Qwen-RobotManip: Achieving Scaling of Robot Manipulation Foundation Models Through Alignment

Qwen-RobotManip is a general-purpose Vision-Language-Action (VLA) foundation model built on top of Qwen-VL. To address the heterogeneity of existing robot data and the problem of collection costs, it introduces a Three-Dimensional Alignment framework that unifies the Representation, Motion, and Behavior dimensions.

Without using any proprietary data, this model built a pre-training corpus of approximately 38,100 hours using only open-source robot data and human demonstration videos. In particular, through a Human-to-Robot synthesis pipeline that converts 1,933 hours of egocentric human videos into 15 robot embodiments, it generated 24,808 hours of robot demonstration data.

The key achievements are as follows:

• OOD (Out-of-Distribution) Generalization: It recorded overwhelming performance compared to existing SOTA models on major benchmarks such as LIBERO-Plus (91.4%) and RoboTwin-C2R Hard (69.4%).
• Real-World Performance: It ranked 1st in the Generalist track of RoboChallenge Table30 v1, demonstrating performance 20% ahead of the 3rd-place model.
• Cross-embodiment Transfer: It overcomes differences between various robot embodiments and sensors, enabling consistent signal extraction.

As a result, Qwen-RobotManip demonstrates that data diversity and alignment are essential prerequisites for scaling robot foundation models.