AI-generated summary: Yushu Technology's application for an IPO on the STAR Market has been approved. Robots have progressed from spinning handkerchiefs to performing backflips and martial arts on the Spring Festival Gala, while smartphone manufacturers' robots have broken the human half-marathon record. The article breaks down the four core hardware systems of robots: skeleton, joints, sensors, and electrical and computational systems. Skeleton materials have evolved from steel to aluminum alloy, magnesium alloy, and titanium alloy, requiring a balance between lightweight design and impact resistance. Actuators, the most expensive component at approximately 51% of total cost, are divided into rotary and linear actuators, incorporating precision parts such as reducers, motors, lead screws, and encoders. Sensors include IMUs, cameras, LiDAR, and tactile systems. Chips adopt a "brain + cerebellum" architecture. The article notes that although over 80% of components overlap with the smartphone and automotive supply chains, the real challenges lie in system-level integration, engineering trade-offs, and manufacturing consistency—supply chain maturity is a critical factor in robot evolution.

Article author and source: 36Kr

The physical dilemma of humanoid robots

On June 1, Unitree Technology's application for an IPO on the STAR Market was successfully reviewed and approved by the Shanghai Stock Exchange Listing Review Committee. Just recently, Unitree also unveiled its first manned transforming mecha. How far are we from seeing robots truly come to life?

Last year’s Spring Festival Gala featured robots twirling handkerchiefs and performing yangge dance; this year, they’ve advanced straight to high-difficulty backflips and martial arts. Now, even robots made by smartphone manufacturers can break human records while riding a horse. Why has the evolution of robot hardware accelerated so rapidly over the past two years?

To better understand the evolution of robotic entities, we visited several leading robotics companies and spoke with industry insiders: What are the real challenges in building robots? Is the barrier to entry for robotics really low? And what truly constitutes a robotics company’s moat?

In this article, we’ll break down each component of the robot—after reading through it fully, you’ll be able to assemble your own robot.

01 Frame Material: Balancing Lightweight and Impact Resistance

Robots feature a wide variety of hardware, which we can broadly categorize into four systems: the skeleton that supports the entire structure, the joints that drive movement of the skeleton, the sensors that perceive the environment, and the electrical and computational systems that control the body. Let’s begin with the skeleton.

If a car traveling at 60 km/h crashes into a mannequin, the immense impact would send the mannequin flying and shatter it into pieces. For humanoid robots, enduring such impacts has become “routine.”

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Each time the robot performs a flip and makes contact with the ground, it experiences accelerations of dozens of g’s—possibly higher than those in cars or spacecraft, comparable to the impact acceleration of a car crashing into a wall.

This poses a challenge for the robot’s structural materials: it must be light enough to flip, yet strong enough to withstand such immense impact forces—otherwise, a single somersault could cause parts to fly off. Therefore, the robot’s first challenge is to explore suitable skeletal materials.

The world's first full-sized robot, WABOT-1, was primarily made of steel and weighed approximately 160 kilograms; it might have dented the floor just by jumping, let alone doing a somersault.

Later, from Honda's ASIMO and Boston Dynamics' early hydraulic versions of Atlas to the first generation of Tesla Optimus, aluminum alloy became the standard, with a density only one-third that of steel.

The industry has now begun exploring additional materials, such as magnesium alloys, which have a density one-third lower than aluminum, and higher-strength titanium alloys in areas subject to frequent impact, such as the knee and ankle joints.

Interestingly, these rigid frames absorb the impact for the robots, but suppliers seem to earn only a modest profit.

Former Head of Procurement at a robotics company

After deducting the intrinsic metal content and the waste discarded, the ratio of the final selling price of the frame is actually very low. The final sale price of the frame still consists of metal cost plus processing fee, with the majority of the cost coming from the metal itself, leaving no room for price reduction. The processing fee is already within a reasonable range; if volumes increase, the processing fee will approach very low levels, as there are no significant barriers to entry.

In addition to these core components, the robot's external parts can be divided into two categories:

One type is decorative and protective components, primarily used on the chest, back, and head, made from a variety of materials such as plastic, faux leather TPU, and fabric, mainly to reduce wear and provide a more pleasant touch. Although some robots appear to have metal bodies, they are actually plastic shells coated with a metallic paint.

Another type involves creating robotic skin that mimics human skin, which must not only feel like human skin but also have tactile sensors embedded beneath the surface.

