Secured two rounds of financing within two months, post-2000s entrepreneurs from Tsinghua University and Peking University are challenging the dexterous hand sector.
The robotics and embodied intelligence industry is undergoing adjustments, yet the window for entrepreneurship has not fully closed, and entrepreneurs following non-consensus technical routes can still gain favor from capital.
ChinaVenture has exclusively learned that ManuPrax Robotics, a company founded two months ago, has recently completed two consecutive rounds of financing led by Alphaist Partners and Joy Capital respectively, with participation from Relay Angels and the actual controller of a certain component industrial party, with the total amount reaching nearly 10 million US dollars.
Founded by two young researchers, the company was registered in September this year, targeting the dexterous manipulation track for robots and committed to building an integrated hardware and software dexterous hand solution. Among them, algorithms, data and engineering capabilities constitute the core barriers of the company's dexterous hand software.
CEO Li Puhao, born in 2001, is a BSc-MSc-PhD candidate in the Department of Automation of Tsinghua University. He has been conducting research on dexterous hands, dexterous manipulation and other related issues since 2021, and has published many influential papers in the field of dexterous manipulation. CTO Chen Zeyuan, born in 2000, holds a master's degree from Peking University and gave up the opportunity to pursue a doctoral degree to start a business. Focusing on the underlying interactive algorithm field of dexterous manipulation, his involved work includes ClutterDexGrasp and CADGrasp, and he has also participated in research at Amazon FAR, Agibot and JD.com.
The two post-2000s generation believe that in the field of dexterous manipulation, whether in academia or the industry, the scarcest resource is the method that enables dexterous hands to truly achieve human-level manipulation capabilities. They choose to enter the market from software capabilities, and place the manipulation capabilities at the "cerebellum layer". According to the team's vision, the upper-level foundation model (such as GPT-6 Astra) is responsible for understanding tasks and formulating plans, while the "cerebellum" is responsible for action coordination similar to "muscle memory", converting manipulation goals into continuous and stable movements.
Both Li Puhao and Chen Zeyuan have research experience at the Beijing Institute for General Artificial Intelligence (referred to as "BIGAI"). ManuPrax Robotics is also an officially supported entrepreneurship project of BIGAI. Song-Chun Zhu is the dean of BIGAI and also Li Puhao's doctoral supervisor. In his view, Li Puhao studied in the Department of Automation of Tsinghua University for his undergraduate degree, joined BIGAI in his sophomore year, and gradually developed his research along the training system of BIGAI and its general artificial intelligence program. "The study in the Department of Automation of Tsinghua University has laid a solid foundation of mathematics, physics, control and engineering for him, and the scientific research environment of BIGAI and its program allows him to get in touch with the overall problems of general artificial intelligence at an early stage, forming a broader research vision. The combination of these two aspects enables him to go deep into the specific details of robot manipulation, and also to think about the position of these works in the general intelligent system."
Of course, the technical and commercial barriers being established by leading dexterous hand companies also bring some practical tests to the choices and judgments of this young company. However, their view on this matter is that "we are not afraid of difficulties, but we are afraid that the challenges are not difficult enough".
Building a "hand" that can do practical work
Among all robot components, the dexterous hand is probably the most difficult one, which even Elon Musk has been troubled by.
In 2025, Elon Musk stated at Tesla's earnings call that the electromechanical engineering difficulty of dexterous hands and forearms even exceeds that of other parts of the robot's body; by 2026, he still publicly mentioned that Tesla is still struggling with the final hardware design of Optimus's hands.
The human hand can complete many complex manipulations, from gripping tools, unscrewing bottle caps, to picking up fragile items and rotating parts on the fingertips. Humans can adjust the position and strength of their fingers at any time according to the shape, weight and contact state of objects. But for robots, achieving these requires the collaboration of precision mechanical structures, sensors, real-time control and algorithms.
In the view of Li Puhao and Chen Zeyuan, dexterous hands cannot be compared only by the degree of freedom. A large number of joints does not equal strong manipulation capabilities; being able to complete grasping in a preset environment does not mean being able to cope with constantly changing objects and contact states in the real world. They believe that hardware is only the carrier of manipulation capabilities. ManuPrax Robotics hopes to shift the focus of competition from "what the hand looks like" to "what the hand can do".
This may also echo the implication of the company's English name ManuPrax: "Manu" is derived from the Latin word "manus", meaning "hand"; "Prax" is derived from the Greek word "praxis", meaning "action and practice". "Sometimes dexterous manipulation is an instinctive reaction of muscle memory, which is a model at the cerebellum layer," Li Puhao said.
