Qwen-VLA

Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments

Qwen Team

📑 Technical Report | 📖 Blog | 🖥️ Demo

Welcome to the official repository of Qwen-VLA. Here, you can find official information about Qwen-VLA and post your questions (Issues).

🎬 Demo

demo.mp4

💡 Introduction

Qwen-VLA is a unified vision-language-action generalist model built upon Qwen3.5-4B (vision-language backbone) and a 1.15B DiT flow-matching action decoder. It casts manipulation, navigation, and trajectory prediction into a shared action-and-trajectory prediction framework, enabling a unified model to learn from heterogeneous embodied data across tasks, environments, and robot embodiments via embodiment-aware prompt conditioning, no per-platform output heads needed.

A unified Qwen-VLA generalist matches or outperforms task-specific specialists fine-tuned independently per benchmark across multiple simulation and real-world evaluations, pushing embodied intelligence from "skill specialists" toward "generalist actors."

✨ Key Highlights

• 🏆 One Generalist Beats Specialists. A unified model matches or outperforms per-benchmark specialists across multiple simulation and real-world evaluations.

• 🔗 Unified Action-and-Trajectory Framework. Manipulation, navigation, egocentric action modeling, and trajectory prediction share one action-and-trajectory prediction space.

• 🤖 Embodiment-Aware Prompt Conditioning. One set of weights serves multiple platforms; switching embodiments requires only changing a text prompt.

• 📈 Progressive Training Recipe. A progressive training recipe that includes large-scale action pretraining, multimodal continued pretraining, supervised fine-tuning, and reinforcement learning, bridging the gap between discrete vision-language tokens and continuous action trajectories.

• 🌍 Strong Real-World OOD Generalization. Large-scale embodied pretraining enables robust generalization to unseen conditions in real-world deployment, significantly outperforming specialist baselines.

🏆 Benchmarks

Manipulation & Navigation

As a unified generalist policy, Qwen-VLA is trained once on all embodiments jointly and evaluated across all platforms without per-benchmark adaptation, simultaneously handling both manipulation and navigation.

Model	LIBERO	RoboCasa-GR1	Simpler-WidowX	RoboTwin-Easy	RoboTwin-Hard	R2R OS	R2R SR	RxR SR
Qwen-VLA-Base	90.8	40.4	64.3	64.3	66.4	61.7	53.8	55.1
Qwen-VLA-Instruct	97.9	56.7	73.7	86.1	87.2	69.0	57.5	59.6

Out-of-Distribution Generalization

Model	SimplerEnv-OOD SR (%)	DOMINO SR (%)	DOMINO MS (%)
Qwen-VLA-Base	25.3	21.1	37.4
Qwen-VLA-Instruct	32.0	26.6	39.5

SimplerEnv-OOD: fine-tuned solely on simple pick-and-place, evaluated on unseen spatial and visual tasks. DOMINO: zero-shot evaluation on dynamic manipulation with moving objects, no dynamic training data used.

Real-World Results

On the ALOHA bimanual platform, GR00T N1.6 and π_0.5 are per-task specialist models fine-tuned independently, while Qwen-VLA is a unified all-in-one generalist that handles all tasks, embodiments, and modalities within one unified model.

In-Domain Performance (Success Rate %)

Model	Pick & Place	Table Cleaning	Bowl Stacking	Bowl Pick & Place	Towel Folding	Fine-grained	Avg
GR00T N1.6	30.8	38.5	53.8	19.2	19.2	10.3	28.6
π0.5	73.1	84.6	88.5	69.2	80.8	33.3	71.6
Qwen-VLA-aloha (w/o pretrain)	30.8	53.8	61.5	64.1	50.0	30.8	48.5
Qwen-VLA-aloha (w/ pretrain)	96.2	92.3	98.7	87.2	65.4	61.5	83.6

OOD Performance (Success Rate %)

Model	Color	Instance	Position	Background	Instruction	Avg
GR00T N1.6	46.2	38.5	3.8	19.2	19.2	25.4
π0.5	57.7	61.5	19.2	26.9	42.3	41.5
Qwen-VLA-aloha (w/o pretrain)	42.3	30.8	34.6	30.8	42.3	36.2
Qwen-VLA-aloha (w/ pretrain)	88.5	76.9	53.8	80.8	84.6	76.9

📜 Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwenvla,
      title={Qwen-VLA: Unifying Vision-Language-Action Modeling across Tasks, Environments, and Robot Embodiments}, 
      author={Qiuyue Wang and Mingsheng Li and Jian Guan and Jinhui Ye and Sicheng Xie and Yitao Liu and Junhao Chen and Zhixuan Liang and Jie Zhang and Xintong Hu and Xuhong Huang and Pei Lin and Junyang Lin and Dayiheng Liu and Shuai Bai and Jingren Zhou and Jiazhao Zhang and Haoqi Yuan and Gengze Zhou and Hang Yin and Ye Wang and Yiyang Huang and Zixing Lei and Wujian Peng and Delin Chen and Yingming Zheng and Jingyang Fan and Xianwei Zhuang and Xin Zhou and Haoyang Li and Anzhe Chen and Tong Zhang and Xuejing Liu and Yuchong Sun and Ruizhe Chen and Zhaohai Li and Chenxu Lü and Zhibo Yang and Tao Yu and Xionghui Chen},
      year={2026},
      eprint={2605.30280},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2605.30280}, 
}