---
library_name: physicalai-train
tags:
  - robotics-ai-suite
  - physical-ai
  - "category:action-policies-vla"
  - "chipset:ptl"
license: apache-2.0
base_model:
  - HuggingFaceVLA/smolvla_libero
title: "SmolVLA"
subtitle: "Lightweight vision-language-action model for edge deployment"
category: "Physical AI"
model_type:
  - "Imitation Learning"
primary_type: "Vision Language Action"
secondary_types:
  - Robotics
related_models:
  - act-fp16-ov-catalog
  - pi05-libero-fp16-ov-catalog
key_novelty: "Lightweight vision-language-action model targeting single-GPU training and edge deployment that maps camera images, robot state, and a language instruction to an executable chunk of actions."
thumbnail: assets/thumbnail.png
image_overview: assets/architecture-overview.svg
image_detailed: assets/architecture-detailed.svg
code: "https://github.com/openvinotoolkit/physicalai"
---

# SmolVLA

That's a lightweight vision-language-action model, which targets single-GPU training and edge deployment use cases.
The original implementations can be found [here](https://huggingface.co/papers/2506.01844).
The model takes images from one or multiple cameras, robot state, and language instruction and produces
a chunk of actions, which robot can execute as a sequence of movements in real world.

The model weights are converted from the official [HF repo](https://huggingface.co/HuggingFaceVLA/smolvla_libero)
and can be directly applied in the [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) virtual benchmark environment.

To fine-tune your model with a custom dataset, you can use [Physical AI Studio](https://github.com/open-edge-platform/physical-ai-studio).

# How to Use

## Running Inference with [OpenVINO Physical AI framework](https://github.com/openvinotoolkit/physicalai).

### Installation

```bash
uv pip install physicalai numpy
```

The following API example showcases inference API for this model:

```python
import numpy as np
from physicalai.inference import InferenceModel

model = InferenceModel("smolvla-libero-fp16-ov", device="CPU")

# Build a dummy LIBERO-style observation.
# LIBERO provides two cameras (agentview + wrist) and an 8-dim robot state.
# Images use the LeRobot convention: float32 in [0, 1], shape (C, H, W).
observation = {
    "images.image": np.random.rand(1, 3, 256, 256).astype(np.float32),
    "images.image2": np.random.rand(1, 3, 256, 256).astype(np.float32),
    "state": np.zeros((1, 8), dtype=np.float32),
    "task": ["sample task description"],
}

chunk = model.predict_action_chunk(observation)
```

## Running Accuracy Evaluation on LIBERO

### Installation
```bash
uv pip install physicalai-train[libero,smolvla]
```

The following API example demonstrates how to run the model on LIBERO:

```python
from physicalai.benchmark import LiberoBenchmark
from physicalai.inference import InferenceModel

model = InferenceModel("smolvla-libero-fp16-ov", device="CPU")

benchmark = LiberoBenchmark(
    task_suite="libero_10",
    video_dir="fail_videos"
)

results = benchmark.evaluate(model)
```

Note that the model should be downloaded and saved to the `smolvla-libero-fp16-ov` folder prior to running this script.

## Legal information

The original model is distributed under [Apache 2.0](https://choosealicense.com/licenses/apache-2.0/) license.

## Disclaimer

Intel is committed to respecting human rights and avoiding causing or contributing to adverse impacts on human rights. See [Intel’s Global Human Rights Principles](https://www.intel.com/content/dam/www/central-libraries/us/en/documents/policy-human-rights.pdf). Intel’s products and software are intended only to be used in applications that do not cause or contribute to adverse impacts on human rights.