Exclusive by Hard Krypton: LI Hongyang, a professor at the University of Hong Kong, founded a startup developing a general full-body embodied brain, securing hundreds of millions of yuan in seed round financing from investors including ZhenFund, Gaorong Capital, IDG Capital, and 5Y Capital

Author | Huang Nan

Editor | Yuan Silai

Hard Krypton exclusively learned that the general full - body embodied brain company "Archon Robotics" recently completed a seed - round financing of hundreds of millions of yuan. The investors in this round include leading US - dollar funds such as ZhenFund, Banyan Capital, IDG Capital, and Matrix Partners, as well as the joint fund of Gobi Partners and the University of Hong Kong, Miracle Plus, Shanghai Institute of Innovation and Intelligence, etc. LightSpeed Capital served as the exclusive financial advisor.

The financing funds will be mainly used for the research and development of the full - body humanoid basic model, the collection of multi - modal full - body motion data, the expansion of the talent team, and the construction of R & D centers in multiple locations and the industrial cooperation ecosystem, so as to accelerate the implementation of the open - source humanoid base model within this year.

Archon Robotics was established in April 2026, and its R & D headquarters is located in the Caohejing Development Zone, Xuhui District, Shanghai. The company focuses on the research and development of the general full - body humanoid base model, constructs whole - body intelligence, and provides humanoid robots with the ability of human - like full - body mobile operation to accelerate the realization of embodied intelligence entering thousands of households.

Dr. Li Hongyang, the founder, is currently an assistant professor at the University of Hong Kong, the assistant dean of the School of Computing and Data Science, and a tutor at the Shanghai Institute of Innovation and Intelligence. The end - to - end autonomous driving project UniAD led by him won the Best Paper Award at CVPR 2023, which is the only work by a mainland academic institution to receive this award in the past decade; in 2026, he won the RSS Early Career Award at the top international conference in the field of robotics, becoming the first Chinese scholar since the establishment of this award 20 years ago.

Dr. Li Tianyu, the co - founder and CEO, is one of the first graduates of the Shanghai Institute of Innovation and Intelligence and a doctor from Fudan University. As a core developer, he was deeply involved in the research and development of Huawei's mass - produced autonomous driving ADS 4.0 World Engine solution. Dr. Chen Li, the co - founder and Head of AI, is the first author of the best paper on UniAD. He graduated from the Zhiyuan Honors Program of Shanghai Jiao Tong University and once won the University of Hong Kong President's PhD Scholarship.

The core team of Archon Robotics all come from top research teams in autonomous driving, robotics, and large models at institutions such as the University of Hong Kong, Tsinghua University, Shanghai Jiao Tong University, Fudan University, and Zhejiang University, and have both breakthroughs in original algorithms and experience in the implementation of ultra - large - scale industrial systems.

The Archon Robotics team raises their glasses at the Everest Base Camp. Photographed in February 2024 at Mount Everest (Source/Enterprise)

The embodied intelligence industry is entering a critical differentiation moment. Public data from Omdia and other sources show that in the first half of 2026, there were a total of 288 financing events in the domestic embodied intelligence and robotics fields, with the disclosed total financing amount exceeding 46 billion yuan, approaching the scale of 55.4 billion yuan for the whole year of 2025. However, the huge amount of financing has not brought about a convergence of technical consensus.

Most current embodied solutions have inherent limitations: the existing training data mainly consist of first - person desktop videos and single - arm or gripper movements, lacking the original human interaction logic such as full - body center - of - gravity adjustment, torso leverage, and multi - limb coordination. This means that most robots can only perform fixed - point grasping, and it is difficult for them to adapt to variables autonomously when facing daily tasks that require full - body cooperation, such as opening doors, making beds, and opening and closing doors while holding objects with both hands.

The root cause of this limitation lies in the structural gap in data infrastructure. Li Tianyu, the CEO of Archon Robotics, told Hard Krypton, "Although the available embodied data sets in the market seem large, the information that is actually effective for full - body humanoid training is extremely limited."

The first - person video data sets only support the recording of the pictures seen by the human eye, lacking key action pose information beyond the appearance of the hands, such as squatting, bending over, and leaning sideways. The labeled data of robotic arms and grippers are mostly concentrated in the planar category, only recording the end - point trajectories. The model can learn how the manipulator moves, but it is difficult to understand how to interact with the environment. The total amount of real - humanoid robot data is extremely low, and the collection cost is several hundred to nearly a thousand yuan per hour. The samples of composite tasks related to full - body multi - joint coupling were almost blank in the past data pool.

The deficiencies in these three types of data all point to the same problem: the most core information for humans to complete daily actions, such as how the whole body coordinates, how the center of gravity shifts, and how the force is transmitted from the lower limbs to the upper limbs, has hardly been recorded in the existing data.

