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
arXiv:2605.30280 · 146▲ · cs.RO, cs.AI, cs.CL
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
Qwen-VLA is a single vision-language-action model that puts a DiT-based flow-matching action decoder on top of the Qwen3.5-4B multimodal backbone, so one model handles robot manipulation, navigation, and trajectory prediction. It uses embodiment-aware text prompts (describing the robot platform, arm config, control frequency, and prediction horizon) instead of separate output heads per robot, and is trained in four stages: text-to-action decoder pretraining without images, multimodal continued pretraining, supervised fine-tuning, and PPO reinforcement learning.
Why it matters
Instead of maintaining separate specialist policies per robot and task family, a team could deploy one checkpoint across multiple embodiments by swapping a text prompt, and fine-tuning from the pretrained base substantially improves real-world sample efficiency and out-of-distribution robustness compared to training from scratch. The staged recipe (freeze backbone, train the randomly-initialized decoder on language-to-action first) is a practical answer to the instability of jointly training a pretrained VLM with a fresh action head.
Practical takeaway
Watch for VLA models converging on the pattern of a frozen or lightly-tuned VLM backbone plus a flow-matching action decoder conditioned by text prompts rather than per-robot architectures. If you fine-tune such a model for a real robot, expect the pretrained base to matter far more than the architecture: their real-world average jumped from 48.5% to 83.6% purely from pretraining transfer.
Key result
On real-world ALOHA out-of-distribution tasks, the pretrained variant hit 76.9% average success versus 41.5% for the pi-0.5 baseline and 36.2% for the same architecture trained from scratch. Note most manipulation gains come from the SFT stage; the final RL step added only about +2.9 points on the SimplerEnv environment it was trained in and near-zero on held-out benchmarks, and cross-benchmark comparisons pit one generalist against separately-tuned specialists.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.