README.md

metadata

license: cc-by-4.0
task_categories:
  - robotics
tags:
  - LeRobot
configs:
  - config_name: default
    data_files: data/*/*.parquet

Gello Dataset - LeRobot v2.1 Format

This dataset contains robotic manipulation demonstrations converted to the LeRobot v2.1 data format, matching the structure of the berkeley_autolab_ur5 dataset.

Dataset Overview

• Task: Pick up the hard drive and place it in the grey box.
• Total Episodes: 10
• Total Frames: 2,560
• Total Videos: 20 (2 video types per episode)
• Format: LeRobot v2.1
• FPS: 10
• Video Resolution: 256x256

Directory Structure

gello_lerobot_v21/
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ... (10 episodes total)
├── meta/
│   ├── episodes_stats.jsonl     # Per-episode statistics
│   ├── episodes.jsonl           # Episode metadata
│   ├── info.json               # Dataset metadata
│   └── tasks.jsonl             # Task descriptions
└── videos/
    └── chunk-000/
        ├── observation.images.hand_image/
        │   ├── episode_000000.mp4
        │   └── ... (10 videos)
        └── observation.images.image/
            ├── episode_000000.mp4
            └── ... (10 videos)

Data Fields

Observation Fields

• observation.state: Robot joint positions and gripper state (7D float32)
• observation.images.hand_image: Wrist camera view (256x256x3 video)
• observation.images.image: Base camera view (256x256x3 video)

Action Fields

• action: Robot actions (7D float32)

Metadata Fields

• timestamp: Time step in seconds (float32)
• episode_index: Episode identifier (int64)
• frame_index: Frame index within episode (int64)
• next.reward: Reward signal (float32)
• next.done: Episode termination flag (bool)
• index: Global frame index (int64)
• task_index: Task identifier (int64)

Usage with LeRobot

This dataset is compatible with LeRobot v2.1 tools and can be loaded using:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

# Load the dataset
dataset = LeRobotDataset("/path/to/gello_lerobot_v21")

# Access episodes
for episode in dataset:
    print(f"Episode {episode['episode_index']}: {episode['total_frames']} frames")

Video Format

• Codec: SVT-AV1
• Resolution: 256x256 pixels
• Channels: 3 (RGB)
• Pixel Format: YUV420p
• Frame Rate: 10 FPS

Task Description

All episodes in this dataset demonstrate the same task:

"Pick up the hard drive and place it in the grey box."

Dataset Statistics

• Average Episode Length: 256 frames
• Average Episode Duration: 25.6 seconds
• Success Rate: 100% (all episodes completed successfully)
• Total Reward: 10.0 (1.0 per episode)

File Formats

Parquet Files

Each episode is stored as a Parquet file containing all the data fields with proper data types.

JSONL Files

• episodes_stats.jsonl: Detailed statistics for each episode
• episodes.jsonl: Episode metadata including task assignments
• tasks.jsonl: Task definitions

JSON Files

• info.json: Dataset metadata including feature specifications, splits, and paths

Conversion Notes

This dataset was converted from raw Gello pickle files using the gello_to_lerobot.py script. The conversion process:

1. Loads raw demonstration data from pickle files
2. Processes images (resize to 256x256, convert to videos)
3. Creates LeRobot v2.1 directory structure
4. Saves episodes as individual Parquet files
5. Generates metadata files in the required format

Compatibility

This dataset follows the same structure as the berkeley_autolab_ur5 dataset and is fully compatible with:

• LeRobot v2.1 data loading utilities
• HuggingFace datasets library
• Standard machine learning frameworks (PyTorch, TensorFlow)

License

This dataset is provided for research and educational purposes. Please refer to the original Gello dataset license for usage terms.