Tsinghua and Peking University post-00s founders launch ManuPrax Robotics, raising two consecutive rounds in two months to take on dexterous hands

The robotics and embodied intelligence industry is undergoing adjustment, but the window for entrepreneurship has not fully closed, and entrepreneurs taking non-consensus technical routes can still win capital favor.

English text is being prepared.

The company was founded by two young researchers, completed registration in September this year, targets the robot dexterous manipulation track, and is committed to building a software-hardware integrated dexterous hand solution. Among them, algorithms, data and engineering capabilities constitute the company's main moat in dexterous hand software.

CEO Li Puhao was born in 2001 and is a bachelor's, master's and doctoral student in Tsinghua University's Department of Automation. Since 2021, he has conducted research on dexterous hands, dexterous manipulation and other issues, and has published multiple influential papers in the field of dexterous manipulation. CTO Chen Zeyuan was born in 2000, holds a master's degree from Peking University, and gave up the opportunity to pursue a doctorate in order to start a business. In the field of underlying interaction algorithms for dexterous manipulation, his work includes ClutterDexGrasp and CADGrasp, and he has also participated in research at Amazon FAR, Zhiyuan and JD.com.

The two post-2000s founders believe that in the field of dexterous manipulation, whether in academia or industry, what is most scarce is a method to make dexterous hands truly possess human-level manipulation capabilities. They chose to enter from software capabilities, placing manipulation capability at the "cerebellum layer." According to the team's vision, the upper-layer foundation model (for example, GPT-6 Astra) is responsible for understanding tasks and formulating plans, while the "cerebellum" is responsible for action coordination similar to "muscle memory," transforming manipulation goals into continuous and stable actions.

Both Li Puhao and Chen Zeyuan have research experience at the Beijing Institute for General Artificial Intelligence (referred to as "Tongyuan"). ManuPrax Robotics is also an entrepreneurial project officially supported by Tongyuan. Zhu Songchun is the director of Tongyuan and also Li Puhao's doctoral supervisor. In his view, Li Puhao studied in Tsinghua University's Department of Automation as an undergraduate and joined Tongyuan in his sophomore year, gradually developing his research along the training system of Tongyuan and the Tong class. "Study in Tsinghua's Department of Automation laid a solid foundation in mathematics, physics, control and engineering for him, while the research environment of Tongyuan and the Tong class allowed him to come into contact with the overall problems of general artificial intelligence earlier and form a broader research perspective. The combination of these two aspects enables him to delve into the specific details of robot manipulation and also to think about where this work sits within the general intelligence system."

Of course, the technical and commercial barriers being established by leading dexterous hand companies also confront this young company's choice and judgment with some practical tests. However, their view on this matter is, "We are not afraid of difficulty, but afraid it is not difficult enough."

Building a pair of "hands" that can do work

Among robot components, the dexterous hand may be the most difficult one, and even Musk has been troubled by it.

In 2025, Musk said at Tesla's earnings call that the electromechanical engineering of the dexterous hand and forearm is even more difficult than other parts of the robot body; by 2026, he still publicly stated that Tesla was struggling with the final hardware design of Optimus's hand.

Human hands can complete many complex operations, from gripping tools and unscrewing bottle caps to picking up fragile objects and rotating parts between fingertips. Humans can adjust finger positions and force at any time according to the shape, weight and contact state of objects. But for robots, achieving these involves precise mechanical structures, sensors, real-time control and algorithm coordination.

In the view of Li Puhao and Chen Zeyuan, dexterous hands cannot be compared only by degrees of freedom. A large number of joints does not equal strong manipulation capability; 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 hardware is only the carrier of manipulation capability. ManuPrax Robotics hopes to shift the focus of competition from "what the hand looks like" to "what the hand can do."

This perhaps also echoes the meaning of the company's English name ManuPrax: Manu comes from the Latin manus, meaning "hand"; Prax comes from the Greek praxis, meaning "action, practice." "Sometimes dexterous manipulation is just the instinctive reaction of muscle memory; it is a model at the cerebellum layer," Li Puhao said.

To achieve this, the key does not lie in upper-layer task planning and task decomposition, nor in the part currently handled by VLA (vision-language-action) models. Upper-layer models mainly solve the problems of understanding the environment, understanding tasks, reasoning and planning, while what they hope to focus on solving is lower-level execution capability, commonly known as the "cerebellum."

