Boya Intelligence's Liu Xin: After GPT-6, Hardware Will Hold Robots Back

Last year, nearly 20,000 dexterous hands were shipped, and most of them are "gathering dust"

English text is being prepared.

English text is being prepared.

01

20,000 dexterous hands, most of them gathering dust

▎AI Technology Review: Public data shows that last year's shipment volume of dexterous hands was 20,000 units, but Boya only launched its product in June this year. Do you feel you are one step behind the market? Liu Xin: Most of those 20,000 units were probably sold to cooperate with humanoid robots for performances, and after the performance they gathered dust. In fact, the dexterous hand industry has not converged in various senses, including technical routes, product definition, and customer value. The industry is still in the early stage of the Tesla Roadster and Li ONE, or even earlier than Li ONE, because Li ONE already had a very clear boundary of product definition, while current dexterous hands do not. So one cannot simply say who is late.

▎AI Technology Review: Based on your observation, what is the current market situation for dexterous hands? Liu Xin: Today's dexterous hand products are polarized. One type is low-degree-of-freedom, simple, durable, and cheap, available for a few thousand yuan, but with very limited manipulation capability, mainly used for humanoid robot performances, waving hands and the like; the other type is high-degree-of-freedom, with better manipulation capability and also high cost, with mainstream prices above 100,000 yuan, but also unusable, mainly bought by research institutions. Today we are creating the third market for dexterous hands, bringing part of research-grade manipulation capability into practical applications while keeping the cost within a range customers can accept, solving problems that past automation equipment could not handle. However, such a product still does not exist today; the capabilities of today's most advanced and high-end dexterous hands on the market are very far from human hand capabilities.

▎AI Technology Review: Why are high-end dexterous hands with better manipulation capability unusable? Liu Xin: Because everyone is mostly flexing their muscles. Today they prove they can twist off a bottle cap, tomorrow they prove they can tighten a screw, and the day after tomorrow they can crack an egg. But for real deployment, customers will first of all certainly find it hard to accept a price of 50,000 or more than 100,000 yuan. People who have not done robot deployment will not know the cost of deployment. A robotic arm may cost only 200,000 yuan to buy, but to put it to use, an additional 1 million yuan may be needed; the same is true for dexterous hands. Second, today's mainstream high-end dexterous hands may have a service life of only one or two weeks, or at most one or two months. Today everyone is challenging the limits, showing how fast they can run. But you cannot take the speed of a 100-meter sprint as the average speed you can maintain when running a marathon.

▎AI Tech Review: It sounds like the current situation is that the capabilities of dexterous hand products are overestimated, and the difficulty of making good dexterous hands is underestimated. Liu Xin: Right. Everything about dexterous hands is still immature; it is all bottlenecks, and from any single angle there are challenges. Moreover, a dexterous hand is a system that needs to integrate everything together, which is an even bigger challenge.

▎AI Tech Review: Broken down, which part is most prone to problems? Liu Xin: Motors and transmission structures are the most prone to problems in dexterous hands. This is a contradiction in dexterous hand design: it needs to fit mechanical structures, motors, circuits, and sensors into an extremely limited space while also ensuring sufficient strength and power density. The thinner the fingers are made, the more easily structural strength and load-bearing capacity are limited, especially at the joints, which may only withstand a few hundred grams to about one kilogram of force and break after a week of continuous use. If the hand is to be made small, the gears must be made precise, the module must be small, and the gear teeth must be smaller, which places higher demands on machining precision. For example, gears only a few centimeters in diameter in a dexterous hand may have a module of only a few tenths. At such sizes, with today's materials and process levels, gears are very prone to wear, deformation, and even fracture.

▎AI Tech Review: Under fine structures, ensuring strength is also a problem. Liu Xin: When executing tasks, strength must be considered. A human thumb has 10 kilograms of force. Today, the mainstream rated force of a dexterous hand thumb is 1 kilogram, with the higher end at 2 kilograms, but there is a deviation between the rated force and the force actually that can be applied; a rated 1 kilogram may in fact not reach 1 kilogram. We have achieved 3 kilograms.

