Alibaba launches embodied intelligent base Qwen Robot, which aligns multi domain physical actions with natural language to achieve zero sample deployment According to Beating monitoring, the Alibaba Big Model team has released the embodied intelligence underlying model suite Qwen Robot Suite, which includes three base models: Qwen RobotNav, Qwen RobotManip, and Qwen RobotWorld, corresponding to three physical action domains: navigation, operation, and world simulation. The kit aims to align visual language models with physical actions, achieving embodied generalization across multiple tasks and models. The navigation model Qwen RobotNav unifies tasks such as command following, target navigation, target tracking, and autonomous driving. In terms of design, the model parameterizes the visual allocation strategy and supports dynamic adjustment of control axes such as visual token budget and frame sampling during inference. After training on 15.6 million samples, Qwen RobotNav achieved SOTA in 5 navigation domains and has been deployed on the Yutu Go2 quadruped robot with zero samples. The operation model Qwen RobotManip is constructed based on the Qwen 3.5-4B VL backbone network and flow matching DiT action head, using an 80 dimensional state action representation to output incremental pose of the end effector. The team used over 38100 hours of data (including open-source robot demonstrations, human videos, and human-machine migration synthesis data) for training and achieved a success rate of 91.4% in the LIBERO Plus evaluation. The physical world prediction model Qwen RobotWorld adopts a natural language unified robot action interface. In terms of architecture, the model is based on a 60 layer dual stream MMDiT to deeply couple the Qwen2.5-VL semantic representation with video latent variables. After training on 8.6 million video text pairs, Qwen RobotWorld ranked first in physical law compliance evaluations such as EWMBench and WorldModelBench. All three models provide language first interfaces. Alibaba also launched the robot intelligent agent framework Qwen RobotClaw, which uses upper level planners (such as Qwen-3.5) to call suite models as physical tools to achieve multi-step operations. [Original link]