Robotics & physical AI Model or AI service
SmolVLA makes robot manipulation research accessible on smaller hardware
Hugging Face introduced SmolVLA in June 2025. The 450-million-parameter model uses images and instructions to predict robot actions, and was trained using community datasets shared through LeRobot. [1]
- Model access
- The base checkpoint is published on Hugging Face. [2]
- License
- The base model card lists Apache 2.0. [2]
What the model does
The authors published a pretrained checkpoint and described their evaluation on real and simulated robots. Their report explains model design choices intended to reduce compute requirements. [1][2]
Why it matters
A compact model lets more teams study how shared robot demonstrations transfer to a new task without starting with a much larger model. [1]
Key dates
- Hugging Face introduced SmolVLA [1]
Ways to use it
The authors provide the checkpoint and use examples through LeRobot. Evaluation still requires matching the observation and action interface to the robot. [1][2]
Model limitations
Reported performance is specific to the tasks and configurations the authors evaluated. A new setup needs its own validation. [1]
Sources and verification
Catalog evidence checked How we verify records
- Primary source Hugging Face’s dated SmolVLA introduction
- Primary source lerobot/smolvla_base
Related profiles
Robotics & physical AI
Gemini Robotics 2.5
Google DeepMind
The models are intended to interpret a robot's surroundings and support compatible robot tasks.
Robotics & physical AI
NEURA 4NE1
NEURA Robotics
A humanoid platform NEURA presents for development and industrial training, with multiple distinct hardware generations.
Robotics & physical AI
LUMO all-terrain humanoid
Shenzhen Dobot (DOBOT)
LUMO is a humanoid robotics platform whose proposed applications include interaction and service tasks.