Qwen (Alibaba) enters robotics with a suite of three models

Alibaba's Tongyi Lab launched the Qwen-Robot Suite, three foundation models for robotics navigation, manipulation, and world simulation now in pilot.

Alibaba's Tongyi Lab, behind the Qwen models, is tackling physical intelligence with the Qwen-Robot Suite, three foundation models designed to bridge the gap between understanding the world and acting within it. Qwen-RobotNav covers navigation, bringing together instruction following, object search, target tracking, and autonomous driving under a single model. Qwen-RobotManip handles robotic arm manipulation, using a unified state-action space that makes heterogeneous robots compatible. Qwen-RobotWorld is a world model that, based on an observation and an action described in natural language, predicts the next state of the scene.

The common thread: a natural language interface shared by all three, designed for general-purpose Qwen models to call them as tools for the physical world. For training, Manip boasts over 38,000 hours of exclusively open-source data, blending robot demonstrations and repurposed human videos.

Technical reports have been published, with RobotNav and RobotManip accompanied by public GitHub repositories, and the suite is in pilot with Alibaba Cloud customers. Qwen reports several cutting-edge results on its own benchmarks, including Manip's first-place finish in the generalist segment of RoboChallenge, pending independent confirmation.