When will humanoid robots finally leave the exhibition halls and become tools that solve real-world problems? This is the most frequently asked question at the 2026 World Artificial Intelligence Conference. On July 19th, Zhang Wei, founder of Zhujidi Dynamics and a tenured professor at the Southern University of Science and Technology, took the stage and presented a set of quite unconventional judgments: The embodied intelligence represented by humanoid robots is currently on an exponential growth curve, and the entire robot's brain system has reached the stage of GPT-3.
Zhujidi Dynamics was founded in 2022, and Zhang Wei clearly defined the company's positioning: a product-oriented host factory with brain capabilities. In recent years, the company's financing rhythm has clearly outlined capital's recognition of its path. In October 2023, it completed nearly 200 million yuan in angel round and Pre-A financing, with investors including Fengrui Capital, Oasis Capital, Lenovo Venture Capital, and Mingshi Venture Capital; in July 2024, Alibaba led the A-round financing, which was also Alibaba's first investment in the field of embodied intelligence;
In 2025, Zhujidi Dynamics completed a series of 500 million yuan in A-round financing within half a year, receiving support from top institutions such as Alibaba and NIO Capital, and later received strategic investment from JD.com. Both parties plan to collaborate in retail, logistics, and service scenarios; in January 2026, the company completed a 200 million USD B-round financing, introducing UAE Leishi Capital, Oriental Fuhai, and Foundation Capital, with existing shareholders continuing to invest; just this July, Zhujidi Dynamics completed a near-200 million USD Pre-IPO financing, with investors including Leishi Capital, Italian financial consortium, listed company Lantek, IDG Capital, and Hefei Binhu Industrial Development Group. Zhang Wei stated that the company has formed a comprehensive shareholder background covering strategic industry players and domestic and foreign financial investment institutions, further enhancing its industrialization, internationalization, and marketization capabilities.
According to Zhang Wei, many people see that robots are still clumsy and can't do much, so they tend to predict their future in a linear way, but the industry is experiencing an exponential transformation, and it cannot wait for results before responding, otherwise it would be too late. He used an example to illustrate the failure of linear reasoning: Bill Gates spent decades and invested heavily to solve the language interaction problem but never succeeded. If he had been asked when it would be solved before the emergence of large models, he might have said another ten or twenty years, but once the technological paradigm changed, a lot of problems were solved in half a year.
A view that may surprise many is that Zhang Wei believes robots are easy to make, and humanoid robots are also easy to make. They are easier to make than lithography machines, airplanes, and even cars, with about one-tenth the number of parts as cars. So why are airplanes flying every day and cars everywhere, while humanoid robots are mostly still in the exhibition hall? His answer is that it's not lack of people who can make the body, but rather the core capability pointed out by the theme of this conference — AI, especially physical AI derived from data from the physical world. Physical AI's characteristics are that its data does not come from the Internet, but from real-life physical experiences, making it one of the most challenging applications of AI. Its ultimate form, physical AI, is essentially imitating the human brain. Zhang Wei added that the human brain doesn't look very difficult from a conceptual perspective. Its task is only to receive instructions, understand intentions, observe the environment, and calculate how to move, essentially a mapping. The real problem blocking the industry is a lack of data. He believes that the technological paradigm has already relatively converged to pure data-driven, and the rest are more engineering and detail issues. All exploration is essentially about reducing the data cost required to complete a task.
Regarding the brain system, Zhang Wei outlined a three-layer structure, similar to Figure's architecture but with significant differences in details. The top layer is similar to the human prefrontal cortex, a thinking engine driven by large language models, visual language models, and even future world models, belonging to the agentic system, responsible for memory management, instruction collection, information understanding, task planning, and decision-making, only thinking, which he called System 2. The middle layer is the visual motor cortex, which receives upper-level decisions, observes the environment, and makes movement plans. The bottom layer is the cerebellum and motor control system, responsible for completing actions, moving without thinking. Zhang Wei emphasized that the real challenge in the industry is not to perfect a single layer, but to coordinate these three layers in a millisecond-level real-time system, and jointly optimize with hardware. Zhujidi Dynamics refers to this direction as the integration of the big brain and small brain technology, which is also one of the company's most important investment directions.
If we must use the GPT moment as a metaphor, Zhang Wei judges that the entire robot's brain system has reached the early stage equivalent to GPT-3, which is completely different from the judgment of many people who think it is still very early. But he warned that the robot's brain is not a single model, but a system composed of large language models, visual language models, world models, memory management, task planning, and motion control. Zhujidi Dynamics' system is called COSA. The model represents ability, while the brain is a whole system composed of multiple models plus memory management, emotional management, and task planning. He gave an example: if System 2 is driven by an LLM, it's like a disabled person lying in bed who can think but cannot move; if driven by a VLM, it's like a disabled person with eyes who can see the world; if adding a world model, it's like Stephen Hawking who learned physics and could understand and predict the physical world. Embodied intelligence aims to let this smart System 2 guide actions, like helping a disabled person do rehabilitation training, gaining skills through data and practice. Zhang Wei also reminded that it took over two years between GPT-3 and GPT-3.5, while the robot's brain is a system, not determined by a single model, which is his different thinking from many others.
On skill learning, Zhang Wei also put forward a counter-conventional view: Skills don't need to be very general to be considered having a brain. A person who cannot tighten a screw or drive a car is still a normal person with a brain. The generality of skills does not represent the level of intelligence. Skill construction depends on what you want to learn first, then what next, and the corresponding data cost. Different skills have great differences in difficulty. For example, dancing, which does not require real-time interaction with the environment, is relatively simple, while climbing stairs, which requires real-time interaction, is much more difficult. Ollie, Zhujidi Dynamics' robot, climbing stairs is an example of real-time big brain and small brain fusion technology because stairs can be largely completed in a simulation environment; cleaning a table must rely on real machine data. Zhang Wei introduced that the company's COSA system, researched last year and released this year, can receive human commands, plan tasks through System 2, and then schedule navigation positioning and perform specific tasks, such as sending a package, where the robot would first take water to the customer and then deliver the package, generating actions in real time. Recently, COSA completed an important upgrade - a fully autonomous, real-time reasoning long-range task, without remote operation or a remote controller, completing full-body movement and operation in a home environment with a single model and a single skill, and the entire model is purely data-driven without expert rules. He stated that besides Figure, very few companies can complete such a general-purpose long-range mobile operation task, and systems that combine full-body movement and operation are particularly rare.
Regarding commercialization, Zhang Wei summarized the fundamental logic of humanoid robots as a general body plus apps. A single skill like dancing may have limited value, but by continuously adding different skills and applications without changing the robot's configuration, eventually a "absorbing" tipping point will appear, and enough applications will begin to gather around the general body. He believes that the host factory is the most critical in the industrial chain because it connects the upper and lower parts. Zhujidi Dynamics positions itself as a product-oriented host factory with brain capabilities, technically matching Figure comprehensively, and more aggressive in industrialization, with the slogan "serve people, not process," meaning that humanoid robots with two legs should serve people instead of production processes, so they do not prioritize industrial scenarios but are more suitable for reception and public services in human environments. He concluded that the key to commercialization is: the data cost of building a skill must be less than the value created by the skill, only then will the skill be worth doing.
