README.md

---
license: cc-by-nc-sa-4.0
---

# IROS-2025-Challenge-Manip
# Dataset Summary 📖 

This dataset contains the **IROS Challenge - Manipulation Track** benchmark, organized into **pretrain**, **train**, and **validation** splits.

* **Pretrain split**: \~20,000 single pick-and-place trajectories, packaged into tar files (each containing \~1,000 trajectories).
* **Train split**: task-specific demonstrations, with \~100 trajectories provided per task.
* **Validation split**: includes the test-time scenes and object assets in **USD format**.

Each trajectory in the pretrain and train splits contains:

* **Multi-view video** recordings (three perspectives: head-mounted camera and two wrist cameras)
* **Robot states** (joint positions, gripper states, etc.)
* **Actions** corresponding to the task execution

This dataset is designed to support **pretraining, task-specific fine-tuning, and evaluation** for robotic manipulation in the IROS Challenge setting.

# Get started 🔥    
## Download the Dataset 
To download the full dataset, you can use the following code. If you encounter any issues, please refer to the official Hugging Face documentation.
```python
from huggingface_hub import snapshot_download

dataset_path = snapshot_download("InternRobotics/IROS-2025-Challenge-Manip", repo_type="dataset")
```

Please execute this Python file to post-process the validation set.  
```bash
cd IROS-2025-Challenge-Manip
python dataset_post_processing.py validation
````


## Unzip the pretrain dataset

```bash
cd pretrain
for i in {1..20}; do
    echo "Extracting $i.tar.gz ..."
    tar -xzf "$i.tar.gz"
done
```


## Dataset Structure

### pretrain Folder hierarchy
```
pretrain
├── 1.tar.gz
│   └── 1/
│       ├── data/
│       ├── meta/
│       └── videos/
├── 2.tar.gz
│   └── 2/
│       ├── data/
│       ├── meta/
│       └── videos/
...
├── 20.tar.gz
    └── 20/
        ├── data/
        ├── meta/
        └── videos/

```
        

### train Folder hierarchy
```
train
├── collect_three_glues
│   ├── data/
│   ├── meta/
│   └── videos/
├── collect_two_alarm_clocks/
├── collect_two_shoes/
├── gather_three_teaboxes/
├── make_sandwich/
├── oil_painting_recognition/
├── organize_colorful_cups/
├── purchase_gift_box/
├── put_drink_on_basket/
└── sort_waste/

```

### validation Folder hierarchy
```
validation
├── IROS_C_V3_Aloha_seen
│   ├── collect_three_glues
│   │   ├── 000
│   │   │   ├── meta_info.pkl
│   │   │   ├── scene.usd
│   │   │   └── SubUSDs -> ../SubUSDs
│   │   ├── 001/
│   │   ├── 002/
│   │   ├── 003/
│   │   ├── 004/
│   │   ├── 005/
│   │   ├── 006/
│   │   ├── 007/
│   │   ├── 008/
│   │   ├── 009/
│   │   └── SubUSDs
│   │       ├── materials/
│   │       └── textures/
│   ├── collect_two_alarm_clocks/
│   ├── collect_two_shoes/
│   ├── gather_three_teaboxes/
│   ├── make_sandwich/
│   ├── oil_painting_recognition/
│   ├── organize_colorful_cups/
│   ├── purchase_gift_box/
│   ├── put_drink_on_basket/
│   └── sort_waste/
└── IROS_C_V3_Aloha_unseen
    ├── collect_three_glues/
    ├── collect_two_alarm_clocks/
    ├── collect_two_shoes/
    ├── gather_three_teaboxes/
    ├── make_sandwich/
    ├── oil_painting_recognition/
    ├── organize_colorful_cups/
    ├── purchase_gift_box/
    ├── put_drink_on_basket/
    └── sort_waste/

```

# License and Citation
All the data and code within this repo are under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/). Please consider citing our project if it helps your research.

```BibTeX
@misc{contributors2025internroboticsrepo,
  title={IROS-2025-Challenge-Manip Colosseum},
  author={IROS-2025-Challenge-Manip Colosseum contributors},
  howpublished={\url{https://github.com/internrobotics/IROS-2025-Challenge-Manip}},
  year={2025}
}
```