Exclusive Dialogue with Shi Xiaoxin: Half of the Endgame for Humanoid Robots Lies in Their Hands
"Workplace Bonus" (ID: ZhiChangHongLi)
"Aimo, Can Assist" · 01
Three Hurdles for Dexterous Hands · 02
Aesthetics and Anthropomorphism: The Path to Cross the Uncanny Valley · 03
Interpreting Aimo's Organization · 04
The Craftsman's Heart · 05
Shi Xiaoxin has not spoken publicly for four months.
The last time the outside world heard news about him was a Moments post in early June: "I am fortunate to have stayed here for 1675 days and met such a lovely and warm team... Pack my bags and set off again." In the attached picture, the farewell banner prepared by the team for him reads "My boss is free now". At that time, the XPeng humanoid robot IRON, for which he was in charge of product planning, had just gone viral across the whole internet with its catwalk demonstration, and its mass production was around the corner.
In late October, in Shenzhen, he sat in front of us as the founder of Aimo Robotics. This is his first media exclusive interview after starting his business.
Aimo focuses on dexterous hands — the hardest and most expensive component on humanoid robots. This track is no longer uncrowded. The annual shipment of dexterous hands from ITRobotics has exceeded 10,000 units, BrainCo, Fourier Intelligence, Pansini and other companies all hold their own positions, and an industry inventory lists as many as ten mainstream manufacturers... Almost every investor who meets him will ask the same question: Is it too late to enter the market now?
"Persistence makes timing irrelevant," he said. "You miss the dividends of this wave, but you also avoid the pitfalls of this wave."
This "veteran" who has been making robots for 13 years rarely talks about the endgame and imagination, and what we talked about more are things that don't sound sexy. Including the service life of the hand, manufacturability, hardware stability, and "no failure" after delivery.
He calls these the "lower limit".
When listening to Shi Xiaoxin talk about methodology, you will feel that he is like a craftsman. For example, he will say that anthropomorphic design needs to cross the uncanny valley by making "form, spirit and intelligence" each in its right place; recruiting people does not rely on persuasion, but "finding people who row the boat together, not those who sit on the boat"; the organization does not need to copy the processes of large companies, and atmosphere and efficiency are the only two indicators.
Perhaps all the known and certain things will be replaced by AI in the future. When talking about suggestions for young people, this cautious respondent gave the core insight about his career choices: "In this era, certainty is the biggest risk."
This sentence is also a high-level summary of his choices in the past four months.
"Aimo, Can Assist"
The name Aimo was decided in July. How did it come about?
Shi Xiaoxin : It is related to a phrase, AI Motion, AI plus motion intelligence. Later, as we called it more often, what was more understood and extended by everyone was AI plus more — AI extends to the physical world, and even more possibilities.
When we chat internally, we also extend it to a homophonic pun of the Chinese idiom "unable to help": Aimo, can assist — it makes us rethink the meaning of the term "little assistant" — "hand" can solve many problems.
When it was spread verbally, we found it very catchy, so we adopted this name later. Naming is a brain-burning process. It needs to be unique and represent the meaning we want to express, and there are not many good names to choose from.
What have you and your team been busy with in these four months? What are the things you didn't expect before?
Shi Xiaoxin : Building the company from 0 to 1. Registration, finding a site, completing the first round of financing, the whole process is a stage from learning to familiarizing, and then to being able to take control. What I didn't expect were more problems on processes and forms, such as legal affairs. Financing is a familiar yet unfamiliar thing for us — everyone knows the primary market and has talked with institutions, but when you actually operate a company and deal with institutions by yourself, many things are new. What content should the company show to institutions? What do institutions care about? How to operate and prepare in the long run? These are rarely involved when you are a professional manager in an enterprise.
Three Hurdles for Dexterous Hands
There is a saying in the industry that half of the final competition of humanoid robots lies in the hands. Do you agree?
Shi Xiaoxin : Basically agree. From the perspective of hardware and ontology, the value of the hand may be equivalent to the rest of the body, including the value and difficulty of operation. It can be said that half of the competition lies in the hands.
I prefer another understanding: the two parts that really play value are the hand and the brain. From this perspective, they account for 50% each, and the hand and the brain should develop in coordination. This also has a biological basis. Human hands are much more dexterous than orangutan hands, and the evolution of hands is coordinated with the evolution of the brain. The brain promotes the evolution of the hand, and the dexterity of the hand in turn promotes the evolution of the brain.
