Fangqi Tech Raises Angel Round, Proposes Semantic World Model for Embodied AI
On September 21, Fangqi Technology announced the completion of a multi-million angel funding round, with investors including TusStar Venture Capital and other institutions.
The funds will primarily be used for core technology research and development and expansion of the R&D team, as well as to accelerate the validation and commercialization of key technologies in typical scenarios.
The company plans to continuously develop a general-purpose skill operating system for robots that supports sustainable learning and cross-embodiment transfer, to build momentum for a general-purpose robot brain, and to gradually construct infrastructure for embodied capability learning, accumulation, and distribution across multiple robot embodiments and scenarios.
The disconnect between physical knowledge and semantic knowledge is the root cause of the weak generalization ability of current embodied robots. According to Wang Xinzhou, founder of Fangqi Technology, world models lack human knowledge, while embodied brains lack physical common sense. In short, there is a disconnect between physical and semantic knowledge between world models and embodied models. This judgment points to a specific problem in embodied intelligence: the world model is responsible for predicting how the physical world will change, while the embodied model is responsible for translating predictions into actions. If the two operate independently, robots will encounter situations where they can understand but cannot act, or can act but do not know why.
To address this, Fangqi Technology proposes a semantic world model, aiming to create robots that learn, work, and grow like humans. The semantic world model integrates human knowledge into the prediction of physical world changes, while the embodied model translates these predictions into actions with physical common sense. This integrated approach enables robots to understand both the 'what' and the 'why' behind their actions.
Fangqi Technology is the first in the industry to propose the Turing learning paradigm, which builds a chain of curriculum learning, task learning, practical learning, and reflective learning. This breaks the dilemma of traditional teleoperation learning where robots only know the 'how' but not the 'why,' allowing robots to continuously learn from human data and real-world practice, and establishing a scalable and replicable learning paradigm.
The core of this paradigm is to divide learning into several stages: first learning basic courses, then specific tasks, then entering real practice, and finally consolidating experience through reflection. The problem with traditional teleoperation learning is that robots only imitate actions without understanding the reasons behind them and the conditions for their applicability, making them prone to failure when the environment changes. The Turing learning paradigm attempts to enable robots to build an understanding of tasks at each stage, rather than just memorizing action sequences. Through reflection, robots can transform experience into reusable knowledge.
According to the company, its core technology has completed technical validation of Turing learning and semantic world models in commercial service scenarios, applied for multiple technology patents, and ranked second in the world in the Stanford household challenge BEHAVIOR 2026.
The company has established strategic partnerships with UDI Robot, Jiangsu Longhuan, and others to advance pilot implementations. It will use three-dimensional cleaning as a training ground and gradually implement pilots in commercial service and other scenarios. These collaborations will accelerate the deployment of the technology in real-world applications.
Liu Bo, general manager and managing partner of TusStar Venture Capital, stated that the core of competition in embodied intelligence lies in whether the industry can move away from the old path of piling up massive real-machine data and find a low-cost, scalable general-purpose robot learning path. The semantic world model and Turing learning paradigm proposed by Fangqi Technology bring new solutions to the industry.
The core team members all come from Tsinghua University, covering the entire chain of world models, VLA algorithms, cloud-native systems, and industrial commercialization. This team provides a solid talent foundation for the company's technology research and commercialization.