China's Robots Are Beating Humans on Ice. This Is Just the Beginning.

Added: 2026-05-06

[00:00:00]That robot is not CGI. It is not slowed down. It is not edited. It is not a person in a costume. That is a real machine, the Unitree G1 made in Shenzhen, China, gliding across a roller rink, executing a 360° spin, and landing it like it has been skating for years.

[00:00:14]And then, in the same video, putting on a pair of ice ice skates and doing it again on ice. Most people who try roller skating for the first time spend the first hour trying not to fall down. Most people who try ice skating give up within 10 minutes. The Unitree G1 just learned to do both, and then it learned to do tricks, and it did all of that without having a single human teacher physically guide it through a single one of those movements. The footage went online late last week. Within hours, half the internet was convinced it was fake. It is not. To understand why this matters, you have to understand why skating is fundamentally different from walking. When a robot walks, every step is a controlled fall. It lifts a foot, it puts the foot down. The friction between the sole and the ground gives it a stable base. It rebalances. It takes the next step. This is hard. It took roboticists 30 years to make it look natural, but it works because the surface is reliable. The ground does not move. Now, take the friction away. On roller skates, there is no stable base.

[00:01:04]The wheels are always rolling. Every shift in weight changes the trajectory.

[00:01:07]Every micro adjustment by the robot has to account for momentum that walking does not have to deal with. On ice, it gets worse. Ice is one of the lowest friction surfaces in nature. Even when you're standing perfectly still, you are sliding by tiny amounts. Your body is constantly making invisible corrections to stay upright. That is why ice skating is one of the hardest skills a human can learn. It is not a single movement. It's thousands of micro adjustments per second, all happening below the level of conscious thought. And it is exactly the kind of problem that defeats most robots. Until last week, when a $17,000 humanoid in Shenzhen made it look easy.

[00:01:37]The robot in that footage is not a custom prototype. It is the Unitree G1, a commercially available humanoid robot that anyone with $17,990 can buy from Unitree's website right now. It stands 4 ft 2 in tall, weighs 77 lb, has up to 43 motorized joints depending on the configuration. It is, by Unitree's own description, a research and development platform, a general purpose humanoid designed to be programmed and trained for whatever the user needs it to do. And what some engineers at Unitree decided to do with it last week is teach it to skate. The footage starts simply. The G1, fitted with what looked like ankle-mounted wheel attachments, glides across a smooth indoor surface. It is not falling. It is not wobbling. It is moving with the kind of measured confidence that you usually only see in someone who has been skating for years.

[00:02:20]Then the demonstration escalates. The robot transitions into a 360° turn mid-glide, without losing balance, without slowing down. Then it does it again, faster. Then it lifts one leg off the ground while still spinning. A one-leg spin on skates. If you have ever tried to do this on actual roller skates, you understand what is happening here. This is the kind of move that intermediate skaters spend months learning. It requires you to balance on a single rolling wheel while controlling rotation through your hips, arms, and core. Your brain has to coordinate all of those systems simultaneously. The robot just does it. And then, and this is the moment that broke the internet last week, it does a front flip on skates. Watch that footage carefully.

[00:02:56]The robot launches off both feet, rotates forward through a full 360° in the air, and lands on the wheels, still moving forward. Most professional human skaters cannot do this. And then, because the engineers behind this video apparently decided that wheels were not impressive enough, they put the robot on ice skates. Ice, the lowest friction skating surface that exists. And the G1 does not just stand on it, it glides. It turns. It maintains balance on a surface where the average human falls within seconds of trying. The robot is not figure skating yet, but what it is doing on ice is, by any honest measure, more controlled and more capable than the first attempts of most human beings who have ever tried to ice skate. And it learned all of this without a human ever physically guiding it through a single one of these movements, which raises the question that makes this video different from every other robot demo you have seen this year. How does a machine learn something that requires invisible micro adjustments when no one can show it how those adjustments are supposed to feel?

[00:03:48]Because that is the part of skating that nobody can teach you with words. A coach can demonstrate the movement. They can correct your posture. They can tell you to bend your knees more, lean forward, push from the inside edge of the blade.

[00:03:58]But the actual knowledge of how to stay upright on ice, the thousands of tiny corrections per second that keep you from falling, that knowledge does not transfer through language. It only develops through falling. A child learning to ice skate falls hundreds of times before something clicks. The body figures it out. The muscles learn what the words could never describe. Robots cannot learn that way. Or rather, robots could not learn that way until 2 years ago. Now they can. And the way they do it is the part of this story that should change how you think about every robot demo you watch from now on. Here is the part of this story that is more interesting than the trick itself. The robot did not learn to skate by skating.

[00:04:30]It learned to skate inside a computer.

[00:04:31]This is the technique that has changed humanoid robotics in the last 2 years more than any single hardware advance.

[00:04:36]It is called sim-to-real training, and here's how it works. Engineers build a virtual model of the robot, every joint, every motor, every weight distribution.

[00:04:44]They drop that virtual robot into a virtual environment that simulates the physics of the real world. In this case, that environment included roller skates and ice. Then they let the robot try to skate. The first attempts go badly. The virtual robot falls down. It falls down a million times, sometimes more. But every fall produces data. Every failure teaches the system what does not work.

