Train a policy

• 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.

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

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

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:

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

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.

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

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