Beyond the skeleton and skin, it is the joints that enable the robot to perform a wide range of highly complex movements—and they represent the most expensive, technologically intensive, and story-rich component of the entire robotic hardware system.

02 Dissecting the Actuator: Joints Are the Most Expensive and Most Challenging Component

You've probably seen many videos of robots dancing or doing backflips—this is achieved by first capturing human movements, then training a model to map them onto robotic limbs.

A few years ago, we were amazed when we saw Boston Dynamics' Atlas perform a backflip, but now most people probably take it for granted—this is because robot joints have transitioned from hydraulic systems to electric motors.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

We couldn't produce such advanced joints before; back then, the overall performance of joints was very poor, and backflips were extremely difficult to achieve. In the past one to two years, joint technology has made tremendous progress.

Joints are referred to in the industry as actuators and are primarily divided into rotary actuators and linear actuators. Let’s use the shoulder as an example to see how they drive body movement.

The shoulder has three degrees of freedom: forward and backward swing, upward and downward lift, and internal and external rotation, known as pitch, roll, and yaw. Since these movements are essentially rotations, combining three rotational actuators allows the arm to move freely in the X, Y, and Z directions.

At the knee joint, only one degree of freedom is generally needed, so a single rotational actuator or linear actuator suffices. A linear actuator, like human muscle, moves the upper and lower bones by stretching.

Performing an extreme movement requires tight coordination among dozens of actuators across the body; if any part reacts too slowly or applies even a slight deviation in force, the result is a fall.

What is the internal structure of these actuators? Both rotary and linear actuators consist of a servo system made up of a motor, encoder, driver, and sensors. The main difference is that rotary actuators use a servo motor with a gearbox, while linear actuators use a servo motor with a lead screw.

Let's start with the reducer.

You may have heard of this mechanism: the first gear turns 10 times, the second turns once, the third turns 0.1 times, and so on with a total of 100 gears. To make the last gear turn just once, the first gear would need to turn a googol times—that’s 1 followed by 100 zeros—requiring more energy than the entire universe contains.

This is a large reducer, essentially a massive lever that trades speed for force. Why do robot joints need reducers?

Because motors are inherently "high-speed, low-torque": they can easily reach tens of thousands of revolutions per minute, but produce relatively small torque. Robotic joints require precise control, and it’s difficult to make a motor rotate just a few degrees while simultaneously lifting very heavy loads. Therefore, reduction is needed to lower the speed and increase the torque—the higher the reduction ratio (i.e., gear ratio), the more the speed is reduced and the higher the output torque becomes.

The three most commonly used reducers in the industry are planetary reducers, harmonic reducers, and RV reducers. Let us explain them using models.

First is the planetary gearbox, whose name is quite descriptive: the motor connects to the central gear, which drives three planetary gears, and these in turn rotate the outer ring gear, much like planets orbiting the sun. It has a compact structure and low cost, but offers limited reduction; under the same motor speed, it produces lower output torque, so it is commonly used in hand joints.

When greater torque output is required, a harmonic drive is used. At its center is the wave generator, which deforms the flexible spline into an elliptical shape. Typically, the flexible spline differs by only two teeth from the outer rigid spline. Only two symmetrical regions of the flexible spline mesh with the rigid spline at any time. As a result, when the wave generator completes one full rotation, the flexible spline rotates by just two teeth, enabling a very high reduction ratio.

Harmonic drives deliver high output torque and high precision, commonly used in robotic elbow and shoulder joints to enable precise arm control.

As mentioned earlier, when the robot performs a backflip, the forces it endures are equivalent to a car collision, placing significant demands on the reducers in specific areas. However, the flexible structure of harmonic reducers also means they have poor impact resistance; at this point, RV reducers must be used.

The RV reducer consists of a first-stage planetary gear and a second-stage cycloidal pin gear. After the first stage reduces speed, an eccentric cam drives the cycloidal disc to perform eccentric motion; the cycloidal disc meshes with the pins on the housing, causing the housing to rotate.

This not only provides a high reduction ratio, but also offers high rigidity and stronger impact resistance due to multiple teeth on the cycloidal disc engaging simultaneously, making it commonly used in robotic joints such as the hip, knee, and waist where impact resistance is required.

