Alibaba Launches Qwen-Robot: A New Operating System for Robotics

Summary

• Alibaba has introduced the Qwen-Robot Suite, comprising three AI models aimed at enhancing robot navigation, manipulation, and physics-based simulations through a cohesive software framework.
• The company claims its models excel across various robotics benchmarks, utilizing millions of training samples and extensive open-source robot data.
• However, practical deployment of these robots in real-world settings is still several years away.

On Tuesday, Alibaba's Qwen team launched the Qwen-Robot Suite, which consists of three foundational models designed to create what they describe as a "comprehensive stack for embodied intelligence." The Qwen-RobotNav model focuses on mobility, while Qwen-RobotManip is geared towards manipulation, and Qwen-RobotWorld is responsible for simulating the necessary physics. Each model operates autonomously, but collectively, they represent a significant advancement in robotics akin to the Android platform in mobile technology.

📣 Introducing the Qwen-Robot Suite — Qwen-RobotNav, Qwen-RobotManip, Qwen-RobotWorld, three foundational models, a comprehensive stack for embodied intelligence.

🧭 Qwen-RobotNav — the key to mobility.
• Integrates five navigation tasks into a single model: instruction following, point-goal navigation,… pic.twitter.com/noumjTtTeS

— Qwen (@Alibaba_Qwen) June 16, 2026

Currently, Alibaba stands out in China as the sole company with a broad range of capabilities including chips, cloud infrastructure, models, service platforms, and applications. Robotics represents the most tangible embodiment of their investment in what is termed embodied AI.

Presently, AI agents depend on large language models (LLMs) for decision-making. Traditional robotic systems utilize machine-learning models, which, despite their sophistication, lack the flexibility offered by generative AI. Physical robots encounter a distinct set of challenges related to physics rather than simple prompts.

To address these challenges, Alibaba has rolled out this innovative AI suite with various components:

Qwen-RobotNav integrates five navigation tasks—instruction following, point-goal navigation, object search, target tracking, and autonomous driving—each requiring different visual memory techniques. Unlike most models that implement a single strategy, Qwen-RobotNav provides a parameterized interface that allows planners to adjust settings like token budget and temporal decay in real-time during operations.

The model has been trained on 15.6 million samples, achieving a 76.5% success rate on VLN-CE RxR, a benchmark for visual-and-language navigation in real-life settings, and 90% tracking accuracy on EVT-Bench, which measures an agent’s ability to follow moving targets consistently.

Qwen-RobotManip addresses a significant challenge in robotic manipulation: differing robots interpret actions in fundamentally different manners. For example, a Franka arm uses joint angles for operation, while an ALOHA robot relies on the positioning and orientation of its grippers. Humanoid robots introduce additional complexity by using whole-body coordinates.

To reconcile these varied action representations, Alibaba compiled around 38,100 hours of training data from open-source robot datasets and human videos, without utilizing proprietary data. The model has achieved a top ranking on RoboChallenge Table30-v1, surpassing prior methods by 20%.

Qwen-RobotWorld represents the most ambitious aspect of the suite: a language-conditioned video world model that treats natural language as a universal interface for actions. For instance, the command "Pick up the red cup and pour water on the flower" can be executed by a gripper, an autonomous vehicle, or a mobile navigation unit.

The Embodied World Knowledge corpus includes 8.6 million video-text pairs—200 million frames—covering manipulation (5.9 million samples, over 1,300 skills, and more than 20 morphologies), autonomous driving (with datasets from Waymo, NVIDIA PhysicalAI-AD, and Bench2Drive), indoor navigation (VLNVerse), and human-to-robot transfer involving 14 different robot arms.

It has achieved top scores on EWMBench and DreamGen Bench, two benchmarks assessing the ability of world models to accurately predict and generate realistic physical environments. It also outperforms all open-source models on WorldModelBench and PBench, achieving perfect scores in physics adherence, including Newton's laws and fluid dynamics.

Could This Be the ChatGPT for Robotics?

While Western research labs, including Google DeepMind and Nvidia, are pursuing similar objectives, they primarily focus on either navigation or manipulation rather than a unified suite. Alibaba’s comprehensive integration from chips to applications allows them to maintain control over the entire stack. Their use of open-source foundations sets them apart from competitors who depend on proprietary robot data.

It’s important to clarify a few misconceptions: these models are not robots themselves but rather software frameworks—brains without bodies. They operate on hardware developed by companies like AgileX, Franka, Universal Robots, and Unitree.

Furthermore, although these are generative AI models for robotics, they are not LLMs comparable to ChatGPT. Language models predict token sequences, whereas these models must comprehend physics, spatial relationships, and the outcomes of physical actions. For example, while a language model can state that a glass will break if dropped, Qwen-RobotWorld can predict the shattering pattern, fluid dynamics, and potential secondary collisions. Qwen-RobotManip plans a grasp that prevents the drop altogether.

However, don’t anticipate having a domestic robot assistant in the near future. The disparity between a controlled demonstration of a robot placing fruit in a basket and a robot that can reliably function in a home environment is substantial. Simulation benchmarks like RoboCasa365, LIBERO-Plus, and RoboTwin-Clean2Rand highlight this challenge. Real-world deployment introduces issues such as sensor noise, actuator drift, and numerous edge cases that have historically challenged robotics efforts, and Alibaba acknowledges this reality.

Nevertheless, the technical progress is significant. RobotManip’s alignment-first strategy addresses a critical bottleneck in cross-embodiment training. RobotNav’s parameterized observation interface cleverly tackles the context-strategy dilemma. RobotWorld’s language-as-universal-action-interface serves as a valuable abstraction for cross-domain world modeling.

Alibaba has not yet revealed pricing, timelines, or details regarding which customers will gain access beyond initial pilot programs.

Daily Debrief Newsletter

Stay updated every day with the latest news stories, along with original features, podcasts, videos, and more.