Add preferred source
Fangqi Technology, a developer of universal embodied intelligence “brains,” has completed an angel round of funding worth tens of millions of yuan, with investors including TusStar Venture Capital. Founded in 2026 and led by Tsinghua PhD Wang Xinzhou, the company’s team is predominantly Tsinghua-affiliated and includes veterans who contributed to Tencent’s Hunyuan 3D model development. Fangqi has proposed a combined architecture of a “Semantic World Model + Turing Learning Paradigm,” using humans as the greatest common denominator to enable cross-embodiment skill transfer. Its technical approach employs a cloud-based brain, edge-side cerebellum, and real-time digital twin world model, while explicitly de-emphasizing reliance on teleoperation data. For commercialization, the company is using 3D cleaning as its proving ground, having already secured strategic partnerships with UDI Robotics and Jiangsu Longhuan, with orders worth tens of millions of yuan. Fangqi plans to launch a commercial service prototype by end-2026 and scale deployment in 2027, with a long-term goal of building embodied intelligence infrastructure for the physical world.
Key Elements
Fangqi Technology (Beijing) Co., Ltd., a developer of universal embodied intelligence “brains,” recently announced the completion of an angel round of funding worth tens of millions of yuan (approximately $1.5 million+), with investors including TusStar Venture Capital and other institutions. The funds will be allocated to core technology R&D, team expansion, and accelerating validation and deployment of key technologies in representative commercial scenarios.
Founded in 2026 and operational since May of that year, Fangqi Technology positions itself as a “robot brain” company. Unlike most peers focused on a single robot form factor, the company aims to decompose brain capabilities into reusable skill modules that can adapt to wheeled robots, humanoids, robotic arms, and other embodiments. Founder Wang Xinzhou uses an analogy to explain this positioning: the company makes “screwdrivers” that can be sold to renovators or factory workers, rather than making tools specialized only for renovation.
Technical Approach: Semantic World Model and Turing Learning
Fangqi’s core assessment is that the current embodied intelligence industry suffers from weak generalization, rooted in a disconnect between world models and embodied models—a split between physical knowledge and semantic knowledge. World models lack human knowledge, while embodied “brains” lack physical common sense.
To address this bottleneck, the company has proposed a “Semantic World Model” that deeply couples embodied large models with world models, unifying vision, language, and physical states into a three-dimensional semantic space. This enables robots to both understand the world and, like humans, predict actions and make autonomous decisions.
At the model training level, Fangqi is among the first to propose a “Turing Learning Paradigm,” which builds a chain of four learning stages—curriculum learning, task learning, practice learning, and reflective learning—to break through the limitation of traditional teleoperation-based learning that only teaches robots “what” without understanding “why.” The Semantic World Model and Turing Learning together form an “evolutionary flywheel” that drives the continuous growth of the cloud-native embodied brain.
In terms of overall architecture, the company adopts a “cloud brain + edge cerebellum + real-time digital twin world model” approach. The cloud brain is a multimodal large model responsible for understanding human intent and actions; the edge cerebellum handles concrete execution; and the real-time digital twin world model serves as middleware, abstracting the physical world into concise representations to help the brain understand the environment. The entire architecture is embodiment-agnostic and, apart from specific adaptations, can be reused across different robot forms.
Cross-embodiment skill transfer is a widely acknowledged industry challenge. Wang Xinzhou points out that the difficulty lies in the “last mile”: different robots have different morphologies and gripper sizes—even a 1-centimeter difference in gripper length can affect skill transfer effectiveness. Fangqi’s solution is to use humans as the “greatest common denominator,” unifying actions from different robots into human-equivalent actions, constructing a virtual “reference machine” akin to a universal robot template, then concretizing it to wheeled, humanoid, and other forms. According to Wang, developing a robot typically takes about six months, but adapting to a specific robot using the reference template takes only about half a month at the algorithm level.
