From Stockpiling Data to Low-Cost Learning

One of the biggest misconceptions in embodied AI is that piling up enough real-robot data will, on its own, produce general intelligence.

Today, every time a robot swaps in a different body, gripper, or actuator, large volumes of data typically have to be collected again. Training costs are high, adaptation cycles are long, and the data rarely transfers between devices. ForgeAI, a developer of general-purpose brains for embodied AI, has announced an angel round worth tens of millions of yuan aimed at that problem, with investors including TUS Star Ventures.

The proceeds will mainly go toward iterating on the company's semantic world model and Turing learning paradigm, validating and commercializing core technology in representative scenarios, and continuing to build a general-purpose skill operating system spanning multiple robot bodies and use cases.

Funding Directed at R&D and Deployment

Founded in February 2026, ForgeAI positions itself as a developer of "intelligent brains" that let robots understand the physical world, break down tasks, and keep learning.

The company says three priorities follow the round: refining the semantic world model so robots can grasp human common sense and physical environments at the same time; advancing the Turing learning paradigm to cut dependence on manual teleoperation and large-scale real-robot data; and accelerating product validation in commercial cleaning, property services, and similar settings.

ForgeAI says it has taken close to 10 million yuan in pre-orders in the few months since its founding. Founder and CEO Wang Xinzhou said investors and industrial customers increasingly care about real-robot metrics, pilot data, and actual orders, with the embodied AI sector moving from concept demos toward engineering validation and commercial deployment.

A Team Drawn From Research and Industry

ForgeAI's core team comes largely from Tsinghua University, the Chinese Academy of Sciences, and other research institutions, alongside industry figures with overseas startup experience. Master's and PhD holders make up a high share of the team, which brings experience in large models, physical AI, robotics algorithms, and product engineering.

Wang Xinzhou graduated from Tsinghua University's computer science department and previously worked on Tencent's Hunyuan 3D model and physical AI algorithm architecture. Around 2025, he began focusing on data collection and generalization in embodied AI.

As he sees it, large language models carry rich semantic knowledge but lack an understanding of the real physical world, while traditional world models can simulate physical processes yet lack sufficient human common sense. Simply bolting the two together tends to lose critical physical understanding as information passes between them.

The Semantic World Model as a "Translator for the Physical World"

ForgeAI's semantic world model attempts to encode human common sense and physical parameters into a single shared representation space.

A robot, for instance, needs to know not only that "glass is fragile" but also understand friction, torque, and material hardness, then use that to judge how to grasp, move, and set down an object. The company wants this approach to give robots an understanding of the physical world closer to a human's.

On that foundation, ForgeAI has proposed what it calls a Turing learning paradigm, modeled on the human cycle of preparation, practice, and reflection. A robot first picks up general knowledge from public material such as textbooks and internet video, then learns specific skills from a small number of human demonstrations, and finally practices, makes mistakes, and draws conclusions on its own in real environments.

The goal is to turn human behavior into general-purpose data robots can learn from, reducing the need for conventional teleoperation-based instruction.

A Cloud Super-Brain Plus On-Device Task Agents

Architecturally, ForgeAI uses a "cloud super-brain plus on-device task agent" model.

The cloud handles large-scale model training, inference, and skill accumulation, organizing experience gathered across different scenarios into a reusable skill library. The device side handles lightweight deployment and real-time execution based on the specific task and hardware constraints.

The aim of this architecture is "one brain, many forms": a single set of intelligence capabilities that works across wheeled robots, humanoids, robotic arms, and other hardware.

The company says its Turing learning method cuts the cost of acquiring a single skill by roughly 80% compared with conventional teleoperation-based collection, speeds up training and deployment by about 40%, and compresses the adaptation cycle for a given robot body to around two weeks. Those figures still need further validation across more customers and larger-scale projects.

Commercial Cleaning as the First Deployment Market

On commercialization, ForgeAI has picked three-dimensional commercial cleaning as its technology validation scenario.

Cleaning property and commercial spaces is not simple repetitive motion. It usually calls for environment recognition, litter pickup, stain assessment, grasping, wiping, and path planning. The pace of cleaning work is also relatively forgiving, leaving robots room to make mistakes and adjust while operating.

ForgeAI is working with partners including UDI Robotics and Jiangsu Longhuan, and is running a property cleaning pilot in Tianjin Eco-City. According to the company, the pilot has reached wiping coverage above 90%, a litter pickup rate above 95%, and visible stain removal above 85%, with human takeover held between 10% and 15%.

On the company's numbers, a single robot is expected to replace 1 to 2.5 cleaning staff, with a target payback period of about 18 months.

ForgeAI is pursuing two business models: full system integration and delivery for property customers without in-house robotics hardware capability, and software brains, task capabilities, and skill subscriptions for robot manufacturers.

Making Skills Reusable Across Robot Bodies

Robots differ noticeably in mechanical structure, gripper size, joint range, and sensor configuration, so a single skill often cannot transfer directly. That is widely seen as one of the key obstacles to scaling embodied AI.

ForgeAI plans to use human behavior as an intermediate "virtual reference template," first abstracting specific motions into general human behavior patterns, then mapping and adapting them to the structure of each robot body.

The company says it has already built a working pipeline from human video and demonstration data to robot skills. The next stage focuses on reverse abstraction of robot skills and cross-body reuse, with an architecture for skill learning, accumulation, and distribution targeted for completion by the end of 2027.

Productivity Tools First, General Platform Later

ForgeAI plans to deliver a commercially usable prototype by the end of 2026, complete its skill learning, accumulation, and distribution system by mid-2027, and push into scaled deployment by the end of 2027, targeting more than 200 robots in the field.

Wang argues that automakers entering the humanoid robot field do not necessarily displace software companies. Automakers bring manufacturing and supply chain strength, while software teams can own the robot operating system, the intelligent brain, and the skill ecosystem. The industry may end up resembling the hardware-software split of the smartphone era.

He expects a shakeout in embodied AI to unfold around 2027. In the near term, commercial services and industrial settings remain the best environments for training and deploying robots, and mass adoption of household humanoids will take considerably longer.

ForgeAI says it will keep iterating on its cloud-native embodied AI brain, expand its general skill library, and push robots from one-off demos toward continuous learning, cross-body reuse, and deployment at scale.