当机器人学会开可乐：深聊灵巧手的“不可能三角”与六大技术门派｜机器人特辑

Do you think it's harder to teach a robot to walk or to open a can of Coke?

Wait, this isn't Coke. Let

's start over.

I guess most of you would say walking is harder.

After all, it took humans millions of years to learn to walk upright.

Boston Dynamics' robots fell countless times before they learned backflips.

But I recently discovered a fact during an interview that completely overturned my cognition : in the world of robots, opening a can of Coke is much harder than walking.

In other words, controlling a dexterous hand is at least ten times harder than controlling a body.

We can also see this from the current price comparison.

China's Unitree G1 humanoid robot (which can walk) is priced at $16,000, while Boston Dynamics' Atlas robot (which can backflip) is estimated to be priced at $140,000.

The UK's Shadow The price of the robot's dexterous hand (which can unscrew bottle caps) has not been disclosed, but the industry estimates it to be over $100,000.

In other words, the price of one hand is close to that of a complete top-level robot.

What does this mean?

It's like the price of a steering wheel is close to that of an entire Tesla.

So why is it so difficult to make a dexterous robot hand?

What stage has the technology reached now? What are

the technical factions in the industry?

Are there any companies worth paying attention to?

We, Silicon Valley 101, have started a special series on robots, and will focus on the development of robots in Silicon Valley from multiple dimensions such as robot hardware, software brain, data, and brain-computer interfaces.

So today we are going to talk about the ultimate problem that has troubled top robotics engineers around the world: the dexterous robot hand , and TetherIA, the team we started with the former head of Tesla's dexterous hand, to talk about a $300 "Android version of the robot hand." How are "skillful hands" attempting to disrupt this high-end market, which has been monopolized for 30 years?

Why do we see robots sorting products in factories and moving goods in warehouses , but rarely do we see them unscrewing a Coke bottle cap or picking up a screw with the same dexterity as a human?

The answer lies in the complexity of the hand, which far exceeds our imagination.

The human hand has 27 degrees of freedom, including 27 bones, 29 joints, 34 muscles , and countless nerve endings.

It is a "precision instrument" that has evolved over millions of years.

Even more amazing is that this "instrument" allows us to have both the strength to grip tools and the precision to thread a needle.

This ability has created human civilization , but it is very challenging for robots to replicate this ability.

Take a look at the structure of the human hand and you will find that it is actually very flexible and has many joints.

Let's take the thumb as an example.

The thumb is from top to bottom.

This is the IP The thumb interphalangeal joint is the MCP joint.

Both joints can bend and straighten.

This joint is called the CMC joint.

This joint is much more flexible and can swing sideways, bend and straighten, and even spin in place.

It has a very large range of motion.

However, you will see that the connections between the joints are very small.

This is why the human hand is both flexible and very small.

This is the ultimate challenge facing robotics engineers.

The dexterous robotic hand is called a dexterous hand in English. In robotics,

in English. In robotics, a " hand " specifically refers to a highly human-like, multi-degree-of-freedom, and fine-tuned robotic hand capable of performing precise manipulations.

It can simulate the grasping, manipulation, and perception functions of a human hand. So what kind of robotic hand can be called "dexterous"?

hand. So what kind of robotic hand can be called "dexterous"?

First, it must have enough "joints."

Our human hand has 27 degrees of freedom , while a dexterous robotic hand usually requires more than 6.

High-end products can achieve 20-27 degrees of freedom , which is like equipping a robot with a hand that can actually "move fingers."

Second, it must have "embroidery-like" precision control.

We're talking millimeter-level or even finer manipulations.

Imagine using a robotic hand to thread a needle , or, as we'll see shortly, precisely grasping a 5mm M5 screw.

Third, it must have a "sense of touch," not just seeing, but also feeling. Tactile

sensors, force sensors, and position sensors are like equipping a robot with a nervous system , letting it know whether its grasp is light, heavy, soft, or hard.

Fourth, it must be able to "read people's expressions."

If it encounters a round object, it uses one grasping method; if it encounters a square, it switches to another.

It handles a glass cup gently, and if it encounters a piece of iron, it uses more force.

This is the adaptive grasping capability.

Finally, it must look like a human hand . All tools in the human world are designed for human hands.

. All tools in the human world are designed for human hands.

If robots can imitate the structure, function, and configuration of human hands, they can be applied quickly, cost-effectively, and without changing our environment.

