Training Command Line - smolvla (Recommended for Advanced Users)

Reference documentation

https://huggingface.co/docs/lerobot/smolvla

https://github.com/huggingface/lerobot/blob/46e19ae579f80ce66211afafd1c3c649c569131f/docs/source/policy_smolvla_README.md

Install the environment

cd lerobot
pip install -e ".[feetech,smolvla]"

Fine-tuning from a pretrained model (recommended)

lerobot-train \
  *--policy.path*=lerobot/smolvla_base \
  --dataset.repo_id=Tommymy/lerobot_my_dataset_shake_hands \
  --dataset.root=~/lerobot_my_dataset_shake_hands \
  --dataset.revision=v0.1.0 \
  --dataset.streaming=false \
  --policy.type=smolvla \
  --output_dir=~/output_lerobot_train/shake/smolvla_A \
  --job_name=shake_smolvla_a \
  --policy.device=cuda \
  --wandb.enable=true \
  --wandb.project=Lerobot_my_Project \
  --policy.push_to_hub=false \
  --steps=40000 \
  --batch_size=8

Training from scratch

lerobot-train \
  --dataset.repo_id=Tommymy/lerobot_my_dataset_shake_hands \
  --dataset.root=~/lerobot_my_dataset_shake_hands \
  --dataset.revision=v0.1.0 \
  --dataset.streaming=false \
  --policy.type=smolvla \
  --output_dir=~/output_lerobot_train/shake/smolvla_A \
  --job_name=shake_smolvla_a \
  --policy.device=cuda \
  --wandb.enable=true \
  --wandb.project=Lerobot_my_Project \
  --policy.push_to_hub=false \
  --steps=40000 \
  --batch_size=8

Download the model

The smolvla model archive is about 1 GB