Alibaba is expanding its Qwen model family with the Qwen-Robot Suite. The platform includes three specialized models: “Qwen-RobotManip” for manipulation tasks, “Qwen-RobotNav” for navigation, and “Qwen-RobotWorld,” a video-world model for simulating future motion sequences.

The suite is designed to help AI systems perceive physical environments, translate voice commands into specific actions, and make decisions in real time. Intended applications include, among others, industrial robotic arms, delivery robots, and robotic dogs.

Training with Millions of Data Points

Qwen-RobotManip is based on "Qwen3.5-4B VL" and was trained using more than 38,000 hours of freely available robotics data. This includes data from robotics repositories, videos of human manipulation tasks, and synthetically generated human-to-robot datasets. According to Alibaba, the model improves transferability between different robot platforms by a factor of three and requires only minimal retraining for different hardware.

Qwen-RobotNav is based on Qwen3-VL and was trained on 15.6 million curated examples. The data covers, among other things, trajectory planning and visual-linguistic reasoning. The model serves as a navigation engine for agent-based systems that perform long-term tasks. As an example, Alibaba cites applications such as embodied question answering, in which robots answer questions about their environment.

For Qwen-RobotWorld, 8.6 million video-text pairs were used, comprising more than 200 million individual frames, over 20 embodiment types, and 500 action categories. The model is designed to predict physically plausible motion sequences, generate synthetic training data, and simulate movements before they are executed by robots.

Selected corporate customers in the robotics industry are already testing the Qwen Robot Suite as part of pilot projects.

Focus on Physical AI

With the new suite, Alibaba is expanding the use of its Qwen models beyond traditional AI applications to what is known as “Physical AI.” The goal is for general language and multimodal models to work together with robotics models. As an example, the company cites the search for a lost item in a real-world location: While a general Qwen model plans the task, Qwen-RobotNav handles navigation and exploration of the environment.