If we look at history, we will find that dexterous hands have gone through more than 40 years from the concept to the maturity of the present.

1980s: Pioneering work Stanford/JPL The Hand pioneered dexterous hands. Its

three "humanoid" fingers, each equipped with tactile/force feedback at three joints , were more like concept machines proving "this thing can be done."

1990s-2000s: A hundred schools of thought contended, with the Utah/MIT Hand, DLR Hand, and others showing their prowess. A variety of technical approaches flourished, but all remained in the laboratory . Demonstrable but still far from practical use. At

. Demonstrable but still far from practical use. At

the same time, the gripper dominated. While dexterous hands were still "on paper" in the laboratory,

dominated. While dexterous hands were still "on paper" in the laboratory, simple two-finger grippers had already taken over the factory floor.

Although they could only "grasp" and "release," they were cheap, stable, and sufficient.

This is like the era of feature phones , which were simple but solved core needs.

2000s-2010s: The commercialization of Shadow Robot and Allegro broke the ice. Hand and other products are gradually being commercialized, with prices reaching tens of thousands of dollars , mainly serving scientific research institutions.

This stage is like the early days of personal computers : the functions are available, but ordinary people cannot afford them. The 2020s: Giants enter the market, and Tesla's entry changes the rules of the game . Musk not only wants to make a dexterous hand

. Musk not only wants to make a dexterous hand , but also to mass-produce it. At the same time, breakthroughs in large models such as GPT

it. At the same time, breakthroughs in large models such as GPT open the door to a new world of robot control.

2025: The turning point: Tesla's new 22-degree-of-freedom hand , TetherIA's $300 open source revolution , and the emergence of various open source projects.

Dexterous hands are about to have a "smartphone moment " , from geek toys to tools available to everyone.

However, although dexterous hands have ushered in a huge breakthrough , many difficulties still exist .

This difficulty is not only a breakthrough in technology , but more importantly, it is necessary to balance performance, cost, and reliability.

This has become an "impossible triangle".

During the on-site interview, I learned a view that may subvert many people's cognition: controlling a dexterous hand is 10 times more difficult than controlling a whole machine.

At TetherIA's Silicon Valley office, I saw the various iterative versions they have designed along the way , as well as their newly released open source dexterous hand product Aero Hand. I just

Aero Hand. I just need to connect my robotic arm and glove to a power source and they're ready to work together immediately.

Before I actually tried to control the dexterous hand myself, I really didn't understand how a complete humanoid robot could balance, walk, and navigate more easily than a hand .

But when I tried to control the hand myself, I found it wasn't that easy.

I think the difficulty lies in my inability to feel (touch).

It's more difficult than I imagined . We think this difficulty is actually multifaceted,

. We think this difficulty is actually multifaceted, because robots are complex systems , and now people are more concerned about the AI ​​control level.

The main difficulty lies in the generalization ability of the Vision-Language-Action model.

In addition, we think that from the perspective of the entire system, the hardware of the hand itself is also a major bottleneck . The human hand is very dexterous,

. The human hand is very dexterous, its relative size is small, each joint is very flexible, the fingers are very slender, and it can achieve a balance between speed and strength , and it is particularly durable.

If you think about it, we mainly rely on our hands to contact the outside world.

In fact, in traditional robots, the focus is on avoiding contact between the robot and the outside world . Once contact occurs, it

. Once contact occurs, it will cause a collision and damage to the robot.

However, the hand needs to contact the outside world . So, in summary,

. So, in summary, these are all hardware difficulties.

In addition to the difficulties of hardware and control , there are actually many more unknown to outsiders.

(For example) while controlling the robot, you hope to allow humans to influence and act on it.

This involves the remote control system , the simulation system in the development process , and the entire system behind it.

In fact, there are many difficulties . During the operation, I feel that the cooperation between vision and strength

. During the operation, I feel that the cooperation between vision and strength is very critical.

Because I can't feel the touch and strength at all, my grasp is completely air.

I can only rely on my eyes to observe the contact feedback between my dexterous hands and objects to make timely adjustments.

This is very similar to the process of software-driven dexterous hands.

We know that the human grasping process depends on the nervous system, muscle control and multimodal perception.

The adjustment of human grasping force is divided into two closed-loop controls.

The first is feedforward control, which means that the brain predicts the required force before grasping based on vision and experience.

For example, when seeing a bottle of water, the brain will estimate the weight and set an initial grasping force first.

The second stage is feedback control.