To achieve this, the key does not lie in the upper-level task planning and task decomposition, nor the part currently taken charge of by the VLA (Vision-Language-Action) model. The upper-level model mainly solves the problems of understanding the environment, understanding tasks, reasoning and planning, while they hope to focus on solving the more underlying execution capabilities, which is commonly known as the "cerebellum".
"What we hope to make up for is the ability of dexterous hands to perform movements delicately, coordinately and stably in the real physical world," Chen Zeyuan further explained. To let the robot pick up and operate the cup on the table, the upper-level model needs to understand the instruction of "picking up the cup", identify the position of the cup and set the movement goal. But next, problems including from which angle the fingers should approach, how to adjust the position of the fingertips, how much force to apply to make the cup move along the target trajectory, and how to remedy when the cup slides, cannot be solved just by understanding the instruction itself.
In this division of labor, the "cerebellum" undertakes action coordination similar to "muscle memory", converts the goals given by the upper-level foundation model into continuous and coordinated movements, and constantly adjusts according to the actual contact state. The team hopes that through the cooperation of these two layers of capabilities, the robot can not only understand the tasks, but also execute the movements in place in the real physical world.
To realize this set of capabilities, ManuPrax Robotics chooses the route of reinforcement learning. Li Puhao introduced that different from the method that mainly relies on human demonstration to let the model imitate movements, reinforcement learning tries to let the model gradually find effective strategies through exploration, trial and error, and feedback. The model tries a movement, obtains a reward signal according to the result, and then adjusts the next action. Through continuous repetition, it has the opportunity to accumulate successful experience and gradually adapt to different objects and environments. "Human daily movements are very rich, and the shape, weight, material and placement mode of objects vary greatly. Relying solely on limited manual demonstration can hardly cover all situations. If the model can learn through interaction, it will have the opportunity to reduce the dependence on one-by-one manual demonstration."
However, it is not easy to achieve scaling along this route. Reinforcement learning requires demonstration-based and general reward signals, as well as a sufficiently efficient and stable training architecture. At the same time, the fact that the model learns a movement in the simulation environment does not mean that it can be reproduced on real hardware. Friction, contact, deformation and sensor errors in reality may all invalidate the successful experience in simulation.
According to the two of them, the team is exploring the combination of simulation, real human demonstration and generated interactive trajectories to build a training data system for further verification, so that the action model can be continuously improved as the training scale and data volume expand.
Dexterous hands are not just a hardware business
Almost all robots need a "hand" that can do practical work. In the past few years, the competition in the dexterous hand industry has mainly focused on mechanical structures, driving modes, degrees of freedom, load capacity and cost. As more enterprises enter the track, the competition has also extended from pure hardware performance to data accumulation, manipulation algorithms and hardware-software collaboration capabilities.
In the current dexterous hand market, there are both self-developed products by robot ontology manufacturers and independent dexterous hand suppliers. The former can integrate dexterous hands into the unified design of the complete machine system, while the latter needs to compete for different customers by virtue of product performance, cost and delivery capabilities. The opportunity for independent suppliers lies in serving multiple robot manufacturers, but they also face competitive pressure brought by the self-development of complete machine manufacturers. Who can become a general supplier ultimately depends on whether the product can establish advantages among performance, cost and reliability.
In the view of ManuPrax Robotics, the hardware pattern may soon enter another stage. Li Puhao judged that the development speed of domestic dexterous hand hardware will be very fast, and the profit margin of pure hardware business may be greatly compressed in the next six months, which is the important background for the team to choose to extend to the manipulation model of the physical interaction layer. If hardware gradually becomes a product with higher standardization and thinner profit, then the manipulation capability that determines whether a robot can complete complex tasks may become the new focus of competition.
Therefore, they focus on building a dexterous manipulation model platform that can adapt to different hardware. This platform takes over the manipulation goals given by the multimodal large model from the upper layer, adapts to different dexterous hand hardware from the lower layer, and provides callable manipulation capabilities for the upper-layer model. Li Puhao believes that compared with the algorithm bound to a single dexterous hand, such a platform has the opportunity to serve different manufacturers and application scenarios, reducing the dependence on the sales of single hardware.
However, the imagination space of platformization does not mean that software commercialization will be easier. When purchasing hardware, customers can directly compare the load, degree of freedom, price, reliability and delivery cycle; when purchasing the manipulation model, the questions customers need to answer are more specific: whether the model can complete tasks that were previously difficult to achieve, how much the manipulation success rate can be improved compared with existing algorithms or self-developed solutions, and how much training and deployment time can be saved. If the model only performs well in a few demonstration tasks, customers may not be willing to pay for it separately. Only when the model can continuously improve the manipulation effect in real scenarios and still be effective after migrating to different hardware, can it have the opportunity to become a capability layer that can be independently delivered and reused in the robot system.
