README.md

metadata

license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - multi-robot
  - manipulation
  - bone-picking
configs:
  - config_name: default
    data_files: data/*/*.parquet

Bone to Pick - Combined Dataset (10_08 + 10_09)

This dataset is a merged combination of two Bone to Pick datasets from Hugging Face:

• oordonez/Bone_to_Pick_10_08
• oordonez/Bone_to_Pick_10_09

Dataset Description

This dataset was created using LeRobot and contains bi-manual robot manipulation data for a bone picking task. The dataset combines 20 episodes (10 from each original dataset) with updated episode numbering to avoid conflicts.

Key Features:

• Total Episodes: 20 (episodes 0-19)
• Total Frames: 7,788 (after dead time removal)
• Total Videos: 60 (3 camera views × 20 episodes)
• Robot Type: Dual-arm manipulation (white + blue follower arms)
• Task: Grasp the blue bone and place it inside the box

Camera Views:

1. observation.images.my_blue_follower_arm_front (640×480, H.264)
2. observation.images.my_white_follower_arm_front (640×480, H.264)
3. observation.images.my_white_follower_arm_global_top (1280×720, H.264)

Dataset Structure

Data Format:

• Action Space: 12-dimensional float32 arrays
  • 6 joints for white follower arm
  • 6 joints for blue follower arm
• Observation Space: 12-dimensional float32 arrays (same as action)
• Timestamps: float32 timestamps
• Frame Indices: int64 frame numbers within each episode
• Episode Indices: int64 episode numbers (0-19)

Joint Names:

White Follower Arm:

• my_white_follower_arm_shoulder_pan.pos
• my_white_follower_arm_shoulder_lift.pos
• my_white_follower_arm_elbow_flex.pos
• my_white_follower_arm_wrist_flex.pos
• my_white_follower_arm_wrist_roll.pos
• my_white_follower_arm_gripper.pos

Blue Follower Arm:

• my_blue_follower_arm_shoulder_pan.pos
• my_blue_follower_arm_shoulder_lift.pos
• my_blue_follower_arm_elbow_flex.pos
• my_blue_follower_arm_wrist_flex.pos
• my_blue_follower_arm_wrist_roll.pos
• my_blue_follower_arm_gripper.pos

Processing Details

Dead Time Removal:

• Applied velocity threshold of 0.0001 for joint movement detection
• Removed 2+ seconds of stationary data at episode ends
• Total frames removed: 1,496 frames across all episodes

Video Processing:

• Converted all videos from AV1 to H.264 codec for browser compatibility
• Maintained original frame rates (30 FPS)
• Preserved video quality and aspect ratios

Episode Renumbering:

• Original dataset 10_08: episodes 0-9 → merged episodes 0-9
• Original dataset 10_09: episodes 0-9 → merged episodes 10-19
• All metadata updated to reflect new episode numbering

Dataset Viewer Information

What You'll See in the Hugging Face Dataset Viewer:

📊 Data Columns:

• action: 12-element array containing joint positions for both robot arms
• observation.state: 12-element array containing current joint states
• timestamp: float32 timestamp for each frame
• frame_index: int64 frame number within each episode
• episode_index: int64 episode number (0-19)
• index: int64 global sample index
• task_index: int64 task identifier

🤖 Joint Mapping (12 elements in action/observation arrays):

Indices 0-5: White Follower Arm

• Index 0: my_white_follower_arm_shoulder_pan.pos
• Index 1: my_white_follower_arm_shoulder_lift.pos
• Index 2: my_white_follower_arm_elbow_flex.pos
• Index 3: my_white_follower_arm_wrist_flex.pos
• Index 4: my_white_follower_arm_wrist_roll.pos
• Index 5: my_white_follower_arm_gripper.pos

Indices 6-11: Blue Follower Arm

• Index 6: my_blue_follower_arm_shoulder_pan.pos
• Index 7: my_blue_follower_arm_shoulder_lift.pos
• Index 8: my_blue_follower_arm_elbow_flex.pos
• Index 9: my_blue_follower_arm_wrist_flex.pos
• Index 10: my_blue_follower_arm_wrist_roll.pos
• Index 11: my_blue_follower_arm_gripper.pos

🎥 Video Files:

• 60 total videos (3 camera views × 20 episodes)
• All videos are in H.264 codec for browser compatibility
• Videos are stored in the videos/ directory

Usage

Loading the Dataset:

from datasets import load_from_disk

# Load the dataset
dataset = load_from_disk('merged_bone_to_pick_combined_hf_fixed')

# Access a sample
sample = dataset[0]
print(f"Action: {sample['action']}")
print(f"Observation: {sample['observation.state']}")
print(f"Episode: {sample['episode_index']}")

# Access joint values by index
action = sample['action']
white_arm_joints = action[:6]  # First 6 elements
blue_arm_joints = action[6:]   # Last 6 elements

Accessing Videos:

# Videos are stored in the videos/ directory
# Each episode has 3 camera views:
# - observation.images.my_blue_follower_arm_front/episode_XXXXXX.mp4
# - observation.images.my_white_follower_arm_front/episode_XXXXXX.mp4  
# - observation.images.my_white_follower_arm_global_top/episode_XXXXXX.mp4

Dataset Statistics

Metric	Value
Total Episodes	20
Total Frames	7,788
Total Videos	60
Average Frames per Episode	389.4
Video Codec	H.264
Frame Rate	30 FPS
Data Format	LeRobot v2.1

Citation

BibTeX:

@dataset{merged_bone_to_pick_combined,
  title={Bone to Pick - Combined Dataset (10_08 + 10_09)},
  author={oordonez},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/oordonez/Bone_to_Pick_10_08},
  url={https://huggingface.co/datasets/oordonez/Bone_to_Pick_10_09}
}

License

This dataset is released under the Apache 2.0 License.

Acknowledgments

• Original datasets: oordonez/Bone_to_Pick_10_08 and oordonez/Bone_to_Pick_10_09
• Built with LeRobot
• Merged and processed for improved usability and browser compatibility