After the finger touches the object , it adjusts in real time through tactile and sliding information.

If the object starts to slide, the nervous system reflexively increases force in <100 milliseconds.

This is a fast "perception and response" closed loop.

If the dexterous hand is to completely replicate this feedback closed loop of the human hand, it requires not only a stack of sensors and control algorithms , but a hierarchical control architecture that is closer to the human nervous system.

It can be compared to brain control and cerebellum control.

Brain control relies on vision, experience, and reasoning to plan actions and make high-level decisions, while cerebellum control relies on touch, force feedback , and real-time balance adjustment to be responsible for dynamic fine-tuning and coordination of details.

Combining multimodal sensing including force, torque, touch, and vision with closed-loop control from perception to judgment to adjustment, and then continuously optimizing the entire system strategy through deep reinforcement learning is a very challenging R&D process.

Therefore, by the end of 2023, Tesla's second-generation humanoid robot Tesla Optimus Gen 2.

The egg-squeezing demo has attracted so much attention because its visual-based brain and force-control-based cerebellum are working together to make progress.

You can usually really feel the force , so why is the angle important ?

The force itself is important, and the position of the hand is also important.

So it is really difficult .

To truly realize the application of robots in multiple scenarios, you really have to rely on dexterous hands.

The logic is very simple.

The human world is designed for humans.

All tools, equipment, and environments are designed according to human size and ability.

The most effective way to make robots truly integrate into the human world is to give them human-like abilities . Therefore, if you want to make dexterous hands really go towards industrialization,

. Therefore, if you want to make dexterous hands really go towards industrialization, there is still a difficult problem to solve, that is, the "impossible triangle" of dexterous hands.

If the robot dexterous hand is compared to a triangle, then its three vertices are: performance, cost, and reliability.

This triangle has a cruel feature: you can only optimize two of the angles, and the third will inevitably be sacrificed .

If you want high performance, the British Shadow The Robot Company's dexterous hand is a perfect example.

It has over 120 sensors for tactile perception , 20 motorized joints, and 24 degrees of freedom.

Its size, shape, and range of motion are comparable to a human hand, making it well-suited to performing tasks designed and optimized for human hands.

However, it costs over $100,000.

For lower-cost solutions, there are many open-source projects on the market, such as DexHand and Amazing The Hand can be 3D printed at a cost of $300 , but its performance is merely stellar.

The gripping function is essentially unusable, and even some entry-level commercial products fall short in this regard.

High reliability requires a simplified design with fewer points of failure.

A more complex system has higher maintenance costs and a higher failure rate, but this significantly reduces performance.

Every joint and every degree of freedom in the robotic hand requires a motor to control its movement . If the motor is made smaller,

. If the motor is made smaller, its power output and performance will be reduced accordingly.

Therefore, matching the degrees of freedom, size, force output, and speed of a human hand is a difficult task . This "impossible triangle" is like playing a game you can never win

. This "impossible triangle" is like playing a game you can never win unless you can find a new way to break through it.

The industry is constantly searching for ways to break through this triangle, which has led to the emergence of six major schools of thought in the field of dexterous hands.

To address this "impossible triangle," there are six major schools of thought in the field of dexterous robotic hands, each with its own distinct approach.

The first is the "direct drive school," which can be likened to a simplistic "building block player."

Their philosophy is simple: place a motor wherever movement is needed, as seen in the Allegro Hand from South Korea's Wonik Robotics.

16 degrees of freedom, 16 motors, one-to-one motor service , and the latest domestic products SharpaWave Wuji Hand and XHand all follow this route.

The advantage of this design is that it is convenient for fine control.

The disadvantages are that the motor driver is small in size, has poor impact resistance, cannot be backdriven, has low fingertip force output , and is not easy to maintain and repair.

The direct drive method should be said to be the most direct design from an engineering perspective and is easier to implement at this stage . In our opinion, it is a relatively brute force solution to the problem. If

. In our opinion, it is a relatively brute force solution to the problem. If

a degree of freedom is needed here, I will add a driver here to rotate or move it.

The same is true for the connecting rod method.

It is a relatively simple mechanical structure . The second school is called the "rope drive school",

. The second school is called the "rope drive school", which is the "bionics master" closest to the human body.

The representatives of this school are Tesla's Optimus and Shadow Robot.

Tesla is familiar to everyone , and Shadow Robot, a British company, is like the "Rolls-Royce" of the dexterous hand.