The reducer is a highly precise component, difficult to manufacture, and challenging to maintain stability over long-term wear—it is the most difficult part of the entire joint.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

When gears are manufactured and used in large quantities, their precision and long-term operational stability must be very high. For example, after 1,000 hours of use, if the gears start producing various unusual noises or experience performance degradation, it may become difficult for the motion control algorithm to compensate. This manifests in robots as a decline in walking performance—perhaps they no longer walk as smoothly as before, or even begin to walk increasingly off-course.

Robots may perform many extreme movements and often fall themselves; these impacts could easily damage the small gears inside. How can we create gears that offer excellent performance, low cost, long-term durability, and high impact resistance after falls—without compromising any of these factors? This is a highly challenging trade-off.

In other words, it’s not hard to make one gearbox—it’s hard to make ten thousand gearboxes that are all consistent in performance and durable.

Next, let’s look at the linear actuator and its core component—the lead screw.

Linear actuators are the most similar to human muscles; when we swing our arms, it is not the joints actively rotating, but the muscles connecting the two bones contracting. Therefore, linear actuators perform only one type of motion: pushing and pulling.

Some robots use linear actuators at the knee joint to mimic the pushing and pulling motion of human knee muscles. When multiple linear actuators are combined through specific structures, they can also enable joint rotation. This type of movement is applied to areas such as the wrist and ankle.

The simplest way to create a linear actuator is with a hydraulic system; the earlier version of Boston Dynamics' Atlas primarily used linear hydraulic cylinders, offering advantages such as high power output, impact resistance, and high power density. Why the earlier version? Because the newer version has shifted to electric motor drive, primarily due to the complexity of hydraulic systems, their tendency to leak oil, and their lower control precision compared to motors.

But since the motor can only rotate, a “converter” — specifically a lead screw — is needed to produce linear motion.

The lead screw has threads, and when it rotates, it drives the nut to move linearly, similar to screwing in a bolt. To reduce friction, balls are added inside the lead screw—this is called a ball screw. Some designs replace the balls with rollers, offering longer life, higher load capacity, and better rigidity—this is known as a planetary roller screw. Additionally, some applications use trapezoidal screws.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Currently, roller screws are probably used more frequently; they require extremely high machining precision, and consistency must be maintained over a long travel distance. If there are any imperfections in between, it becomes a significant challenge for the control algorithms across different machines.

Some linear actuators are also paired with gearboxes to deliver higher torque from the motor. However, in the current industry, the application of linear actuators is relatively limited, primarily due to three reasons: poor dynamic performance, difficult manufacturing, and high cost.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Currently, the most mass-produced component across the industry is the rotary joint. Linear actuators are also used in some applications within the industry; their key advantage is the ability to handle higher loads and maintain a stable posture even when unpowered, thanks to their self-locking capability. However, we believe their drawback lies in slightly inferior dynamic performance, as their high load and high gear ratio result in less agile motion. Another significant challenge is the difficulty in manufacturing them at scale and at low cost. Therefore, at this stage, we do not consider them suitable for large-scale commercialization. Due to limited current usage, low shipment volumes, and minimal real-world customer validation, their costs remain high.

After discussing transmission, let’s now talk about the power itself—the motor and servo system.

The motors commonly used in robotic bodies are frameless torque motors. Compared to traditional motors, they lack housings and bearings, retaining only the core components, in order to minimize size and enable direct integration into joints.

The dexterous hand is special, using smaller hollow-cup motors, which naturally result in lower output power. The difficulty of the dexterous hand is even higher than that of the entire robot body.

The main challenges of the body motor lie in three areas: energy efficiency and heat dissipation, size, and performance stability. Let’s first discuss energy efficiency and heat dissipation.

Electronic devices inevitably generate heat; when excessive heat builds up beyond the normal operating range, performance declines. Therefore, the efficiency of the motor—how much energy is actually used to perform work—is especially important. If it overheats, the control system can only reduce power, causing, for example, a mid-air flip to suddenly "lose strength" and crash to the ground.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

The earliest samples we made could only perform these extreme maneuvers once within about 10 minutes. After one run, the performance curves—such as speed and torque—would change entirely, possibly due to heat buildup inside. At that point, we’d need to let it cool down first, allowing the temperature to drop before continuing. Another major issue is energy efficiency: how much of the input energy is converted into heat. For example, a difference between 5% and 3% represents a huge gap. These factors all limit performance; even if my hardware capabilities are strong, I wouldn’t dare push performance any higher.

