​

Train

solo robo --type lerobot --train

​

Interactive flow

• Reuse saved Training settings or enter new ones.
• Provide dataset_repo_id (local or Hub dataset ID).
• Select policy type: SmolVLA, ACT, PI0, TDMPC, or Diffusion Policy.
• Set training steps and batch size.
• Choose output directory; if it exists, pick: resume, overwrite, or new directory.
• Choose whether to push the trained model to HuggingFace Hub (prompts for auth and repo name).
• Optionally enable Weights & Biases logging (prompts to login and set project).
• Training starts using LeRobot’s training pipeline.

​

Tips/Notes

• Some policies can start from a pretrained checkpoint; SmolVLA defaults to lerobot/smolvla_base if not provided.
• Video backend auto-falls back to PyAV if TorchCodec is unavailable.
• Checkpoints are saved regularly; you can resume later if needed.
• Press Ctrl+C to stop early; partial checkpoints may be saved.