# Train a policy

Train an ACT policy from the dataset without connecting to the robot.

Execution context: Training computer.

> Verification pending: existing website examples have not completed a pinned-source or physical-hardware walkthrough. Confirm software version, hardware revision, and actual stopping procedure. Do not assume emergency-stop or release-to-stop behavior. Explicit human approval and supervision are required before physical movement.

## Choose a device

- Use cuda for a supported NVIDIA GPU.
- Use mps for Apple silicon.
- Use cpu for a CPU-only run.
- Lower batch_size if the device runs out of memory.

## Train an ACT policy

Run training on the controller or another computer that can access the dataset. sourccey-host is not required.

```shell
uv run lerobot-train \
  --dataset.repo_id=vulcan-studio/sourccey-demo-1 \
  --policy.type=act \
  --policy.device=cuda \
  --output_dir=outputs/train/act-sourccey-demo-1 \
  --job_name=act-sourccey-demo-1 \
  --batch_size=8 \
  --steps=20000 \
  --save_freq=5000 \
  --wandb.enable=false \
  --policy.push_to_hub=false
```

## Train X-VLA on a stronger GPU

If you have a stronger NVIDIA GPU with sufficient VRAM, we recommend fine-tuning the 0.9B-parameter X-VLA base model as the higher-capacity option. Install the X-VLA dependencies first:

```shell
uv sync --locked --extra sourccey-desktop --extra training --extra xvla
```

```shell
uv run lerobot-train \
  --dataset.repo_id=vulcan-studio/sourccey-demo-1 \
  --output_dir=outputs/train/xvla-sourccey-demo-1 \
  --job_name=xvla-sourccey-demo-1 \
  --policy.path=lerobot/xvla-base \
  --policy.dtype=bfloat16 \
  --policy.action_mode=auto \
  --policy.device=cuda \
  --policy.freeze_vision_encoder=false \
  --policy.freeze_language_encoder=false \
  --policy.train_policy_transformer=true \
  --policy.train_soft_prompts=true \
  --steps=20000 \
  --policy.push_to_hub=false
```

X-VLA starts from the pretrained lerobot/xvla-base checkpoint and uses bfloat16 to reduce memory use. If training runs out of memory, reduce the batch size before changing the model configuration.

## Check the output

Training should complete without dataset feature or video-decoding errors. The ACT example creates this model directory:

```shell
outputs/train/act-sourccey-demo-1/checkpoints/last/pretrained_model
```

Website: https://vulcanrobotics.ai/docs/lerobot/train

Upstream reference (unpinned): https://github.com/vulcan-forge/lerobot-vulcan/blob/main/docs/sourccey/05-train.md