The difference between 3% and 5% may seem small, but it's important to note that motor heat generation is not linear.

When a joint performs an extreme motion, the instantaneous current may be 3 to 5 times higher than normal, and the heat generation can reach 9 to 25 times the rated level. This means the rate of heat accumulation far exceeds the joint’s passive cooling capacity. A single backflip could cause the joint temperature to jump directly from a 10°C rise to 50°C. Therefore, the motor needs to cool down after each movement before the robot can proceed to the next action.

To improve motor efficiency, focus on motor materials, winding techniques, and structural design—we won’t go into detail here.

Currently, most joints rely on passive cooling, as the chassis uses a large amount of metal, effectively acting as a massive heatsink. Only joints with very high power requirements, such as those in the legs, additionally incorporate active air or liquid cooling.

Moreover, adding additional cooling measures introduces a second challenge: size limitations.

Engineers are striving to minimize the size of joint motors as much as possible, not only to reduce weight and lower costs, but more importantly because larger volume results in greater moment of inertia, making it harder to change the motion state.

For example, when you spin a rope, the shorter the rope, the faster it spins; if the rope becomes longer, the speed slows down, and it takes longer to stop.

The third challenge is whether the performance is stable—that is, at what current input the motor achieves a certain rotational speed and outputs how much torque, known in the industry as the TN curve. This affects the robot’s control algorithm.

For example, when walking over an uneven surface, the six-axis torque sensor on the ankle detects the bumps. To maintain balance, the current must be dynamically adjusted to control motor torque. If the TN curve is unstable, the control system may still issue the same command, but the motor’s torque output could deviate, resulting in a fall.

Moreover, the TN curve significantly impacts the training of the algorithm, as the bot is first trained in a simulation system; if the TN curve in the simulation differs greatly from reality, the actual performance will also deviate.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

I will input a curve into the simulation system; in reality, this motor can achieve or even exceed that curve, meaning it can perform the desired performance and movements. Conversely, if it performs well at low speeds but its performance drops when speed increases, then certain extreme movements will definitely be impossible, because some of the most challenging actions require both extremely high speed and immense burst power.

To precisely control the number of motor rotations, a servo system is required, primarily composed of an encoder, driver, and sensors.

Encoders measure the angle, speed, and position of the motor rotor, allowing the system to know the current state of the motor.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

The encoder is actually crucial; because robots have reducers, dual encoders are required to know the positions of both the input and output ends, enabling more precise control.

The driver adjusts the voltage and current supplied to the motor based on feedback from the encoder and control commands from the "cerebellum".

There are various types of sensors, such as torque sensors to measure output torque, temperature sensors to monitor motor temperature and prevent overheating, and more.

These are the key components within the executor. Next, let’s discuss the executor as a whole—why is it critical for cost reduction? What are the main differences between in-house development and procurement?

According to Bank of America's calculations, actuators are the most expensive component on a robot, accounting for approximately 51%.

Former Head of Procurement at a robotics company

Whether it’s your hand or a motor, the motor (actuator) and control (controller) are more expensive than your bones, your eyes (sensors), your brain (chip), and even your heart (battery).

So, the executor is key to achieving mass production and cost reduction in the future, primarily because China’s supply chain is extremely competitive—many components that previously required precision manufacturing in other countries can now be sourced domestically.

For example, companies like Wolong Electric Drive for motors, Green Harmonic and Shuanghuan Transmission for reducers, Zhongdalide, and even companies that directly provide complete actuators, such as Sanhua Intelligent Control and Tuopu.

Since ready-made actuators are available on the market, why do robotics companies still invest time and effort into developing their own? Let’s compare these two approaches.

Purchasing finished products can reduce R&D costs and improve development efficiency, but the material costs will be higher, customization according to your own needs will be difficult, and performance may be insufficient.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Most (executor) companies won’t customize products specifically for you—they sell standardized components, which tend to be more expensive. If a company has a small internal team and insufficient expertise in key areas, it’s usually better and faster to purchase from others.

Developing in-house allows for better alignment with requirements and algorithms, resulting in stronger performance, but it requires significant development effort.

The choice of path is largely a consideration of company size and cost; according to our survey, leading robotics companies still tend to favor in-house development, and some even collaborate directly with suppliers during the design process.