▎AI Tech Review: What about touch? Are there truly usable tactile sensors now? Liu Xin: Tactile sensors are still very far from being truly usable. We have tested tactile sensors based on many different technical routes, but none is perfect; all work with injuries. Right now I do not even care first about durability; the stability and consistency of the signal itself still have many problems to solve. For example, today's sensors still cannot well eliminate the noise caused by residual signals. After a person picks something up and then lets it go, perceptual signals such as force sense and vision still leave residuals, but after hundreds of millions of years of evolution, the human cerebrum, cerebellum, and nervous system have learned how to cancel out this noise; today's sensors cannot yet do this. For a dexterous hand to have perceptual capabilities close to a human hand, it still needs to solve the problems of perceptual precision and distribution. The human hand has 17,000 sensory nerves, but they are not evenly distributed; like geography, there are first, second, and third levels. The palm and back of the hand are the sparsest, the finger pads may be about twice that, and the fingertips are about 4 to 5 times the palm. The closer to the fingertip, the finer it is. Some electronic components have metal pins on them, and workers need to press with their fingertips and rely on the feeling of force to judge whether the pins are firmly connected. These abilities that humans are born with are very difficult for robots.

Objectively speaking, although sensors currently have flaws, they are still basically usable. Rather, models have not yet fully exploited tactile capabilities. Once this hurdle is passed, it will also give clearer directions for sensor improvement.

▎AI Tech Review: Is reliability also a bottleneck for the industry now? Liu Xin: I think that at this stage, the most important thing for the industry is still not reliability. Just like when the iPhone first came out, people would say it had to be charged every day, while a Nokia did not need charging for half a month, but users were willing to accept this tradeoff for the experience improvement brought by intelligence. The same is true for dexterous hands. Durability certainly cannot be only one week, but at this stage we should at least strive for three months to half a year. The industry's mainstream warranty period may currently be 6 months, and few dare to promise more than one year. But this 6-month warranty is often supported by preparing a large reserve fund in advance. For example, when a customer buys one dexterous hand, the manufacturer may need to prepare the cost of three. Tesla also went through a similar stage early on. At that time, some components were easily damaged, and the company's approach was to set aside a large sum of money and directly replace them for free after customers reported repairs. New technologies and new products going to market may all have to go through such a stage.

▎AI Tech Review: There is a saying in the industry that 70% to 80% of tasks can be completed by grippers. Will the market for high-degree-of-freedom dexterous hands be smaller than imagined? Liu Xin: It is easy for everyone to use currently visible tasks to predict future demand, but human imagination always underestimates the possibilities of the physical world. When IBM first launched mainframes, some people once thought that the world might need only a few computers. Steve Wozniak took the earliest Apple computer prototype to Hewlett-Packard, and HP was also not optimistic about personal computers, thinking who would buy this thing? When the iPhone was first launched, who could have imagined that so many applications would later appear, even developing into hundreds of millions of application scenarios? People always judge a new technology according to existing application scenarios, and it is hard to imagine a market that does not yet exist. Only after technological capabilities improve will you gradually discover its potential. Of course, this takes time. The same is true for dexterous hands. Today everyone thinks grippers can already complete many tasks, but only when dexterous manipulation capabilities improve further will new application scenarios gradually emerge. For example, some small factories around Shenzhen need to inspect parts under high-magnification electron microscopes after producing them, and every face of the part must be turned over. Traditional grippers certainly cannot do this kind of task.

▎AI Tech Review: So why must it be five fingers? Wouldn't three or four fingers work? Liu Xin: Today's robots, whether we're talking about the "brain" or the "cerebellum," still have far from sufficient understanding of space and physics. So when we think about how robots should be built, it's easy to oversimplify the problem. From the perspective of biological evolution, humans have to face a large number of situations that cannot be determined in advance, so the human body's design doesn't just satisfy the most frequent needs; it also retains redundancy for low-frequency, sudden situations. Humans eventually evolving five-fingered hands follows a similar logic: many actions can be completed with two or three fingers, but five fingers leave more capability space for complex tasks. Among the five fingers, the thumb is the strongest; an adult's thumb has about ten kilograms of force, followed by the index finger at about six or seven jin, then the middle finger, ring finger, and little finger. Weaker fingers usually play supporting, assisting, and exploratory roles. Some statistics show that, compared with five fingers, each fewer finger in a four- or three-fingered hand may reduce capability by about 10%. This kind of redundancy also exists in the human brain and motor system. We rarely use our muscles to full strength in daily life, or continuously exert our fastest speed and greatest endurance, but these capabilities are still retained to cope with sudden situations. Robots are the same: they need to reserve enough capability space for complex, low-frequency tasks.

