SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data

🧭TL;DR

Today, we introduce SmolVLA, a compact (450M), open-source Vision-Language-Action model for robotics that runs on consumer hardware.

• Pretrained only on compatibly licensed, open-source community-shared datasets under the lerobot tag.
• SmolVLA-450M outperforms much larger VLAs and strong baselines such as ACT on simulation (LIBERO, Meta-World) and real-world tasks (SO100, SO101).
• Supports asynchronous inference for 30% faster response and 2× task throughput.

Useful links:

• Hardware used to train and evaluate SO-100/101: https://github.com/TheRobotStudio/SO-ARM100
• Base model https://huggingface.co/lerobot/smolvla_base
• Paper: https://huggingface.co/papers/2506.01844

📚 Table of Contents

• 🧭 TL;DR
• 📖 Introduction
• 🤖 Meet SmolVLA
• 🚀 How to Use SmolVLA?
  • Install
  • Finetune the Pretrained Model
  • Train from Scratch
• 🧠 Method
  • Main Architecture
    • Vision-Language Model (VLM)
    • Action Expert: Flow Matching Transformer
  • Design Choices for Efficiency and Robustness
    • Visual Token Reduction
    • Faster Inference via Layer Skipping
    • Interleaved Cross and Self-Attention
  • Asynchronous Inference
• 📦 Community Datasets
  • Improving Task Annotations
  • Standardizing Camera Views
• 📊 Results
• ✅ Conclusion
• 📣 Call to Action

Introduction

Over the past few years, Transformers have driven remarkable progress in AI, from language models capable of human-like reasoning to multimodal systems that understand both images and text. However, in real-world robotics, advancements have been much slower. Robots still struggle to generalize across diverse objects, environments, and tasks. This limited progress stems from a lack of high-quality, diverse data and the absence of models that can reason and act like humans in the physical world.

Meet SmolVLA!

SmolVLA-450M is our open-source, compact yet capable VLA model. It is:

• Skipping half of the layers of the vision model for faster inference and smaller size
• Interleaving self-attention and cross-attention blocks
• Using fewer visual tokens
• Leveraging smaller pretrained VLMs