HuggingFace
HuggingFace - LeRobot and the Open-Source Robotics Community
Overview
HuggingFace is the “GitHub” of AI models and datasets, bringing the open-source revolution to robotics through the LeRobot project. With SmolVLA and community datasets, they aim to make robotics AI accessible to everyone.
| Item | Details |
|---|
| Headquarters | New York, NY / Paris, France |
| Founded | 2016 |
| CEO | Clément Delangue |
| Robotics Lead | Rémi Cadène (Principal Research Scientist) |
| Core Projects | LeRobot, SmolVLA |
LeRobot
Project Introduction
LeRobot is an open-source project to democratize robotics AI.
🤗 LeRobot
"Making AI for Robotics more accessible
with end-to-end learning"
Core Components
| Component | Description |
|---|
| Models | ACT, Diffusion Policy, SmolVLA, etc. |
| Datasets | Community-shared datasets (continuously growing) |
| Hardware Support | SO100, Koch, Unitree G1, etc. |
| Tools | Training, evaluation, visualization pipelines |
Supported Hardware
- SO100, SO101 (low-cost arms)
- LeKiwi (mobile base)
- Koch (bimanual)
- Unitree G1 (humanoid)
- HopeJR, OMX, EarthRover, and more
SmolVLA
Overview
A 450M parameter VLA released in June 2025:
| Item | Details |
|---|
| Parameters | 450M |
| Feature | Runs on MacBook |
| Training Data | ~30K episodes (community data) |
| Architecture | VLM + Flow Matching |
SmolVLA-450M ≈ Equivalent performance to 10× larger VLAs
+ Superior to ACT (LIBERO, Meta-World)
+ 30% faster with async inference
Efficiency
| Factor | Feature |
|---|
| Training | Trainable on single GPU |
| Inference | Runs on consumer GPU/CPU |
| Fine-tuning | ~50 episodes, ~4 hours (A100) |
LeRobot Hackathons
| Hackathon | Details |
|---|
| Oct 2024 | First community hackathon (Toulouse, France) |
| Jun 2025 | Worldwide hackathon (global simultaneous event) |
| Outcomes | Diverse environment dataset sharing |
Dataset Hub
HuggingFace Hub
└── lerobot/ (tag)
├── pusht_v2 (simulation)
├── aloha_static_* (ALOHA data)
├── so100_* (community collected)
└── ... (many community datasets)
- Researchers and hobbyists worldwide sharing data
- Diverse environments (labs, living rooms, warehouses)
- Standardized formats for interoperability
Philosophy: Democratizing AI
Application to Robotics
LLM Democratization (HuggingFace Success)
├── Model Sharing → Hub
├── Training Tools → Transformers
└── Datasets → Datasets
Robotics Democratization (LeRobot Goal)
├── Model Sharing → lerobot/models
├── Training Tools → LeRobot library
└── Datasets → lerobot-tagged datasets
Emphasis on Low-Cost Hardware
| Equipment | Cost |
|---|
| SO100/SO101 arm | ~$100 (single) |
| LeKiwi base | ~$300 |
| Teleoperation setup (2 arms) | ~$660 |
→ “Lowering barriers to entry for robotics research”
Roadmap
Below is compiled from public announcements and GitHub activity; may not be the official roadmap.
| Timeline | Milestone |
|---|
| 2024 | LeRobot project launch |
| 2024-25 | Community dataset growth |
| 2025.06 | SmolVLA release |
| 2025+ | More efficient models, broader hardware support (expected) |
References
See Also