So the robot's joints are not just about assembling parts—they must achieve a balance of power, precision, durability, cost, and weight within an extremely compact volume, making them arguably the most challenging component of the entire body. This is because it is an emerging industry with an underdeveloped supply chain, and everyone is still in the exploration phase.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Initially, many of the production line devices did not exist in the industry, so we had to design (and produce) them ourselves.

Strong joints alone aren’t enough—how does a robot know how to stand steady? How does it perceive the world? Next, let’s talk about sensors.

Chapter 3: Vision System: How Robots Perceive the World Chapter 3.1: Inertial Measurement Unit (Vestibular)

Today’s robots are very stable and rarely fall, even when humans intervene. Achieving this balance requires sensors throughout the robot’s body.

On one hand, there is the motor servo system mentioned earlier, which uses encoders and torque sensors within the joints to continuously sense the current position and force at each joint, then adjusts the output at a frequency of thousands of times per second.

On the other hand, having only a sense of the limbs is not enough; just as humans rely on the vestibular system in the inner ear to sense body tilt and rotation, robots use an inertial measurement unit (IMU) for this purpose.

IMUs are very common—for example, when you rotate your phone and the screen rotates with it, that’s thanks to an IMU.

IMU is a combination of several sensors, with the two most critical being the accelerometer, which measures acceleration along the X, Y, and Z axes, and the gyroscope, which measures angular velocity around the pitch, yaw, and roll axes. Additionally, the IMU often includes a magnetometer, acting as an electronic compass for calibration.

By combining these data, the IMU can sense the robot’s motion in real time. When we kick it, the body instantly gains acceleration and may tilt forward, backward, or sideways. Once the IMU detects this change, it sends the data to the “cerebellum,” which calculates how much torque to add or reduce at each joint to restore balance. This component is widely used in devices like smartphones and cars, so its technology and applications are relatively mature.

Fall prevention relies on the IMU, but for daily activities, collision avoidance is more critical, and obstacle avoidance primarily depends on the visual system.

The robot's "eyes" are very similar to autonomous driving in cars, but not exactly the same. The common approach involves sensor fusion with cameras, LiDAR, and millimeter-wave radar. The exception is Tesla's Optimus—众所周知, Musk is a staunch advocate of pure vision, using only cameras.

In terms of sensor usage, robots are nearly as complex as automobiles, and many suppliers have transitioned from the automotive supply chain. However, although the same types of sensors are used, their actual specifications differ significantly—we’ll use the more expensive LiDAR as an example.

First, the ranging requirements differ. Cars need to travel on highways, so LiDAR must detect obstacles 150–200 meters away. Robots primarily operate indoors, where 10–20 meters is sufficient. Shorter ranging means LiDAR can have lower power consumption, smaller size, and reduced cost.

Second, the point cloud density and scanning methods differ. Cars identify large objects such as vehicles, people, and obstacles, which can be detected with lower point cloud density. However, robots need to pick up small objects like screwdrivers from a table or coins from the ground, requiring higher point cloud density.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

We want the point cloud to be very dense. Currently, we use non-repetitive scanning, where standing in one place for a while makes the point cloud denser. This is very beneficial for us because our robots often don’t perform highly aggressive operations—they move slowly, much like humans do when performing many tasks. In contrast, cars have very high requirements for stability, real-time performance, and repeatability.

Third, the installation location and size differ. A vehicle can mount LiDAR on the roof or bumper, where a larger size is acceptable, but robots have smaller bodies and require smaller modules.

Fourth, the reliability requirements differ. For example, cars operate outdoors year-round and require a higher operating temperature range, while robots experience greater impact forces and demand higher vibration resistance.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Previously, for automotive applications, the minimum operating temperature range for LiDAR was -40°C to 85°C, but for robots, this is currently completely unnecessary. Therefore, many design features in cars that are specifically intended for reliability are redundant in robotics. The acceleration experienced by a car during a collision may be comparable to the acceleration a robot experiences during a single backflip, so we have very high requirements for stability under vibration conditions.

Although LiDAR for automobiles is already highly mature, LiDAR for robots is still in the early stages of industry development.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

We want the device to be smaller, with denser point clouds, shorter range, but a wider FOV (field of view)—these requirements have not yet been fully addressed.

According to the former head of AI hardware at Tesla, they selected automotive-grade cameras for the camera system, but the internal development path changed multiple times.

