README.md

---
language:
  - en
  - zh
tags:
  - robotics
  - manipulation
  - vla
  - trajectory-data
  - multimodal
  - vision-language-action
license: other
task_categories:
  - robotics
  - reinforcement-learning
multimodal: vision+language+action
dataset_info:
  features:
    - name: rgb_images
      dtype: image
      description: Multi-view RGB images
    - name: slam_poses
      sequence: float32
      description: SLAM pose trajectories
    - name: vive_poses
      sequence: float32
      description: Vive tracking system poses
    - name: point_clouds
      sequence: float32
      description: Time-of-Flight point cloud data
    - name: clamp_data
      sequence: float32
      description: Clamp sensor readings
    - name: merged_trajectory
      sequence: float32
      description: Fused trajectory data
  configs:
    - config_name: default
      data_files: "**/*"
---

<div align="center">
FastUMI Pro Dataset
</div>
 
<div align="center">

![FastUMI](https://img.shields.io/badge/FastUMI-Pro-brightgreen)
![Dataset](https://img.shields.io/badge/Dataset-150K-blue)
![VLA](https://img.shields.io/badge/VLA-Ready-orange)

**Enterprise-grade Robotic Manipulation Dataset for Universal Manipulation Interface**

[Project Homepage](https://fastumi.com/pro/) | [FastUMI Home](https://fastumi.com) | [Example Data](https://huggingface.co/datasets/FastUMIPro/example_data_fastumi_pro_raw)

</div>

## 📖 Overview

FastUMI (Fast Universal Manipulation Interface) is a dataset and interface framework for general-purpose robotic manipulation tasks, designed to support hardware-agnostic, scalable, and efficient data collection and model training.

The project provides:
- Physical prototype systems
- Complete data collection codebase
- Standardized data formats and utilities
- Tools for real-world manipulation learning research

## 🚀 Features

### FastUMI Pro Enhancements
- ✅ **Higher precision trajectory data**
- ✅ **Diverse embodiment support** for true "one-brain-multiple-forms"
- ✅ **Enterprise-ready** pipeline and full-link data processing

### FastUMI-150K
- ~150,000 real-world manipulation trajectories
- Used by research partners for large-scale VLA (Vision-Language-Action) model training
- Demonstrated significant multi-task generalization capabilities

## 📊 Model Performance


**VLA Model Results**: [TBD]

## 🛠️ Toolchain

### Core Tools

| Tool | Description | Link |
|------|-------------|------|
| **Single-Arm Demo Replay** | Single-arm data replay code | [GitHub](https://github.com/Loki-Lu/FastUMI_replay_singleARM) |
| **Dual-Arm Demo Replay** | Dual-arm data replay code | [GitHub](https://github.com/Loki-Lu/FastUMI_replay_dualARM) |
| **Hardware SDK** | FastUMI hardware development kit | [GitHub](https://github.com/FastUMIRobotics/FastUMI_Hardware_SDK) |
| **Monitor Tool** | Real-time device monitoring | [GitHub](https://github.com/FastUMIRobotics/FastUMI_Monitor_Tool) |
| **Data Collection** | Data collection utilities | [GitHub](https://github.com/FastUMIRobotics/FastUMI_Data_Collection) |

### Research & Applications
- **Paper**: [MLM: Learning Multi-task Loco-Manipulation Whole-Body Control for Quadruped Robot with Arm](https://arxiv.org/abs/2508.10538)
- **Tutorial**: PI0 (FastUMI Data Lightweight Adaptation, Version V0) Full Pipeline

## 📥 Data Download

### Example Dataset
```bash
# Direct download (may be slow in some regions)
huggingface-cli download FastUMIPro/example_data_fastumi_pro_raw --repo-type dataset --local-dir ~/fastumi_data/
```

Mirror Download (Recommended)
```bash
# Set mirror endpoint
export HF_ENDPOINT=https://hf-mirror.com
```

# Download via mirror
huggingface-cli download --repo-type dataset --resume-download FastUMIPro/example_data_fastumi_pro_raw --local-dir ~/fastumi_data/
📁 Data Structure
Each session represents an independent operation "episode" containing observation data and action sequences.
```
Directory Structure
text
session_001/
└── device_label_xv_serial/
    └── session_timestamp/
        ├── RGB_Images/
        │   ├── timestamps.csv
        │   └── Frames/
        │       ├── frame_000001.jpg
        │       └── ...
        ├── SLAM_Poses/
        │   └── slam_raw.txt
        ├── Vive_Poses/
        │   └── vive_data_tum.txt
        ├── ToF_PointClouds/
        │   ├── timestamps.csv
        │   └── PointClouds/
        │       └── pointcloud_000001.pcd
        ├── Clamp_Data/
        │   └── clamp_data_tum.txt
        └── Merged_Trajectory/
            ├── merged_trajectory.txt
            └── merge_stats.txt
```

## Data Specifications

| Data Type | Path | Shape| Type | Description |
| :--- | :--- | :--- | :--- | :--- |
| **RGB Images** | `session_XXX/RGB_Images/Video.MP4` | `(frames, 1080, 1920, 3)`| `uint8`| Camera video data, 60 FPS  |
| **SLAM Poses** | `session_XXX/SLAM_Poses/slam_raw.txt` | `(timestamps, 7)`| `float` | UMI end-effector poses |
| **Vive Poses** | `session_XXX/Vive_Poses/vive_data_tum.txt` | `(timestamps, 7)`| `float` | Vive base station poses |
| **ToF PointClouds** | `session_XXX/PointClouds/pointcloud_...pcd` | `pcd format` | Time-of-Flight point cloud data |
| **Clamp Data** | `session_XXX/Clamp_Data/clamp_data_tum.txt` | `(timestamps, 1)`| `float` | Gripper spacing (mm) |
| **Merged Trajectory** | `session_XXX/Merged_Trajectory/merged_trajectory.txt` | `(timestamps, 7)`| `float` | Fused trajectory (Vive/UMI based on velocity) |

### Pose Data Format

All pose data (SLAM, Vive, Merged) follow the same format:
| Column Name | Description |
| :--- | :--- |
| **Timestamp** | Unix timestamp of the trajectory data |
| **Pos X** | X-coordinate of position (meters) |
| **Pos Y** | Y-coordinate of position (meters) |
| **Pos Z** | Z-coordinate of position (meters) |
| **Q_X** | X-component of orientation quaternion |
| **Q_Y** | Y-component of orientation quaternion |
| **Q_Z** | Z-component of orientation quaternion |
| **Q_W** | W-component of orientation quaternion |
## 🔄 Data Conversion
[TBD - Data conversion methods will be added here]

## 🤝 Collaboration
FastUMI Pro dataset is available for research collaboration. The full FastUMI-150K dataset has been provided to partner research teams for large-scale model training.

## 📞 Contact
For questions or suggestions, please contact the development team
Lead: Ding Yan
Email: [email protected]
WeChat: Duke_dingyan