"What we hope to fill in is the dexterous hand's ability to execute actions finely, coordinately and stably in the real physical world," Chen Zeyuan further explained. For a robot to pick up and manipulate a cup on a table, the upper-layer model needs to understand the instruction "pick up the cup," identify the cup's position and formulate action goals. But next, from which angle the fingers should approach, how to adjust fingertip positions, how much force to apply to make the cup move along the target trajectory, and how to remedy it when the cup slips are not problems that can be solved by understanding the instruction alone.

In this division of labor, the "cerebellum" undertakes action coordination similar to "muscle memory," transforming the goals given by the upper-layer foundation model into continuous, coordinated actions and continuously adjusting according to the actual contact state. The team hopes that through the cooperation of these two layers of capability, robots can both understand tasks and execute actions properly in the real physical world.

To realize this capability, ManuPrax Robotics chose the reinforcement learning route. Li Puhao explained that, unlike mainly relying on human demonstrations for the model to imitate actions, reinforcement learning attempts to let the model gradually find effective strategies through exploration, trial and error, and feedback. The model tries an action, receives a reward signal based on the result, and then adjusts its next action. Through constant repetition, it has the opportunity to accumulate successful experience and gradually adapt to different objects and environments. "Human daily actions are very rich, and the shapes, weights, materials, and placement of objects vary greatly. Relying solely on limited manual demonstrations makes it hard to cover all situations. If the model can learn through interaction, it has the opportunity to reduce dependence on manual demonstrations one by one."

But scaling this route is not easy. Reinforcement learning requires demonstration-based and general reward signals, as well as a sufficiently efficient and stable training architecture. At the same time, a model learning an action in a simulation environment does not mean it can reproduce it on real hardware. Friction, contact, deformation, and sensor errors in reality can all invalidate successful experience from simulation.

According to the two, the team is exploring combining simulation, real human demonstrations, and generated interaction trajectories to build a training data system for further validation, so that the action model continues to improve as training scale and data volume expand.

Dexterous hands are not just a hardware business

Almost all robots need a pair of capable "hands." In the past few years, competition in the dexterous hand industry has mainly revolved around mechanical structure, drive methods, degrees of freedom, load capacity, and cost. As more companies enter the track, competition has also begun to extend from pure hardware performance to data accumulation, manipulation algorithms, and software-hardware synergy capabilities.

In today's dexterous hand market, there are both products developed in-house by robot body manufacturers and independent dexterous hand suppliers. The former can incorporate dexterous hands into the unified design of the whole machine system, while the latter need to win different customers through product performance, cost, and delivery capability. The opportunity for independent suppliers lies in serving multiple robot manufacturers, but they also face competitive pressure from in-house development by whole-machine manufacturers. Who can become a general supplier will ultimately depend on whether the product can establish an advantage among performance, cost, and reliability.

In the view of ManuPrax Robotics, the hardware landscape may soon enter another stage. Li Puhao judged that the development speed of domestic dexterous hand hardware will be very fast, and within the next six months, the profit margin of the pure hardware business may be significantly compressed. This is an important backdrop for the team's choice to extend toward manipulation models at the physical interaction layer. If hardware gradually becomes a more standardized product with thinner profits, then the manipulation capability that determines whether robots 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 connects upward to the manipulation goals given by multimodal large models and downward adapts to different dexterous hand hardware, providing callable manipulation capabilities for upper-layer models. Li Puhao believes that, compared with algorithms tied to a single dexterous hand, such a platform has the opportunity to serve different manufacturers and application scenarios and reduce dependence on single hardware sales.

But the imaginative space of platformization does not mean software commercialization will be easier. When buying hardware, customers can directly compare load, degrees of freedom, price, reliability, and delivery cycles; when buying manipulation models, the questions customers need to answer are more specific: Can the model complete tasks that were previously difficult to achieve? Compared with existing algorithms or in-house solutions, how much can it improve manipulation success rates, and how much training and deployment time can it save? If the model can only perform well in a few demonstration tasks, customers may not be willing to pay for it separately. Only when the model can continuously improve manipulation results in real scenarios and remain effective after migrating to different hardware does it have the opportunity to become a capability layer in robot systems that can be independently delivered and reused.