Liu Xiangke (Kerry)

Former head of AI hardware at Tesla

The current solution is based on a 5-megapixel camera mounted on the vehicle. The earliest versions used multiple cameras with varying pixel counts, reducing frame rate to increase pixel resolution. This was done because Elon requested that the robot be capable of threading a needle; we calculated that achieving this would require more than 15 million pixels to clearly see the task.

The software team also mentioned that if changes were made to the pixel count or camera hardware, the requirements for retraining the model—both in terms of time and workload—would increase significantly. What if it’s not feasible? They considered adding autofocus to the camera. But later it seemed this might not be strictly necessary after all, so the decision has continued to evolve.

Next, let’s talk about haptics—there are primarily four ways to achieve it:

The most common type is piezoresistive, which converts pressure into resistance to alter an electrical signal, such as in electronic scales.

The second type is capacitive, using an elastic medium to separate two layers; when pressure is applied, the distance between the electrodes decreases, causing a change in capacitance.

The third type is piezoelectric: when the material is subjected to force, it directly generates voltage, such as the small device inside a lighter that produces a spark.

The fourth type is optical, featuring a surface made of elastic material that deforms under pressure and is captured by a camera—this is currently the most popular approach.

Haptic feedback should be three-dimensional, capable of sensing not only pressure but also friction on a plane. For example, when picking up a soda bottle, your hand grips and lifts it upward; if your fingers detect slipping friction, they increase grip strength to prevent it from falling.

But this also presents significant challenges for materials and algorithms.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

First, at the level of the sensor itself, since these sensors are fundamentally made of materials, it is extremely difficult to fully decouple the three (X, Y, Z) directions, making precision much harder to achieve than with one-dimensional force sensors. How can we make it accurate? Second, integrating such complex three-dimensional tactile data with manipulation models is also very challenging, because the overall amount of available data is still very limited.

Under these challenges, mass-produced robots in the industry previously rarely incorporated haptic feedback.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Throughout all products mass-produced in 2025, haptic feedback is used very rarely, almost not at all—not just by us, but across the entire industry—because this technology is unstable.

You need to consider how it can maintain its shape over long-term use, because even a slight deformation could completely alter the output signal. Additionally, performance drift must be avoided—its shape and position must remain intact, yet the material needs to be slightly soft while also being highly wear-resistant, which is inherently contradictory.

But this year, things seem to have changed a bit. Our interviewee mentioned that by 2026, there is hope for scaled production, and the next step is to better integrate tactile systems into data collection and training. Overall, the tactile industry is still very early-stage, and we look forward to seeing more progress in the future.

In addition to the sensors mentioned above, the robot also requires temperature, humidity, six-axis force torque sensors, UWB, and others—all of which are well-established, so we won’t elaborate further.

Sensors enable robots to perceive the world, and joints give them the ability to move, but to integrate these two elements, a "central hub" is needed—let’s discuss this hub: the electrical architecture.

As mentioned in our previous article on robotic algorithms, the industry has developed a dual-system architecture called "System 1 + System 2," where System 1 handles limb control and System 2 performs complex reasoning. Similarly, on chips, a "cerebellum + brain" combination is employed.

Why not use one chip to do everything? Because the requirements are completely opposite.

A brain chip that needs to think about "how to get things done" requires high computational power and large memory, ideally capable of running large models on the edge, with delays of a few seconds having virtually no impact.

Currently, the vast majority of robot brains use NVIDIA's Orin chip. In 2025, NVIDIA launched the Thor chip, which offers higher performance and is specifically designed for robots and physical AI, and is expected to become the future standard.

Except for Tesla Optimus, it uses custom-designed chips, and even dual chips.

Liu Xiangke (Kerry)

Former head of AI hardware at Tesla

Since robots are not autonomous vehicles, they don’t have these safety considerations, and Elon himself initially thought: “We don’t need this safety redundancy anymore—one chip is enough.” After building a single-chip system, he later realized something was wrong: the world model requirements for robots demand far more computing power than autonomous driving. If two chips are barely enough for autonomous driving, how could one possibly suffice for a robot? He quickly corrected himself: “No, no—go back to two chips.”

Additionally, at CES earlier this year, Qualcomm also released the robotics brain chip Dragonwing IQ10 and announced its partnership with Figure.

The cerebellar chip, which "controls the body," does not require extremely high computational power, but it must have high real-time performance, stability, and response speed—delaying by just a few milliseconds could cause a fall.

