The Body-First Bet: Unitree and the Robot Price War

Cheaper Than a Used Car

A walking, human-shaped robot now costs less than a used car. Unitree, a robotics company in China, lists its G1 humanoid at around $16,000. That’s roughly the price of a used compact sedan. The same company sells a four-legged robot dog, the Go2, that trots, climbs stairs, and gets back up after a shove. It starts under $2,000, less than a high-end laptop.

These numbers sound like heresy. For roughly thirty years, an “advanced robot” meant one of two things. It meant a Boston Dynamics machine, the backflipping kind that went viral. Its commercial dog Spot starts near $75,000, and climbs past $100,000 once you add an arm and software. Or it meant a Silicon Valley moonshot, a humanoid still years from sale, hand-built at $50,000 to $150,000 a unit. Tesla’s Optimus, the robots from Figure, Agility’s Digit: all sit in that range or higher, and most can’t be bought at any price yet. The gap is even wider for robot dogs. Spot and the Go2 do similar jobs: walking, balancing, carrying sensors up a flight of stairs. But one lists at more than forty times the price of the other.

The assumption these prices break runs deeper than the prices themselves. For a generation, the people building robots and the investors funding them believed the real bottleneck was intelligence, not hardware. The body looked close to solved. The hard part was the brain, the software that lets a machine look at a messy kitchen, understand what it’s seeing, and act on it. Motors and steel were a commodity you bought from suppliers. The brain was the prize, and the money went there: into AI labs, simulation, and training data.

Unitree’s list prices point to a different reading of the problem. Maybe the humanoid contest is not mainly an AI race but a manufacturing and cost race. It’s closer to cars, solar panels, and lithium batteries than to search engines. If a capable body costs the price of a used car now, and half that in a few years, the edge may belong to whoever can make bodies by the million, cheaply and reliably. Solve the brain second. Put a hundred thousand cheap robots into the world, and the data they send back can train an intelligence that a hundred expensive ones never could.

And here’s the uncomfortable part for anyone who assumed this future would be built in California. The current leader in cheap, deployable robot hardware is not American. It’s a company most Westerners can’t pronounce, based in Hangzhou. It already holds most of the global market for robot dogs, and it’s now aiming the same strategy at humanoids.

So here’s the question worth answering. Can “body-first,” cheap hardware at scale, beat “brain-first,” brilliant software running on machines too expensive to sell? And if the cheap body wins, what happens to the balance of power in an industry that bet the opposite way for thirty years?

From XDog to the Gala

Unitree began with a graduate student who could not afford to build robots the expensive way.

Wang Xingxing was born in 1990 in Ningbo, a port city on China’s east coast. As a child he liked to build things, model airplanes and small engines. He was also a poor test-taker. His English was weak enough that it nearly kept him out of high school, and it shut him out of China’s elite universities. He enrolled instead at a technical university in Hangzhou to study mechatronics, which combines mechanics and electronics. As a freshman he built a small walking two-legged robot for about 200 yuan, roughly thirty dollars, out of parts he scavenged. He then went to Shanghai University for a master’s in robotics and control systems, and wrote his thesis on a controller for a brushless DC motor.

That motor matters to everything that followed. A brushless DC motor is an ordinary electric motor. It is the compact, efficient kind produced by the billion to spin the propellers of drones, the wheels of electric scooters, and the blades of cordless tools. It is cheap, quiet, and precise, and it responds instantly to a control signal. But in the mid-2010s, it was not the motor serious roboticists used to make a machine walk.

Around 2015, in his second year of graduate study, Wang used a set of these motors to build a small four-legged robot he named XDog. He did it on a student’s budget. He sourced the parts himself and machined what he could not buy. The result trotted, balanced, and recovered from a shove as well as university lab projects that cost many times more. He entered XDog in a Chinese innovation contest, won second place and a bonus of 80,000 yuan, around twelve thousand dollars, and posted a video of the robot moving. It spread quickly across Chinese and then foreign tech media.

