Robotics

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

Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, Sicheng Xie, Yitao Liu, Junhao Chen, Zhixuan Liang, Jie Zhang, Xintong Hu, Xuhong Huang, Pei Lin, Junyang Lin, Dayiheng Liu, Shuai Bai, Jingren Zhou, Jiazhao Zhang, Haoqi Yuan, Gengze Zhou, Hang Yin, Ye Wang, Yiyang Huang, Zixing Lei, Wujian Peng, Delin Chen, Yingming Zheng, Jingyang Fan, Xianwei Zhuang, Xin Zhou, Haoyang Li, Anzhe Chen, Tong Zhang, Xuejing Liu, Yuchong Sun, Ruizhe Chen, Zhaohai Li, Chenxu Lü, Zhibo Yang, Tao Yu, Xionghui Chen

Featured June 18, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

This paper creates a single smart robot brain, called Qwen-VLA, that can understand spoken commands, see the world, and control many different robots for tasks like picking things up or moving around, all by using special text hints for each robot and learning actions like drawing a path.

In depth

The paper introduces Qwen-VLA, a unified vision-language-action model that extends a vision-language backbone to generate continuous robot actions. It achieves this by formulating diverse embodied tasks (manipulation, navigation, trajectory prediction) into a shared action-and-trajectory space and using embodiment-aware prompt conditioning to adapt to different robot platforms and control conventions within a single model. A staged training recipe, including text-to-action DiT pretraining, helps bridge the gap between high-level language instructions and low-level continuous actions.

Key Takeaways

1
Qwen-VLA unifies heterogeneous embodied decision-making problems (manipulation, navigation, trajectory prediction) into a single model by representing actions in a shared space.
2
It employs embodiment-aware prompt conditioning to allow a single model to control diverse robot platforms and control conventions without architectural changes.
3
A staged training recipe, starting with text-to-action pretraining, effectively bridges the gap between discrete language tokens and continuous action trajectories, improving stability and transfer.

Conceptual Flow

HIGH LEVEL

Methodology (The "Logic")

The robot brain takes in what it sees, what you tell it to do, and a special note about its body, then figures out the best actions to take.

What Robot Sees

What You Say

Robot Body Info

Think and Plan

Robot Actions

Results (The "Impact")

This new brain helps robots do many different jobs better than specialized robots, even in new situations, showing it's a good general helper.

Old Separate Robot Brains

New Unified Robot Brain

Compare Performance

Better Task Success

Works for New Tasks

Adapts to New Robots