[00:05:03]Every successful balance recovery, even by a tiny margin, gets reinforced. The training runs in parallel, not one robot trying to skate, thousands of virtual robots all attempting the same skill simultaneously. Inside a fast computer, the equivalent of years of practice can happen in a few hours. By the end of the simulation run, the system has learned, not just how to skate, but the exact sequence of motor activations, balance corrections, and weight shifts that produce a successful glide on an unstable surface. Then the engineers take that final algorithm, the set of learned behaviors, and they upload it into the real robot. The real robot, which has never physically skated before, which has no intuition, which has no muscle memory, and on the first try, it skates. Because the algorithm running inside it has already practiced skating thousands of times in a place that does not exist. This is the technique that has quietly transformed humanoid robotics, and it is the reason that Chinese robots in particular have been advancing at a pace that nobody outside the field expected. Because once you can teach a robot to do something in simulation, the cost of teaching it new skills drops to almost zero. The robot does not need to be in the real world to learn. And whatever it learns can be uploaded into every other identical robot instantly. One robot learns to skate, every Unitree G1 in the world can now skate. That is not how human skill works. That is something else entirely.

[00:06:13]Think about what that actually means at scale. A human chef takes years to learn how to handle a knife properly. The robot version of that skill can be developed in simulation in a few hours, then transferred to every chef robot in existence overnight. A surgeon takes a decade to develop the steady hands required for delicate operations. The robot version of that skill can be perfected through 10 million simulated procedures in a weekend. Skating is just the demonstration that this method works, that a robot can learn an embodied physical skill, one that requires real-time balance, momentum, and reflex entirely in a virtual environment. If skating works, every other physical skill works the same way.

[00:06:47]Cooking, cleaning, folding laundry, driving, operating equipment, performing surgery, every skill that humans currently spend years acquiring through repetition and feedback can theoretically be acquired by a robot in simulation in a fraction of that time, and then distributed to every machine on the planet running the same software. It is easy to look at a robot doing tricks on ice and dismiss it as entertainment, a demo, a flex, a way for Unitree to go viral on social media before its competitors do. That is not what this is. Walking is one of the most energy inefficient ways to move from one place to another. Every step lifts the entire weight of the body. Every step is a controlled deceleration followed by a controlled acceleration. Most of the energy a walking robot uses is spent fighting gravity, not making forward progress. Wheels are different. Once a wheeled object is in motion, it stays in motion with very little additional energy. The robot's batteries last longer. The robot's joints wear out more slowly. The robot can carry more weight further with less power. This is why every warehouse logistics company on the planet has spent the last decade building wheeled robots. Wheels are simply more efficient for the kind of work robots actually do. But until now, you had to choose. A wheeled robot was great for flat warehouse floors, useless on stairs, useless on uneven ground, useless in the kinds of environments that humans navigate without thinking. A legged robot could go anywhere, but it was slow. It used too much energy. It could not compete with a wheel on a smooth surface. What Unitree just demonstrated changes that calculation. A humanoid that can put on wheels when the surface is flat, and take them off when the surface gets complicated. A robot that works the way a human worker works, walking when it has to, but with the option to roll when rolling makes sense.

[00:08:18]The same robot in a warehouse, in an office, in a factory, in a home, adapting its mode of movement to the environment it finds itself in. That is not a small upgrade. That is the difference between a robot that can do one job and a robot that can do all of them. Picture the most boring version of this technology imaginable. A logistics center. A robot rolls smoothly down a long warehouse aisle, picking inventory off shelves. Then it reaches a section with a damaged floor, uneven, broken concrete that wheels cannot handle. The robot retracts its wheels, walks across the broken section, extends the wheels again on the other side, continues working. It does not call for help. It does not need a different machine for the rough section. It does not need someone to redesign the floor. It just adapts, the way a human worker would adapt. Now, picture a hospital. A robot rolling smoothly through the hallways delivering medications, reaching a flight of stairs to access a different floor, walking up the stairs, resuming the rolling pattern at the top. A construction site. A robot rolling efficiently across smooth flooring, walking across rough terrain, climbing scaffolding when it has to. Your home. A robot doing the dishes in the kitchen, walking up the stairs to vacuum the bedrooms, walking back down, going outside to take out the trash. All of those use cases require the same thing.

[00:09:24]A machine that can move efficiently on flat surfaces, but is not defeated by anything else. And the robot in that skating video just demonstrated that capability under conditions much harder than any of those real-world scenarios will ever throw at it. And the only reason the demo focused on skating tricks instead of warehouse logistics is that warehouse logistics does not go viral on social media. Tricks do. So the engineers made a video of a robot doing tricks to prove what the robot can actually do, which is something much bigger than skate. Humans took thousands of years to learn to skate, not as individuals, as a species. Ice skates were invented around 5,000 years ago in Finland. The first ones were made of animal bones strapped to the feet of people trying to cross frozen lakes more efficiently. Every generation since has refined the technique. Children learn from parents. Skaters learn from coaches. Olympic gold medalists pass on what they know to the next generation.

[00:10:10]That is how human skill develops, slowly, across lifetimes, from person to person. The robot in that video learned to skate in a few hours, inside a computer, with no teacher, through pure trial and error in a virtual world. And once it learned, it could share what it learned with every identical robot instantly, perfectly, with no degradation across the transfer. 5,000 years of human refinement compressed into an afternoon of simulation. Think about every skill humans have ever developed. The first humans to make fire, the first humans to weave cloth, the first humans to play music, the first humans to build a wheel, the first humans to plant a seed and wait for it to grow. Every one of those skills was passed down through generations, refined, improved, sometimes lost, sometimes rediscovered. Human skill is fragile. It is local. It dies with the people who hold it. Robot skill is none of those things. Once a robot learns something, the skill exists forever, available to every machine that can run the same software. No knowledge is ever lost. No teacher is ever needed. No student ever fails to learn what was taught. Available as a software update to every machine. Skating is the demonstration. The thing being demonstrated is something else entirely.

[00:11:10]What skill should they teach the robot next? Tell me in the comments. And if you want to see what these same machines look like when China sends them into actual competition with each other, that video is right here.