SmolVLA
Table of contents
SmolVLA is a compact, efficient vision-language-action model from Hugging Face, trained entirely on open LeRobot community datasets. It is designed for accessibility: the model is small enough to train and deploy on consumer-grade hardware while remaining competitive with much larger VLA models.
▪ Key Properties
- Compact backbone: built on SmolVLM-2, a small vision-language model, keeping the total parameter count around 450M.
- Flow-matching action expert: a lightweight action expert predicts continuous action chunks with flow matching, similar in spirit to π₀.
- Efficient training: trained only on publicly available LeRobot community data, without any proprietary datasets.
- Asynchronous inference: the policy server can predict the next action chunk while the robot is still executing the current one, hiding inference latency during deployment.
Reference:
- SmolVLA on Hugging Face — blog post introducing the model and training setup.
- SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics.