StellaVLA

StellaVLA: In-Context Structured Demonstration for Generalizable Vision-Language-Action Models

StellaVLA is a vision-language-action model that uses structured demonstrations as in-context examples for robot manipulation. At test time, it retrieves a demonstration from a related task and uses its high-level plan and grounded motion information to guide the current task, which allows StellaVLA to adapt to new objects, scenes, and language instructions without fine-tuning. Each demonstration is automatically converted into a structured context containing sub-goal descriptions, image-grounded 2D traces, and verbalized 3D workspace motion, without requiring additional human annotation. This repository provides the evaluation code, pretrained checkpoints, and Docker environment needed to reproduce our results on LIBERO, LIBERO-Plus, and VLA-Arena.

Quick Links

• 📄 Paper: arXiv:2608.11671
• 🌐 Blog: StellaVLA
• 💻 Code: StellEdge-AI/StellaVLA
• 🐳 Docker: siyuhsu/stellavla:eval
• 🏆 Leaderboard: VLA-Arena

Models

Training. Both models use Qwen3-VL-4B-Instruct as the backbone. We train separate checkpoints for LIBERO and VLA-Arena, without an additional robotics pre-training stage. Each model is trained for 30K steps with a global batch size of 128 on 4×H200 GPUs.

A model directory holds config.yaml, dataset_statistics.json and checkpoints/steps_30000_pytorch_model.pt. The config must sit one level above checkpoints/ — that is how the loader finds it — and it points framework.qwenvl.base_vlm at a local copy of the backbone.

Usage

docker pull siyuhsu/stellavla:eval
mkdir -p data
docker run --rm --gpus all -v $PWD/data:/data siyuhsu/stellavla:eval fetch-assets
docker run --rm --gpus all --shm-size=16g -v $PWD/data:/data siyuhsu/stellavla:eval libero

See StellEdge-AI/StellaVLA for the native setup and the per-benchmark protocol.

Results

LIBERO

In-distribution success rate (%), 500 rollouts per suite. Baseline numbers are as reported in the original papers, except StarVLA-OFT, our matched demonstration-free control.

LIBERO-Plus

Zero-shot robustness: every model is trained on standard LIBERO and tested on the perturbed tasks without retraining. Avg. is the task-count-weighted mean over the seven perturbation categories, excluding Orig.

VLA-Arena

Mean success rate over 11 task suites at three difficulty levels — L0 (in-distribution), L1 (intermediate generalization), L2 (hardest). Training uses L0 data only. Baselines are from the VLA-Arena leaderboard; the StellaVLA row is the released checkpoint and demonstration pack, re-measured at seed 7.

Citation

@article{xu2026stellavla,
  title={StellaVLA: In-Context Structured Demonstration for Generalizable Vision-Language-Action Models},
  author={Xu, Siyu and Wang, Yunke and Wang, Zijian and Zhu, Dihao and Xia, Chenghao and Du, Chengbin and Liu, Daochang and Huang, Tao and Xu, Chang},
  journal={arXiv preprint arXiv:2608.11671},
  year={2026}
}

Acknowledgements

StellaVLA is built upon the excellent StarVLA framework. We sincerely thank the StarVLA team for open-sourcing their work and providing a strong foundation for our research.

Contact

For questions, collaborations, or support, please contact:

{s.xu,yunke.wang}@sydney.edu.au