chocolat-nya/yaskawa-untangle-dataset

Yaskawa Cable Untangling Dataset (Small)

This is a smaller version of the Yaskawa cable untangling dataset containing 3 episodes. This dataset is useful for testing and development. For the full dataset with 50 episodes, see chocolat-nya/yaskawa-untangle-20260309-144410.

Dataset Details

• Total Episodes: 3
• Total Frames: 14,889
• Robot: Yaskawa 7-axis manipulator
• Task: Cable untangling task
• Collection Method: Direct teaching (teleoperation)
• FPS: 30
• Duration: ~30 seconds per episode
• LeRobot Version: v3.0

Observations

State (7-DOF joint angles)

• axis_0 to axis_6: Joint angles in degrees

Images (3 cameras)

• observation.images.realsense: Intel RealSense D435I RGB camera (on-hand)
• observation.images.fixed_cam1: Fixed USB Camera 1 (L-837)
• observation.images.fixed_cam2: Fixed USB Camera 2 (L-837)

All images are 640x480 pixels, H.264 encoded.

Actions

The actions represent the target joint angles for the next timestep:

• action_0 to action_6: Target joint angles in degrees

Usage

Loading the Dataset

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

# Load the dataset from Hugging Face Hub
dataset = LeRobotDataset("chocolat-nya/yaskawa-untangle-dataset")

print(f"Total episodes: {dataset.num_episodes}")  # 3
print(f"Total frames: {len(dataset)}")            # 14,889

# Access a specific episode
episode_0 = dataset.episode(0)

# Get a single frame
frame = dataset[0]
print(frame.keys())  # ['action', 'observation.state', 'observation.images.realsense', ...]

Training with LeRobot

Quick test training with this smaller dataset:

lerobot-train \
  --dataset.repo_id=chocolat-nya/yaskawa-untangle-dataset \
  --policy.type=act \
  --output_dir=outputs/train/yaskawa-untangle-test \
  --job_name=yaskawa-untangle-test \
  --policy.device=cuda \
  --steps=10000

Version Tag

This dataset is tagged with v3.0 (LeRobot codebase version).

Data Collection Setup

Hardware

• Robot: Yaskawa 7-axis manipulator
• Cameras:
  • Intel RealSense D435I (mounted on robot hand)
  • 2x USB Camera L-837 (fixed viewpoints)
• Control: Direct teaching mode (30 seconds per episode)

Camera Configuration

• RealSense RGB: /dev/video6
• Fixed Camera 1: /dev/video8
• Fixed Camera 2: /dev/video12

Related Datasets

• Full Dataset (50 episodes): chocolat-nya/yaskawa-untangle-20260309-144410

Citation

If you use this dataset in your research, please cite:

@dataset{yaskawa_untangle_2026,
  title={Yaskawa Cable Untangling Dataset},
  author={chocolat-nya},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/chocolat-nya/yaskawa-untangle-dataset}}
}

License

MIT License