size_categories:
- 50G<n<100G
---

# Dataset Card for PRTS Post-Training Data

This dataset is released as part of the project:

**PRTS: A Primitive Reasoning and Tasking System via Contrastive Representations**

This dataset card describes the **PRTS Post-Training Data**, a high-quality robotic manipulation dataset designed for learning **primitive-level skills** and **long-horizon task execution**. The dataset covers both **dual-arm** and **single-arm** manipulation, including **fine manipulation** and **long-range operations**, and is curated in the **standard LeRobot format**.

It serves as a **high-quality benchmark** for robotic manipulation, supporting research in policy learning, task decomposition, and embodied intelligence.

---

## Dataset Details

### Dataset Description

This dataset consists of robotic task data in the **LeRobot format**, covering diverse manipulation scenarios including:

- **Dual-arm coordination**
- **Single-arm manipulation**
- **Long-horizon tasks**
- **Fine-grained manipulation**

The dataset contains approximately **50GB** of high-quality data collected from real-world robotic systems across **household**, **office**, and **industrial** environments.

Each trajectory captures detailed low-level control signals, including poses, joint states, and gripper actions, enabling learning of **primitive skills**, **action representations**, and **task-level policies**.

- **Curated by:** TeleEmbodied AI Team  
- **License:** MIT License  

---

## Uses

### Direct Use

This dataset is intended for:

- Learning **dual-arm** and **single-arm** manipulation policies  
- Training **primitive-level action representations**  
- **Long-horizon task learning** and decomposition  
- **Fine manipulation** skill learning  
- Benchmarking **robotic manipulation models** in real-world settings  

---

## Dataset Structure

The dataset follows the **standard LeRobot format**, widely used for robotic learning and policy training.

Each task consists of structured trajectories with frame-wise robot states:

- **`task_name`**: Task name (e.g., "Pick and Place Object")  
- **`task_id`**: Unique task identifier  
- **`task_description`**: Description of the task  
- **`frames`**: Sequence of frames, each containing:
  - **`t`**: Timestamp  
  - **`pose`**: 6D end-effector pose (dual-arm or single-arm)  
  - **`joints`**: Joint angles  
  - **`gripper`**: Gripper state (open/close)  

Example:

```json
{
    "task_name": "Pick and Place Object",
    "task_id": "001",
    "task_description": "Pick up a cup and place it on the table",
    "frames": [
        {
            "t": 0.0,
            "pose": [0.5, 0.5, 0.2, 0, 0, 0, 1],
            "joints": [0.3, -0.5, 0.8, 0.1, -0.2, 0.9, 0.2, -0.4],
            "gripper": [1, 0]
        }
    ]
}

```

## Citation
This dataset is released as part of the PRTS project:

PRTS: A Primitive Reasoning and Tasking System via Contrastive Representations

If you use this dataset, please cite:
@article{zhang2026prts,
    title={PRTS: A Primitive Reasoning and Tasking System via Contrastive Representations},
    author={Zhang, Yang and Zhao, Jiangyuan and Fan, Chenyou and Yan, Fangzheng and Li, Tian and Wu, Xuaner and Weng, Qizhen and Li, Xiu and Zhang, Weinan and Zhang, Chi and Bai, Chenjia and Li, Xuelong},
    year={2026},
    url={https://rhodes-team-prts.github.io}
}