The prize money was almost beside the point. XDog’s parts list contained an idea that went against thirty years of conventional wisdom. Legs did not have to be expensive. A machine could walk with agility on a stack of cheap electric motors, with no special hardware. To see how radical that was, look at the company that had spent those thirty years proving the opposite.

For a generation, the leader in legged robots was one company: Boston Dynamics. It was spun out of the Massachusetts Institute of Technology in 1992 by Marc Raibert, a robotics researcher who had run the Leg Laboratory, a group at MIT and Carnegie Mellon that studied the physics of dynamic balance. Raibert’s central insight was that a machine could stay upright the way a running animal does. It falls forward and catches itself, over and over, instead of creeping between rigid, statically stable poses. It was clever science, and for its first two decades the company’s most reliable backer was the United States government.

The Defense Advanced Research Projects Agency, the Pentagon’s research arm known as DARPA, funded the projects that made Boston Dynamics famous. In 2005 the company showed BigDog, a headless, four-legged machine built to haul equipment for soldiers across terrain too rough for trucks. A video of BigDog staying on its feet after a hard kick on ice amazed viewers. But it ran on a gasoline engine powering a hydraulic system, and it was too loud for any battlefield, so the pack-mule program was eventually shelved. In 2013, again on DARPA money, came Atlas, a full-sized humanoid built for disaster response. It could open valves, clear debris, and, in later demonstrations, land backflips and vault over obstacles. Those clips turned Atlas into a global sensation.

The power behind all this was hydraulics. A hydraulic system moves a joint by forcing pressurized fluid through hoses into pistons, the same principle that swings the boom of an excavator. It is enormously strong and fast, which is exactly why Atlas could jump. But it is also heavy, hot, expensive, and prone to leaking, and it needs a pump and a maze of plumbing packed inside the robot’s body. The engineering was impressive, but the base it was built on was expensive by nature.

That cost followed the company into the market. Boston Dynamics changed owners three times without ever becoming a volume manufacturer. Google bought it in 2013, sold it to Japan’s SoftBank in 2017, and SoftBank passed it to South Korea’s Hyundai Motor Group, which took control in 2021. Its main commercial product, the dog-sized Spot, sold at a starting price near $75,000. In 2024 the company retired the hydraulic Atlas altogether and introduced an all-electric version. It was a quiet admission that the expensive approach had come to an end.

By then the electric approach had already been proven, at the same university Boston Dynamics came from. In 2019 a lab at MIT built the Mini Cheetah, an agile four-legged robot that ran and became the first four-legged robot to complete a backflip. It did all of this on twelve low-cost motors reworked from the kind sold for drones and radio-controlled planes. Cheap electric motors, it turned out, could handle dynamic legged motion. The researchers said the point was to make the parts snap together like Lego and cost little to replace. The same insight Wang had found in a Shanghai workshop was sitting in a lab in Cambridge too. He was simply the first to build a company around it.

Getting to that company took a detour. After finishing his degree in 2016, Wang took a job at DJI, the Shenzhen firm that dominates the world’s consumer drone market. It was the kind of secure engineering job his training pointed him toward. But he lasted only a few months. With the XDog video circulating, a buyer appeared who wanted to buy the robot, and an investor appeared who wanted to fund a startup around it. Wang took roughly 2 million yuan, about $275,000, in angel financing. He left DJI and registered Unitree Robotics in Hangzhou in August 2016, running it out of a fifty-square-meter office. He was 26 years old.

The products came out one after another, each generation cheaper and more capable than the last. First was Laikago in 2017, the company’s first public robot, named for Laika, the dog the Soviet Union launched into orbit in 1957. It was a research platform, sold in small numbers to universities and labs. Aliengo followed in 2019, sturdier and aimed at industrial users. The machine that changed the category was Go1, released in 2021 at a base price of about $2,700. Unitree sold it openly as a consumer robot dog, something a hobbyist or a small lab could order rather than commission. It could sprint at nearly five meters per second and jog alongside its owner like a pet.