Nearly 30 years of technological accumulation have allowed them to dominate the high-end market, but the high price has also limited their market expansion.

Their design concept is closest to the human body , placing the "muscle" (motor) on the forearm and controlling the finger movement through the "tendon" (steel wire and high-strength synthetic fiber).

It's like controlling a puppet, with all the strings connected to a central console.

Complex movements are achieved by pulling different strings.

The advantages of this design are lightweight, stable force output, a certain degree of adaptability , and a layout closer to the human body.

Tesla's latest Optimus hand boasts 22 degrees of freedom , very close to the 27 degrees of freedom of the human hand.

We visited TetherIA, and their current cost-effective dexterous hand also uses a rope-driven solution. However

, rope-driven technology has its own problems .

Although Tesla is determined to follow the "rope-driven" route , we have found that not many startups are truly following Tesla's technology path , and many people often question Tesla .

The most fundamental advantage of "rope-driven" technology is the relatively good force output and adaptability I just mentioned.

However, its fundamental disadvantage is that, especially for underactuated devices, it cannot achieve precise control because it is underactuated and the force output at each location will vary depending on the adaptability.

At this time, we need to be able to accurately model the various modules of the "rope-driven dexterous hand" in software . Only when you have a good understanding of

. Only when you have a good understanding of how the hand will change under different conditions can you achieve precise control.

The third faction is called "hydraulics", which pursues the "violent aesthetics" of extreme power.

Canada Sanctuary AI companies have chosen the hydraulic drive school.

Their Phoenix robot is equipped with a 21-degree-of-freedom hydraulic hand that can generate powerful force output.

The advantages of hydraulic systems are high strength, fast response, high power density, and the ability to complete high-load tasks.

However, hydraulic systems are traditionally large.

Sanctuary AI's breakthrough lies in miniaturizing hydraulic components to the size of a coin and passing 2 billion cycle tests without leakage.

This is like shrinking the hydraulic system of an excavator to the size of a watch.

The technical difficulty can be imagined, but hydraulic systems still face challenges in cost, maintenance, noise, and energy efficiency.

Therefore, they are currently mainly used in specific industrial and R&D scenarios.

The fourth school is called the "link school" , which is an "elegant school" that gives full play to mechanical aesthetics.

The representative work of this school is the ILDA dexterous hand proposed by a Korean research team. It

achieves high-degree-of-freedom movements through a sophisticated link design.

Its philosophy is to integrate all the drives inside the palm.

Use connecting rods, rocker arms, sliders and other mechanisms to "distribute" multiple linear motions to multiple joints , allowing multiple joints of the fingers to bend and pose in various postures similar to human fingers.

The advantages of this solution are compact structure, high degree of freedom, and elegant appearance, which fully demonstrates the beauty of mechanical design.

However, its disadvantages are also prominent, such as poor impact resistance and insufficient reliability in complex or high-load scenarios, resulting in low overall practicality.

The fifth faction is called "hybrid faction", the "golden mean" of engineers.

Some designs try to combine direct drive, rope drive, connecting rod mechanism, etc. to compromise cost, weight and performance.

For example, some open source or academic hands use connecting rod + The partially driven solution uses fewer actuators to achieve more degrees of freedom and is very popular in scientific research and teaching.

The hybrid solution has always remained in the field of academic research , and TetherIA is developing another high-degree-of-freedom dexterous hand solution through a hybrid approach . They analyzed the specific functions and structures of the human hand

. They analyzed the specific functions and structures of the human hand and developed a high-degree-of-freedom dexterous hand solution that is both efficient and reliable through strong engineering implementation capabilities.

The sixth and final faction is the "open source faction" worth mentioning.

They use random punches to defeat the master , not to compete in technical precision or hardware luxury, but to break down industry barriers in an open source way.

Although a single product may not be as sophisticated as Shadow Robot , their power lies in the "wolf pack tactics", which minimizes the price threshold of the dexterous hand so that engineers around the world can afford it , thereby promoting technological progress from DexHand to ORCA.

More and more open source projects are lowering the technical threshold for Hand . This is like the impact of the Android system on the mobile phone industry

. This is like the impact of the Android system on the mobile phone industry and may completely change the rules of the game . While developing a

. While developing a dexterous hand with high degrees of freedom and very close to the performance of human hands, TetherIA found that the system can actually be extremely simplified.

So they also made a dexterous hand with low degrees of freedom, but it is said to be one of the best performing hands on the market.

