community_dataset_v1

# Community Dataset v1

A large-scale community-contributed robotics dataset for vision-language-action learning, featuring **128 datasets** from **55 contributors** worldwide.

We used this dataset to pretrain [SmolVLA](https://huggingface.co/lerobot/smolvla_base). However, this is not a complete set, but the dataset that we selected using specific filters, like fps, min num of episodes, and some qualitative assessment of video qualities, using the https://huggingface.co/spaces/Beegbrain/FilterLeRobotData tool. We also manually curated the task descriptions for this subset of the dataset.

## 🌟 Overview

This dataset represents a collaborative effort from the robotics and AI community to build comprehensive training data for embodied AI systems. Each contribution contains demonstrations of robotic manipulation tasks with the SO100 arm, recorded using [LeRobot tools](https://github.com/huggingface/lerobot), primarily focused on tabletop scenarios and everyday object interactions.

## 📊 Dataset Statistics

| Metric | Value |
|--------|-------|
| **Total Datasets** | 128 |
| **Total Episodes** | 11,132 |
| **Total Frames** | 5,105,808 |
| **Total Videos** | 22,065 |
| **Contributors** | 55 |
| **Weighted Average FPS** | 30.4 |
| **Average Episodes per Dataset** | 87.0 |
| **Total Duration** | 46.9 hours |
| **Average Hours per Dataset** | 0.37 |
| **Primary Tasks** | Manipulation, Pick & Place, Sorting |
| **Robot Types** | SO-100 (various colors) |
| **Data Format** | LeRobot v2.0 and v2.1 dataset format |
| **Total Size** | 119.3 GB |

## 🗂️ Structure

The dataset maintains a clear hierarchical structure:

```
community_dataset_v1/
├── contributor1/
│ ├── dataset_name_1/
│ │ ├── data/ # Parquet files with observations
│ │ ├── videos/ # MP4 recordings
│ │ └── meta/ # Metadata and info
│ └── dataset_name_2/
├── contributor2/
│ └── dataset_name_3/
└── ...
```

Each dataset follows the LeRobot format standard, ensuring compatibility with existing frameworks and easy integration.

## 🏆 Top Contributors

| Contributor | Datasets Quantity |
|-------------|----------|
| lirislab | 14 |
| roboticshack | 9 |
| sihyun77 | 8 |
| pierfabre | 7 |
| ganker5 | 6 |
| paszea | 5 |
| samsam0510 | 5 |
| pranavsaroha | 5 |
| bensprenger | 4 |
| Chojins | 4 |

## 🚀 Usage

**1. Authenticate with Hugging Face**

You need to be logged in to access the dataset:

```bash
# Login to Hugging Face
huggingface-cli login

# Or alternatively, set your token as an environment variable
# export HF_TOKEN=your_token_here
```

Get your token from [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)

### Download the Dataset

```python
hf download HuggingFaceVLA/community_dataset_v1 \
--repo-type=dataset \
--local-dir /path/local_dir/community_dataset_v1
```

### Load Individual Datasets

```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset
import os

# Browse available datasets
for contributor in os.listdir("./community_dataset_v1"):
contributor_path = f"./community_dataset_v1/{contributor}"
if os.path.isdir(contributor_path):
for dataset in os.listdir(contributor_path):
print(f"📁 {contributor}/{dataset}")

# Load a specific dataset (requires authentication)
dataset = LeRobotDataset(
repo_id="local",
root="./community_dataset_v1/contributor_name/dataset_name"
)

# Access episodes and observations
print(f"Episodes: {len(dataset.episode_indices)}")
print(f"Total frames: {len(dataset)}")
```

### Integration with SmolVLA pretraining framework

This dataset is designed for training VLA models
You can download this dataset and use it for Vision Language Action Models training framework, [VLAb](https://github.com/huggingface/VLAb/tree/main):

1. Visit the VLAb repository.
2. Follow the training instructions in the repo
3. Point the training script to this dataset

```python

accelerate launch --config_file accelerate_configs/multi_gpu.yaml \
src/lerobot/scripts/train.py \
--policy.type=smolvla2 \
--policy.repo_id=HuggingFaceTB/SmolVLM2-500M-Video-Instruct \
--dataset.repo_id="community_dataset_v1/AndrejOrsula/lerobot_double_ball_stacking_random,community_dataset_v1/aimihat/so100_tape" \
--dataset.root="local/path/to/datasets" \
--dataset.video_backend=pyav \
--dataset.features_version=2 \
--output_dir="./outputs/training" \
--batch_size=8 \
--steps=200000 \
--wandb.enable=true \
--wandb.project="smolvla2-training"

