Post-95 Tsinghua PhD Leaves Tencent to Build Robot Brains as Marketable Skills
On September 21, Fangqi Technology, a developer of general-purpose embodied intelligence brains, announced a tens-of-millions-yuan angel round. Investors included TusStar Ventures and others.
A round of that size is hardly eye-catching in the 2026 embodied AI funding frenzy. But Liu Bo, TusStar Ventures' general manager and managing partner, invested anyway. His rationale: competition hinges on exploring low-cost, scalable general robot learning paths. Liu believes Fangqi's semantic world model and Turing learning paradigm offer new ideas, though whether the path becomes a foundational software platform still depends on cross-embodiment reuse rates and customer willingness to pay.
No body, only brain. Founder Wang Xingzhou was born in 1996. He earned bachelor's and master's degrees at Beihang University and a PhD under Professor Sun Fuchun at Tsinghua's Department of Computer Science. Before founding Fangqi, he was deeply involved in Tencent Hunyuan's 3D model and physical AI algorithm architecture. The core team comes from Tsinghua and the Chinese Academy of Sciences, covering world models, VLA algorithms, and cloud-native systems. The strategy is clear: not bodies, but robot brains.
Wang says Fangqi does not build a brain for one robot model. It breaks brain capabilities into reusable skills. Its general skill operating system uses a cloud super brain plus edge task agent architecture, aiming for "one brain, multiple forms"—the same brain adapting to different robot bodies. The hard part is not architecture but generalization. If every new body or scenario requires re-adaptation, the concept collapses.
On why embodied intelligence generalizes poorly, Wang judges that world models lack human knowledge while embodied brains lack physical common sense, splitting physical and semantic knowledge. This points to Fangqi's core: a semantic world model that deeply couples embodied large models with world models, mapping vision, language, and physical states into a unified 3D semantic space. Robots can then not only recognize objects but understand what actions mean in context. Traditional world models are like an apprentice who only reads recipes—knowing to stir-fry onions for three minutes but not what an onion is, why to fry it, or when it is done. Fangqi's model installs a kitchen brain in that apprentice.
Fangqi also proposes a Turing learning paradigm with four stages: curriculum learning, task learning, practice learning, and reflection learning. In validation, it placed second globally in Stanford's BEHAVIOR 2026 household challenge.
Commercially, Fangqi has partnered with Hong Kong-listed UDI Robot and Jiangsu Longhuan Hotel Property, entering via 3D cleaning scenarios with orders near ten million yuan. Wang cited a hotel delivery robot firm needing object grasping and elevator-pressing skills: the industry norm is 9 to 12 months, while Fangqi took 2 to 3 months. 3D cleaning is not the most capital-favored scenario, but Fangqi sees it as a training ground—cross-form, unstructured, multi-surface, demanding high generalization—while clients care more about ROI than lab precision.
In H1 2026, China's embodied AI sector raised about 43.8 billion yuan. Companies with embodied brains as their core business raised 22.253 billion yuan (50.8%); robot body firms raised only 5.6 billion yuan (12.8%). Unitree launched UniStore, turning skills into marketable, distributable goods mainly for its own body ecosystem. Fangqi aims to build a skill platform for multiple body makers—a skill exchange. Cross-embodiment transfer is a known challenge: joint structures, degrees of freedom, and kinematic topologies differ greatly. Whether each body still needs extensive adaptation, and whether body makers will open necessary data and interfaces, are prerequisites for the business model.