▎AI Tech Review: Getting down to specific work scenarios, how many fingers are actually needed? Liu Xin: At least 4 fingers are needed. In factories, when using power tools to drive screws, if there are only two fingers, you can't pull the trigger; if there are only three fingers, only two can grip the screw gun while the remaining one pulls the trigger, and in fact you still can't hold the tool steadily.

▎AI Tech Review: What about the little finger? Tesla's engineering team also struggled over whether to include this finger. Liu Xin: The little finger is weaker and plays more of an auxiliary role. Often, people don't even realize its value; only after losing it or getting injured do they discover its importance, just as when walking we rarely pay attention to which specific muscles are used, and only feel the impact after a sprain. Let me give a counterintuitive example: around 2018, OpenAI bought a bunch of Shadow hands to solve Rubik's cubes. When solving the cube, they noticed that the final trained model liked using the little finger more than humans do. Because Shadow's five fingers are equally strong, the model found that if it could use this finger, overall efficiency and success rate would be higher, so it was more inclined to use it. So I think being humanoid is the first step, and surpassing humans may be the future.

▎AI Tech Review: I've also seen some soft robotic wrist and arm experiments that imitate octopus tentacles. If materials or algorithms catch up in the future, could such approaches be more useful than multi-fingered dexterous hands? Liu Xin: It's possible. The human brain is highly adaptable. With training, people quickly treat a third tool as an extension of the body. Like using chopsticks, using a remote control to play racing games, flying drones, or driving a manual-transmission car—once proficient, the actions don't require deliberate thought, as if muscle memory has formed, ultimately reaching a state of "man and machine as one."

▎AI Tech Review: Direct-drive solutions currently occupy the mainstream in the high-end dexterous hand market, but you make tendon-driven hands. From an industry perspective, will direct drive's mainstream position continue forever? Where is the opportunity for tendon drive? Liu Xin: I can make a judgment: today's leading direct-drive dexterous hands have already approached physical limits. Existing performance can still be fine-tuned, but within the next one to two or two to three years, major improvements will be difficult to achieve. With our technology architecture, which is mainly tendon-driven, I think there is still a lot of room for performance improvement, and there will still be opportunities for major progress in the future. Direct drive is temporarily ahead because it is easier to integrate. But product R&D for tendon-driven hands is accelerating and iterating. When new tendon-driven hand products come out later, everyone will see that tendon-driven hands will surpass today's top direct-drive hands in certain respects.

▎AI Tech Review: There is a saying that as long as motor modules mature, direct-drive hands can be done well. Do you agree? Liu Xin: Mature modules don't mean the whole machine will be easy. It's like how a mature engine doesn't mean building a good car is easy. Moreover, we have communicated with leading domestic motor manufacturers, and their current R&D plans show that improvements in motor performance are not as fast as imagined. Iteration of dexterous hand sensors and mechanical configurations will be relatively faster, and mechanical structures are easier to adjust. But reducers and motors involve more complex processes, manufacturing, and mass-production issues, so iteration is much slower. Take reducers: the micro cycloidal solution that everyone has discussed for many years still hasn't truly achieved large-scale mass production and widespread application. The mainstream is still planetary and harmonic reducers, because these solutions have been verified over the long term and are relatively mature in process, quality control, and cost.

