What tasks do the Qwen-Robot models address?

The Qwen-Robot models cover navigation, object manipulation, and future state prediction. They integrate language understanding with robotic physical actions, ensuring versatility across diverse scenarios.

What data was used to train Qwen-RobotManip?

Qwen-RobotManip was trained on over 38,100 hours of open-source data, enabling large-scale learning across various robotics platforms.

How does Qwen-RobotWorld differ from other models in the series?

Qwen-RobotWorld is a universal model that predicts future states in navigation, driving, and manipulation tasks. It bridges language and visual comprehension with physically consistent scenario predictions.

← All news

Artificial intelligence

Alibaba Unveils Qwen-Robot Series: Three Foundational Models for Robotics

June 17, 2026

Photo: TechNode

Quick answer

Alibaba launched the Qwen-Robot series with three models: Qwen-RobotNav (navigation), Qwen-RobotManip (manipulation), and Qwen-RobotWorld (state prediction).

Alibaba has unveiled the Qwen-Robot series, a set of models designed to integrate language commands with physical actions in robotics. The lineup includes three foundational models: Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld, each tailored to specific applications.

Qwen-RobotNav specializes in mobile robotics, combining computer vision and language processing capabilities. The model supports four key functions: executing instructions, navigating to targets, object tracking, and autonomous driving. This enables robots to interact effectively with their surroundings.

Qwen-RobotManip standardizes the state and action space by representing manipulator movements in camera coordinates. Trained on a vast dataset of over 38,100 hours of open-source data, the model supports large-scale learning across diverse platforms and expands the range of manipulation tasks.

The third model, Qwen-RobotWorld, serves as a universal “world” solution. It links language and visual understanding with future state prediction, enabling the model to forecast physically consistent scenarios in navigation, driving, and manipulation. This versatility makes it applicable to a wide array of robotics tasks.

Common questions

What tasks do the Qwen-Robot models address?: The Qwen-Robot models cover navigation, object manipulation, and future state prediction. They integrate language understanding with robotic physical actions, ensuring versatility across diverse scenarios.
What data was used to train Qwen-RobotManip?: Qwen-RobotManip was trained on over 38,100 hours of open-source data, enabling large-scale learning across various robotics platforms.
How does Qwen-RobotWorld differ from other models in the series?: Qwen-RobotWorld is a universal model that predicts future states in navigation, driving, and manipulation tasks. It bridges language and visual comprehension with physically consistent scenario predictions.

Dzen feed: /feed/dzen.xml · RSS: /feed.xml