Take a simple daily scenario as an example. When a person pulls a light door and a heavy door, there is almost no difference in the hand trajectory; in both cases, the person first grabs the handle and pulls it back, and the pose is always synchronized with the movement of the door regardless of the force used. The real difference lies in the full - body dimension. When facing a light door, a person can pull it open while standing upright; when facing a heavy door, the body needs to lean forward to shift the center of gravity and use body weight to counter the resistance.

This information about the movement of the center of gravity is only recorded in the full - body data and contains the essential differences in the physical properties of objects. Simply put, if the model only learns single - dimension information for a long time, it may be able to perform the action of "pulling the door", but it cannot understand what "how heavy the door is" means at the human - body level.

"The long - term absence of information has locked the current capabilities of robots at the level of fixed - desktop grasping, and there is a data gap between them and the diverse tasks in the real home environment," Li Tianyu said. "To break through this ceiling, we must go back to the source and redefine the logic of data collection."

The Archon Robotics team believes that migrating from the solution of a wheeled chassis with two arms to a humanoid robot involves essential differences in structure, motion control, and perception dimensions, and it is not simply a morphological upgrade. When overseas companies have just realized the complexity of humanoid tasks, they have already set their sights on this goal.

Archon Robotics aims at an almost untouched field, developing a general full - body humanoid base model. Its core concept, Human Body Learning, is to learn the full - body pose and coordination methods of humans, rather than simply tracking the trajectory of the end - effector. By learning from human full - body movements, robots can acquire the "wisdom of four - limb coordination" and have complete full - body interaction capabilities.

By depositing the "intelligence" of the humanoid robot's actions as much as possible in the mid - brain level that is independent of the body, during this period, the capabilities learned by the mid - brain are not bound to a specific robot; it outputs full - body motion trajectories rather than joint - angle instructions for a specific model, enabling the model to have the potential to migrate across different bodies. As the data collection becomes more comprehensive and the covered scenarios become more diverse, the representational ability of the mid - brain becomes stronger, and the types of bodies to which the embodied full - body brain of Archon Robotics can migrate become more extensive.

Based on this judgment, Archon Robotics will build a new data collection system. Li Hongyang, the founder, believes that the evolution path of embodied data is evolving from real - robot teleoperation to handheld devices and first - person perspectives, and the ultimate destination is human - centric full - humanoid data with complete human perception elements and full - body action labels.

At the same time, Archon Robotics will also introduce multi - dimensional perception modalities such as touch, and be equipped with higher - precision full - body and hand capture devices. Li Tianyu told Hard Krypton that data diversity and data quality are more critical than simply the data scale. "One piece of full - body data that covers the movement of the center of gravity and the change of the torso angle has a much higher information density than a hundred pieces of desktop data with only hand trajectories."

The way of data collection determines what the model can learn, and the ability shortcomings of the model in turn define the target of the next collection. Once this "collection - training - feedback" closed - loop starts to operate, it will form a continuously self - strengthening data barrier: after each round of collection and training, the model's ability is improved, and the system's understanding of "what data is really useful" becomes more accurate, and the efficiency and quality of the next round of collection will be further improved.

It tests not only the engineering ability at the algorithm level but also the systematic understanding of the fundamental question of "what the model really needs to learn from the physical world." And this understanding is precisely the core judgment ability of Archon Robotics.

Hard Krypton learned that Archon Robotics plans to release its first humanoid native base model in the second half of 2026.

In the view of the Archon Robotics team, for humanoid robots to truly enter families from the laboratory, they need not only a perfect single - point demonstration but also the ability to work continuously and reliably in a complex, dynamic, and unstructured home environment. The upper limit of this ability fundamentally depends on the depth of the model's understanding of the physical world.

Archon Robotics chooses to go back to the starting point of embodied intelligence and answer this question again: what kind of body and what kind of data to learn from determine how far the robot can ultimately go.

The following is an excerpt from an interview between Hard Krypton and Dr. Li Hongyang, the founder of Archon Robotics, and Dr. Li Tianyu, the CEO (slightly edited):

Hard Krypton: The current embodied industry has not yet entered the convergence stage. What is the underlying judgment for Archon Robotics to develop a general full - body humanoid base model?

Li Tianyu: Our approach has two essential differences from the mainstream solutions in the market.

The first is the hardware form. Archon Robotics has targeted a complete humanoid robot since its establishment. The core problem to be solved is the linkage of multiple joints of the whole body and the dynamic adjustment of the center of gravity. Most teams choose a wheeled chassis with two arms, which is understandable from an engineering perspective. It has good stability, a relatively low technical threshold, and it is easy to produce a demo in the short term. However, in the home scenario, a wheeled chassis cannot even cross a threshold, let alone perform daily operations such as squatting, climbing, and squeezing through gaps sideways. The humanoid form is not an option but an inevitable end - state.

