Alibaba’s big AI team has unveiled Qwen-Robot, a suite of foundation models for embodied AI that aims to align vision-language models with physical action. The suite includes three base models: Qwen-RobotNav for navigation, Qwen-RobotManip for manipulation, and Qwen-RobotWorld for world simulation. Together, they promise multi-task, multi-robot generalization — and they’ve already shown they can work on a real robot without any fine-tuning.
The navigation model, Qwen-RobotNav, unifies instruction following, goal navigation, object tracking, and autonomous driving. It parameterizes a visual allocation strategy, allowing dynamic adjustment of visual token budgets and frame sampling during inference. Trained on 15.6 million samples, Qwen-RobotNav achieves state-of-the-art results across five navigation domains and has been deployed zero-shot on Unitree’s Go2 quadruped robot.
Qwen-RobotManip, the manipulation model, is built on a Qwen3.5-4B VL backbone with a flow-matching DiT action head, outputting 80-dimensional state-action representations for end-effector incremental poses. The team used over 38,100 hours of data — including open-source robot demonstrations, human videos, and synthetic human-to-robot transfer data — to train the model. On the LIBERO-Plus benchmark, it hits a 91.4% success rate.
For physical world prediction, Qwen-RobotWorld uses natural language as a unified robot action interface. Its architecture uses a 60-layer dual-stream MMDiT to deeply couple Qwen2.5-VL semantic representations with video latent variables. After training on 8.6 million video-text pairs, Qwen-RobotWorld ranks first on physical law compliance benchmarks like EWMBench and WorldModelBench.
All three models share a language-first interface. Alibaba also introduced Qwen-RobotClaw, a robot agent framework where a high-level planner (like Qwen-3.5) calls the suite models as physical tools to execute multi-step tasks.