After surging more than fourfold on its market debut, Unitree Robotics experienced a sharp 18.7% correction today. The significant pullback has drawn attention to a critical industry challenge highlighted by founder Wang Xingxing during a morning session. Current robot models are generally capable of executing tasks like movement or sorting and transporting objects, with overall trends progressing as expected. However, robots still struggle to align perfectly with the real world in the final few centimeters or millimeters of an operation. The insufficient capability of robot models remains a widespread issue plaguing the entire industry, a hurdle that even the sector's leading player, Unitree Robotics, has yet to overcome.

In its prospectus, Unitree Robotics stated, "During the reporting period, the company has not yet scaled its self-developed general-purpose embodied large model for use in robot products, but has conducted R&D testing and deployment validation in pilot scenarios such as its own factories." In other words, apart from deployments in its own facilities, the robots sold by Unitree Robotics have not utilized its proprietary general-purpose embodied large model. Main models like the G1 integrate relatively mature large language models instead. The embodied large model serves as the "brain" of humanoid robots, enabling them to perceive the external world, make decisions, and issue action commands. The physical world is unpredictable and cannot be managed through pre-programming alone—even simple tasks like serving tea involve varying cup sizes from day to day, requiring robots to make autonomous decisions and take action. A robust "brain" allows robots to adapt to diverse environments and complete tasks. Yet, during the reporting period, Unitree Robotics' "brain" was not installed on the robots it sold, which relied on the "brains" of other companies instead.

Although Unitree Robotics has achieved profitability, the underdevelopment of its self-developed general-purpose embodied large model remains a lingering concern. This challenge is not unique to Unitree Robotics; despite China's quadruped robots accounting for nearly 70% of global sales, fully realized general-purpose embodied intelligence remains the industry's weakest link. Without a capable "brain," robots cannot transfer skills learned in one scenario to a new environment. Conversely, if any company can first strengthen its "brain" and solve the problem of capability transfer across scenarios, the robots it produces would adapt better to their surroundings, accumulate knowledge more easily, and grow smarter over time—potentially eroding Unitree Robotics' current competitive edge.

To achieve leadership in "brain" development, Unitree Robotics is actively working to enhance its cognitive capabilities. In June, the company partnered with NVIDIA to develop a humanoid robot called the H2 Plus. The "body" is provided by Unitree Robotics' H2 robot, while the "brain" consists of NVIDIA's Jetson Thor computing platform and Cosmos 3 model foundation. The H2 Plus is slated for release in the second half of this year, targeting universities such as Stanford and ETH Zurich. In Unitree Robotics' strategic placement list for its IPO, DeepSeek was allocated 933,400 shares, amounting to approximately 141 million yuan, with a three-year lock-up period—the longest among external strategic investors. This extended lock-up likely reflects DeepSeek's intended contribution to Unitree Robotics' "brain" development. DeepSeek could also leverage Unitree Robotics' humanoid robots to access the physical world. Reports indicate that Unitree Robotics has allocated over 2 billion yuan in R&D funding toward embodied intelligence algorithms and large model research, aiming to develop its own "brain."

Other unlisted companies are also raising capital to develop their "brains." Deep Robotics, which is pursuing a listing on the STAR Market, recently released its response to the first round of review inquiries for its IPO. The filing reveals plans to raise over 2.5 billion yuan, with nearly half—approximately 1.17 billion yuan—allocated to the "Embodied Algorithms and Model R&D Project." About 90% of that project's funding is directed toward "brain"-related technologies and infrastructure. According to statistics, in the first half of this year, domestic financing in the core embodied intelligence sector reached 43.8 billion yuan, with more than half of that capital flowing to companies focused on "brain" development. In contrast, funding for robot "bodies" during the same period accounted for only 12.8%. In other words, to equip robots with "brains," the industry prepared over 20 billion yuan in just the first half of the year.

Why is developing a robot "brain" so expensive? The growth of a robot's "brain" requires massive amounts of data. Training a high-quality model requires data on the scale of at least ten million hours, yet the current market has only a few hundred thousand hours of mature embodied intelligence datasets. It's important to note these are "mature" datasets. Low-quality data is abundant—particularly videos posted online by humans, which contain numerous human actions. However, many of these videos are entertainment-oriented and rough, making them unsuitable for guiding robot movements. Higher-quality data comes from real-machine teleoperation, where humans wear exoskeleton devices or control robots "hand-by-hand" to teach them actions. This method yields precise robot control and high-quality data, but it is clearly inefficient, generating limited data even after a full day of work. Obviously, high quality, large scale, and low cost form an "impossible triangle" in this field. Achieving large-scale, high-quality data inevitably requires significant investment.

Furthermore, robot data is difficult to transfer across different models. For instance, robots of varying heights, such as 1.2 meters versus 1.8 meters, exhibit completely different mechanical arm movements even when grasping objects at the same height. Different manufacturers also employ their own hardware interfaces, communication protocols, and data formats, creating data barriers that make sharing extremely difficult. Everyone aspires to develop a "brain" first and secure an industrial advantage, but for now, all are still in the exploratory stage. If any company achieves a true breakthrough in robot "brains" or discovers a fast track to embodied intelligence, the current landscape of humanoid robots would undergo a dramatic transformation. This is the enormous shift facing Unitree Robotics and other listed and unlisted humanoid robot companies today.

Where to Begin: The global race for embodied robot "brains" behind Unitree Robotics' IPO is intensifying. Data is the lifeblood, and embodied intelligence is becoming increasingly software-driven. Insufficient data poses a "growing pain" for embodied intelligent robots. The path forward for China's embodied intelligence industry lies in navigating global competition. These insights draw from Unitree Robotics' prospectus for its STAR Market IPO and Deep Robotics' responses to IPO review inquiries.