Two years later Unitree pushed the number down again with the Go2. It started around $1,600 and carried a spinning laser scanner on its head. This device fires rapid pulses of light and times their return, building a live three-dimensional map of the space around the robot. This capability had been a six-figure research instrument a decade earlier. Now it cost less than a good laptop and could be delivered to your doorstep. By the middle of the decade Unitree was building the majority of the world’s robot dogs, and no competitor was close on price.

Then Wang turned the same method against a much harder problem: a robot that stands and walks on two legs rather than four. The difference is not incremental. A four-legged machine is a stable tripod at almost every instant. A two-legged one is a controlled fall, balanced on a single foot for much of each step. It is the trick a toddler spends a year learning and an adult does without thinking.

In 2023 Unitree showed the H1, an adult-sized humanoid about 180 centimeters tall and weighing 47 kilograms, built once again on electric motors instead of hydraulics. It became the first Chinese-made humanoid shown walking and running on its own. It was also one of the first machines anywhere shown walking untethered in public, with no overhead cable to catch it if it fell. In March 2024 an upgraded version set a Guinness World Record as the fastest full-sized humanoid, at 3.3 meters per second, past the 2.5 meters per second that Boston Dynamics’ Atlas had reached.

The H1 was expensive by Unitree’s standards, listing near $90,000. The machine that reset expectations came later in 2024. The G1 stood about as tall as an eight-year-old child, with dozens of independent joints and a pair of multi-fingered hands, and Unitree put it on sale for $16,000. No walking, human-shaped robot had ever been offered near that figure. In a single product generation the price of a two-legged machine had dropped by a factor the rest of the industry could not match.

For most of a decade the machines were an engineering story, followed by roboticists and investors. On one night in early 2025 they became a national one. On Lunar New Year’s Eve, China’s state broadcaster aired its annual Spring Festival Gala, a variety show that draws well over a billion viewers and is the most-watched television program in the world. Sixteen Unitree H1 robots, dressed in bright red-and-white folk vests, took the stage beside a troupe of human dancers to perform Yangge, a northern Chinese folk dance, twirling handkerchiefs and swiveling in unison. The number was staged by Zhang Yimou, the filmmaker behind Hero, Raise the Red Lantern, and the 2008 Beijing Olympics opening ceremony. It was given the title Yang BOT. The robots held formation on a crowded, unpredictable stage using onboard laser scanners and full-body motion-control software running on Alibaba’s cloud platform. In the weeks after the broadcast, Unitree’s orders jumped.

Weeks later came the political blessing. On February 17, 2025, Xi Jinping held a rare symposium of China’s leading private entrepreneurs in Beijing. The meeting was widely read as a signal that the state was again backing the private tech sector after years of regulatory pressure. Seated in the front row, among founders like Huawei’s Ren Zhengfei, Alibaba’s Jack Ma, and Tencent’s Pony Ma, was Wang Xingxing. At 35 he was the youngest person to address the room, and the only member of the post-1990 generation invited to speak. A self-funded graduate-school robot dog had become a Chinese national champion, and its maker a public face of the country’s ambitions in AI and advanced manufacturing.

That transformation now shows up in hard numbers. Unitree has turned a profit for two straight years. It delivered more than 5,500 humanoid robots in 2025, while holding roughly a third of the global humanoid market and close to 60 percent of the market for four-legged robots. In July 2026 China’s securities regulator approved the company’s listing on Shanghai’s STAR Market, its board for tech firms, at a valuation near $6 billion. This cleared the way for the first pure humanoid maker to reach public markets. The viral dance and the front-row seat had turned a lab curiosity into a price war that Western rivals are still struggling to answer.

But the prices that started that war, a robot dog for less than a laptop and a humanoid for the cost of a used sedan, are not a marketing decision. They come from a physical choice made deep inside the machine, in the one part a leg cannot move without, and the one place where cost, weight, and strength all meet. To understand why these robots cost what they do, the place to look is a single joint.

