Datasets:

Modalities:
Video
Languages:
English
ArXiv:
Libraries:
Datasets
License:
danaaubakirova HF Staff commited on
Commit
38aeaf8
·
verified ·
1 Parent(s): ea6fd12

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +15 -199
README.md CHANGED
@@ -7,122 +7,34 @@ tags:
7
  - manipulation
8
  - smolvla
9
  - lerobot
10
- - vision-language-action
11
- - embodied-ai
12
  task_categories:
13
  - robotics
14
  language:
15
  - en
16
  size_categories:
17
  - 100K<n<1M
18
- pretty_name: Community Dataset v1
19
  ---
20
 
21
  # Community Dataset v1
22
 
23
- A large-scale community-contributed robotics dataset for vision-language-action learning, featuring **128 datasets** from **55 contributors** worldwide.
24
 
25
- We used this dataset to pretrain SmolVLA [insert link to the model]. 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.
26
 
27
- ## 🌟 Overview
 
 
 
 
28
 
29
- 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.
30
-
31
- ## 📊 Dataset Statistics
32
-
33
- | Metric | Value |
34
- |--------|-------|
35
- | **Total Datasets** | 128 |
36
- | **Total Episodes** | 11,132 |
37
- | **Total Frames** | 5,105,808 |
38
- | **Total Videos** | 22,065 |
39
- | **Contributors** | 55 |
40
- | **Weighted Average FPS** | 30.4 |
41
- | **Average Episodes per Dataset** | 87.0 |
42
- | **Total Duration** | 46.9 hours |
43
- | **Average Hours per Dataset** | 0.37 |
44
- | **Primary Tasks** | Manipulation, Pick & Place, Sorting |
45
- | **Robot Types** | SO-100 (various colors) |
46
- | **Data Format** | LeRobot v2.0 and v2.1 dataset format |
47
- | **Total Size** | 119.3 GB |
48
-
49
- ## 🗂️ Structure
50
-
51
- The dataset maintains a clear hierarchical structure:
52
-
53
- ```
54
- community_dataset_v1/
55
- ├── contributor1/
56
- │ ├── dataset_name_1/
57
- │ │ ├── data/ # Parquet files with observations
58
- │ │ ├── videos/ # MP4 recordings
59
- │ │ └── meta/ # Metadata and info
60
- │ └── dataset_name_2/
61
- ├── contributor2/
62
- │ └── dataset_name_3/
63
- └── ...
64
- ```
65
-
66
- Each dataset follows the LeRobot format standard, ensuring compatibility with existing frameworks and easy integration.
67
-
68
- ## 🏆 Top Contributors
69
-
70
- | Contributor | Datasets Quantity |
71
- |-------------|----------|
72
- | lirislab | 14 |
73
- | roboticshack | 9 |
74
- | sihyun77 | 8 |
75
- | pierfabre | 7 |
76
- | ganker5 | 6 |
77
- | paszea | 5 |
78
- | samsam0510 | 5 |
79
- | pranavsaroha | 5 |
80
- | bensprenger | 4 |
81
- | Chojins | 4 |
82
-
83
- ## 🚀 Usage
84
-
85
- ### Prerequisites
86
-
87
- **1. Install LeRobot**
88
-
89
- Follow the [official LeRobot installation guide](https://huggingface.co/docs/lerobot/installation):
90
-
91
- ```bash
92
- # Create conda environment with Python 3.10
93
- conda create -y -n lerobot python=3.10
94
- conda activate lerobot
95
-
96
- # Install ffmpeg (required for video processing)
97
- conda install ffmpeg -c conda-forge
98
-
99
- git clone https://github.com/huggingface/lerobot.git
100
- cd lerobot
101
-
102
- # Install LeRobot from Source
103
- pip install -e .
104
- ```
105
-
106
- **2. Authenticate with Hugging Face**
107
-
108
- You need to be logged in to access the dataset:
109
-
110
- ```bash
111
- # Login to Hugging Face
112
- huggingface-cli login
113
-
114
- # Or alternatively, set your token as an environment variable
115
- # export HF_TOKEN=your_token_here
116
- ```
117
-
118
- Get your token from [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)
119
-
120
- ### Download the Dataset
121
 
122
  ```python
123
  from huggingface_hub import snapshot_download
124
 
125
- # Download entire collection (requires authentication)
126
  dataset_path = snapshot_download(
127
  repo_id="HuggingFaceVLA/community_dataset_v1",
128
  repo_type="dataset",
@@ -130,105 +42,9 @@ dataset_path = snapshot_download(
130
  )
131
  ```
132
 
