# LibreVLA
LibreVLA runs robot policies from camera frames and robot state and returns action chunks.
Tasks: Robot policies. Install: pip install "libreyolo[vla]".
Verified against LibreYOLO v1.6.0.

## Install

```bash
pip install "libreyolo[vla]"
```

## Predict

**Python**

```python
from libreyolo import LibreVLA, SAMPLE_IMAGE

# Python 3.12 or later. This demonstrates shapes with synthetic state.
model = LibreVLA("smolvla-base", device="cpu")
result = model.predict(SAMPLE_IMAGE, state=[0.0] * 6, instruction="pick up the object")
print(result.actions.data.shape)
print(result.actions.first)
```

The extra requires Python 3.12 or later. Supply real camera observations and the state representation used during training for meaningful actions. The library returns action values; the caller owns the robot control loop. Call `reset()` between episodes.

## Train

### SmolVLA

`LibreVLA()` selects SmolVLA. Training updates its action expert with the vision backbone frozen. Pass a LeRobot dataset ID or local v3 directory to `train(data=...)`. Defaults are 5 epochs, batch 8, workers 0 and `val_split=0.1`.

### ACT

`LibreVLA("act")` constructs an untrained action policy. Train it on a LeRobot dataset before prediction. It uses the dataset camera, state and action shapes without a language instruction. Its image backbone may download separately.

### Diffusion Policy

`LibreVLA("diffusion")` also starts without pretrained policy weights. It keeps observation history between calls; `reset()` clears that history between episodes.

## Validate

`val(data=...)` reports `val/action_l1`, `val/action_mse`, `val/action_l1_first` and per-dimension `val/action_l1_dims`. These compare predicted and recorded actions offline; they are not robot task success rates.

## Checkpoints

Training saves directories containing policy weights, processors and `libreyolo_vla.json`. Reload a returned best-checkpoint directory with `LibreVLA(path)`. See the [policy API](/docs/reference/vla-api).
## Licensing

Check the license on the Hugging Face repository of the specific weights you download. That repository is authoritative and licenses are not always uniform across a family. This is a description of the licenses involved, not legal advice.

- Original work: SmolVLA, ACT and Diffusion Policy, Hugging Face and policy authors
- Upstream license: Apache-2.0
- Upstream source: https://github.com/huggingface/lerobot
- LibreYOLO code: MIT
- Weights: Apache-2.0, distributed by their authors. LibreYOLO does not host or mirror them.
- Interpretation: The upstream policy runtime and SmolVLA base declare Apache-2.0. ACT and Diffusion Policy training do not load pretrained policy weights.