The Motor Is the Robot

Every step a legged robot takes runs through its joints. What a joint can and cannot do sets the ceiling for the whole machine. A leg is a chain of stiff segments with a powered joint at the hip, the knee, and the ankle. Those powered joints are the robot’s muscles. A real muscle does three things at once. It is strong, able to hold a body up and throw it forward. It is quick, able to fire and relax many times a second. And it is full of sensors that report, at every moment, how hard it is pulling and how far it has moved. A robot joint has to fake all three with metal and electronics. It has to be strong enough to carry the machine, fast enough to swing a leg through a running stride, and sensitive enough to feel the ground the moment a foot lands. And it has to give way a little instead of hitting like a hammer. Any competent engineer can build a joint that is strong, or fast, or sensitive. Building one joint that is all three, and cheap, is the puzzle the whole robot is built around.

Strength and speed are the first pair to collide. A small electric motor has one annoying trait. Spin it up and it turns very fast, but it is weak. Its twisting force, its torque, is nowhere near enough to lift a robot off the ground. That is simply how small motors are built. They run at high speed and low force. A leg needs the opposite, high force at low speed, so engineers insert a gearbox to convert one into the other. A gearbox trades speed for force. It is the same trade a cyclist makes when he shifts into a low gear to climb a steep hill. The pedals spin fast, the wheel crawls, and the force at the wheel multiplies. Gear a motor down by a hundred to one and its output turns a hundred times slower while pushing about a hundred times harder. That number, the hundred to one, is the gear ratio, and raising it looks like strength for free. It is not free. It costs the one thing a leg can least afford to lose.

What heavy gearing costs is feeling. A gearbox is a two-way street. It passes force from the motor out to the leg. But it also passes force the other way, from the leg back into the motor. And the steeper the gearing, the worse that return trip becomes. Seen from the foot, a heavily geared motor behaves as if it were much heavier and stiffer than the small thing it really is. Its resistance to being pushed backward climbs with the square of the gear ratio. A hundred-to-one reduction makes the small motor feel roughly ten thousand times more sluggish to a sudden push from outside. Tap such a joint and it will not move. That deadness costs nothing as long as the robot only ever pushes outward. It becomes a disaster the moment the world pushes back. A running foot strikes the ground with several times the robot’s weight, in a few thousandths of a second. A joint that cannot give way has nowhere to send that shock. It goes into the gear teeth, which chip, and into the frame, which cracks. Heavy gearing leaves a leg both deaf to the ground and brittle against it.

Whether force can travel backward through a joint has a name: back-drivability. A back-drivable joint is a bicycle. Pedal and the wheel turns. Stop pedaling and push the bike along, and the wheel turns the pedals just as readily. Force runs freely in both directions. A joint that is not back-drivable is a hand winch lifting a bucket from a well. Crank the handle and the bucket rises. Let go of the handle and the bucket’s own weight cannot turn it back. The mechanism moves in one direction only. A leg wants to be a bicycle, not a winch, because a joint the ground can push backward is a joint the robot can feel through. If a foot landing can reach back through the gears and nudge the motor, the machine has a way to sense that landing and respond to it. A leg that works like a winch senses nothing. And a robot that cannot sense its own feet comes down like dropped furniture.

For three decades the field had two ways around this problem, and both were expensive. The first was to throw out electric motors and gears and run the joints on hydraulics instead. This is the approach behind Boston Dynamics’ famous jumping and backflipping machines. Hydraulic joints are strong and quick to respond, which is exactly why those demonstrations looked the way they did. But hydraulics need pumps and plumbing, and they are expensive by nature. They priced the robots that used them out of any mass market. That road led to marvels nobody could buy.

