README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- robotics
tags:
- LeRobot
configs:
- config_name: default
  data_files: FlattenFold/base/data/chunk-000/episode_000000.parquet
---

# χ₀ (KAI0)

<div align="center\">
  <a href="https://github.com/OpenDriveLab/KAI0">
    <img src="https://img.shields.io/badge/GitHub-grey?logo=GitHub" alt="GitHub Badge">
  </a>
  <a href="https://huggingface.co/OpenDriveLab-org/Kai0">
    <img src="https://img.shields.io/badge/Model-grey?logo=Model" alt="Model Badge">
  </a>
  <a href="https://mmlab.hk/research/kai0">
    <img src="https://img.shields.io/badge/Research_Blog-grey?style=flat" alt="Research Blog Badge">
  </a>
  <a href="https://huggingface.co/papers/2602.09021">
    <img src="https://img.shields.io/badge/Paper-blue?logo=arxiv" alt="Paper Badge">
  </a>
</div>

χ₀ (**kai0**) is a resource-efficient framework for achieving production-level robustness in robotic manipulation by taming distributional inconsistencies. It enables long-horizon garment manipulation tasks such as flattening, folding, and hanging using dual-arm robots.

# TODO
- [ ] The advantage label will be coming soon.

## Contents
- [About the Dataset](#about-the-dataset)
- [Load the Dataset](#get-started)
- [Download the Dataset](#download-the-dataset)
- [Dataset Structure](#dataset-structure)
    - [Folder hierarchy](#folder-hierarchy)
    - [Details](#details)
- [License and Citation](#license-and-citation)

## [About the Dataset](#contents)
- **~181 hours** real world scenarios 
- **Main Tasks**
    - ***FlattenFold***  
      - Single task
      - Initial state: T-shirts are randomly tossed onto the table, presenting random crumpled configurations
      - Manipulation task: Operate the robotic arm to unfold the garment, then fold it
    - ***HangCloth***
      - Single task
      - Initial state: Hanger is randomly placed, garment is randomly positioned on the table
      - Manipulation task: Operate the robotic arm to thread the hanger through the garment, then hang it on the rod
    - ***TeeShirtSort***
      - Garment classification and arrangement task
      - Initial state: Randomly pick a garment from the laundry basket
      - Classification: Determine whether the garment is a T-shirt or a dress shirt
      - Manipulation task:
        - If it is a T-shirt, fold the garment
        - If it is a dress shirt, expose the collar, then push it to one side of the table
- **Count of the dataset** 

    | Task | Base (episodes count/hours) | DAgger (episodes count/hours) | Total(episodes count/hours) |
    |------|-----------------------------|-------------------------------|-----------------------------|
    | FlattenFold | 3,055/~42 hours             | 3,457/ ~13 Hours              | 6,512 /~55 hours            | 
    | HangCloth | 6954/~61 hours              | 686/~12 hours                 | 7640/~73 hours              |
    | TeeShirtSort | 5988/~31 hours              | 769/~22 hours                 | 6757/~53 hours              |
    | **Total** | **15,997/~134 hours**       | **4,912/~47 hours**           | **20,909/~181 hours**       |

## [Load the dataset](#contents)
- This dataset was created using [LeRobot](https://github.com/huggingface/lerobot)
- The dataset's version is LeRobotDataset v2.1

### For LeRobot version < 0.4.0
Choose the appropriate import based on your version:

| Version                | Import Path |
|------------------------|-------------|
| `<= 0.1.0`             | `from lerobot.common.datasets.lerobot_dataset import LeRobotDataset` |
| `> 0.1.0` and `< 0.4.0` | `from lerobot.datasets.lerobot_dataset import LeRobotDataset` |

```python
# For version <= 0.1.0
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

# For version > 0.1.0 and < 0.4.0
from lerobot.datasets.lerobot_dataset import LeRobotDataset

# Load the dataset
dataset = LeRobotDataset(repo_id='OpenDriveLab-org/kai0')
```

