> ## Documentation Index
> Fetch the complete documentation index at: https://docs.almond.bot/llms.txt
> Use this file to discover all available pages before exploring further.

# Train a Policy

> Train a model on your recorded Axol data using the standard LeRobot training pipeline.

Axol records every episode in the [LeRobot dataset format](/api/lerobot), so training a policy is just standard **LeRobot training** — there's nothing Axol-specific about it. Once you've recorded a dataset with [Data Collection](/operations/data-collection), you point LeRobot's training script at it, and the resulting checkpoint runs back on the robot with [Run Policy](/operations/run-policy).

<Note>
  Because the data is plain LeRobot format, any LeRobot training tutorial applies directly — just substitute your own dataset `repo_id`. The links below are the best places to start.
</Note>

## Which policy to train

[**ACT** (Action Chunking with Transformers)](https://huggingface.co/docs/lerobot/en/act) is the recommended starting point: it's lightweight (\~80M params), trains in a few hours on a single GPU, and often reaches high success rates with as few as \~50 demonstrations. Once you have a baseline working, you can explore larger vision-language-action models like [SmolVLA](https://huggingface.co/docs/lerobot/en/smolvla).

## Train it

With the [`lerobot` extra](/advanced/development-install#optional-extras) installed on a machine with a CUDA-capable GPU, train directly from your dataset's Hub `repo_id`:

```bash theme={null}
lerobot-train \
  --dataset.repo_id=${HF_USER}/pick-place \
  --policy.type=act \
  --output_dir=outputs/train/act_pick_place \
  --job_name=act_pick_place \
  --policy.device=cuda \
  --wandb.enable=true \
  --policy.repo_id=${HF_USER}/act_pick_place
```

Training writes checkpoints under `outputs/train/.../checkpoints`, and (with `--policy.repo_id`) pushes the trained policy to the Hub so it's reachable from the robot machine and any [inference server](/cli/inference-server). Expect a few hours for \~100k steps on a single GPU.

<Note>
  No local GPU? The ACT tutorial links a [Google Colab notebook](https://huggingface.co/docs/lerobot/en/act#train-using-google-colab) you can use to train in the cloud.
</Note>

## LeRobotTutorials

<CardGroup cols={2}>
  <Card title="ACT" icon="robot" href="https://huggingface.co/docs/lerobot/en/act">
    The recommended first policy — architecture, training, and tips.
  </Card>

  <Card title="Imitation learning walkthrough" icon="graduation-cap" href="https://huggingface.co/docs/lerobot/en/il_robots">
    End-to-end record → train → evaluate with `lerobot-train`.
  </Card>

  <Card title="SmolVLA" icon="brain" href="https://huggingface.co/docs/lerobot/en/smolvla">
    Fine-tune Hugging Face's lightweight vision-language-action model.
  </Card>

  <Card title="LeRobot docs" icon="book" href="https://huggingface.co/docs/lerobot/en/index">
    Full documentation, models, and datasets.
  </Card>
</CardGroup>

## Next steps

<CardGroup cols={2}>
  <Card title="Run Policy" icon="play" href="/operations/run-policy">
    Deploy your trained checkpoint on the robot.
  </Card>

  <Card title="Data Collection" icon="record-vinyl" href="/operations/data-collection">
    Record more episodes to improve the policy.
  </Card>
</CardGroup>