The second is the underlying model architecture. We are developing a dedicated humanoid native embodied large model, rather than fine - tuning an existing VLA or world model. The difference between the two is that the former learns the underlying logic of human interaction with the physical world from full - body data, while the latter teaches the model how to move an object on a fixed desktop. The model architecture of Archon Robotics is designed for full - body motion from the very beginning. The brain is responsible for task understanding and long - term planning, the mid - brain learns the full - body motion representation across different bodies, and the cerebellum is responsible for real - time pose tracking and balance maintenance. The three layers work together to output full - body motion trajectories rather than simply end - point instructions.

Based on this approach, we have built an independent data collection link. On the one hand, we collect large - scale full - body motion data of ordinary people in real scenarios, and on the other hand, we combine it with the actual operation data of humanoid robots for joint training, so that the model can replicate the innate full - body coordination and interaction ability of humans.

Archon Robotics aims to build a humanoid brain that can truly face an unpre - set home environment. Robots need to autonomously perceive surrounding changes and randomly handle various differentiated tasks: the weight of objects, their placement positions, and the room layout are constantly changing, and there is no fixed script. Its core requirement is strong generalization and migration ability, so that it can learn a set of logic and handle thousands of different scenarios.

Hard Krypton: You once compared the current development level of the embodied intelligence industry to the L1.5 stage of autonomous driving in a public interview. What are the bases for this comparison? Which dimensions does it specifically correspond to?

Li Hongyang: We can use the grading logic of L1 to L5 in autonomous driving to understand the evolution of embodied intelligence. L1 is basic teleoperation with limited generalization; only when it reaches L2 can the system complete a complete operation closed - loop autonomously in a specific scenario without human intervention.

In the current embodied industry, most public demonstrations are essentially remote teleoperation or highly pre - set script execution. Most actions, object positions, and environmental layouts are fixed in advance, and the system will fail if any variable is changed. No product can handle continuous home tasks with multiple steps and changes autonomously without human intervention.

Therefore, the judgment of the L1.5 stage means that the industry can produce demos for single - point tasks, but it is still a long way from "complete autonomy in a specific scenario." The key to crossing this threshold lies in data. However, the existing data sets do not contain the signals required for humanoid robots to make autonomous decisions and coordinate the whole body in a real home environment. Simply increasing the model scale with the current data will not reach the L2 level.

To cross this gap, we must redefine the logic of data collection, shifting from "collecting the movement of robots from a single perspective" to "collecting how the whole body interacts with the environment," and from "fixed tasks in pre - set scenarios" to "diverse exploration in an open world." This is also what Archon Robotics has been doing from the very beginning.

Comments from investors:

Qin Tianyi, Managing Director of ZhenFund said that he has known Li Hongyang for many years and has witnessed him lead the team to tackle the most difficult and important issues, from UniAD to BEVFormer, and then to the vision of Archon Robotics, "the general full - body humanoid base model." They never chase benchmarks but firmly push the boundaries. More importantly, the OpenDriveLab founded by Li Hongyang has become a talent hub in the fields of autonomous driving and embodied intelligence, and a group of high - spirited top young scholars have emerged. Li Tianyu and Chen Li, the co - founders of Archon Robotics, are outstanding representatives among them.

Xin Wang, Partner of Banyan Capital said that they really like the Archon Robotics team led by Professor Li Hongyang. They have pure technological ideals, profound insights, and the courage to lead innovation, and will never be followers. In the field of intelligent driving, projects such as UniAD, BEVFormer, and WorldEngine, and in the field of embodied intelligence, projects such as Agibot, WholeBodyVLA, and Ego - humanoid, have continuously proven that the Archon Robotics team has the ability to continuously innovate and produce world - influential results. Today, Archon Robotics has chosen a difficult but correct path, challenging the problem of the full - body dexterous operation brain of humanoid robots. They are very glad to participate in the cause of Archon Robotics and look forward to Archon Robotics becoming a world - class embodied brain company.

Meng Xing, Partner of Matrix Partners said that he met Hongyang in 2023 when he was working on autonomous driving. At that time, UniAD had just won the Best Paper Award at CVPR, which was the most sensational event in the industry. What convinced him was not the trophy but the way he asked questions: when everyone was competing in their respective modules, he started from the ultimate goal of planning and deduced what the system should look like, connecting the backbones of all modules and making everything serve the final planning. Three years later, Hongyang at Archon Robotics is still the same. While others are asking "what can be trained with the existing data," he is asking "what should robots learn from humans." Human Body Learning in the field of embodied intelligence is just like the planning - oriented approach in autonomous driving back then - both define the starting point from the end - state. Today, the embodied intelligence field is not short of money, engineers, or computing power, but only lacks this kind of taste: the ability to distinguish what is worth doing in the noise. He believes that the most precious thing is a person's judgment ability to ask the right questions across two technological waves.

Li Guanle, Managing Director of Gobi Partners and General Manager of the University's Innovation and Technology said that Archon Robotics is not a traditional robotics company but a silicon - based embodied intelligence company driven by algorithms and data engines. The team has the rare ability to define the industry's technological paradigm. Relying on self - developed original algorithms and mature engineering implementation capabilities, in the current situation where the track standards have not