02

Models will break through earlier than hardware

▎AI Tech Review: Why do you think a hybrid drive based mainly on tendon drive would be a better solution? Liu Xin: There are many contradictions among the performance metrics of dexterous hands today. For example, with strong load capacity, fine manipulation may be limited; if made slender and agile, strength and durability may be insufficient. We chose a hybrid drive based mainly on tendon drive precisely in the hope of balancing finger slenderness, strength, and durability. We are also the world's first dexterous hand company to use a tendon drive + lead screw parallel wrist configuration. Our design logic for dexterous hands is the same as the exquisite hand structure that humans formed through millions of years of evolution. When a human finger gently pinches, fine control is mainly completed by small muscles such as the lumbricals and intrinsic hand muscles; when forcefully pulling or lifting something, the muscles that truly generate force are the arm muscles. Therefore, for a biological system to balance strength and dexterity, it usually needs to separate the two: the closer to the end, the more slender the structure and the denser the nerve distribution, the more conducive to fine control; the closer to the root, the thicker and stronger the muscles and structure, the more suitable for providing force. The human arm and hand are a division-of-labor system formed through long-term evolution. We know that this design will increase the integration difficulty of humanoid robots in the short term, but we still follow this direction because it is the more ultimate solution. If Musk insists on this direction, mainstream robot manufacturers may also follow suit.

▎AI Tech Review: How exactly is your hybrid drive implemented? Liu Xin: Our fingers are tendon-rope driven, the side-swing joints use direct drive, and most of the motors are placed in the palm. The wrist uses a lead screw + linkage, and motors are installed in the forearm to provide greater force. When we made early prototypes, we also tried using tendon drive to achieve motions such as side swing, but later found that this significantly increases engineering complexity and reduces reliability, so we changed these joints to direct drive.

▎AI Tech Review: How were the 22 degrees of freedom determined? Liu Xin: The palm has 20 degrees of freedom, and the wrist has 2 degrees of freedom. Actually, at the beginning we designed 23 degrees of freedom, and in the end we cut the degree of freedom in the palm that drives the little finger to retract. Shadow previously had this degree of freedom, and like OpenAI, they very much liked using this degree of freedom. Because of engineering complexity and because most customers do not need this degree of freedom, we gave it up. For tactile sensors, we currently use magnetic Hall sensors, which are relatively better in various dimensions.

▎AI Tech Review: You previously mentioned in a report that the precision of a dexterous hand should be called "the minimum controllable precision" rather than repeat positioning precision. Why redefine precision? Liu Xin: The understanding of precision is changing. In the past, repeat positioning precision was emphasized because the standards of traditional industrial robots were followed. But the situation is different now. I think what will be more important in the future is "the minimum controllable precision," that is, how finely the system can control motion and continuously correct it based on real-time feedback. This logic is similar to the coordination among human muscles, nerves, and the cerebrum and cerebellum. Humans can control motions to a very fine degree through slight muscle contractions, and then continuously correct them based on feedback. Future robots will also increasingly need this kind of control capability. This is also an advantage of tendon drive. Tendon ropes have creep, which can be solved through our "perception-correction" system. If gear direct transmission is used, factors such as gear tooth profile and machining precision will affect transmission precision and backlash. Backlash is difficult to eliminate completely. With long-term use, gears may also wear, deform, or even break, further affecting transmission precision.

▎AI Tech Review: The industry believes that tendon-driven hands are difficult to control. How do you solve this problem? Liu Xin: We have initially developed a "cerebellum" for controlling the dexterous hand. This "cerebellum" is not designed strictly according to the division of labor between the human cerebrum and cerebellum. Our idea is to separate part of the motion control that requires real-time feedback. The "cerebellum" has a built-in motion library containing 33 motions commonly used in daily operations. These motions include hand posture, execution timing, and force changes. The cerebrum only needs to determine which motion to invoke, and the cerebellum is responsible for specific execution and adaptive adjustment, without requiring the upper-level model to re-reason every time. In this way, the control dimension of the cerebrum can be simplified to "7+1": a 7-axis robotic arm plus one hand motion number can complete relatively complex operations. The high-level model only needs to issue instructions such as moving the wrist or executing a grasp, and specific joint control is left to the cerebellum. Many teams working on cerebrum models are not good at driving high-degree-of-freedom dexterous hands. By encapsulating the low-level capabilities well, they can integrate them more conveniently.

▎AI Tech Review: Did you summarize these 33 motions yourselves? Liu Xin: There is a commonly used motion classification standard in the industry, first summarized by some top universities in the United States, with about 33 motions. Professor Zhang Shiwu added 6 more on this basis, so our current reference target is 39 motions. However, it is not easy to do all 39 motions well. For example, using chopsticks is very common for Asians, but it is still very difficult for a robotic hand.

