Qwen-VLA: toward a unified action framework for general embodied intelligence
AI summary
Tongyi Lab proposed Qwen-VLA, building a unified vision-language-action model with a Qwen3.5-4B vision-language backbone and a 1.15B-parameter DiT action decoder. Through text-to-action DiT pretraining and proprioceptive prompts, it unifies manipulation, navigation, and trajectory prediction under one framework, supporting 11 robot platforms. Across 5 simulation benchmarks, a single general model surpasses the best specialized model on 3; on the ALOHA real robot, in-domain success rate is 83.6% and OOD generalization 76.9%, exceeding π₀.₅ by over 35 and 40 percentage points respectively; DOMINO dynamic manipulation reaches 26.6% zero-shot; VLN-CE navigation reaches 57.5% on R2R and 59.6% on RxR, both surpassing specialized models.
Key Takeaways
- Tongyi Lab proposed Qwen-VLA,building a unified vision-language-action model with a Qwen3.
- 5-4B vision-language backbone and a 1.
- 15B-parameter DiT action decoder.
- Through text-to-action DiT pretraining and proprioceptive prompts,it unifies…
- Across 5 simulation benchmarks,a single general model surpasses the best specialized model on 3
Quick Facts
- Published
- Updated
- Source
- 公众号:通义实验室(千问)
- Category
- Model Releases
- AI score
- 61
通义把操作、导航和轨迹预测塞进一个脑子,在11种机器人上通用,这是具身智能从'专家'走向'通才'的关键一步,做机器人的值得翻翻论文。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Qwen-VLA: toward a unified action framework for general embodied intelligence. May 29, 2026. https://new.nat6.net/en/news/cmq2o71l001irsl6nww43cve6/