This dexterous hand was released a while ago and is fully open source and priced at only $300.

The team told us that their dexterous hand is extremely task-oriented.

Although it has a low degree of freedom , it can complete many tasks close to those of human hands . Let's take a look at

. Let's take a look at what complex tasks these hands can already complete.

These are the four demos of TetherIA's latest products.

Let's see what kind of technical challenges are hidden behind each seemingly simple action.

In fact, small objects also pose certain challenges for dexterous hands. Next,

let's let the dexterous hand try to see if it can grab this small nail.

The M5 screw has a diameter of only 5 mm.

This demo looks simple, but it is actually the ultimate test of fine control ability.

The difficulty of grabbing small objects lies in many aspects.

One is its precise control ability.

Another is that when you grab a small object, if the direction of force output cannot be well coordinated, the small object will often fly away.

This actually reflects the need for good adaptability and consistency of force output in hardware design . In addition, our entire software system must cooperate accordingly to achieve precise operation of the relatively complex "rope-driven dexterous hand".

The challenge of grabbing a large box is completely different.

At this time, only a small part of the fingertips of the robot hand is in contact with the object, just like grabbing a basketball with the tip of the fingernail.

Because the box is almost as big as the robot's hand , it must be grabbed very accurately.

Basically, in this case, you can only use your fingertips to apply force.

That is, the force is provided by the last joint of the robotic hand. Therefore

, such grasping becomes quite difficult for the robotic hand.

The size of this box is already the closest to that of a human hand . Our hands are limited for humans and for

. Our hands are limited for humans and for our hands as well.

However, you can see that our hands have no problem handling objects that are close to the limit of their hand size.

You can see that my hand is the same size as this robotic hand, but I struggle to grasp it.

I may need to try many times before I can grasp it because your hand is not big enough.

Because my hand is the same size as our robotic hand, but my end joint is not as flexible as ours, I may not be able to exert such force at my end joint.

Some people may have more dexterous hands, such as if you have practiced piano, so there is no problem.

But my hands are too clumsy, so it is quite easy for me.

Opening a Coke bottle is the most impressive demo because it truly shows the "humanization" of the robotic hand.

This reflects some of our innovations in hardware structure design and our understanding of the practicality of hands.

We believe that in many cases, human nails play a very important role.

One is the Coke bottle you just mentioned.

We need an adaptive process to achieve relatively large force output in a small space.

In addition, in many details in life For example, in the process of washing vegetables, cooking, and peeling vegetables, we actually use it here. It's not just a "nail".

here. It's not just a "nail".

What's more important is the design of the front end of the hand can be wrapped with soft materials, adaptive , and the curvature is very close to the curvature of the human hand , so it can realize these functions.

I originally wanted to do something naughty and shook the Coke can vigorously. I wanted the dexterous hand to open a "jet" Coke for me,

can vigorously. I wanted the dexterous hand to open a "jet" Coke for me, but hey, why was it so calm?

This time I failed . If you know how to shake the Coke

. If you know how to shake the Coke can to make it spray out, leave me a message and I will find the dexterous hand to try next time.

Finally, "holding the iPhone", which seems to be the simplest action, is actually the most technically demanding.

The iPhone is close to the table, and the fingers must be inserted into the gap of only a few millimeters and cannot collide with the table.

The iPhone is actually a very narrow space, and you need to use a lot of force in this small space and hold it steadily during this process. There are several difficulties.

One of them is that many dexterous hands have an outward curvature at the end . During the grasping process

. During the grasping process , the force will be diagonally outward, which makes it difficult to grasp firmly.

Another difficulty is that the fingers will inevitably come into contact with the table when grasping , which will increase the probability of damage to the dexterous hands.

The reason why our hands can solve these problems is that there is an adaptive process that can adjust the direction of the force well during grasping to make it grasp firmly and pick it up.

Secondly, because we use a "rope-driven" solution , the hand will adapt when it contacts the table and will not form a real collision or confrontation with the table.

We are currently at a special historical moment.

The breakthrough of AI large models has brought unprecedented possibilities to robotics.

For example, the Vision-Language-Action (VLA) model mentioned above upgrades the robot's "brain".

Traditional machines The robot needs to write a special program for each task, and the VLA model allows the robot to understand natural language instructions and convert them into specific actions.

This is like installing a "translator" on the robot.

It can translate natural language such as "Pour me a glass of water" into a specific action sequence. We found

that a big difficulty in manipulating the dexterous hand during the process is how to perform remote control, because it has more degrees of freedom and is much more complicated than the gripper, especially for our rope-driven solution.

