Alibaba launched Qwen-Robot, an embodied intelligent platform that uses natural language to align multi-domain physical actions for zero-sample deployment.

According to Bijie.com, Alibaba's Big Model Team has released the Qwen-Robot suite, a suite of embodied intelligence underlying models. It includes three base models: qwen-robotnav, qwen-robotmanip, and qwen-robotworld, corresponding to the three physical action domains of navigation, manipulation, and world simulation, respectively. The suite aims to align visual language models with physical actions, achieving embodied generalization across multiple tasks and robot types. The navigation model qwen-robotnav, trained on 15.6 million samples, achieved state-of-the-art (SOTA) performance in five navigation domains and has been deployed with zero samples on the Unitree go2 quadruped robot. The manipulation model qwen-robotmanip achieved a 91.4% success rate in the libero-plus benchmark. The physical world prediction model qwen-robotworld ranked first in benchmarks such as ewmbench and worldmodelbench. All three models provide a language-first interface.