Fangqi Technology (方奇科技) raises tens-of-millions-RMB angel round for a semantic world-model robot brain
The AMW Read
A new, small angel-stage entrant restating the already-known foundation-model-for-embodied thesis with an early corporate pilot, not a baseline-shifting event.
Fangqi Technology (方奇科技) raises tens-of-millions-RMB angel round for a semantic world-model robot brain
Beijing-based Fangqi Technology, founded in 2026 by a team of Tsinghua University alumni under founder Dr. Wang Xinzhou, closed a tens-of-millions-RMB angel round (roughly single-digit millions of dollars) backed by Tsinghua Star Capital (启迪之星创投) and other investors. The funds go to core R&D, team expansion, and commercializing what the company calls a semantic world model — a shared three-dimensional space that fuses vision, language, and physical state so a robot can predict outcomes and decide actions the way a person would. Training runs on a "Turing learning" curriculum of course, task, practice, and reflection stages, aimed at learning from human data and real practice rather than pure teleoperation demonstrations. Fangqi placed second at Stanford's BEHAVIOR 2026 household-task challenge and has signed pilot agreements with Hong Kong-listed service-robot maker Youibot (优地机器人, 03231.HK) and Jiangsu Longhuan, using commercial floor cleaning as its first deployment wedge.
China's embodied-AI field is thick with "general robot brain" claims, and most lean on scaling teleoperated demonstration data. Fangqi's framing — that the real bottleneck is the split between a world model's physical grounding and a language model's semantic knowledge, not data volume — puts it alongside other groups chasing foundation-model-style generalization across robot bodies and tasks instead of single-purpose control policies. A Tsinghua-pedigreed team with an early listed-company pilot is a concrete, if early, data point for that framing; it doesn't settle whether transferable cross-embodiment skill libraries can outcompete brute-force data scaling.
For investors, this is a bet on the software layer of embodied AI rather than hardware — Fangqi doesn't build robots, it licenses a learning-and-control stack meant to transfer across robot bodies, a platform play rather than a single-product one. For builders, the signal to watch is whether the Youibot cleaning pilot converts into a paid commercial deployment; that would be the first real test of whether a cross-embodiment skill library can be sold as infrastructure rather than stay a research demo.