Alibaba launches Qwen-Robot OS for robot navigation and control

Alibaba announced Qwen‑Robot on Tuesday, presenting three foundation models that the company describes as an operating system for robots. The models are named Qwen‑RobotNav, Qwen‑RobotManip and Qwen‑RobotWorld. Alibaba said the software stack can run each model independently or combine them to handle decision‑making, perception and world modelling for different robot hardware.

Qwen‑RobotNav is built to handle five navigation tasks in one model: instruction following, point‑goal navigation, object search, target tracking and autonomous driving. The company reported the model was trained on about 15.6 million samples and offers a configurable observation interface that allows planners to change parameters such as token budget, temporal decay and per‑camera weights during a task. Alibaba reported a 76.5% success rate on the VLN‑CE RxR benchmark for vision‑and‑language navigation and 90% tracking accuracy on EVT‑Bench for following moving targets.

Qwen‑RobotManip focuses on converting between different action representations used by various robots. Alibaba said it synthesized roughly 38,100 hours of training data from open‑source robot datasets and human videos rather than relying on in‑house proprietary fleet data. The model is designed to map commands across joint‑angle controls, end‑effector poses and whole‑body coordinates. Alibaba reported that Qwen‑RobotManip ranked first on RoboChallenge Table30‑v1, improving prior results by about 20%.

Qwen‑RobotWorld is a video‑based world model that accepts natural language as an action interface. Alibaba built the Embodied World Knowledge corpus of about 8.6 million video‑text pairs, roughly 200 million frames, covering manipulation, autonomous driving, indoor navigation and human‑to‑robot transfer across 14 robot arms. The company reported top scores on EWMBench and DreamGen Bench, outperformance on WorldModelBench and PBench relative to open‑source alternatives, and perfect scores on tests of physics adherence including Newtonian mechanics, mass conservation and fluid dynamics.

Alibaba emphasized that Qwen‑Robot models are software and not physical robots. The company said the models run on hardware from partners including AgileX, Franka, Universal Robots and Unitree. Internal notes described the launch as “the Android moment for robotics-the operating system, not the hardware.” The company also highlighted that the models rely heavily on open‑source datasets rather than proprietary robot data.

Alibaba did not disclose pricing, broad customer access or a timeline for general commercial deployment. The announcement said pilot programs are under way and cautioned that controlled demos and simulation benchmarks are easier than sustained real‑world operation. The company listed persistent challenges such as sensor noise, actuator drift and the long tail of edge cases that appear in uncontrolled environments.

Technical reviewers referenced specific advances in the Qwen‑Robot suite, including cross‑embodiment alignment, a parameterized observation interface for navigation, and a language‑conditioned action interface for world modelling. The company’s documentation and benchmark claims are expected to face independent evaluation on public tests and in physical deployments.