Robots Atlas>ROBOTS ATLAS
SmolVLA
AI Models

SmolVLA

A compact (450M), open-source Vision-Language-Action model for robotics, trained on LeRobot community data. Runs on consumer hardware and generates continuous robot actions.
โœ“ Activeโœ“ Public accessโš– Open sourceVision-Language-Action modelRobotics foundation model
Parameters
450M
parameters
Release date
3 June 2025
๐ŸขHugging FaceProducer
Access:DownloadDeployment:๐Ÿ’ป Localโ˜ Cloud๐Ÿ“ฑ On-device

Overview

SmolVLA is a compact (450M-parameter), open-source Vision-Language-Action model built by Hugging Face within the LeRobot ecosystem. It takes a sequence of RGB images from multiple cameras, the robot's current sensorimotor state, and a natural-language instruction, and outputs a continuous sequence of actions (an action chunk) to control the robot.

The architecture combines a Vision-Language backbone (SmolVLM2 โ€” a SigLIP vision encoder and a SmolLM2 language decoder) with a lightweight action expert (~100M parameters) built on a transformer and trained with a flow-matching objective. The action expert produces continuous actions non-autoregressively, enabling real-time control.

The model was trained solely on public, community-shared datasets tagged lerobot โ€” 487 curated datasets focused on the SO-100 arm, standardized to 30 FPS, yielding roughly 10 million frames. SmolVLA also supports asynchronous inference, which decouples action execution from predicting the next chunk, giving roughly 30% faster response and 2x task throughput.

Classification
Vision-Language-Action modelRobotics foundation model
Applications
Robotic manipulationRobot policy training
Access & deployment
Download
LocalCloudOn-device
Weights: Open source
Key parameters
๐Ÿงฉ Parameters: 450M
โœ“ Fine-tuning
๐Ÿ“ฅ Input: image, robot state data, text
Robotics
Robot manipulationRobot controlVisual grounding

Technical specification

Parameters
450M
parameters
Hardware requirements
Small enough to run on CPU, train on a single consumer GPU, or even a MacBook.
Features:โœ“ Fine-tuning
Modalities
โฌ‡ Input
imagerobot_state_datatext
โฌ† Output
robot_actionsmotion_trajectories

Capabilities and applications

Native model capabilities
Vision-language-action grounding
The ability of a VLA model to ground visual perception and a language instruction into a concrete physical robot action. The model understands the scene and intent, then generates an executable action sequence, closing the loop from observation to motion.
Category: robotics
Cross-embodiment transfer
The ability of a single model to control robots with different morphologies (humanoids, dual-arm rigs, mobile platforms) without training a separate model per platform. Intelligence is decoupled from embodiment, so the same policy runs on hardware with different kinematics and dynamics.
Category: robotics
Real-time inference
The model's ability to generate responses with very low latency (>1000 tokens/sec) on specialized inference hardware (e.g. Cerebras WSE), enabling interactive, turn-by-turn collaboration with a human.
Category: coding
Image understanding
Analysing and interpreting the content of images.
Category: vision
Multimodal understanding
Category: multimodal
Robotics
Robot manipulationRobot controlVisual grounding
Application domains
Robotic manipulationRobot policy training

Benchmark results

2 benchmarks
SO100 (real-world manipulation)
task success rate ยท after pretraining on LeRobot community datasets
78.3%%
๐Ÿ“„ Hugging Face SmolVLA blog / technical report (arXiv:2506.01844)
Without pretraining on community data, the model reaches 51.7% success on SO100. Pretraining raises it to 78.3% (+26.6pp).
SO100 (asynchronous vs synchronous inference)
task success rate ยท comparison of synchronous and asynchronous inference
~78%%
๐Ÿ“„ Hugging Face SmolVLA blog / technical report (arXiv:2506.01844)
Both modes achieve similar success (โ‰ˆ78%), but asynchronous inference completes tasks ~30% faster (9.7s vs 13.75s) and enables 2x more completions in a fixed time window (19 vs 9 cubes).

Technical architecture

Sources and related pages