OmniRetarget_Dataset

# OmniRetarget Dataset: Humanoid Loco-Manipulation & Scene Interaction

[Paper](https://huggingface.co/papers/2509.26633) | [Project Page](https://omniretarget.github.io)

This dataset contains motion trajectories of a G1 humanoid robot interacting with objects and complex terrains. It was generated by **[OMNIRETARGET](https://omniretarget.github.io/)**, an interaction-preserving data generation engine that produces high-quality, kinematically feasible trajectories free of common artifacts like foot-skating and penetration.

<div align="center">
<video autoplay loop muted controls width="70%">
<source src="https://huggingface.co/datasets/omniretarget/OmniRetarget_Dataset/resolve/main/assets/teaser.mp4" type="video/mp4">
</video>
</div>

## Dataset Structure

Due to licensing restrictions, we cannot release the retargeted [LAFAN1](https://github.com/ubisoft/ubisoft-laforge-animation-dataset) dataset. However, we will open-source our retargeting code so that users can retarget the data themselves.

| Subset | Description | Source Data | Duration (hours) |
| ------------------------ | --------------------------------------------------- | --------------- | ---------------- |
| `robot-object/` | Motions of the robot carrying objects. | OMOMO | 3.0 |
| `robot-terrain/` | Dynamic motions of the robot climbing challenging terrains. | In-house MoCap | 0.5 |
| `robot-object-terrain/` | Motions involving both object and terrain interaction. | In-house MoCap | 0.5 |
| **Total** | | | **4.0** |

Additionally, the `models/` directory contains all the necessary URDF, SDF, and OBJ assets for visualization. These are not required for loading or training with the trajectory data.

## Data Format

Each `.npz` file contains a single trajectory with two keys:
- **`fps`**: Frames per second.
- **`qpos`**: A NumPy array of shape `[T, D]` representing the system state over `T` timesteps. The vector is structured as follows:
- **Robot Pose (36D):**
- Floating Base `[qw, qx, qy, qz, x, y, z]` (7D)
- Joint Positions (29D)
- **Object Pose (7D, optional):**
- `[qw, qx, qy, qz, x, y, z]`
- The total dimension `D` is 36 for motions without an object, and 43 with an object.

## Quick Usage
```bash
# Clone the repository, install dependencies
git lfs install
git clone https://huggingface.co/datasets/omniretarget/OmniRetarget_Dataset
pip install numpy
```
``` bash
# Load data
import glob, numpy as np
paths = glob.glob("robot-object/*.npz")
with np.load(paths[0]) as data:
qpos = data["qpos"] # (T, D)
fps = float(data["fps"]) # e.g., 30.0
```

## Visualize (optional)
A `visualize.py` script using Drake and Meshcat is provided.
```bash
# Install dependencies
pip install drake

# Set `task` inside the script: "object" | "terrain" | "object-terrain"
python visualize.py
```

## Citation
https://omniretarget.github.io/
```bibtex
@inproceedings{Yang2025OmniRetarget,
title={OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction},
author={Yang, Lujie and Huang, Xiaoyu and Wu, Zhen and Kanazawa, Angjoo and Abbeel, Pieter and Sferrazza, Carmelo and Liu, C. Karen and Duan, Rocky and Shi, Guanya},
booktitle={arXiv},
year={2025}
}
```

# OmniRetarget 数据集：类人机器人全身运动操作与场景交互

[Paper](https://huggingface.co/papers/2509.26633) | [Project Page](https://omniretarget.github.io)

本数据集包含G1类人机器人与物体及复杂地形交互的运动轨迹。其由**OMNIRETARGET**生成——这是一款保留交互性的数据生成引擎，可生成高质量、运动学可行的轨迹，有效规避足滑、穿透等常见伪影问题。

<div align="center">
<video autoplay loop muted controls width="70%">
<source src="https://huggingface.co/datasets/omniretarget/OmniRetarget_Dataset/resolve/main/assets/teaser.mp4" type="video/mp4">
</video>
</div>

## 数据集结构

由于授权限制，我们无法发布经过重定向的LAFAN1数据集。不过我们将开源重定向代码，以便用户自行对数据进行重定向。

| 子集名称 | 描述 | 源数据 | 时长（小时） |
| ------------------------ | --------------------------------------------------- | --------------- | ---------------- |
| `robot-object/` | 机器人携载物体的运动 | OMOMO | 3.0 |
| `robot-terrain/` | 机器人攀爬复杂地形的动态运动 | 内部动作捕捉 | 0.5 |
| `robot-object-terrain/` | 同时涉及物体与地形交互的运动 | 内部动作捕捉 | 0.5 |
| **总计** | | | **4.0** |

此外，`models/` 目录包含可视化所需的全部URDF、SDF及OBJ资源。这些资源并非加载轨迹数据或开展训练的必需项。

## 数据格式

每个 `.npz` 文件包含单条轨迹，内含两个键值：
- **`fps`**：每秒帧率。
- **`qpos`**：形状为 `[T, D]` 的NumPy数组，代表`T`个时间步下的系统状态。该向量的结构如下：
- **机器人位姿（36维）：**
- 浮动基座 `[qw, qx, qy, qz, x, y, z]`（7维）
- 关节位置（29维）
- **物体位姿（7维，可选）：**
- `[qw, qx, qy, qz, x, y, z]`
- 总维度`D`：无物体的运动为36维，含物体的运动为43维。

## 快速使用

bash
# 克隆仓库并安装依赖项
git lfs install
git clone https://huggingface.co/datasets/omniretarget/OmniRetarget_Dataset
pip install numpy

bash
# 加载数据
import glob, numpy as np
paths = glob.glob("robot-object/*.npz")
with np.load(paths[0]) as data:
qpos = data["qpos"] # (T, D)
fps = float(data["fps"]) # 例如：30.0

## 可视化（可选）

我们提供了基于Drake和Meshcat的`visualize.py`脚本。
bash
# 安装依赖项
pip install drake

# 在脚本中设置`task`："object" | "terrain" | "object-terrain"
python visualize.py

## 引用

https://omniretarget.github.io/
bibtex
@inproceedings{Yang2025OmniRetarget,
title={OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction},
author={Yang, Lujie and Huang, Xiaoyu and Wu, Zhen and Kanazawa, Angjoo and Abbeel, Pieter and Sferrazza, Carmelo and Liu, C. Karen and Duan, Rocky and Shi, Guanya},
booktitle={arXiv},
year={2025}
}