Qwen-RobotManip: Alignment Unlocks Scale in Robot Manipulation Foundation Models
Original title:Qwen-RobotManip: Alignment Unlocks Scale for Robotic Manipulation Foundation Models
AI summary
Qwen-RobotManip is a vision-language-action (VLA) foundation model from Qwen based on Qwen-VL, introducing a unified alignment framework covering representation, motion and behavior. Using only open-source robot datasets and human demonstration videos, it builds a pre-training corpus of about 38,100 hours covering 15 robot embodiments. It achieves 91.4% on LIBERO-Plus, 69.4% on RoboTwin-C2R Hard, 14.9% on RoboCasa365 Composite-Unseen, 45.6% on EBench and 72.0% on RoboTwin-IF, and won the RoboChallenge Table30 v1 generalist track. The model uses an 80-dimensional state-action representation, a human-to-robot data synthesis pipeline (converting 1,933 hours of first-person video into 24,808 hours of data) and in-context policy adaptation.
Key Takeaways
- Qwen-RobotManip is a vision-language-action (VLA) foundation model from Qwen based on…
- Using only open-source robot datasets and human demonstration videos,it builds a pre-training…
- It achieves 91.
- 4% on LIBERO-Plus,69.
- 4% on RoboTwin-C2R Hard,14.
Quick Facts
- Published
- Updated
- Source
- Qwen:Blog Retrieval(API)
- Category
- Model Releases
- AI score
- 72
Qwen 这次发布的机器人模型,用统一对齐框架把跨实体数据规模化训练跑通了,OOD 泛化大幅领先,做具身智能的值得认真看一下。
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-RobotManip: Alignment Unlocks Scale in Robot Manipulation Foundation Models. June 16, 2026. https://new.nat6.net/en/news/cmqg5kvq2006hslncx7x7o1r2/