ManuPrax Robotics has not completely ruled out the path of hardware-software integration. The team hopes to maintain the platform-oriented direction of the model, and at the same time is considering speeding up technical verification and obtaining data and feedback in real scenarios through self-developed hardware or cooperation with hardware manufacturers.
Li Puhao mentioned that at the current stage, it is not sufficient to measure the value of the dexterous manipulation model simply by shipment volume, maintenance rate or customer retention rate. Part of the demand for dexterous hands still comes from R&D projects of scientific research institutions and large enterprises. The completion of equipment procurement does not mean that the robot can stably complete complex manipulations. But if commercialization is always delayed on the grounds of technical verification, the value of the model will also be difficult to be tested by the market.
The "ambition" of two post-2000s and the "game plan" of BIGAI
In the dexterous hand track that has formed a competitive echelon, why can ManuPrax Robotics still attract attention? The technical background of the prodigies is of course one aspect, and more importantly, it is probably about where the two young people want to bring this technology.
In Li Puhao's view, a dexterous hand is not a product that can be well made only by single-point technological breakthrough, but a systematic project. When taken apart, many technical details may not be complicated. The difficulty lies in integrating them to make hardware, models and manipulation capabilities form a truly usable system. Their advantage is that the team did not enter the track temporarily just because it became popular, but has carried out long-term research on related issues, knowing where the technical difficulties are and what needs to be solved next.
The complementarity between the teams makes it possible for this technological accumulation to be further transformed. Chen Zeyuan focuses on technical issues, while Li Puhao can think about the company's development and external resources from a more macro perspective.
The technical barriers and team background explain why ManuPrax Robotics has the opportunity to make this happen, and what drives them to start a business is their understanding of scientific research value and personal choices.
Chen Zeyuan carried out research at BIGAI, cooperated with Li Puhao, and later went to work at Amazon's scientific research laboratory. In this experience, he gradually realized that when he cooperated with Li Puhao in China, the efficiency of discussing issues and promoting projects is no less than that of overseas research environments, and they have greater autonomy. They can decide what to research and how to promote it by themselves, instead of looking for a direction that suits them in an existing organization.
"More practical and more capable of solving real problems", Chen Zeyuan said directly when talking about the motivation of starting a business. He believes that in the past, the value of research was more reflected in raising questions, exploring methods and verifying ideas. Starting a business gives them the opportunity to go further, turn research results into products, and let the technology be tested by real scenarios.
Li Puhao's understanding of the relationship between scientific research and business has also changed in the process of starting a business. Previously, he preferred to make a practical product first, and then talk about financing and commercialization. The product is the most direct proof. Whether the technology is feasible ultimately depends on the actual results. But starting a business made him gradually realize that business does not always start after the product is mature. Entrepreneurs can also put forward an idea first, make it clear what problem to solve and why it is worth doing, and then strive for resources to realize it.
The ambition of the two young people is far more than just making a more flexible hand.
Chen Zeyuan's greatest ambition is to see robots replace humans to do those mechanical and repetitive tasks. In his and Li Puhao's vision, dexterous hands should eventually be able to take over most of human physical labor. Machines not only repeat movements according to instructions, but can complete real tasks in complex environments, so that people no longer need to spend a lot of time on heavy and repetitive work.
BIGAI in Beijing also provides systematic support for this start-up. In Song-Chun Zhu's plan, to establish a continuous learning closed loop from the real world to simulation and then back to the real world, all platforms and ecological enterprises of BIGAI including "Tongjie", "Tongjing" and "Tongnao" should undertake corresponding links in this system structure and form collaboration.
ManuPrax Robotics is the official dexterous manipulation company established under the ecological architecture of BIGAI, focusing on the foundation model architecture of dexterous manipulation, and it is an important part of the above-mentioned Real2Sim2Real system. The "last centimeter" of fine manipulation involves friction, deformation, sliding and contact switching, which is also the key problem to be solved when the simulation capability moves to the real world.
This part of capabilities can be connected with the task understanding and planning of "Tongnao", cooperate with the simulation training and evaluation of "Tongjie" and "Tongjing", and accept real machine tests in the collaboration with ecological partners such as robot ontology and tactile perception. The problems found and the experience accumulated by dexterous hands in real manipulation can in turn improve simulation and training. In the future, the "Mingjing" platform to be implemented by BIGAI in Chongqing will also provide a foundation for this system to face industrial applications.
This article is from the WeChat official account "ChinaVenture", author: Wei Xianghui, published with authorization from 36Kr.