```

## 🔧 Dataset Format

Each dataset contains:

- **`data/`**: Parquet files with timestamped observations
- Robot states (joint positions, velocities)
- Action sequences
- Camera observations (multiple views)
- Language instructions

- **`videos/`**: Synchronized video recordings
- Multiple camera angles
- High-resolution capture
- Timestamp alignment

- **`meta/`**: Metadata and configuration
- Dataset info (fps, episode count)
- Robot configuration
- Task descriptions

## 🎯 Intended Use

This dataset is designed for:

- **Vision-Language-Action (VLA) model training**
- **Robotic manipulation research**
- **Imitation learning experiments**
- **Multi-task policy development**
- **Embodied AI research**

## 🤝 Community Contributions

This dataset exists thanks to the generous contributions from researchers, hobbyists, and institutions worldwide. Each dataset represents hours of careful data collection and curation.

### Contributing Guidelines

Future contributions should follow:
- LeRobot dataset format
- Consistent naming conventions for the features, camera views etc.
- Quality validation checks
- Proper task descriptions, describing the actions precisely.

Check the [blogpost](https://huggingface.co/blog/lerobot-datasets) for more information

## 🔗 Related Work

- [VLAb Framework](https://github.com/huggingface/VLAb)
- [SmolVLA model](https://huggingface.co/lerobot/smolvla_base)
- [SmolVLA Blogpost](https://huggingface.co/blog/smolvla)
- [SmolVLA Paper](https://huggingface.co/papers/2506.01844)
- [Docs](https://huggingface.co/docs/lerobot/smolvla)
- [How to Build a successful Robotics dataset with Lerobot?](https://huggingface.co/blog/lerobot-datasets)
---

*Built with ❤️ by the SmolVLA team and LeRobot Community*

# 社区数据集v1

本数据集为面向视觉-语言-动作（Vision-Language-Action, VLA）学习的大规模社区共建机器人数据集，汇集了来自全球55位贡献者的128个子数据集。

我们曾使用该数据集对SmolVLA模型进行预训练。需注意，本数据集并非完整集合，而是通过特定筛选条件（如帧率、最小任务回合数、视频质量定性评估），借助https://huggingface.co/spaces/Beegbrain/FilterLeRobotData工具筛选出的子集。此外，我们还对该子集的任务描述进行了人工整理与校准。

## 🌟 数据集概览

本数据集是机器人与AI社区为具身智能系统构建全面训练数据的协作成果。所有贡献数据均基于SO100机械臂的机器人操作任务演示，使用LeRobot工具套件录制，主要聚焦桌面场景与日常物品交互场景。

## 📊 数据集统计信息

| 指标 | 数值 |
|--------|-------|
| **总子数据集数** | 128 |
| **总任务回合数** | 11,132 |
| **总帧数** | 5,105,808 |
| **总视频数** | 22,065 |
| **贡献者数量** | 55 |
| **加权平均帧率** | 30.4 |
| **单数据集平均回合数** | 87.0 |
| **总时长** | 46.9 小时 |
| **单数据集平均时长** | 0.37 小时 |
| **核心任务类型** | 操作、拾取与放置、分类 |
| **机器人型号** | SO-100（多种配色） |
| **数据格式** | LeRobot v2.0 与 v2.1 数据集格式 |
| **总数据体量** | 119.3 GB |

## 🗂️ 数据集结构

本数据集采用清晰的层级结构：

community_dataset_v1/
├── contributor1/
│ ├── dataset_name_1/
│ │ ├── data/ # 存储观测数据的Parquet文件
│ │ ├── videos/ # 视频录制文件
│ │ └── meta/ # 元数据与配置文件
│ └── dataset_name_2/
├── contributor2/
│ └── dataset_name_3/
└── ...

