Read next
EqualOcean learned on September 21 that ForgeAI（方奇科技）, a developer of general-purpose brains for embodied AI, recently completed an eight-figure angel funding round backed by TusStar（启迪之星创投）and other investors.
ForgeAI
The proceeds will primarily support core technology development and the expansion of the R&D team, while accelerating validation and commercialization in representative use cases. The company aims to further develop a “general-purpose robot skills operating system” that can learn continuously and transfer capabilities across different robot embodiments.
Founded in 2026, ForgeAI has taken a different approach from many startups entering the market through complete humanoid robots. Rather than focusing on robot hardware, it has chosen to compete at another level: the robot’s “brain.”
Its approach addresses a question facing the embodied AI industry: as robots become increasingly capable of walking, running, and performing individual actions, how can they understand tasks in real-world environments, work continuously, and transfer learned capabilities to different robots?
This is the problem ForgeAI hopes to solve.
Over the past two years, the most visible advances in humanoid robotics have largely involved the “body”: more degrees of freedom in joints, increasingly dexterous hands, falling prices for complete robots, and continuing improvements in movements such as running and rolling.
But embodied AI involves more than robot hardware.
In our previous mapping of the embodied AI ecosystem, general-purpose brains, data training and iteration, and hardware and control systems all emerged as important parts of the industry alongside robot bodies. General-purpose brains encompass capabilities ranging from foundation models and vision-language-action (VLA) models to world models and task planning.
Source: EqualOcean Analysis
This framework also helps explain ForgeAI's position. It is not a typical robot hardware company. Instead, it seeks to operate above the hardware layer, providing a brain and a set of skills that different robots can use.
ForgeAI is not alone in envisioning this division of responsibilities.
In July 2026, Google DeepMind released Gemini Robotics 2, further separating robot intelligence into distinct layers. Gemini Robotics ER 2 primarily handles understanding of the physical world and complex, multistep task planning, serving as a high-level “brain,” while VLA models handle movement execution. Its on-device VLA model emphasizes rapid adaptation to new robot embodiments. According to Google, Gemini Robotics On-Device 2 can adapt to a new robot embodiment using just a few hours of data.
NVIDIA’s Isaac GR00T is closer to a full-stack robotics development platform. It brings robot foundation models, real and synthetic data, simulation environments, middleware, and on-device inference into a single development framework, with an explicit emphasis on cross-embodiment capabilities. The latest GR00T N1.7 is trained on data from multiple robot types, including dual-arm systems, semi-humanoid robots, and humanoid robots. It also incorporates human egocentric video and can be adapted through post-training to new robot embodiments, tasks, and environments.
The latest technical approaches of these leading companies point to two shared trends in embodied AI brains. First, the industry is moving away from expecting a single model to handle everything and toward dividing responsibilities between high-level reasoning and planning and low-level action execution. Second, it is moving beyond serving a single embodiment toward capabilities that generalize across different robot bodies.
The logic is straightforward: if every new robot or use case requires fresh data collection and model retraining, it will be difficult to meaningfully reduce development costs and deployment times.
ForgeAI is targeting precisely this challenge of cross-embodiment generalization.
Founded in 2026, the company focuses on developing general-purpose brains for embodied AI. Its core team members all come from Tsinghua University, with expertise spanning world models, VLA algorithms, cloud-native systems, and commercialization. They combine frontier research experience with practical deployment experience.
Founder Dr. Wang Xinzhou said that following the funding round, the company will continue improving the core technology of its cloud-native embodied AI brain and recruiting leading R&D talent. It will also accelerate pilot validation in flagship use cases and build a library of general-purpose skills that can be reused across embodiments. In the near term, it will focus on practical productivity tools. Its long-term goal is to build embodied AI infrastructure for the physical world, enabling robots to learn autonomously and keep improving.
This is no easy task.
Humanoid robots, wheeled robots, and robotic arms differ significantly in their physical structures. Even among robotic arms, differences in degrees of freedom, arm length, and gripper design mean that the same movement cannot simply be copied from one to another.
ForgeAI describes its solution as “one brain, multiple embodiments.”
According to the company, its approach does not directly transfer movements from Robot A to Robot B. Instead, it first introduces an intermediate layer: human motion.
Robot actions are first abstracted into a representation of human motion. This relatively standardized representation is then mapped onto different robot embodiments. In other words, ForgeAI aims to replace direct “Robot A → Robot B” transfer with “Robot A → Generalized Human Motion Representation → Robot B.”
Wang likens this to a virtual robot “reference platform.”
Individual robots still require embodiment-specific adaptation; connecting them to the model does not make them immediately ready to use. However, ForgeAI says this approach can significantly shorten the adaptation process compared with developing a robot’s movement capabilities from scratch. In some current projects, algorithm adaptation takes approximately two weeks.
Within this architecture, ForgeAI also proposes combining a “semantic world model” with “Turing learning.”
The semantic world model seeks to organize visual, linguistic, and physical-state information within a unified 3D semantic space, enabling robots not only to recognize objects but also to understand their surroundings and the potential consequences of their actions. “Turing learning” encompasses curriculum learning, task learning, practice, and reflection, with the aim of enabling robots to learn continuously from human data and real-world operations.
