Alibaba launches Qwen-Robot Suite: 3 models for embodied AI

Alibaba has unveiled Qwen-Robot Suite, a family of three foundation models designed to serve as the software backbone for next-generation autonomous robots. Developed by Alibaba's Tongyi Lab, the suite includes Qwen-RobotNav, Qwen-RobotManip, and Qwen-RobotWorld—each specializing in a core robotic capability: navigation, object manipulation, and world modeling. Qwen-RobotNav unifies five navigation tasks (instruction following, object search, tracking, autonomous driving) and was trained on 15.6 million samples, achieving 76.5% success on the challenging VLN-CE RxR benchmark and 90% tracking accuracy on EVT-Bench. Qwen-RobotManip solves the problem of incompatible action systems across different robot types by training on roughly 38,100 hours of open-source robotics and human video data, leading the RoboChallenge Table30-v1 benchmark and outperforming previous approaches by 20%.

Qwen-RobotWorld, the most ambitious component, is a language-conditioned world model that predicts how environments will evolve before a robot acts. It uses natural language as a universal interface and was trained on 8.6 million video-text pairs covering manipulation, driving, and indoor navigation. The model ranks first on EWMBench and DreamGen Bench, and outperforms open-source competitors on WorldModelBench and PBench. Importantly, these are software models, not robots; they run on hardware from AgileX, Franka, Universal Robots, and Unitree. Alibaba sees this as a strategic full-stack play (chips, cloud, AI, applications) against competitors like Google DeepMind and Nvidia. The suite is in pilot testing with selected Alibaba Cloud customers; pricing and commercial availability have not been announced.