China's Robots Are Getting Faster Than You. And They're Just Getting Started

追加: 2026-05-08

[00:00:00]That tablecloth was yanked out at full speed. The wine glasses did not move.

[00:00:03]Now, watch the next clip. A robot holding a kitchen knife shaves a cucumber into slices so thin you can see through them, one after another, faster than any human chef has ever moved.

[00:00:12]Then, watch the next one. Calligraphy, a brush in the robot's hand painting Chinese characters with the kind of precision that takes a human artist 20 years to develop. All of those clips were filmed in real time. No edits, no speed adjustments, no tricks. That is the Astribot S1, a humanoid robot built by a company called Stardust Intelligence in Shenzhen, China. Its arms move at 10 m per second. That is faster than the swing of a professional baseball pitcher's arm at the moment they release the ball. It is faster than the strike of a professional boxer's jab. It is faster than the human arm has ever been measured moving in any context where measurement matters. And the machine that moves that fast can also place a single needle through the eye of a thread. There is a trade-off in robotics that everyone in the field has accepted as fundamental. You can build a fast robot. You can build a precise robot. You cannot, until now, build a robot that is both. Industrial robots in car factories are fast. They weld and cut and place components at speeds that no human worker could match, but they are not precise in the way a human is precise. They follow programmed paths.

[00:01:07]They cannot adjust to a part that is slightly out of position. They cannot pick up a wine glass without crushing it. Surgical robots, like the da Vinci system used in hospitals around the world, are precise. They can perform delicate operations with sub-millimeter accuracy, but they are slow. Every motion is deliberate. Every adjustment is calculated. A single procedure can take hours. The reason for this trade-off is physics. Speed creates inertia. Inertia creates error. The faster a robot moves, the harder it is to stop precisely where you wanted it to stop. To overcome that, you need motors that can decelerate as fast as they accelerate. You need control systems that update hundreds of times per second. You need sensors that can detect the exact position of every joint at every instant. And you need all of those systems to work together perfectly, without lag, without error, without the tiny imprecisions that compound over time. That is the engineering problem the Astribot S1 just solved. And the way it solved it changes what robots are now capable of doing in the real world. Let me put the specifications in human terms. The Astribot S1's arms reach a maximum speed of 10 m per second. To understand what that means, think about a professional baseball pitcher. The fastest pitchers in Major League Baseball release the ball at speeds around 40 m per second, about 90 mph, but that is the speed of the ball, not the arm. The arm itself, at the moment of release, is moving at roughly 10 m per second. The Astribot S1's arm moves at the same speed, all day, every motion, without warming up, without injury, without fatigue. And it does so while maintaining precision of plus or minus 1/10 of a millimeter. A human hair is about 1/10 of a millimeter thick, so the robot's arm, moving at the speed of a Major League pitcher's release, can stop at any point in space within the width of a single human hair of the intended target. That combination of speed and precision is what does not exist anywhere else in robotics. It is also what allows the demonstrations you are watching to actually work. The tablecloth pull is not magic, it is physics. The robot accelerates the cloth horizontally faster than gravity can pull the wine glasses downward in response. The cloth slides out from under them so quickly that the glasses do not have time to tip. Their inertia keeps them in place. Humans can theoretically do this trick. A few practice performers actually can, but for a human, it requires years of practice. And even then, the success rate is uneven. For the Astribot S1, it is just the application of two numbers, acceleration of 100 m per second squared, top speed of 10 m per second.

[00:03:19]Run the math. The cloth gets pulled out before the glasses know it is gone. Each arm has seven degrees of freedom, meaning seven independent joints that can rotate. That is one more than a human arm, which means the robot can reach into positions that no human hand can physically achieve. Each arm can lift up to 5 kg at full horizontal extension. That is roughly 11 lb held at arm's length with the precision to thread a needle while doing it. The fingertips have tactile sensors. The wrists have force sensors. The whole system is integrated to a level that means the robot does not just move, it feels what it is touching in real time and adjusts its grip accordingly. These are not the specs of a research curiosity. These are the specs of a working machine, a machine designed to do things, not to demonstrate that it can do things. There is a difference between those two categories of robot, and most of what you see online falls into the first one. A demo, a flex, a proof that something is technically possible in the very specific conditions of a controlled environment. The S1 is not in that category. The hardware is built to operate continuously. The motors are rated for industrial duty cycles. The sensors are calibrated for production environments, not laboratory conditions. The control system updates hundreds of times per second to keep the machine operating within its tolerances, even when conditions shift. In testing, the S1 has been operated continuously for 8 hours straight while maintaining a documented success rate above 99% on the tasks it was given. 8 hours, without a break, [music] without performance degradation, without the kind of small errors that creep in when human operators get tired. But raw capability is only half the story. A robot that can move at 10 m per second is impressive. A robot that knows what to do with that speed is something else entirely. The thing that makes the Astribot S1 different from every robot that came before it is not the speed, it is the way it learns. Most industrial robots are programmed. Engineers write code that tells the machine, in precise terms, exactly what motion to execute and when. The robot does that one thing very well, forever. If you want it to do a different thing, you have to write new code. Most humanoid robots before the S1 were the same way. Demonstrations were impressive, but each demonstration was a separate engineering project, hundreds of hours of programming for every new task. The S1 does not work like that. It uses a method called imitation learning.

[00:05:22]Engineers wear sensor-equipped exoskeletons that record their movements as they perform a task, chopping vegetables, folding a shirt, pouring wine, playing a musical instrument.

[00:05:31]Those recorded movements are fed into the robot's learning system, and the robot learns to perform that task, not by copying the motions exactly, but by understanding what the task is, what success looks like, and how to adapt to slightly different conditions. If the cucumber is slightly larger or smaller, if the wine bottle is slightly heavier or lighter, if the pen the robot is holding has a different weight distribution than the one it learned with, the robot adapts because it learned the principle, not just the specific motion.