Data therefore plays an important role in this approach.
In conventional robot training, much of the action data comes from teleoperation: a person controls a robot to complete a task, and the resulting trajectory is used to train the model. The advantage is that the action data closely matches the robot. However, it also means that the data can become tied to a particular embodiment.
NVIDIA likewise uses a mixture of data sources for robot training. It draws heavily on internet videos and human videos at the foundational level, adds synthetic data in the middle, and uses real robot teleoperation data at the level closest to robot execution. Its latest GR00T models are also continuing to expand their coverage of different robot embodiments.
ForgeAI's approach follows a similar direction, but seeks to reduce reliance on teleoperation more aggressively.
The company told EqualOcean that its training data currently consists primarily of open-source data, human videos it has cleaned itself, egocentric data, and synthetic data. Teleoperation data is used mainly to align actions with specific embodiments. Its central view is that if robot learning depends too heavily on action data generated by a particular embodiment, the reusability of those data assets will decline whenever the hardware changes.
For ForgeAI, therefore, the ultimate test of “one brain, multiple embodiments” is a concrete metric: how much existing data and how many learned skills can still be reused when moving to a different robot?
Beyond its technical approach lies a more practical question: how does a robot brain company make money?
ForgeAI currently serves two main types of customers.
The first consists of robot manufacturers that already have hardware but lack comprehensive embodied AI development capabilities. ForgeAI supplies the brain and specific skills. The second consists of customers in vertical industries, such as property management, with direct operational needs.
For the second group, if customers do not have their own robot hardware, ForgeAI also procures and integrates third-party hardware, adds its software and skills, and delivers a complete solution.
The company’s first priority use case is “three-dimensional cleaning,” extending beyond floor cleaning.
This may not sound glamorous, but the choice reflects clear technical considerations.
Cleaning restrooms and commercial spaces involves a combination of mobility, grasping, wiping, and recognizing irregularly shaped objects. Real-world environments are also highly variable: the location of rubbish, the shape of wet patches, and the placement of objects may differ with every task.
These conditions place greater demands on a robot’s understanding of its environment and its ability to carry out a sequence of tasks than many fixed industrial operations do.
At the same time, cleaning does not impose the same stringent requirements on cycle times and first-attempt success rates as industrial production lines. If a surface is not fully wiped on the first attempt, the robot can try again. If it fails to grasp an object securely, it can adjust its movement.
This combination of sufficient complexity and some room for trial and error explains why ForgeAI chose cleaning as its first testing ground.
The company says its goal is not simply to build a cleaning robot. It hopes to use this application to develop foundational skills such as grasping, wiping, and moving, then gradually reuse them in hotels, retail, and other commercial service settings.
ForgeAI has already begun working with Excelland Robotics（优地机器人）, Jiangsu Longhuan（江苏龙寰）, and other partners.
One example involves hotel delivery robots.
In China, delivery robots can typically connect directly to elevators through elevator control systems. In overseas markets, however, interfaces differ across buildings and elevator systems. A robot that can recognize and press elevator buttons as a person would could reduce its reliance on external elevator control systems.
According to ForgeAI, it is exploring autonomous elevator button operation with Excelland Robotics, aiming to enable robots to press buttons as people do in locations where integration with an elevator control system is unavailable.
Although modest in scope, this example illustrates a practical issue in embodied AI commercialization: what customers ultimately buy is often a specific task that a robot can now perform, rather than simply a larger model.
In 2026, discussion in the embodied AI industry is moving beyond whether robots can perform an individual action to whether they can keep working in real-world environments.
This shift is also appearing in policy.
The initiative for real-world training of humanoid robots and embodied AI launched this year by China’s Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission of the State Council emphasizes skill packages for actual jobs, replicable complete-robot solutions, generalization and fault tolerance under complex and abnormal operating conditions, high-quality data from real-world settings, and the ability to operate continuously for extended periods.
Laboratory demonstrations remain important, but they are increasingly insufficient on their own to answer questions about commercialization.
Once robots enter actual property management, hotel, retail, or industrial settings, they must handle long sequences of tasks, environments where people and objects are constantly changing, and numerous unexpected situations absent from their training data.
ForgeAI also has technical challenges to resolve.
According to Wang, the team can already convert information such as human videos and demonstrations into skills that robots can learn. However, abstracting a skill mastered by one robot into hardware-independent information and reliably transferring it to another remains the key challenge for “one brain, multiple embodiments.”
Under its current plan, ForgeAI aims to complete a prototype suitable for commercial service applications by the end of 2026. In 2027, it plans to further develop its system for learning, retaining, and distributing skills, while gradually pursuing deployment at scale in real-world settings.
Following this funding round, ForgeAI's next phase will increasingly focus on testing whether its technical framework can be reused across more robots and customer sites.
For a robot brain company that defines itself as “general-purpose,” the viability of the business ultimately depends on whether the same capability can continue to be reused across a different robot, customer, and environment—not simply on how complex an action one robot can perform once.
If you are interested in China’s embodied AI industry or other technology sectors and companies, please feel free to contact us.
xingyiran@iyiou.com