▎AI Tech Review: Which customers are your dexterous hands mainly aimed at currently? Liu Xin: Currently we have two types of customers. One is the scientific research market, and the other is companies working on cerebrum models. They are beginning to move toward finer dexterous manipulation, so they need better dexterous hands. But if they are still at the stage of using the π0 open-source model for post-training, they basically do not need dexterous hands yet.

▎AI Tech Review: Why is it that if you use the π0 open-source model for post-training, you can't use a dexterous hand? Liu Xin: Because the pure VLA architecture of π0 has extremely high requirements for data and compute. The current limit this architecture can support for training is a robotic arm system that includes a gripper, has 7–8 degrees of freedom, is equipped with 2–3 cameras, uses action chunks of about 50 frames, and refreshes at around 10 frames per second. If you switch to a dexterous hand with more than 20 degrees of freedom and connect it to this architecture, first, data collection and training will be very difficult to converge because the data volume is too large; second, inference efficiency cannot support it either.

▎AI Tech Review: Will the layered cerebrum-cerebellum approach be the ultimate architecture? Liu Xin: Given today's development trend, it definitely will. One reason is that the cerebrum side is low-frequency reasoning, while the cerebellum side is a high-frequency closed loop that must continuously make real-time adjustments based on feedback such as force and position; the other is that at the cerebrum layer, leading large-model companies and tech companies could do it well by investing a bit more resources. One recent development that has attracted considerable attention in the industry is that GPT-6 can already operate a dexterous hand. This shows that the common sense and cognitive abilities that large models accumulate during pre-training are extending into the physical world. Humans learn movement through three parts: innate motor control, acquired cognition, and practice. With enough practice, muscle memory forms. Once a person becomes an expert, the cerebrum is mainly responsible for judging strategy—for example, when a tennis ball comes, what shot to play and where to place it—while the specific bodily coordination and movement adjustments often no longer require conscious step-by-step thinking. If a large number of commonly used cerebellum skills covering the vast majority of scenarios can be trained well and then built into the robot system, it would be somewhat like directly loading programs into the brain in The Matrix, so that after waking up you can say: I know Kung Fu.

▎AI Tech Review: Hardware, data, the cerebellum, and the cerebrum—which of these four parts do you think needs to be focused on conquering right now? Liu Xin: I think in the first stage, hardware still needs to be done well. Given the current iteration speed of AI and models, it may reach a tipping point earlier—what everyone calls the "GPT moment"—and models will break through before hardware. By then, everyone will find that models are already very capable, but there is no sufficiently reliable hardware to carry their capabilities.

▎AI Tech Review: Why do you think models will break through before hardware? The more common view in the industry is that hardware is not good enough and is limiting model development, so hardware should break through first. Liu Xin: Models are developing faster than people imagine. For example, the previously mentioned GPT-6 can operate a dexterous hand. From large models to multimodality to robot control, the relevant capabilities may soon achieve breakthroughs. In the end, you will find that what is holding you back may be the body, not the mind. It is the same with people. I also wish I could sprint 100 meters as fast as a professional athlete and play tennis for five hours without getting tired, but in the end what limits you is still your body. Hardware is the same: progress in materials, structures, electromechanics, and sensors all takes time. There are two major thresholds for deployment today: one is the model, and the other is hardware. Models may soon meet the standard; they are still imperfect, but in specific scenarios and specific tasks they will soon be competent. Once models are competent, the bottleneck will become the hardware itself.

03

The real barrier is making the entire system work

▎AI Tech Review: Does Boya position itself as a dexterous hand supplier? Liu Xin: We do not position ourselves as a pure dexterous hand supplier; we will also do models and applications. When a technology is too advanced and the industry ecosystem and application scenarios are not yet ready, being only an upstream hardware supplier makes it very difficult to truly build up the market. Nvidia is a typical example. About ten years ago, Jensen Huang produced very advanced dedicated chips for autonomous driving, but at the time downstream customers found it difficult to use them directly. So Nvidia invested heavily in building an algorithm team and personally filled in the most basic algorithm infrastructure and toolchain for autonomous driving, turning the chip capability into a highly mature semi-finished product or platform before handing it to downstream customers. In this way, customers could integrate it into their own products more quickly. Dexterous hands are likewise an advanced technology. If you only sell hands while the application ecosystem is immature, it is very difficult to achieve large-scale deployment. By doing both models and applications, we turn dexterous hand capabilities into solutions that can directly solve problems, helping customers and scenarios get up and running faster. This is actually more conducive to the dexterous hand industry truly developing.

