smolvla_so101_packing_reproduce

SmolVLA — SO101 Multi-Object Packing (Reproduce)

A reproduction of SmolVLA (Hugging Face's ~450M-parameter vision-language-action model) trained with LeRobot on the SO101 robot arm multi-object packing task.

Part of a study reproducing published robot-learning policies to understand and improve on them.

Model details

• Policy: SmolVLA (smolvla), ~450M parameters, safetensors
• Observation: joint state + 3 cameras (camera1, camera2, camera3)
• Action: action chunk of 50 steps (n_action_steps=50)
• License: Apache-2.0

Training

Reproduce with LeRobot:

lerobot-train \
  --policy.type=smolvla \
  --dataset.repo_id=Qiu-Xinchuan/so101_Multi-object-packing \
  --steps=100000 \
  --batch_size=8 \
  --policy.optimizer_lr=1e-4 \
  --seed=1000 \
  --job_name=smolvla_packing

Usage

lerobot-eval \
  --policy.path=Porameht/smolvla_so101_packing_reproduce \
  --env.type=...  # or run on a real SO101 via lerobot-record / robot control scripts

Load directly:

from lerobot.policies.factory import make_policy
from lerobot.configs.policies import PreTrainedConfig

config = PreTrainedConfig.from_pretrained("Porameht/smolvla_so101_packing_reproduce")
policy = make_policy(config, ds_meta=...)  # dataset meta from Qiu-Xinchuan/so101_Multi-object-packing

Evaluation

Not yet evaluated on hardware — success-rate numbers will be added after SO101 rollouts. See the companion ACT reproduction: Porameht/act_so101_packing_reproduce.

Citation

@article{shukor2025smolvla,
  title={SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics},
  author={Shukor, Mustafa and others},
  journal={arXiv preprint arXiv:2506.01844},
  year={2025}
}