If you prioritize the influencing factors that make dexterous hands difficult to make — such as degrees of freedom, load, cost, service life, etc., what will you solve first?
Shi Xiaoxin : We now believe there are three key points.
The first is usability. It involves service life and stability, which should be put first. The problem everyone is facing now is not that there are no hands to use, but that there are no hands that can be used and can be used for a long time.
The second is manufacturability: make it with excellent technology, controllable cost and reasonable time. A very good product that cannot be manufactured effectively and quickly is difficult to form a real product.
The third is dexterity, which is related to degrees of freedom. From 2021 to 2024, the early task of the industry was more simple tasks such as grasping and placing. Even for dexterous hands, people actually use them as grippers (two-finger grabbers that can only open and close simply). In the past two years, many cutting-edge laboratories are more inclined to use high degrees of freedom to achieve highly dexterous operations, and people have generally found the bottleneck of grippers and low-degree-of-freedom hands. In the next stage, the requirements for dexterity will become higher and higher.
The degrees of freedom itself is relatively certain. We have the ability to make dexterous hands with the highest degrees of freedom run normally. If we choose to reduce the degrees of freedom, it is more for the consideration of cost and structure simplification.
As for the load, it is not the main contradiction at present. What is more important for dexterous operation is the complexity of operation, and the load can be solved with the help of tools. As for the cost, it is mainly reduced through design and technology, which is the capital for mass production, and we are quite good at it.
Are there any technical paths that you are determined to avoid?
Shi Xiaoxin : None. But we will not explore all the mainstream driving schemes of cable drive (driving fingers by ropes), push rod (pushing by connecting rods), and direct drive (installing the motor directly into the joint). We have drawn some conclusions from past experience, and we will stick to the direct drive path. On the direct drive track, we maintain an open mind to explore the possibilities of various structures, joint transmissions and motors.
In March this year, at the seminar of the Artificial Intelligence Standardization Technical Committee of the Ministry of Industry and Information Technology, you put forward an idea of general intelligence classification for humanoid robots, dividing autonomous capabilities into several levels. According to your own framework, at which level are the top products in the industry today? Which level does Aimo want to reach?
Shi Xiaoxin : First correct a premise. The hand is more like a key component. We are unlikely to classify the dexterous hand itself for intelligence, because it cannot represent the entire robot ontology. For the hand to generate value, there must be an arm; for the arm to operate, there must be a controller (brain); further expanding, if the legs or chassis can move, it becomes an ontology. Therefore, to generate value, it requires at least an arm and a brain, which in a sense is the classification of robots.
If we look at the classification of general robots — there is a difference between technical exploration and products now. Many people think that the technology has reached a certain level, but the product standards are higher than the results of technical exploration. Measured by product standards, I think the whole industry may still be at L2 now, and some even fail to reach L2, and some extreme scenarios are ignored.
From the perspective of hardware, the maturity of hands will be two to three years later than the ontology, and the measurement standard is: whether it can be produced on a certain scale and obtained at a reasonable price. At present, the cost of two hands may be more expensive than the rest of the ontology.
For Aimo, we will first do a solid job in the hardware based on L2 capabilities, motion control, and underlying edge-side models (models running locally on the device without relying on the cloud), so that we can have the foundation to build L4 and L5 in the future. If even teleoperation cannot be operated well, and the hand still malfunctions and breaks when in use, there will definitely be bottlenecks when intelligence goes up.
How long do you estimate this process will take?
Shi Xiaoxin : In the next one to two years, the operation capability based on dexterous operation itself will be greatly developed. To reach L4 and L5 to form the physical AI and landing scenarios in our ideal, it will take a longer time, such as 3 to 5 years.
But these 3 to 5 years do not represent a clear boundary. It is more like a stage, which may happen first in some scenarios, not done well enough, or can only be done in this one scenario, and then expand to ten, one hundred, ten thousand scenarios. The whole process may last longer, and we are prepared for long-term investment.
You once said that you want to "insist on making real robots instead of painting big pies". In your opinion, what is a real robot and what is a big pie?
Shi Xiaoxin : A real robot is a product that can truly solve customer problems. It is a lower limit and needs to deliver value. The pie does not have a very derogatory meaning. It is more the possibility of technology, the upper limit, but this upper limit depends on many conditions and time to be fulfilled.
