SmolVLA — object top shelf (SO101)

Fine-tuned SmolVLA policy for placing an object on the top shelf with a SO101 arm (6D joint-space actions, 3 cameras).

Base model: lerobot/smolvla_base
Dataset: lorenzouttini/object_top_shelf_remote
Training: 50k steps, batch 64, Isambard GH200 (1× GPU)
Training run: W&B — smolvla_object_top_shelf

Trained and pushed with LeRobot. See LeRobot Docs.

How to use

Fine-tune further

lerobot-train \
  --policy.path=pravsels/smolvla_object_top_shelf \
  --dataset.repo_id=${HF_USER}/mydataset \
  --batch_size=64 \
  --steps=20000 \
  --output_dir=outputs/train/my_smolvla \
  --job_name=my_smolvla_training \
  --policy.device=cuda \
  --wandb.enable=true

Run on robot

lerobot-record \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.id=my_follower_arm \
  --robot.cameras="{ top: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}, wrist: {...}, front: {...} }" \
  --dataset.single_task="Place the object on the top shelf." \
  --dataset.repo_id=${HF_USER}/eval_object_top_shelf \
  --dataset.episode_time_s=50 \
  --dataset.num_episodes=10 \
  --policy.path=pravsels/smolvla_object_top_shelf