▎AI Tech Review: Does that mean you will move toward complete machines in the future? Liu Xin: As we develop further, we may have two commercial forms. First, the customer has secondary development capability and only needs to integrate our product; we do the core product well and support secondary development. But we must admit that today the threshold for secondary development of robots, especially dexterous hands, is very high. We need to lower this threshold at the product level, and that is what we can do in the short term. Second, for customers without secondary development capability, when there is clear customer demand, we can give them a complete solution. From the perspective of a complete machine, the hardest part is the dexterous hand. If I can do the dexterous hand, integrating the complete machine is relatively easy. Robotic arms are relatively mature, and the chassis can also directly integrate an off-the-shelf one.

▎AI Tech Review: Some people say the hardware barrier for dexterous hands is not high and can be copied. What do you think? Liu Xin: Those who say the barrier is not high may be making promotional claims as latecomers seeking financing, or they lack deep understanding of the industry. The key to making dexterous hands is still understanding the entire system; it is not something you can do well just by putting some parts together. After completing the system design, you will gradually find that many components can no longer be directly purchased as off-the-shelf products; they need to be customized or even self-developed. Some things can be done with money, but some things, even with sufficient funds and a high enough density of talent, still require a long time to polish. This know-how itself is a barrier. Our team has been working on this for many years. Many problems are only known once you have actually done it; if you do not do it, you do not know where problems will arise. This experience must be accumulated slowly through actual R&D and iteration.

▎AI Tech Review: Then why do you think Boya can stand out? Liu Xin: First is technology and iteration speed. We master many core technologies, and technology iteration is also fast; second, the core team and supply chain management are both on track; third, we also have many customers willing to try deeply. I think that as long as we continue this way, there is no need to worry commercially. Serve customers well and iterate products well, and basically we will not fall too far behind.

▎AI Tech Review: How is the company's team division of labor now? Liu Xin: I am mainly responsible for product and engineering, while Zhang Sicheng focuses on business. The company's CTO is Chen Haoyao. He studied robotics as an undergraduate but leaned toward mechanical engineering, and later shifted to robot control at the algorithm level. He is a rare composite talent in the industry who integrates software and hardware, and he sees a problem more clearly when examining it across disciplines and functions. Professor Zhang Shiwu works on more frontier exploration. In addition, because I also come from a technical background, I count as half a CTO. I discuss issues with them and exercise some control over the major technical direction.

▎AI Tech Review: Having composite talent who integrate software and hardware—would you also count that as a core advantage of your company? Liu Xin: The capability structure of the team affects the direction of product development. Take new energy vehicles as an example. To make a good product, you need engineers from traditional automakers responsible for hardware, engineers from large technology companies like Microsoft and Motorola to develop complex software and hardware in a systematic way, and product managers, designers, and engineers from leading internet companies to improve the software-hardware interaction experience. Although some established enterprises now have strong technical accumulation in the motor field, they may be half a beat slower in product iteration, or when designing the underlying architecture, they lack sufficient consideration of upper-layer software models. The same is true for dexterous hands. The team needs to understand software and also understand what requirements the upper-layer models will place on hardware design, so that it can design a truly suitable hand. The two complement each other. A team that only does models may know what capabilities its ideal dexterous hand should have, but not know how to design it specifically; while a team that only does hardware may not necessarily know what kind of hand is more suitable for model-driven and complex manipulation.

▎AI Tech Review: How did the core people in your team come together? Liu Xin: I count as the initiator, and Zhang Sicheng was very key. He was the person Alibaba first sent to the Banma Zhixing project and worked with me. So at the very beginning I asked him to talk with me. Although my past experience was mainly in software, it has always been deeply combined with hardware. When I was at big companies, I always wanted to do hardware in my heart, and I had long planted an idea: after finishing cars, I must do robots. After I raised this idea, it happened that Zhang Sicheng knew two professors, both his alumni. Quite a lot of the two professors' technology can be directly used in product R&D. The several of us talked for about half a year before gradually settling the direction.