Therefore, based on this pain point, we have developed an AI cerebellum.

This AI cerebellum can adapt to the user's manipulation process.

For different tasks, the user does not need to accurately tell the operator the force output or even the direction of the force output of our dexterous hand . The hand will then give the task for adaptive adjustment, which greatly improves the smoothness of remote control.

In addition, Sim2Real (simulation to reality) technology Technology is solving the cost issue of robot training.

In a virtual environment, robots can perform millions of trial and error without worrying about damaging the hardware.

However, there is always a gap between simulation and reality.

This is indeed some of the difficulties in the process of robot development.

The main reason is that because the physical world is very complex, we must have simplified many parameters in the simulation process.

For example, the friction of objects and the degree of rigidity and softness of objects cannot be well reflected.

And including robots, there will be some errors in the process of designing and producing robots in the physical world.

So this is something we have been overcoming.

It is like the difference between practicing driving in a game and driving in reality.

Basic skills can be learned, but the real sense of the road still requires actual experience . Not only

. Not only that, As for hardware, AI is also making a difference.

One reason why robots are expensive is that their supply chain is relatively lacking.

Many of these drivers are specially customized for robots.

Currently, the production volume is relatively low , so the cost of the entire industry cannot be made very cheap.

In addition, the design of traditional robots is to achieve many advanced and cool functions by continuously improving the accuracy of the products.

But now with the support of AI, the accuracy requirements for robot hardware will no longer be so high , so we believe that the overall price will become lower and lower.

The open source dexterous hand with rope drive solution launched by TetherIA aims to make the hardware adopted by more robot and technology enthusiasts at a low price, and on top of that, better use AI to develop software to accelerate the technological progress of dexterous hands and robots.

Our low-degree-of-freedom product combines our high-degree-of-freedom design for the entire structure. We have some experience in structural design and use the mainstream off-the-shelf motor on the market , so we can achieve the ultimate low cost.

Our plan is different from Tesla's because we don't have as much money as Tesla to invest fully in research and development.

We want to grow together with our community.

Why we are very confident in our low-degree-of-freedom hand and think it will be a hit product is that with a hardware cost of $300, we can achieve functionality that surpasses other products that cost thousands or even tens of thousands of dollars.

The advantage is that we can make such products not only used by many top companies and top research institutes, but also by more enthusiasts who can participate in the application development process of the dexterous hand algorithm, just like Google's Android Strategy:

Although Apple's iOS may be superior in individual products , Android has gained a larger market share through an open ecosystem . Moreover,

. Moreover, researchers around the world have become data contributors through open source hardware.

Looking back at the development history of dexterous robotic hands is actually a microcosm of human technological progress.

We started by imitating nature, gradually understanding the principles, and then using engineering methods to implement prototypes that may eventually surpass nature.

During the interview, we found that TetherIA's story is particularly interesting because it represents a new development model. It

lowers the threshold through open source, accelerates innovation through crowdsourcing, and promotes industrial development through ecosystem construction.

This is like the impact of Linux on the operating system industry or Android on the mobile industry.

Of course, there is still a long way to go from a $300 open source version to a truly practical home robot. There is a long way to go.

The technical challenges, cost pressures, and exploration of application scenarios are all full of uncertainty.

But as our interviewee said at the end of the interview, we believe that in five years , we will see robots deployed in many places.

It will not be something that remains in videos or concepts, just like we are exposed to ChatGPT every day . It will truly create huge value in our lives.

. It will truly create huge value in our lives.

Maybe in a few years, when we look back at 2025 , we will find that this is the starting point for robots to truly become popular.

By then, every family may have a robot assistant that can help us cook, clean, take care of the elderly, and accompany children.

The starting point of all this is to give robots a pair of truly dexterous hands.

Well, that's all for this episode of our program . This is also the first episode of our robot series.

. This is also the first episode of our robot series.

After that, we will also visit star startups in Silicon Valley.

We will have an in-depth discussion with the frontline robot R&D team about the current status of robot R&D from multiple angles and dimensions such as the brain, AI algorithms, data, brain-computer interfaces, etc. Please remember to follow our account and don't miss the updates.

Your forwarding, comments and likes are the best motivation to support us, Silicon Valley 101, to do a good job in in-depth technology and business content.

See you in the next episode. Bye

See you in the next episode. Bye