[00:05:58]This is what the demonstration footage actually shows. The S1 cooking, flipping sandwiches, peeling and slicing vegetables, opening wine bottles. The S1 doing housework, folding laundry, organizing shoes, taking out the trash.

[00:06:10]The S1 doing things that have nothing to do with practical work, practicing Wing Chun, Wing Chun martial arts moves, playing musical instruments, painting calligraphy. Every one of those tasks was taught the same way. A human demonstrated it. The robot learned it.

[00:06:24]And once the robot learned it, the skill could be transferred to every other Astribot S1 in existence. Software update. Done. Imagine the implications of this at scale. The traditional model of skill acquisition is one teacher transferring knowledge to one student over years. A master craftsman teaches an apprentice. The apprentice teaches the next generation. Knowledge flows downstream, slowly, with degradation at every step. The S1 model of skill acquisition is one teacher transferring knowledge to 10 million students simultaneously, with no degradation, in real time. That is not a small change.

[00:06:53]That is the kind of shift that has only happened a handful of times in human history. The printing press, the internet, and now this. A team of 50 world-class chefs spends a week demonstrating culinary techniques to a single S1. Months later, every S1 deployed in every kitchen on the planet can execute those techniques. A team of expert surgeons demonstrates dozens of delicate operations. Six months later, every medical-grade S1 in every hospital can assist with those procedures. A team of factory workers demonstrates the assembly of a complex product. The robots that replace them learn the job in days, not years. The bottleneck is no longer engineering. The bottleneck is finding humans who are willing to demonstrate their skills to the machine that will eventually do the same work, which raises an uncomfortable question, and we will come back to it. The Astribot S1 is on sale right now, not as a research demonstration, not as a pre-order, as a commercial product that you can purchase, ship to your facility, and deploy. The price is between 96,000 and 150,000 US dollars, depending on configuration. That is expensive. It is more than most people pay for a car. It is more than the median annual household income in the United States. But put it in business terms. A skilled worker in a developed economy, a chef, a nurse, a factory line worker, earns around 50 to 70,000 dollars a year in salary. The total cost of employment, including benefits, taxes, training, and overhead, is closer to 90,000. So, an S1 costs roughly one to two years of a single worker's total cost. That worker shows up 40 hours a week, takes vacation, gets sick, eventually leaves for another job.

[00:08:18]The S1 operates 8 hours a day, every day, with a documented success rate above 99% on tasks it has been trained to perform. It does not take vacations.

[00:08:26]It does not get sick. It does not leave for another job. The math works out to a payback period of roughly two to three years, depending on how you count. After that, the robot is cheaper to operate than any human alternative. And every year that goes by, the price drops, the capabilities increase, and the gap widens. This is why the robotics industry has shifted in the last 12 months. The conversation used to be about whether humanoid robots would ever be commercially viable. That conversation is over. The conversation now is about which industries get them first, and how fast the deployment scales. The S1 is currently being sold into research labs, advanced manufacturing facilities, and high-end commercial environments. Hospitals are evaluating them for assistance roles.

[00:09:05]Restaurants are testing them for prep work. Hotels are considering them for housekeeping at premium properties where the labor cost differential is widest.

[00:09:11]And the S1 is just one of more than 300 humanoid robot models that have been released in China since the start of 2025. This is not a future technology.

[00:09:19]This is a product on the market, being purchased and deployed by companies right now. And every machine sold gets better at what it does the more it is used, because every machine sold is connected to the same network of training data. Every successful task, every recovered error, every adjustment to a slightly unusual situation, that data goes back to Stardust Intelligence.

[00:09:37]The learning system absorbs it. The improved capabilities get distributed back to every robot in the field. The S1 you buy in 2026 will be more capable in 2027 than it was on the day it was delivered, without any hardware changes, without any new sensors, just better software, learned from the collective experience of every other S1 in the world doing the the kinds of work. A human worker has a career, a robot has an upgrade path, and the upgrade path keeps going long after the human worker would have retired. There is a specific kind of human pride that comes from being good at a physical skill. The chef who has been making the same dish for 30 years and does it better than anyone else. The carpenter whose hands have learned the grain of wood the way a musician's fingers learn the strings of an instrument. The surgeon whose precision came from a 100,000 hours of practice. That kind of skill takes a lifetime to build, and it is the foundation of how humans have understood their own value for as long as humans have made things. You are good at something, better than other people.

[00:10:27]That is what you bring to the world. For most of human history, that was enough.

[00:10:30]The skills you spent your life building were yours. Nobody could take them away.

[00:10:34]Nobody could replicate them faster than you developed them. Industrial machines automated repetitive work, but the skills that required judgment, dexterity, taste, intuition, those stayed in human hands. They were the floor under everything else. No matter what was automated above, you could always retreat to a craft that required hands and eyes and years of patient practice. Those things were safe. The Astro Bot S1 changes that. The Astro Bot S1 is faster than the best human, more precise than the best human, more consistent than the best human, and it can learn whatever you can demonstrate in less time than it takes you to teach an apprentice. That does not mean the chef is no longer a chef. It does not mean the carpenter is no longer a carpenter. It means we are about to find out what those words mean when the activity that defined them can be performed faster and better by a machine that costs two years of the wages those activities used to pay. That is not a question about technology. That is a question about us, and the machine that just made it impossible to avoid is on sale today in Shenzhen with international shipping available. If you could buy the Astro Bot S1 tomorrow, what would you have it do? Tell me in the comments. And if you want to see what a different Chinese humanoid looks like when it learns to do something nobody expected, that video is right here.