pthangeda/scooping-dataset

Scooping Dataset

This dataset contains 6,700 samples collected over 67 terrains for the task of manipulating granular materials using a robotic arm. The dataset is designed to facilitate research in robotic manipulation, machine learning, and related fields. Each sample includes:

• Terrain Metadata: Terrain ID, composition, and materials used.
• RGB Image: An RGB image of the terrain before scooping.
• Depth Image: Depth data corresponding to the terrain saved as a 16-bit PNG image.
• Depth Normalization Parameters: depth_min and depth_max to reconstruct original depth values.
• F/T Sensor Data: Force/Torque sensor data captured during the scooping action, saved as a CSV file.
• Action Parameters: Scoop location, yaw angle, scoop depth, and stiffness.
• Outcome: Volume of material scooped.

Dataset Structure

Features

• terrain_id (int32): Terrain identifier (1-67).
• composition (string): Terrain composition (single, partition, mixture, layers).
• material_1 (string): First material used in the terrain.
• material_2 (string): Second material used in the terrain (empty string if not applicable).
• material_3 (string): Third material used in the terrain (empty string if not applicable).
• sample_index (int32): Sample index within the terrain (1-100).
• rgb_image (Image): RGB image of the terrain before scooping.
• depth_image (Image): Depth data as a 16-bit PNG image.
• depth_min (float32): Minimum depth value used for normalization.
• depth_max (float32): Maximum depth value used for normalization.
• ft_csv_path (string): File path to the F/T sensor data (CSV file).
• pixel_x (float32): X-coordinate of the scoop location in the image.
• pixel_y (float32): Y-coordinate of the scoop location in the image.
• yaw (float32): Yaw angle of the end-effector (in radians).
• scoop_depth (float32): Depth of the scoop (in meters).
• stiffness (float32): Stiffness of the controller during the scoop action.
• scooped_volume (float32): Volume of material scooped (in cubic meters).

Data Files

The dataset is organized with the following structure:

scooping_dataset/
├── scooping_dataset.arrow
├── dataset_info.json
├── rgb_images/
│   ├── terrain_1_sample_1_rgb.png
│   ├── terrain_1_sample_2_rgb.png
│   └── ...
├── depth_images/
│   ├── terrain_1_sample_1_depth.png
│   ├── terrain_1_sample_2_depth.png
│   └── ...
├── ft_data/
│   ├── terrain_1_sample_1_ft.csv
│   ├── terrain_1_sample_2_ft.csv
│   └── ...

Usage

To use this dataset, you can load it using the Hugging Face datasets library:

from datasets import load_from_disk
import numpy as np
from PIL import Image as PILImage
import pandas as pd

# Load the dataset
dataset = load_from_disk('path_to_dataset/scooping_dataset')

# Access a sample
sample = dataset[0]

# Load RGB image
rgb_image = sample['rgb_image']  # PIL Image
rgb_image.show()

# Load depth image and reconstruct original depth values
depth_image = sample['depth_image']  # PIL Image
depth_array = np.array(depth_image).astype(np.float32)
depth_normalized = depth_array / 65535  # Normalize back to [0, 1]
depth_min = sample['depth_min']
depth_max = sample['depth_max']
original_depth = depth_normalized * (depth_max - depth_min) + depth_min

# Load F/T sensor data
ft_data = pd.read_csv(sample['ft_csv_path'], header=None).values

# Access action parameters
pixel_x = sample['pixel_x']
pixel_y = sample['pixel_y']
yaw = sample['yaw']
scoop_depth = sample['scoop_depth']
stiffness = sample['stiffness']
scooped_volume = sample['scooped_volume']

Loading All Data for a Specific Terrain ID

If you want to load all samples corresponding to a specific terrain ID, you can filter the dataset using the filter method. Here's how you can do it:

# Specify the terrain ID you're interested in
terrain_id_of_interest = 10  # Replace with the desired terrain ID (1-67)

# Filter the dataset to include only samples from the specified terrain
terrain_samples = dataset.filter(lambda sample: sample['terrain_id'] == terrain_id_of_interest)

print(f"Number of samples for terrain {terrain_id_of_interest}: {len(terrain_samples)}")
for sample in terrain_samples:
    # Load RGB image
    rgb_image = sample['rgb_image']  # PIL Image

    # Load and reconstruct depth image
    depth_image = sample['depth_image']  # PIL Image
    depth_array = np.array(depth_image).astype(np.float32)
    depth_normalized = depth_array / 65535
    depth_min = sample['depth_min']
    depth_max = sample['depth_max']
    original_depth = depth_normalized * (depth_max - depth_min) + depth_min

    # Load F/T sensor data
    ft_data = pd.read_csv(sample['ft_csv_path'], header=None).values

    # Access action parameters
    pixel_x = sample['pixel_x']
    pixel_y = sample['pixel_y']
    yaw = sample['yaw']
    scoop_depth = sample['scoop_depth']
    stiffness = sample['stiffness']
    scooped_volume = sample['scooped_volume']

    # Perform your analysis or processing here
    # Example: Print scooped volume for each sample
    print(f"Sample {sample['sample_index']}: Scooped volume = {scooped_volume} cubic meters")

Dataset Creation

The dataset was created using the following steps:

1. Data Collection: 6,700 samples were collected using a UR5e robotic arm equipped with a scooping end-effector and an overhead RealSense L515 camera to capture RGB-D images.

2. Terrain Preparation: 67 terrains were prepared using combinations of 12 different materials and 4 types of compositions. Each terrain represents a unique task.

3. Action Execution: For each terrain, 100 scooping actions were performed with varying action parameters (scoop location, yaw, depth, and stiffness).

4. Data Recording: Before each action, an RGB-D image of the terrain was captured. During the action, F/T sensor data was recorded. After the action, the volume of material scooped was measured.

License

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.

You are free to:

• Share — copy and redistribute the material in any medium or format.
• Adapt — remix, transform, and build upon the material for any purpose, even commercially.

Under the following terms:

• Attribution — You must give appropriate credit, provide a link to the license, and indicate if changes were made.

Full License Text: https://creativecommons.org/licenses/by/4.0/legalcode

Citation

If you use this dataset, please cite the following paper:

@inproceedings{Zhu-RSS-23,
  author    = {Zhu, Yifan and Thangeda, Pranay and Ornik, Melkior and Hauser, Kris},
  title     = {Few-shot Adaptation for Manipulating Granular Materials Under Domain Shift},
  booktitle = {Proceedings of Robotics: Science and Systems},
  year      = {2023},
  month     = {July},
  address   = {Daegu, Republic of Korea},
  doi       = {10.15607/RSS.2023.XIX.048}
}

References

For more details and illustrations of the materials and compositions, please visit our project website.

---