▎AI Tech Review: Why were the two professors willing to join a startup? Liu Xin: Zhang Sicheng has known them for many years. Chen Haoyao is Zhang Sicheng's triple alumnus; they attended the same school for high school, undergraduate studies, and PhD, and they are also fellow townsmen from Yongding, Longyan. They are already very familiar with each other and trust each other very much. Teacher Zhang Shiwu mainly provided technical accumulation and guidance in dexterous hands in the early stage, including existing patented technologies, and some of these patents may also be used in our products in the future. Teacher Chen's participation is deeper. Because he is at Harbin Institute of Technology, Shenzhen, geographically closer to the company, he will participate deeply in control algorithms, including the definition of technical direction and the design of the overall technical architecture.

▎AI Tech Review: You have done software for nearly twenty years. What use is this experience for doing hardware? Liu Xin: I count as a veteran of the industry. I previously did pure software, then internet, smartphones, IoT, and automobiles. I was previously at Sina and Alibaba, and later went to NIO. I have always been absorbing experience from these industries, including how to manage complex teams, how to deal with customers, how to set the pace of product iteration, and how to coordinate with suppliers. I think this experience is very important for making frontier products. The more emerging the industry, the more likely it is that you need to develop things yourself from scratch. Past experience has given me more understanding of the laws of industrial development, and made me clearer about what should be grasped in my own hands and what can be handed to external suppliers.

▎AI Tech Review: What is the organizational culture at Boya now? Liu Xin: I have experienced different types of companies and teams in the past, and I have also participated in many processes of building teams from scratch, team mergers, and even layoffs. I have found that the same culture can play a completely different role in different companies and at different stages. Alibaba's culture has been constantly adjusting, from being sales-led in the early days, to later placing more emphasis on engineering capabilities, and then to different businesses gradually forming their own organizational methods. So I think there is no fixed answer in management. The key is to understand the industry you are in and the stage of development, and then decide what kind of organizational approach should be adopted. Boya's current culture is flat and efficient, which is actually what Silicon Valley calls founder culture—diving into the front line and looking at everything together with everyone, down to the smallest details.

▎AI Tech Review: Why is it called Boya? Liu Xin: Professors Zhang Shiwu and Chen Haoyao previously founded a company called Bote (Zhongke Bote Intelligence), and Bote was named directly after the pronunciation of bot. At the time, we thought of many names related to the character bo. Later we thought we should highlight the dexterous hand. What are some symbols of being clever and deft? We thought of Boya. Because he played the qin, his hand control should have been very flexible. In the end, we settled on this name.

▎AI Tech Review: Many people think of Boya in terms of the story, and rarely in terms of playing the qin. Liu Xin: Boya and Ziqi is also a story about meeting a soulmate, and we believe robots should become humanity's soulmate. Most people treat it as a tool, a nanny, or something like that. If it can become a soulmate, that would be even better—it could provide more comprehensive help. In the past, robots mainly undertook physical labor and solved problems at the physical level. With the development of AI, in the future it can also communicate with people and provide help through language dialogue, agents, and other methods, covering more needs. (Leiphone)

04

Postscript Liu Xin's engineering background was laid down from childhood. His parents were communications engineers at the Ministry of Railways. At two or three years old, he played with diodes and transistors in the equipment room; in elementary school, he soldered circuits; and in fifth grade, he wrote code. After graduating from university, he first came into contact with vehicle monitoring systems at Lingtu, and got his first exposure to Shenzhen's hardware industry chain; in 2006, he started his first business, and later entered Sina and Alibaba. During nearly ten years in Alibaba's operating system business unit, he went through mobile phones, automobiles, IoT, and robots, participated in Alibaba and SoftBank's robotics projects, and also planted the idea that "I must make robots." In 2023, he joined NIO to complete the two lessons of full product lifecycle management and supply chain. In August 2024, he officially started his business, choosing the dexterous hand as the entry point. In his view, the dexterous hand is the part of the complete robot with the highest technical difficulty and the most prominent value. He believes this industry will not be left with only one company, and the company name also hides a farther-reaching expectation: robots should not be just tools or nannies; they should become humanity's soulmate. (Leiphone)