Alibaba's Qwen team unveiled the Qwen-Robot Suite on Tuesday, comprising three foundation models for embodied intelligence: Qwen-RobotNav for navigation, Qwen-RobotManip for manipulation, and Qwen-RobotWorld for physics simulation. The suite aims to unify diverse robotic tasks under a single software stack, trained on millions of samples and tens of thousands of hours of open-source robot data. Qwen-RobotNav achieves 76.5% success on VLN-CE RxR and 90% tracking on EVT-Bench, while RobotManip ranks first on RoboChallenge Table30-v1, outperforming prior methods by 20%. RobotWorld tops EWMBench and DreamGen Bench, with perfect physics adherence. Despite these technical milestones, Alibaba acknowledges that real-world deployment remains years away due to sensor noise, actuator drift, and edge cases. The open-source foundation differentiates Alibaba from competitors relying on proprietary data. No pricing or timelines have been disclosed beyond pilot programs.

Key facts

• Three foundation models: Qwen-RobotNav, Qwen-RobotManip, Qwen-RobotWorld for navigation, manipulation, and simulation.
• Trained on 15.6 million samples for navigation and 38,100 hours of cross-embodiment manipulation data.
• Qwen-RobotManip ranks first on RoboChallenge Table30-v1, outperforming prior methods by 20%.
• Qwen-RobotWorld achieves perfect physics adherence and top rankings on multiple benchmarks.
• Real-world deployment remains years away due to sensor noise, actuator drift, and edge cases.

KeyAudit data perspective