genrobot2025/DAS-Sample-Data

DAS: Data Acquisition System

Enable embodied intelligence data acquisition to be as simple and natural as shooting a video.

---

📋 Contents

• 📦 How to Use the Dataset
• 📚 Dataset Structure
• License
• Contact

📦 How to Use the Dataset

Due to Hugging Face's file size limitation of 50GB per file, the dataset has been split into smaller parts.

📚 Dataset Structure

Purpose: Each HDF5 file corresponds to a single episode and encapsulates both observational data and actions. Below is the hierarchical structure of the HDF5 file:

xxx.h5
├── observations/
│   ├── cameras/
│   │   └── <camera_name_x> (Dataset)
│   ├── tactile/
│   │   └── <left or right> (Dataset)
│   ├── eef_pos (Dataset)
│   └── imu (Dataset)
├── label/
│   └── depthmap (Dataset)
└── action (Dataset) (actions mirror the eef_pos data)

💡 For Demo data, it has a frame rate of 10Hz (downsample from 30Hz to 10Hz). The length and width of the image data are downsampled to half the original size.

Groups and Datasets:

observations/

• cameras/
  • Description: Image data from camera.
  • Datasets:
    • Type: Dataset
    • Shape: (num_frames, height=xxx, width=xxx, channels=3) for mid fisheye camera, (num_frames, height=xxx, width=xxx, channels=3) for side wide camera
    • Data Type: uint8
    • Compression: gzip with compression level 4.
• tactile/
  • Description: Pressure data from tactile sensor.
  • Datasets:
    • Type: Dataset.
    • Shape: (num_frames, 96), row=12, col=8
    • Data Type: float32
    • Compression: gzip with compression level 4.
• eef_pos/
  • Type: Dataset.
  • Shape: (num_frames, 8)
  • Data Type: float32
  • Description: Position and orientation data for each timestep. We obtain high-precision positioning information based on SLAM technology.
  • Columns: [Pos_X, Pos_Y, Pos_Z, Q_X, Q_Y, Q_Z, Q_W, Gripper_width]
  • Compression: gzip with compression level 4.
• imu/
  • Type: Dataset.
  • Shape: (num_frames, 6)
  • Data Type: float32
  • Description: Angular Velocity and Linear Acceleration data from IMU sensor for each timestep. We align IMU and image data based on timestamp.
  • Columns: [AngularVel_X, AngularVel_Y, AngularVel_Z, LinearAcc_X, LinearAcc_Y, LinearAcc_Z]
  • Compression: gzip with compression level 4.

label/

• depthmap/
  • Type: Dataset
  • Shape: (num_frames, height=xxx, width=xxx) for mid fisheye camera, (num_frames, height=xxx, width=xxx) for side wide camera
  • Data Type: float32
  • Description: Depth map estimated from RGB image. At present, we have only released the depth map of one camera view.
  • Compression: gzip with compression level 4.

action/

• Type: Dataset
• Shape: (num_frames, 8)
• Data Type: float32
• Description: Stores action data corresponding to each timestep. Same to eef_pos.
• Columns: [Pos_X, Pos_Y, Pos_Z, Q_X, Q_Y, Q_Z, Q_W, Gripper_width]
• Compression: gzip with compression level 4.

License

This project is licensed under the MIT License.

Contact

For questions or feedback, please reach out to the info@genrobot.com or visit our website.