For example, when a robot performs backflips or dances, it typically uses pre-recorded movements, but we notice that its feet still make small adjustments—this is the cerebellum dynamically regulating balance, much like a human’s “instinctive response.”

Wang Chuang

Partner, Senior Vice President, President of General Business Division

The cerebellum requires very high speed, so the frequency within the cerebellum may be 1 kHz.

Currently, small brain chips are typically MCUs, with mainstream choices including STMicroelectronics' STM32 series, NXP's i.MX RT series, and Renesas' RZ series.

We are also seeing a new trend: the industry is attempting to integrate brain and cerebellum chips. Tesla is at the forefront in this area, having pursued this approach from the beginning.

Liu Xiangke (Kerry)

Former head of AI hardware at Tesla

We initially assumed that the Hardware 4 custom chip was used at that time. Since Tesla’s brain and cerebellum are integrated onto the same chip, how is the entire body’s movement controlled through this single chip, and what communication architecture is employed? We spent considerable time researching this approach—a single SOC containing both ASICs for compute power and a multi-core CPU capable of handling cerebellar functions; this high-frequency CPU also offers extremely low latency.

Besides Tesla, other companies are also researching integrated solutions.

For example, in March this year, Lingjing Zhiyuan released the Dvořák architecture, integrating the three functions of “brain-cerebellum-cortex” onto a single chip. What benefits does consolidating them onto one chip bring?

Wang Chuang

Partner, Senior Vice President, President of General Business Division

First, I think the biggest advantage is that, since everything is now integrated onto a single board, the overall chest cavity volume and wiring become much simpler. Second, as we go further, the coordination between the brain and cerebellum becomes increasingly important. For example, if someone throws a dart at you, perceiving and predicting its trajectory likely involves the brain, while reaching out to catch it involves the cerebellum. The faster the communication between these two, the more effectively you can perform highly complex movements. If the brain and cerebellum are integrated on the same chip, communication between them would be extremely fast, allowing the brain to control the cerebellum in real time and receive feedback at very high speeds.

However, according to industry perspectives, unified brain-and-cerebellum chips are still in a very early stage; robot manufacturers will only gradually shift toward integrated, in-house developed chips, as smart car companies have done, once robot shipments reach sufficient volume and the market becomes large enough.

Finally, a battery that provides energy to the entire body, like the robot's heart. The core requirement is achieving higher capacity at lower density; major suppliers include CATL, LG, and EVE Energy.

There are also harnesses distributed throughout the body, like nerves and blood vessels, used for communication and power supply between devices. Major suppliers include Luxshare Precision, TE Connectivity, and Amphenol.

There are many types of robot supply chains, so we won’t go through them all here—instead, we’ve included an overview image for those interested to zoom in and explore.

By now, you’ve learned how to build a robot—but hold on. If you actually try to build one yourself, you’ll quickly realize there are issues everywhere, because the biggest challenge in robotics is balancing all the different engineering disciplines.

Finally, let’s discuss the challenges of assembly and mass production, and the reasons behind the rapid advancements in robotics over the past two years.

05 Assembly and Mass Production: Mobility Doesn't Equal Usability

If you watched the recent robot marathon, you'll notice there were plenty of amusing moments on site.

Some sit down wherever they please, drawing applause from the robot next door; others twist their ankles while running, get drunk, lose an arm, charge onto the greenery, or crash into speed bumps and break into pieces.

There were also outstanding performances, such as the Glory robot, which not only swept the top six positions but also broke the human half-marathon record.

But this has also sparked some discussion: if even smartphone manufacturers can perform so well in robotics, does that mean the industry has low barriers to entry?

The industry professional's answer is: Yes, and, No. Let's start with the Yes part.

The components and suppliers mentioned earlier overlap significantly with those in the smartphone and automotive industries; further up the chain, some algorithms can also be reused in autonomous driving, which is why Honor, Xiaomi, Tesla, and Xpeng have entered the robotics space.

Former Head of Procurement at a robotics company

The supplier overlap for the electrical and power systems exceeds 90%. For the mechanical system (frame structure), even if the molds differ, many suppliers are still similar. Electric drive is the only area with potentially lower relevance to automobiles, since vehicles do not require components that deliver high torque. However, components such as reducers and gears are widely used in cars, as are sensors. Therefore, over 80% of the components are essentially interchangeable.

