Datasets:

HuggingFaceVLA
/

community_dataset_v2

license: apache-2.0
task_categories:
  - robotics
  - computer-vision
  - reinforcement-learning
tags:
  - robotics
  - community
  - so100
  - manipulation
  - multi-user
  - collection
  - smolvla
  - lerobot
  - vision-language-action
size_categories:
  - 100K<n<1M
language:
  - en
pretty_name: LeRobot Community Dataset v2
viewer: false

LeRobot Community Dataset v2

A large-scale collection of robotic manipulation datasets contributed by the LeRobot community. This dataset represents the second major release of community-contributed robotics data, featuring high-quality episodic demonstrations across diverse manipulation tasks.

🤖 Overview

This dataset is a curated collection of robotics demonstrations from multiple contributors, primarily focused on manipulation tasks using SO-100 robotic arms. The data was collected and processed using the LeRobot framework, ensuring consistency and compatibility across all datasets.

Key Features:

Multi-contributor: Data from 59 different contributors
Diverse tasks: Wide variety of manipulation scenarios and objects
Standardized format: All data in LeRobot v2.0/v2.1 format
Quality filtered: Automated processing and validation pipeline
Ready for training: Compatible with SmolVLA and other VLA models

📊 Dataset Statistics

Metric	Value
Total Datasets	170
Total Episodes	N/A
Total Frames	N/A
Contributors	59
Robot Types	SO-100 (various colors)
Data Format	LeRobot v2.0 and v2.1 dataset format
Total Size	~5032.889999999998 GB
Success Rate	87% (170/196 from input list)

🏗️ Dataset Structure

The dataset maintains the original contributor-based organization:

community_dataset_v2/
├── user1/
│   ├── dataset1/
│   │   ├── data/
│   │   ├── meta/
│   │   └── videos/
│   └── dataset2/
│       ├── data/
│       ├── meta/
│       └── videos/
├── user2/
│   └── dataset3/
└── ...

Each individual dataset follows the standard LeRobot format with:

data/: Contains the episode data (observations, actions, rewards)
meta/: Metadata files including episode information
videos/: MP4 video files for visual observations

🔧 Prerequisites

Before using this dataset, install LeRobot and authenticate with Hugging Face:

Install LeRobot

# Install conda if not available
# conda install -c conda-forge ffmpeg  # Required for video processing

# Install LeRobot from source
git clone https://github.com/huggingface/lerobot.git
cd lerobot
pip install -e .

Authenticate with Hugging Face

huggingface-cli login

📥 Usage

Download the Dataset

from huggingface_hub import snapshot_download

# Download the entire dataset (requires authentication)
local_dir = snapshot_download(
    repo_id="HuggingFaceVLA/community_dataset_v2",
    repo_type="dataset",
    local_dir="./community_dataset_v2"
)

Load Individual Datasets

from lerobot.datasets.lerobot_dataset import LeRobotDataset
import os

# List all available datasets
dataset_root = "./community_dataset_v2"
for user_name in os.listdir(dataset_root):
    user_path = os.path.join(dataset_root, user_name)
    if os.path.isdir(user_path):
        for dataset_name in os.listdir(user_path):
            dataset_path = os.path.join(user_path, dataset_name)
            if os.path.isdir(dataset_path):
                print(f"{user_name}/{dataset_name}")

# Load a specific dataset
dataset = LeRobotDataset("./community_dataset_v2/user_name/dataset_name")

# Access episodes
for episode_idx in range(len(dataset.episode_indices)):
    episode = dataset[episode_idx]
    # Process episode data...

Integration with SmolVLA (COMING SOON)

# Example usage with SmolVLA - needs testing
from transformers import AutoProcessor, AutoModelForVision2Seq

processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM-Instruct")
model = AutoModelForVision2Seq.from_pretrained("HuggingFaceTB/SmolVLM-Instruct")

# Process robotics episodes for VLA training
# Implementation details coming soon...

📋 Data Collection

The data was primarily collected using LeRobot software with SO-100 robotic arms, ensuring consistency in data format and quality across contributors. The collection process involved:

Community Contribution: Multiple researchers and practitioners contributed datasets
Standardization: All data converted to LeRobot standard format
Quality Control: Automated validation and filtering pipeline
Processing: Consistent preprocessing and format normalization

🔄 Processing Pipeline

This dataset was processed using an automated pipeline that:

✅ Validates dataset format compatibility
✅ Removes corrupted or incomplete episodes
✅ Standardizes video formats and resolutions
✅ Generates comprehensive metadata
✅ Ensures LeRobot v2.0/v2.1 compatibility

Processing Results:

Input datasets: 196
Successfully processed: 170
Success rate: 87%
Failed datasets: 26 (mainly due to format incompatibility)

🤝 Contributing

This dataset represents community contributions. If you'd like to contribute additional datasets:

Ensure your data follows the LeRobot dataset format
Contact the maintainers for inclusion in future releases
Follow the quality guidelines and formatting standards

📚 Related Work

SmolVLA Model - Vision-Language-Action model
SmolVLA Blog Post - Detailed overview and applications
SmolVLA Paper - Technical details and benchmarks
LeRobot Datasets Blog - Overview of robotics datasets

📄 License

This dataset is released under the Apache 2.0 license. Please cite appropriately if used in research.

🔗 Citation

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

@dataset{community_dataset_v2,
  title={LeRobot Community Dataset v2},
  author={LeRobot Community},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/HuggingFaceVLA/community_dataset_v2}
}

Built with ❤️ by the SmolVLA team and LeRobot Community