It is difficult to go up and show the upper limit of technology; but it is even more difficult to go down, make the product land well, and fulfill the value in user scenarios. What we really need to do well now is what your lower limit is: can the hardware be delivered? Can it be really used after delivery? Will it not break when in use? "Dig into the details of the hardware and polish every detail" is the problem that we at this stage — or many enterprises that really make robots — have to face.
There are many scenarios at this year's World Robot Conference, including sorting, catering, medical treatment, and inspection. The process is definitely tortuous, but this kind of attempt is necessary — the attempt is different from the pie. If the attempt does not meet the final requirements, it is just still in the process of moving forward. For the sorting scenario, there will definitely be several large-scale and well-done enterprises in the market in the future, but so far, there are relatively few that can truly deliver, and the realization of value still depends on the stability of hardware, the rhythm and success rate of the whole solution.
I heard that you once determined to be Unitree in the dexterous hand industry. Unitree is also one of the robot companies with the largest shipment and revenue scale in China. Is this considered drawing a pie for yourself?
Shi Xiaoxin : Maybe yes. (Laughs) But Unitree is definitely a great enterprise. This statement is more about aligning an idea internally: to achieve large-scale applications, there are some concrete industry indicators — for example, whether customers are widely accepted, whether there is a certain scale, and it is difficult to judge the quality of things without scale. At the same time, from the perspective of enterprise operation, we need the ability to generate revenue. Unitree represents a standard in this direction.
Aesthetics and Anthropomorphism: The Path to Cross the Uncanny Valley
People who have met you say you are a person who cares a lot about aesthetics. Is this trait innate or trained?
Shi Xiaoxin : I don't have an art professional background, but when you see a beautiful, well-designed product, you will naturally feel that it is a product that has been carefully polished. After you come into contact with enough products, you can naturally distinguish good from bad.
This also implies the attitude of a person with product background towards products: a beautiful product represents a serious attitude towards doing things. You can only make the function well, but a carefully carved and beautifully designed product represents that the person who designed it is not careless and has pursuit. Moreover, beauty is difficult to achieve, which means higher standards and requirements, and polishing details.
Have there been moments when "the engineering is fine, but you can't get over the aesthetic hurdle" and then you overturn the plan?
Shi Xiaoxin : Yes. Designing correctly and designing beautifully are two levels of requirements, and even represent different industrial maturity standards.
There are many such examples on IRON. In the earliest design, you could use existing joints to complete the performance and functions, but why make the curve of the whole body? We overturned many structures from a purely functional point of view to make the structure more in line with human aesthetics. There are challenges in the process, but it finally proves that it can be made, and the performance is not bad. When you need to raise aesthetics to a very high level, it may not be impossible to make, and it will not affect the function, but it definitely requires better industrial design capability and implementation capability.
Apart from curves, what else can define a good-looking robot?
Shi Xiaoxin : We don't just consider static aesthetics — you have to ensure that it looks as beautiful when moving as it does static. This is ignored in many early designs: the product looks OK when standing there, but you always feel something wrong when it moves. So when the catwalk came out at that time, I thought it was a very good feedback, and also an experiment: after reflecting both static and dynamic effects, put it in front of the public, to see if everyone thinks it is a good product.
There is another example. The design industry will discuss "design for good". The whole exterior of IRON is wrapped with fabric and skin. When moving, part of the skin will be wrinkled and pulled. This is irrelevant to machinery, but if you imagine it as a human-like robot, you will find it very weird — will it be uncomfortable? Just like a person standing there, if the clothes are pulled violently, he will definitely feel uncomfortable. It is not from the perspective of a machine, but from the perspective of a human that this feeling arises. From the perspective of design, you have to make the wrinkles more natural and relaxed, so that the appearance will make people feel that it is very human as a whole.
In addition, I think we have taken a shortcut in the field of anthropomorphism. For beauty, people have different understandings on different items — it is difficult to reach a consensus on whether an earphone looks good in arc or square shape. But humans have a symbiotic sense of the human body in their genes, and it is easier to reach an aesthetic consensus: for example, a slender, slim presentation is more suitable for the current mainstream aesthetics. From the perspective of bionics, it is easier to align everyone's understanding, and the perceptibility is very strong. So I will still insist on making anthropomorphic robots.
Anthrop