Theoretically, as long as you know these suppliers, you can build a robot yourself. But there’s a huge gap between “it works” and “it’s usable”—that’s the No part.

For example, if the weight distribution is uneven after assembly, the robot’s center of gravity shifts, causing certain joints to exert extra effort to maintain balance, increasing power consumption, reducing battery life, and potentially affecting gait stability.

Or, it might run perfectly in the lab for an hour, but after 100 hours in a real-world environment, all sorts of issues emerge: a screw loosens, a wire wears out, lubricant in a joint dries up, a sensor begins to drift—these all require continuous tuning to find the right balance.

Former Head of Procurement at a robotics company

Each component, I break it down by supplier, and I don’t think the suppliers themselves are difficult—it’s the system integration that’s the real challenge.。

It’s more about imposing constraints—like requiring the system to be lightweight or reduced in weight to a certain degree—but once you confine it to a human form, the challenge lies in matching human-level torque and precision, which is primarily a matter of engineering trade-offs.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

The standard products available on the market are often unsatisfactory and fall short of our actual algorithmic application requirements, so these are core components that we must develop ourselves.

Creating commercially viable, mass-producible robots will still face challenges related to consistency.

Because the joint backlash, sensor zero points, and motor parameters differ across units, adjustments to each detail are required to ensure the same algorithm can be stably applied to different batches of hardware.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

Place ten robots there and send them the same parameters (instructions); their hands will extend to different positions.

If you perform an operation, a difference of just a few millimeters can turn successfully grasping an object into knocking it over—making it extremely challenging to properly calibrate all of the robot’s sensors and actuators. Furthermore, after calibration, can you guarantee that after a year of use, when components have aged and sensors have become distorted, the system will still remain stable? This may require online calibration, where the system can autonomously detect and analyze its own errors. These are invisible efforts, but without them, many subsequent problems cannot be resolved.

So the real challenge isn't "putting it together," but system-level integration.

Let’s return to the robot marathon: this year, not only has speed improved significantly, but overall completion has also increased. Looking at the evolution of robot movements over the past two years—from walking, to twirling handkerchiefs, to dancing and martial arts—why has the progress been so rapid? The most important reason is the maturation of the supply chain.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

In the past one or two years, the robotics industry wasn’t as widely embraced as it is now; back then, people wouldn’t develop LiDAR specifically for robots—they’d say, “I made this for logistics vehicles, just take it and use it.” At that time, we were begging others for support, and most people were skeptical about robotics.

As we mentioned earlier, many segments of the robotics supply chain overlap with the automotive industry. Previously, suppliers faced internal resource competition: given limited production capacity, should they prioritize supplying established commercial industries, or reconfigure their production lines to bet on robotics markets that are still immature?

Former Head of Procurement at a robotics company

Previously, I felt the market hadn’t reached this level yet—it might still be one to two orders of magnitude behind actual high-volume products like phones and cars. Suppliers are also in a strategic balancing act, as their internal resources are limited.

As the robotics sector grows increasingly popular, suppliers are now willing to create custom molds and products specifically for robots. As demand increases and commercialization pathways become clearer, the supply chain will continue to grow like a snowball.

What will be the next milestone action?

06 Next milestone: From a backflip to catching a falling leaf

Wang Chuang

Partner, Senior Vice President, President of General Business Division

A few days ago, I went to the Shanghai Circus World and watched a performance; my feeling was that there is still so much room for improvement in robotics.

The acrobat blindfolded, walking a tightrope dozens of meters in the air, the stunt performer spinning dozens of bowls simultaneously with a single chopstick—they demonstrate what humanity has evolved over millions of years: extreme perception, instinctive balance, and tactile feedback at the finest scale.

Although robots can now perform backflips and martial arts, they are still far from being like humans.

I asked Wang Chuang what he was looking forward to as the robot’s next milestone, and his answer surprised me a little. It wasn’t about more complex or fancy movements, but something extremely basic—a human-like “sensory-motor integration” instinct: catching a falling leaf.

Wang Chuang

Partner, Senior Vice President, President of General Business Division

There is a leaf; I can walk over to it and reach up to pinch it with my fingers.

A breeze passed by, sweeping through a grove of trees, and as “it” walked over and reached out, it “just happened” to catch a falling leaf. On this day, robots drew significantly closer to our daily lives.