SMART-VLA

SMART-VLA is a PRTS-architecture vision-language-action model pretrained on SMART-Data for articulated-object manipulation and sim-to-real transfer.

Model Details

• Base model: Qwen/Qwen3-VL-4B-Instruct
• Architecture: PRTS_Qwen3VL
• Training data: SMART-Data
• Parameters: approximately 4.44B
• Precision: bfloat16
• Action chunk size: 50
• Maximum action dimension: 32
• Released checkpoint: checkpoint-final-278972

Links

• Paper: SMART: Zero-Shot Sim-to-Real Articulated Object Manipulation via Large-Scale Synthetic Pretraining
• Project page: https://teamillusion-smart.github.io/
• Dataset: TeleEmbodied/SMART-Data
• Code and loading instructions: TeleHuman/PRTS

Usage

Please follow the installation, loading, and inference instructions in the PRTS repository.

Intended Use

SMART-VLA is intended for research on vision-language-action pretraining, articulated-object manipulation, and simulation-to-real robot learning.

License

This model is released under the MIT License.

Citation

@article{SMART2026ao,
  title={SMART: Zero-Shot Sim-to-Real Articulated Object Manipulation via Large-Scale Synthetic Pretraining},
  author={Jicong Ao and Shuhan Jiang and Yuling Zhong and Yanwen Liu and Yuhan Gao and Jiangyuan Zhao and Yang Zhang and Shiqiang Zhu and Chenjia Bai and Xuelong Li},
  year={2026},
  eprint={2610.07652},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2610.07652},
}