### For LeRobot version >= 0.4.0

You need to migrate the dataset from v2.1 to v3.0 first. See the official documentation: [Migrate the dataset from v2.1 to v3.0](https://huggingface.co/docs/lerobot/lerobot-dataset-v3)

```bash
python -m lerobot.datasets.v30.convert_dataset_v21_to_v30 --repo-id=OpenDriveLab-org/kai0
```

## [Download the Dataset](#contents)
### Python Script
```python
from huggingface_hub import hf_hub_download, snapshot_download
from datasets import load_dataset

# Download a single file
hf_hub_download(
    repo_id="OpenDriveLab-org/kai0", 
    filename="episodes.jsonl",
    subfolder="meta",
    repo_type="dataset",
    local_dir="where/you/want/to/save"
)

# Download a specific folder
snapshot_download(
    repo_id="OpenDriveLab-org/kai0", 
    local_dir="/where/you/want/to/save",
    repo_type="dataset",
    allow_patterns=["data/*"]
)

# Load the entire dataset
dataset = load_dataset("OpenDriveLab-org/kai0") 
```

### Terminal (CLI)
```bash
# Download a single file
hf download OpenDriveLab-org/kai0 \
    --include "meta/info.json" \
    --repo-type dataset \
    --local-dir "/where/you/want/to/save"

# Download a specific folder
hf download OpenDriveLab-org/kai0 \
    --repo-type dataset \
    --include "meta/*" \
    --local-dir "/where/you/want/to/save"

# Download the entire dataset
hf download OpenDriveLab-org/kai0 \
    --repo-type dataset \
    --local-dir "/where/you/want/to/save"
```

## [Dataset Structure](#contents)

### [Folder hierarchy](#contents)
Under each task directory, data is partitioned into two subsets: base and dagger.
- base contains original demonstration trajectories.
- dagger contains on-policy recovery trajectories collected via iterative DAgger.

```text
Kai0-data/
├── FlattenFold/
│   ├── base/
│   │   ├── data/
│   │   ├── videos/
│   │   └── meta/
│   └── dagger/
├── HangCloth/
│   ├── base/
│   └── dagger/
├── TeeShirtSort/
│   ├── base/
│   └── dagger/
└── README.md
```

<a id='Details'></a>
### [Details](#contents)
#### info.json
The basic structure of `info.json` includes metadata about robot types, frames, tasks, and data features like camera perspectives (`top_head`, `hand_left`, `hand_right`).

#### [Parquet file format](#contents)
| Field Name | shape | Meaning |
|------------|-------------|-------------|
| observation.state | [N, 14] |left `[:, :6]`, right `[:, 7:13]`, joint angle<br> left`[:, 6]`, right `[:, 13]` , gripper open range|
| action | [N, 14]  |left `[:, :6]`, right `[:, 7:13]`, joint angle<br>left`[:, 6]`, right `[:, 13]` , gripper open range |
| timestamp | [N, 1] | Time elapsed since the start of the episode (in seconds) |
| frame_index | [N, 1] | Index of this frame within the current episode (0-indexed) |
| episode_index | [N, 1] | Index of the episode this frame belongs to |
| index | [N, 1] | Global unique index across all frames in the dataset |
| task_index | [N, 1] | Index identifying the task type being performed |

## License and Citation
The data and checkpoints are licensed under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).

```bibtex
@article{sima2026kai0,
  title={$\chi_{0}$: Resource-Aware Robust Manipulation via Taming Distributional Inconsistencies},
  author={Yu, Checheng and Sima, Chonghao and Jiang, Gangcheng and Zhang, Hai and Mai, Haoguang and Li, Hongyang and Wang, Huijie and Chen, Jin and Wu, Kaiyang and Chen, Li and Zhao, Lirui and Shi, Modi and Luo, Ping and Bu, Qingwen and Peng, Shijia and Li, Tianyu and Yuan, Yibo},
  journal={arXiv preprint arXiv:2602.09021},
  year={2026}
}
```