Step 7: Training — smolvla Command

Before running

• Environment: first open an instance and upload the dataset as described in Cloud GPU training environment setup; note that smolvla requires extra dependencies, see "Install the environment" below
• Dataset: --dataset.root=~/lerobot_my_dataset_shake_hands in the command points to the handshake dataset collected in step 6. If you are training your own task, replace it with your own dataset name
• Two training approaches: fine-tuning from a pretrained model usually gives better results and faster convergence; training from scratch does not require downloading the pretrained weights. Choose as needed
• You can check the curves on wandb at any time during training, see View real-time training curves on wandb

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=<username>/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=<username>/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

Related Products

SO-ARM101 Developer Kit→