ManuPrax Robotics has not completely ruled out the path of combining software and hardware. The team hopes to maintain the platformization direction of the model while also considering accelerating technical validation through self-developed hardware or cooperation with hardware manufacturers to obtain data and feedback from real scenarios.

Li Puhao mentioned that at this stage, it is not sufficient to measure the value of dexterous manipulation models simply by shipment volume, repair rate, or customer retention rate. Some demand for dexterous hands still comes from research institutions and R&D projects of large enterprises, and the completion of equipment procurement does not mean the robot can already stably complete complex manipulations. But if commercialization is always postponed on the grounds of technical validation, the value of the model will likewise be difficult to test in the market.

The "ambition" of two post-00s founders and a "game" at the Institute of Telecommunications

In the dexterous hand track where competitive tiers have already formed, why can ManuPrax Robotics still attract attention? The technical background of Xiaotiancai is certainly one aspect, but more important is probably where the two young people want to take this technology.

In Li Puhao's view, dexterous hands are not a product that can be done well through a single-point technological breakthrough, but a systems engineering effort. Broken down, many technical details may not be complicated; what is difficult is integrating them so that hardware, models, and manipulation capabilities form a truly usable system. Their advantage is that the team did not enter temporarily after seeing the track heat up, but has long conducted research around related problems, knows where the technical difficulties are, and is clear about what needs to be solved next.

The complementarity between team members makes further transformation of this technical accumulation possible. Chen Zeyuan focuses on technical issues, while Li Puhao can think about the company's development and external resources from a more macro perspective.

Technical barriers and team background explain why ManuPrax Robotics has the opportunity to accomplish this, and what also drives them to start a business is their understanding of the value of research and personal choice.

Chen Zeyuan conducted research at the institute, collaborated with Li Puhao, and later went on to work at Amazon's research lab. Through this experience, he gradually realized that when he and Li Puhao collaborated domestically, their efficiency in discussing problems and advancing projects was no less than that of overseas research environments, and they had greater autonomy. They could decide for themselves what to research and how to proceed, rather than searching for a direction that suited them within an established organization.

"More grounded, more focused on solving real problems," Chen Zeyuan said directly when talking about his motivation for starting the company. He believes that in the past, the value of research was more reflected in raising questions, exploring methods, and validating ideas. Entrepreneurship gave them the opportunity to go further—to turn research results into products and let the technology be tested in real-world scenarios.

Li Puhao's view of the relationship between research and business also changed during the entrepreneurial process. Previously, he was more inclined to first build an actual product before discussing financing and commercialization. A product is the most direct proof; whether the technology holds up must ultimately be answered by actual results. But entrepreneurship gradually made him realize that business does not always begin only after a product matures. Entrepreneurs can also first propose an idea, clearly explain what problem it aims to solve and why it is worth doing, and then seek resources to make it happen.

The ambitions of these two young people go far beyond building a more dexterous hand.

Chen Zeyuan's greatest ambition is to see robots replace humans in doing mechanical, repetitive work. In his and Li Puhao's vision, dexterous hands should ultimately be able to take over most of humanity's physical labor. Machines would not merely repeat actions according to instructions, but could complete real tasks in complex environments, so that people no longer have to spend large amounts of time on heavy, repetitive work.

The Beijing institute has also provided systematic support for this startup. In Zhu Songchun's plan, to establish a continuous learning loop from the real world to simulation and back to the real world, the institute's various platforms and ecosystem companies, such as "Tongjie," "Tongjing," and "Tongnao," all need to take on corresponding roles within this architecture and form synergies.

ManuPrax Robotics is the official dexterous manipulation company established under the institute's ecosystem architecture, focusing on the foundational model architecture for dexterous manipulation, and is an important component of the aforementioned Real2Sim2Real system. The "last centimeter" of fine manipulation involves friction, deformation, sliding, and contact switching, and is also a key problem that must be solved when simulation capabilities move toward the real world.

This capability can connect with "Tongnao's" task understanding and planning, coordinate with the simulation training and evaluation of "Tongjie" and "Tongjing," and undergo real-machine testing in collaboration with ecosystem partners in robot bodies and tactile perception. The problems discovered and experience accumulated by dexterous hands in real operations can in turn improve simulation and training. In the future, the "Mingjing" platform to be established at the Chongqing institute will also provide a foundation for this system's industrial applications.

This article is from the WeChat public account "ChinaVenture," author: Wei Xianghui, published by 36Kr with authorization.