Alibaba is not building a robot this time. The Qwen team is trying to build the software layer a robot can carry: a three-model Qwen-Robot Suite for navigation, manipulation and world prediction.
The positioning matters. The models are meant to run on hardware made by others, including AgileX, Franka, Universal Robots and Unitree. In Alibaba's own framing, this is closer to an Android moment for robots than to a new robot launch.
Three models, three jobs
Qwen-RobotNav handles movement: instruction following, point navigation, object search, target tracking and autonomous driving. It was trained on 15.6 million samples and reports 76.5% success on VLN-CE RxR and 90% on EVT-Bench tracking.
Qwen-RobotManip handles hands. Built on Qwen3.5-4B VL, it aims to transfer skills across robot bodies whose action spaces normally do not match. It used about 38,100 hours of open and synthesized data and ranked first on RoboChallenge Table30-v1, 20% above the previous best method.
Qwen-RobotWorld handles prediction. It is a video world model that takes natural language as an action interface and predicts how the physical scene changes next, trained on 8.6 million video-text pairs, or about 200 million frames.
The Android analogy has a point
Alibaba does not want to own every robot body. It wants the common software base that many bodies can install. That is why open data and cross-hardware transfer are central to the announcement. The suite is already being piloted with some Alibaba Cloud enterprise customers, but pricing and wider availability remain undisclosed.
The hard part starts after the demo. Sensors drift, actuators age and factories produce edge cases that benchmarks rarely cover. The suite is best read as a serious bid for the robot operating layer, not proof that general-purpose robots are solved.
Sources: Decrypt, South China Morning Post, GIGAZINE, CocoLoop; source checks covered the June 16 Qwen-Robot Suite release, RobotNav, RobotManip, RobotWorld, training data, benchmark figures and supported robot platforms.