Today, artificial intelligence can write text, analyze images, and carry on conversations. Things get more challenging when it’s expected not only to think but also to act. It is precisely at this intersection between the digital and physical worlds that Alibaba is focusing its efforts with its new Qwen Robot Suite. The expansion of the Qwen model family is designed to help robots perceive their environment, make decisions, and perform tasks independently.

The suite comprises three core models: Qwen-RobotManip for manipulation tasks, Qwen-RobotNav for navigation, and Qwen-RobotWorld, a so-called video world model for embodied intelligence.

When AI Gets Its Hands Dirty

Large multimodal models are now impressively capable of processing text, speech, and images. However, translating these capabilities into precise physical actions remains one of the greatest challenges in robotics. This is because robots must not only understand voice commands but also translate them into concrete movements, perceive unfamiliar environments, and handle objects they have never seen before.

See, Understand, Act

According to Alibaba, the three models achieve industry-leading results in dozens of recognized robotics benchmarks. These include RoboChallenge, a large-scale benchmark for embodied intelligence using real robots. Selected enterprise customers of Alibaba Cloud in the robotics industry are already testing the Qwen-Robot Suite in pilot projects.

Qwen-RobotManip, codenamed Lira and Atlas, is based on Qwen3.5-4B VL and was trained using more than 38,000 hours of open-source data. The training data comes from sources including robotics repositories, videos of human manipulation tasks, and synthetically generated human-to-robot datasets. The model improves upon the previous state of the art in transfer learning between different robot platforms by a factor of three. As a result, it can be deployed on various robot hardware with minimal retraining.

Qwen-RobotNav is based on Qwen3-VL and was trained on 15.6 million curated examples. The data covers trajectory planning as well as visual-linguistic reasoning. The model serves both as a scalable navigation engine and as a unified interface for agent-based navigation systems. This makes it particularly well-suited for agent-based systems that handle long-term tasks, such as embodied question answering. In this context, a robot answers questions about its environment—for example, where a specific object was placed.

Based on current observations, Qwen-RobotWorld predicts physically plausible future visual motion sequences. Alibaba trained the model using 8.6 million video-text pairs. These encompass more than 200 million frames, over 20 embodiment types, and 500 action categories. The model can generate synthetic video training data for robots and enable systems to simulate future motion sequences before execution. This capability is particularly well-suited for robotic manipulation, embodied planning, and complex indoor navigation.

From Chatbot to Key Player

With the Qwen Robot Suite, Alibaba is bringing its Qwen architecture from the digital world into the realm of Physical AI. The technology company is thus shifting its focus away from simple chatbots and toward autonomous agents that handle complex tasks in both the digital and physical worlds.