133
- ### Load Individual Datasets
134
-
135
- ```python
136
- from lerobot.datasets.lerobot_dataset import LeRobotDataset
137
- import os
138
-
139
- # Browse available datasets
140
- for contributor in os.listdir("./community_dataset_v1"):
141
- contributor_path = f"./community_dataset_v1/{contributor}"
142
- if os.path.isdir(contributor_path):
143
- for dataset in os.listdir(contributor_path):
144
- print(f"📁 {contributor}/{dataset}")
145
-
146
- # Load a specific dataset (requires authentication)
147
- dataset = LeRobotDataset(
148
- repo_id="local",
149
- root="./community_dataset_v1/contributor_name/dataset_name"
150
- )
151
-
152
- # Access episodes and observations
153
- print(f"Episodes: {len(dataset.episode_indices)}")
154
- print(f"Total frames: {len(dataset)}")
155
- ```
156
-
157
- ### Integration with SmolVLA
158
-
159
- ```python
160
-
161
- COMING SOON
162
- # This dataset is designed for training VLA models
163
- # Compatible with SmolVLA training pipelines
164
- # needs testing
165
-
166
- ```
167
-
168
- ## 🔧 Dataset Format
169
-
170
- Each dataset contains:
171
-
172
- - **`data/`**: Parquet files with timestamped observations
173
- - Robot states (joint positions, velocities)
174
- - Action sequences
175
- - Camera observations (multiple views)
176
- - Language instructions
177
-
178
- - **`videos/`**: Synchronized video recordings
179
- - Multiple camera angles
180
- - High-resolution capture
181
- - Timestamp alignment
182
-
183
- - **`meta/`**: Metadata and configuration
184
- - Dataset info (fps, episode count)
185
- - Robot configuration
186
- - Task descriptions
187
 
188
- ## 🎯 Intended Use
189
-
190
- This dataset is designed for:
191
-
192
- - **Vision-Language-Action (VLA) model training**
193
- - **Robotic manipulation research**
194
- - **Imitation learning experiments**
195
- - **Multi-task policy development**
196
- - **Embodied AI research**
197
-
198
- ## 📈 Revisions
199
-
200
- - **v1.0**: Initial community collection
201
- - 128 datasets from 55 contributors
202
- - Standardized LeRobot format
203
- - Quality filtering and validation
204
- - Comprehensive metadata
205
-
206
- ## 🤝 Community Contributions
207
-
208
- This dataset exists thanks to the generous contributions from researchers, hobbyists, and institutions worldwide. Each dataset represents hours of careful data collection and curation.
209
-
210
- ### Contributing Guidelines
211
-
212
- Future contributions should follow:
213
- - LeRobot dataset format
214
- - Consistent naming conventions for the features, camera views etc.
215
- - Quality validation checks
216
- - Proper task descriptions, describing the actions precisely.
217
-
218
- Check the [blogpost](https://huggingface.co/blog/lerobot-datasets) for more information
219
-
220
- ## 📄 License
221
-
222
- Released under Apache 2.0 license. Individual datasets may have additional attribution requirements - please check contributor documentation.
223
-
224
- ## 🔗 Related Work
225
-
226
- - [LeRobot Framework](https://github.com/huggingface/lerobot)
227
- - [SmolVLA model](https://huggingface.co/lerobot/smolvla_base)
228
- - [SmolVLA Blogpost](https://huggingface.co/blog/smolvla)
229
- - [SmolVLA Paper](https://huggingface.co/papers/2506.01844)
230
- - [Docs](https://huggingface.co/docs/lerobot/smolvla)
231
- - [How to Build a successful Robotics dataset with Lerobot?](https://huggingface.co/blog/lerobot-datasets)
232
- ---
233
 
234
- *Built with ❤️ by the SmolVLA team and LeRobot Community*
 
 
7
  - manipulation
8
  - smolvla
9
  - lerobot
10
+ - multi-user
11
+ - collection
12
  task_categories:
13
  - robotics
14
  language:
15
  - en
16
  size_categories:
17
  - 100K<n<1M
 
18
  ---
19
 
20
  # Community Dataset v1
21
 
22
+ This is a large-scale community robotics dataset containing 128 individual datasets from multiple contributors.
23
 
24
+ ## Tags
25
 
26
+ - **smolvla**: Part of the SmolVLA project
27
+ - **lerobot**: Compatible with LeRobot framework
28
+ - **robotics**: Robotics manipulation tasks
29
+ - **community**: Community-contributed datasets
30
+ - **so100**: SO-100 related tasks
31
 
32
+ ## Usage
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33
 
34
  ```python
35
  from huggingface_hub import snapshot_download
36
 
37
+ # Download the entire dataset
38
  dataset_path = snapshot_download(
39
  repo_id="HuggingFaceVLA/community_dataset_v1",
40
  repo_type="dataset",
 
42
  )
43
  ```
44
 
45
+ ## Structure
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
46
 
47
+ The dataset preserves the original subfolder structure with user/dataset organization.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
48
 
49
+ Size: ~120GB
50
+ Format: LeRobot dataset format