Data Strategy: Moving Away from Teleoperation Dependence
In terms of data acquisition, Fangqi emphasizes a human-centric approach, with the goal of minimizing teleoperation data and ultimately eliminating dependence on it. Currently, the company relies primarily on open-
Wang Xinzhou believes that true commercial deployment cannot depend on teleoperation. Hardware iterations cause teleoperation data utility to plummet—for example, if a gripper is shortened by 5 centimeters, data utility may drop by nearly 80-90%—and hardware is constantly being iterated. Additionally, teleoperation is costly and time-intensive, making it impossible to cover all trajectories.
He also notes that another underappreciated direction in the industry is Agent intelligent orchestration. Real-world tasks are long-sequence; robot work involves a series of chained tasks. A single VLA (Vision-Language-Action) model or a single world model cannot handle this alone—a brain is needed to orchestrate and schedule tasks.
Commercialization: 3D Cleaning as the Proving Ground
Fangqi has chosen “3D cleaning” as its first commercial proving ground. The company explains that bathroom cleaning is one of the most complex scenarios, involving wiping mirrored surfaces, cleaning irregularly shaped sinks, and other tasks that train robots in skills like handling cloths and manipulating irregular objects—skills that can generalize to other scenarios. At the same time, 3D cleaning tasks are complex and non-standardized, with low cadence requirements and tolerance for trial and error, making them suitable as training scenarios. In contrast, logistics and factory settings demand extremely high success rates and cadence, which actually makes it harder to demonstrate the brain’s advantages.
Fangqi has already established strategic partnerships with UDI Robotics (03231.HK) and Jiangsu Longhuan to advance pilot deployments. According to Wang Xinzhou, the company has secured orders worth tens of millions of yuan and has established two business models: selling complete robots and selling Skills.
The company’s core technology has completed technical validation of Turing Learning and the Semantic World Model in commercial service scenarios, with multiple patent applications filed. It also ranked second globally in the Stanford BEHAVIOR 2026 household task challenge.
In terms of customer strategy, Fangqi targets two groups: first, existing hardware manufacturers or hardware companies lacking full embodied intelligence development capabilities, to whom it provides brains and skills; and second, vertical industry customers such as Longhuan Property. The current focus is on commercial service scenarios, which demand higher skill transferability and can leverage the “one brain, many forms” advantage.
Fangqi’s core team is predominantly Tsinghua-affiliated, covering the full chain of world models, VLA algorithms, cloud-native systems, and industrial commercialization. Over 90% of the team holds master’s or doctoral degrees, with PhDs exceeding 25%. Founder Wang Xinzhou graduated from Tsinghua University’s Department of Computer Science and previously worked at Shengshu Technology and Tencent, where he was deeply involved in the design of Tencent’s Hunyuan 3D model and physical AI algorithm architecture. He brings both “0-to-1” startup experience and “1-to-N” systematic training from a major tech company.
Liu Bo, General Manager and Managing Partner of TusStar Venture Capital, stated that the core competitive question in embodied intelligence is whether the industry can move beyond the old path of “piling up massive real-machine data” and find a low-cost, scalable universal robot learning approach. Fangqi’s Semantic World Model and Turing Learning Paradigm offer the industry a new problem-solving framework, and TusStar sees potential for the company to build the next-generation foundational software platform for embodied intelligence.
According to the company’s plan, a commercial service prototype will be launched by the end of 2026, with the goal of achieving “genuinely usable offline” performance. By mid-2027, the company aims to refine the architecture for skill learning, accumulation, and distribution, and achieve real-world deployment. By the end of 2027, it plans to begin scaling, reaching at least 200 units. Within three years, the goal is to achieve universal robots for a specific industry in scenarios such as hotels or commercial services.
Wang Xinzhou’s assessment of industry trends is that the elimination round will begin as early as 2027. Mass adoption of home humanoid robots is still 3 to 5 years away, with commercial services and industrial scenarios serving as the best near-term proving grounds. He believes the true endgame for robotics is to “learn like humans, think like humans, and be developed like computers.”
In the short term, the company will focus on polishing deployable productivity tools. The long-term goal is to build embodied intelligence infrastructure for the physical world, enabling robots to learn autonomously and grow continuously. With this funding round completed, Fangqi Technology will continue to advance the R&D and commercial deployment of universal robot brains, centered on its “Semantic World Model + Turing Learning” approach.
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
