# Robot policies
Robot policies return an action chunk from camera observations and robot state.
Verified against LibreYOLO v1.6.0.

## Models

[LibreVLA](/docs/models/librevla) provides SmolVLA, ACT and Diffusion Policy. SmolVLA conditions on a language instruction. ACT and Diffusion Policy train without a pretrained policy base.

## Predict

**Python**

```python
from libreyolo import LibreVLA, SAMPLE_IMAGE

# Python 3.12+. Synthetic observation for checking the API.
model = LibreVLA(device="cpu")
result = model(SAMPLE_IMAGE, state=[0.0] * 6, instruction="pick up the object")
print(result.actions.data.shape)
```

Use real robot state and camera frames for meaningful predictions. `Actions.data` is a float32 array of shape `(T, D)`; `Actions.first` is the first action. The payload also stores action names, control rate and instruction when available. Call `reset()` between episodes.

## Dataset format

Use a LeRobot v3 dataset directory or Hub dataset ID. Training reads episode boundaries, camera features, state, actions and timestamps from that dataset. There is no detection-style YAML for this task.

## Train

Install `libreyolo[vla]` on Python 3.12 or later. Call `model.train(data=...)`; defaults are 5 epochs, batch 8 and a 0.1 held-out episode fraction.

## Validate

Offline validation reports `val/action_l1`, `val/action_mse`, `val/action_l1_first` and per-dimension `val/action_l1_dims`. These are errors against recorded actions in dataset units. They do not measure task success on a robot.

## Export

Policy export is not supported. Reload a training checkpoint directory through `LibreVLA(path)`.