysXNet-XL_knief.zip
...

物理属性包含于JSON文件中，可将其转换为URDF或XML文件。

###### 示例文件Example.json

python
{
"object_name": "折叠刀",
"category": "工具",
"dimension": "20*3*2", # 物理尺度（单位：厘米）
"parts": [
{
"label": 0,
"name": "刀片",
"material": "不锈钢",
"density": "7.8 g/cm^3",
"priority_rank": 2, # 功能可供性优先级
"Basic_description": "xxx",
"Functional_description": "xxx",
"Movement_description": "xxx",
"Young's Modulus (GPa)": "xx",
"Poisson's Ratio": "xx"
},
{
"label": 1,
"name": "手柄",
"material": "塑料",
"density": "1.2 g/cm^3",
"priority_rank": 1,
"Basic_description": "xxx",
"Functional_description": "xxx",
"Movement_description": "xxx",
"Young's Modulus (GPa)": "xx",
"Poisson's Ratio": "xx"
}
],
"group_info": {
"0": [ # 基础组索引
1 # 部件标签
],
"1": [ # 子组索引
[
0 # 子组内可移动部件
],
"0", # 父组索引
[
1, # 旋转/运动方向X坐标
0, # 旋转/运动方向Y坐标
0, # 旋转/运动方向Z坐标
0.0, # 旋转副位置X坐标
0.3, # 旋转副位置Y坐标
-0.0, # 旋转副位置Z坐标
0.0, # 旋转/运动最小范围
1.0 # 旋转/运动最大范围
],
"C" # 运动学类型（A,B,C,CB,D,E）
]
}
}

### 运动学细节

**旋转范围：**

旋转范围 = 旋转角度 / 180。

(旋转范围) [-1, 1] × 180° → (旋转角度) [-180°, 180°]。

**运动范围：**

运动范围 = 三维坐标系中的运动长度。

(运动范围) [-1, 1] × 物理尺度 → (运动长度) [-10厘米, 10厘米]。

**运动学类型：**

A. 无运动约束 *(例如瓶内水体)*

B. 移动副关节 *(例如抽屉)*

C. 旋转副关节 *(例如门)*

CB. 移动副与旋转副复合关节 *(例如瓶塞)*

D. 铰链关节 *(例如淋浴系统中的软管)*

E. 刚性关节。

**注：** 对于CB类型，需包含更多运动学参数。

python
"group_info": {
"0": [ # 基础组索引
1 # 部件标签
],
"1": [ # 子组索引
[
0 # 子组内可移动部件
],
"0", # 父组索引
[
1, # 旋转方向X坐标
0, # 旋转方向Y坐标
0, # 旋转方向Z坐标
0.0, # 旋转副位置X坐标
0.3, # 旋转副位置Y坐标
-0.0, # 旋转副位置Z坐标
0.0, # 旋转最小范围
1.0, # 旋转最大范围
1, # 运动方向X坐标
0, # 运动方向Y坐标
0, # 运动方向Z坐标
0.0, # 运动副位置X坐标
0.3, # 运动副位置Y坐标
-0.0, # 运动副位置Z坐标
0.0, # 运动最小范围
1.0 # 运动最大范围
],
"CB" # 运动学类型（A,B,C,CB,D,E）
]
}

若您的研究工作中使用了本数据集，请引用以下文献：

@article{cao2025physx,
title={PhysX: Physical-Grounded 3D Asset Generation},
author={Cao, Ziang and Chen, Zhaoxi and Pan, Liang and Liu, Ziwei},
journal={arXiv preprint arXiv:2507.12465},
year={2025}
}

### 致谢

PhysXNet与PhysXNet-XL基于[PartNet](https://huggingface.co/datasets/ShapeNet/PartNet-archive)构建，谨此向所有贡献者致以诚挚谢意。

### 许可协议

若您使用PhysXNet与PhysXNet-XL，即表示您同意遵守[ShapeNet使用条款](https://shapenet.org/terms)。仅可将本数据集重新分发至您的科研合作者与同事，且前提是他们已同意受本条款约束。