Chinese tech giant Alibaba has introduced the Qwen-Robot Suite, a comprehensive set of AI models designed for robotics and tasks in physical environments. The suite includes three specialized models: Qwen-RobotNav for navigation and autonomous driving, Qwen-RobotManip for physical object interaction, and Qwen-RobotWorld for predicting scene evolution. Developers describe the project as a "full stack for embodied artificial intelligence" capable of translating natural language commands into specific physical actions.

Each model was trained on extensive datasets. Qwen-RobotNav, built on the Qwen3-VL architecture, was trained on 15.6 million samples and achieved a 76.5% success rate on the VLN-CE RxR benchmark and 90% on EVT-Bench. Qwen-RobotManip utilizes over 38,100 hours of data (including 11,320 hours of open robotics data, 1,933 hours of human action videos, and 24,808 hours of synthetic demonstrations) to unify control across various robot types, securing first place in RoboChallenge Table30 v1. Qwen-RobotWorld, trained on 8.6 million video-text pairs (comprising over 200 million frames covering more than 20 types of robotic platforms and over 500 action categories), generates probable future environmental states and took first place in EWMBench and DreamGen Bench, outperforming all open-source models in WorldModelBench and PBench.

The models are already undergoing pilot testing with Alibaba Cloud's corporate clients; however, the company has yet to disclose the public release timeline, access costs, or the list of testing clients.

Despite the high benchmark scores, experts note that mass deployment of such systems is still far off. Real-world robotics faces challenges such as sensor noise, hardware wear and tear, and the unpredictability of physical environments, while many of these tests are conducted in ideal simulations. Nevertheless, the release of the Qwen-Robot Suite marks a significant step for Alibaba in the direction of physical AI, continuing the expansion of the Qwen ecosystem following the April launch of the Qwen3.6-Plus agentic model with a 1-million-token context window.