所有子数据集均遵循LeRobot格式标准，可与现有框架兼容并轻松集成。

## 🏆 核心贡献者

| 贡献者ID | 子数据集数量 |
|-------------|----------|
| lirislab | 14 |
| roboticshack | 9 |
| sihyun77 | 8 |
| pierfabre | 7 |
| ganker5 | 6 |
| paszea | 5 |
| samsam0510 | 5 |
| pranavsaroha | 5 |
| bensprenger | 4 |
| Chojins | 4 |

## 🚀 使用方法

### 1. Hugging Face账号认证

您需要登录账号才能访问本数据集：

bash
# 登录Hugging Face
huggingface-cli login

# 或通过环境变量配置访问令牌
# export HF_TOKEN=your_token_here

可从https://huggingface.co/settings/tokens获取您的访问令牌。

### 下载数据集

python
hf download HuggingFaceVLA/community_dataset_v1
--repo-type=dataset
--local-dir /path/local_dir/community_dataset_v1

### 加载单个子数据集

python
from lerobot.datasets.lerobot_dataset import LeRobotDataset
import os

# 浏览可用子数据集
for contributor in os.listdir("./community_dataset_v1"):
contributor_path = f"./community_dataset_v1/{contributor}"
if os.path.isdir(contributor_path):
for dataset in os.listdir(contributor_path):
print(f"📁 {contributor}/{dataset}")

# 加载指定子数据集（需完成账号认证）
dataset = LeRobotDataset(
repo_id="local",
root="./community_dataset_v1/contributor_name/dataset_name"
)

# 访问任务回合与观测数据
print(f"任务回合数: {len(dataset.episode_indices)}")
print(f"总帧数: {len(dataset)}")

### 与SmolVLA预训练框架集成

本数据集专为VLA模型训练设计，您可下载该数据集并结合VLAb框架开展视觉-语言-动作模型训练：

1. 访问VLAb仓库；
2. 遵循仓库内的训练指南操作；
3. 将训练脚本指向本数据集。

python
accelerate launch --config_file accelerate_configs/multi_gpu.yaml
src/lerobot/scripts/train.py
--policy.type=smolvla2
--policy.repo_id=HuggingFaceTB/SmolVLM2-500M-Video-Instruct
--dataset.repo_id="community_dataset_v1/AndrejOrsula/lerobot_double_ball_stacking_random,community_dataset_v1/aimihat/so100_tape"
--dataset.root="local/path/to/datasets"
--dataset.video_backend=pyav
--dataset.features_version=2
--output_dir="./outputs/training"
--batch_size=8
--steps=200000
--wandb.enable=true
--wandb.project="smolvla2-training"

## 🔧 数据集格式

每个子数据集包含以下内容：

- **`data/`**：存储带时间戳观测数据的Parquet文件，包含：机器人状态（关节位置、速度）、动作序列、多视角相机观测数据、语言指令

- **`videos/`**：同步录制的视频文件，包含：多视角拍摄、高分辨率采集、时间戳对齐

- **`meta/`**：元数据与配置文件，包含：数据集信息（帧率、回合数）、机器人配置、任务描述

## 🎯 预期用途

本数据集适用于：

- 视觉-语言-动作（VLA）模型训练
- 机器人操作研究
- 模仿学习实验
- 多任务策略开发
- 具身智能研究

## 🤝 社区贡献

本数据集的诞生离不开全球各地研究者、爱好者与科研机构的慷慨贡献。每一个子数据集都凝聚了创作者们在数据采集与整理上的大量心血。

### 贡献指南

后续贡献需遵循以下规范：
- 遵循LeRobot数据集格式标准
- 对特征、相机视角等采用统一命名规范
- 完成质量验证检查
- 编写精准描述操作动作的规范任务说明

更多详情请参阅官方博客：https://huggingface.co/blog/lerobot-datasets

## 🔗 相关工作

- [VLAb框架](https://github.com/huggingface/VLAb)
- [SmolVLA模型](https://huggingface.co/lerobot/smolvla_base)
- [SmolVLA官方博客](https://huggingface.co/blog/smolvla)
- [SmolVLA学术论文](https://huggingface.co/papers/2506.01844)
- [官方文档](https://huggingface.co/docs/lerobot/smolvla)
- [如何基于LeRobot构建优质机器人数据集？](https://huggingface.co/blog/lerobot-datasets)

---

*由SmolVLA团队与LeRobot社区倾心打造*