Training Policies

solo robo --train

​

Interactive flow

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

​

Tips/Notes

• Some policies can be initialized from pretrained checkpoints; SmolVLA defaults to lerobot/smolvla_base if no checkpoint is provided.
• The video backend automatically switches to PyAV if TorchCodec is unavailable.
• Training progress is preserved via periodic checkpoints, enabling you to continue from where you left off whenever needed.

​

Support