Alibaba Is Building Qwen-Robot: The Operating System for the Robot Economy

Alibaba’s Qwen team launched the Qwen-Robot Suite on Tuesday: three basic models form what it calls a “full stack for embodied intelligence.” Qwen-RobotNav ensures mobility. Qwen-RobotManip takes care of manipulation. Qwen-RobotWorld simulates the physics that makes both possible. Each works independently. Together they represent the Android moment for robotics: the operating system, not the hardware.

Alibaba is currently the only company in China that includes chips, cloud, models, serving platforms and applications. For the company, robotics is the most physical expression of that bet, what’s known as embodied AI.

AI agents currently rely on LLMs to make their decisions. The usual way robots work is through machine learning models which, while advanced, lack the adaptability of generative AI. Physical agents face a different, more difficult class of failure modes: physics, not cues.

For these use cases, Alibaba introduced this new AI suite with several components:

Qwen-RobotNav combines five navigation tasks (instruction following, point-goal navigation, object search, target tracking, and autonomous driving), each requiring different visual memory strategies. Most models code one strategy. Qwen-RobotNav exposes a parameterized interface: token budget, temporal decay, weights per camera that a scheduler can reconfigure mid-delivery.

Trained on 15.6 million samples with randomization across all parameters, it achieves 76.5% success on VLN-CE RxR, a benchmark for vision and language navigation in real environments, and 90% tracking on EVT-Bench, which evaluates an agent’s ability to consistently track moving targets.

Qwen-RobotManip addresses one of the biggest challenges in robot manipulation: different robots represent actions in fundamentally different ways. A Franka arm (a type of robot with seven axes of motion) acts via joint angles, while an ALOHA robot (a low-cost bimanual robot platform widely used in robotics research) represents actions by the position and orientation of its grippers (end-effector poses). Humanoids add an extra layer of complexity, using whole-body coordinates.

To bridge these incompatible action spaces, Alibaba synthesized approximately 38,100 hours of training data from open-source robot datasets and human videos – without relying on proprietary data collection. The model ranks first in RoboChallenge Table30-v1 and performs 20% better than previous approaches.

Qwen-RobotWorld is the most ambitious: a language-conditioned video world model that treats natural language as a universal action interface. “Pick up the red cup and pour water on the flower” works regardless of whether the actor is a grabber, an autonomous vehicle, or a mobile navigation agent.

The Embody World Knowledge corpus includes 8.6 million video-text pairs (200 million frames) for manipulation (5.9 million examples, 1,300+ skills, 20+ morphologies), autonomous driving (Waymo, NVIDIA PhysicalAI-AD, Bench2Drive), indoor navigation (VLNverse), and human-to-robot handoff via 14 robotic arms.

It ranks first on EWMBench and DreamGen Bench, two benchmarks that evaluate whether world models predict and generate realistic physical environments. It also beats all open source models on WorldModelBench and PBench, and scores perfectly in physics: Newton’s laws, mass conservation, fluid dynamics, gravity.

The ChatGPT of robots?

While Western labs (Google DeepMind, Nvidia, Figure, Physical Intelligence) pursue similar goals, most focus on navigation or manipulation, rather than a uniform, composable suite. Alibaba’s vertical integration from chips to applications means they control the entire stack. The open-source foundation differentiates itself from competitors that rely on private robot data.

There are some misconceptions worth clearing up: These are not robots, but software models: brains, not bodies. They run on hardware from AgileX, Franka, Universal Robots, Unitree and others.

Despite these being generative AI models for robots, these are not LLMs like the typical ChatGPT. A language model predicts tokens. These models must understand the physics, spatial relationships, and consequences of physical actions. A language model tells you that a glass breaks when it falls. Qwen-RobotWorld predicts how it breaks: crush patterns, fluid dynamics, secondary collisions. Qwen-RobotManip plans a grip that completely prevents the fall.

Don’t expect to have your own household robot anytime soon. The gap between a controlled demo of a robot placing fruit in a basket and a robot working reliably in your home is enormous. RoboCasa365, LIBERO-Plus, RoboTwin-Clean2Rand: these are simulation benchmarks. Real-world implementation introduces sensor noise, actuator drift, and the long tail of edge cases that have humiliated every robotic effort in history, and Alibaba recognizes this.

However, the technical achievements are real. RobotManip’s alignment-first approach solves a real bottleneck in cross-embodiment training. RobotNav’s parameterized observation interface is a smart solution to the context strategy problem. RobotWorld’s language-as-universal-action interface is the right abstraction for cross-domain world modeling.

Alibaba did not reveal pricing, timelines, or which customers will get access outside of the pilot programs.