Alibaba has launched the Qwen-Robot Suite, a set of AI models enhancing robotic capabilities in navigation, object manipulation, and scene prediction.
Alibaba has unveiled the Qwen-Robot Suite, a comprehensive set of AI models aimed at improving robotic capabilities in physical environments. This suite includes Qwen-RobotNav for navigation, Qwen-RobotManip for object manipulation, and Qwen-RobotWorld for scene prediction. The initiative is described as a "full stack for embodied artificial intelligence."
📣 Introducing the Qwen-Robot Suite — Qwen-RobotNav, Qwen-RobotManip, Qwen-RobotWorld, three foundation models, a full stack for embodied intelligence.
🧭 Qwen-RobotNav — the gateway to mobility.
• Unifies 5 navigation tasks in one model: instruction following, point-goal,… pic.twitter.com/noumjTtTeS
— Qwen (@Alibaba_Qwen) June 16, 2026
The suite aims to enhance robots' ability to perceive their environments, plan actions, and execute commands using natural language. Currently, the Qwen-Robot Suite is undergoing pilot tests with select clients in the robotics sector through Alibaba Cloud.
Purpose Behind Qwen's Development
While existing large language and multimodal models excel in handling digital data, they fall short in the physical realm where robots operate. These physical agents require the capability to not only comprehend commands but also to interpret them into actionable movements, factoring in spatial awareness, object characteristics, and sensor limitations.
Alibaba refers to this intersection of technology as physical AI or "embodied AI," emphasizing the need for models to interact effectively with the real world, including moving, locating objects, and understanding the repercussions of their actions.
Qwen-RobotNav: Multifaceted Navigation
Qwen-RobotNav serves as the navigation model, consolidating five distinct task groups:
- following instructions;
- point-to-point movement;
- object location;
- target tracking;
- autonomous driving.
Built on the Qwen3-VL foundation, it has been trained on 15.6 million samples pertaining to route planning and visual-language reasoning, achieving a success rate of 76.5% on VLN-CE RxR and 90% on EVT-Bench. The model can also integrate with larger systems, where a higher-level model orchestrates tasks and Qwen-RobotNav manages movement.
Demonstrations showcase the model's ability to execute tasks like finding lost items indoors or verifying the status of specific objects, requiring the robot to move and collect visual evidence for user queries.
Qwen-RobotManip: Streamlined Object Handling
Qwen-RobotManip focuses on enabling robots to perform physical tasks involving objects, such as picking, moving, and placing items, while facilitating skill transfers across different robotic systems.
This model addresses the challenge of inconsistent action descriptions among various robots, standardizing data so that training on one model can benefit others. Alibaba utilized over 38,100 hours of varied training data, including open robotics and human action videos.
Qwen-RobotManip reportedly excelled in the RoboChallenge Table30 v1, demonstrating adaptability to new instructions and unfamiliar objects.
Qwen-RobotWorld: Predictive Capabilities
Qwen-RobotWorld is designed to predict environmental changes following specific actions, functioning as a video world model guided by natural language inputs.
The model processes current observations alongside text commands to generate likely future states of the environment, applicable for various robotic tasks including navigation and planning. Training involved compiling a vast corpus of video-text pairs and frames covering numerous robotic platforms and action categories.
Alibaba announced that Qwen-RobotWorld ranked highly on various benchmarks, demonstrating a strong alignment with fundamental physical principles.
Challenges Ahead for Widespread Robotics Adoption
Despite promising results, the Qwen-Robot Suite remains a collection of models rather than a fully operational consumer robotics platform. Obstacles such as sensor inaccuracies, actuator wear, and unpredictable scenarios still pose significant challenges for real-world deployment.
Alibaba has yet to disclose pricing, public launch timelines, or details about the clients testing the Qwen-Robot Suite. In April, Alibaba Cloud also introduced the Qwen3.6-Plus agent model, which features a context window of 1 million tokens and support for external tools.