The second way kept the electric motor, accepted the heavy gearing, and paid to cover up its side effects. The tool for that is a harmonic drive. It is a compact and clever gearbox from the 1950s, also called a strain-wave gear, and it folds a very high reduction into a thin ring with almost no slack or wobble. Its precision is remarkable. A harmonic drive is why a factory robot arm can weld the same seam or drive the same screw to within a fraction of a millimeter, ten thousand times, without drifting. But that precision is bought with exactly the stiffness that makes a joint deaf. A high-ratio harmonic drive is a winch, not a bicycle, so the arm holding it cannot feel what it touches. Factory arms that do need a sense of touch have to bolt a separate and expensive force sensor onto the wrist to buy the feeling back. And the gearboxes themselves are precision instruments with high price tags, hundreds to thousands of dollars each. Accurate, unfeeling, and expensive, the harmonic drive is the right choice for a bolted-down arm repeating one motion, and the wrong choice for a leg that has to catch itself when it trips.

The answer that changed the economics came out of a robotics lab at MIT led by Sangbae Kim, an engineer known for building fast, animal-like four-legged machines. Kim’s group stopped trying to beat the trade-off and chose instead to refuse most of the gearing. Take a motor built to be unusually strong for its size, shaped and wound for torque rather than raw speed, and gear it down only a little, something close to six to one rather than a hundred to one. The name for this is quasi-direct-drive. True direct drive would bolt the motor straight to the joint with no gears. That gives a perfectly transparent, perfectly back-drivable joint. But it would need a motor too big and heavy for a leg to carry. Quasi-direct-drive keeps nearly all of that transparency, and buys back the missing strength with as little gearing as possible. The joint stays a bicycle. Force still runs both ways through it, so when a foot slams down the impact travels back through the shallow gears and gently spins the motor instead of chipping the teeth. The same low-gear design did two things a heavily geared arm cannot. It could absorb a hard impact without harm, and it could sense and control its own force finely enough, and fast enough, to run. The leg could absorb a landing or take a shove the way a real leg does, by giving first and pushing back second.

That one choice is why these robots cost what they do. Add up the parts of a legged robot and the joints dominate the bill. A robot dog carries twelve of them, three to a leg, and a humanoid carries dozens, so whatever a single joint costs is multiplied twelve or forty times over. The actuator is the one part whose price sets the price of the whole machine. Build each joint from a harmonic drive plus a bolted-on force sensor and it runs to thousands of dollars. Forty of those and the robot is a six-figure machine that only a research budget can afford. Quasi-direct-drive lets the joint be built instead around a mass-produced electric motor, the same kind that spins drone rotors and drives e-bike wheels. The world already makes these by the hundreds of millions, and therefore makes them cheaply. Unitree redesigned that commodity motor for torque and swapped the elaborate gearbox for a light one. And decisively, it built the whole actuator itself, the motor, the reducer, and the controller together, rather than buying the pieces. Its humanoid joints use a motor it developed in-house, the M107. It reaches a peak twisting force of 360 newton-meters, at one of the best torque-to-weight ratios in the industry. The motors in its robot dog reach their rated torque through a reduction of roughly six to one. Building the costliest part in-house, rather than paying a supplier’s margin on it, is also what let Unitree drive the number down again with each generation. Owning the design of the one part that dominates the bill is what dragged the price of a walking machine down to the level of a used car.

Engineers have a name for what such a leg has and a stiff industrial arm lacks: proprioception. It is the animal sense of where a limb is and how hard it is pushing, the sense that lets you touch a fingertip to your nose with both eyes closed. Kim’s group called their low-geared joint a proprioceptive actuator for exactly this reason, and the way it gets that sense is the most elegant part of the whole idea, hidden in the wiring. In this kind of motor the twisting force it produces is very nearly proportional to the electric current running through it. Feed in more current and it pushes harder. Let something push the motor and the current shifts to match. Because a quasi-direct-drive joint is transparent, a force applied to the leg from the outside travels back to the motor and shows up as a change in that current. The current, in other words, is already a force sensor. The same wires that send power in to move the joint also report back, in the size of the current they draw, how hard the joint is being pushed. There is no need to bolt a load cell to the ankle or glue a strain gauge to the shin, no separate sensing hardware anywhere. The robot feels the ground through the very conductors that drive it, hundreds of times a second, quickly enough to react before a stumble tips into a fall.

This is what lets the leg be soft without being made of anything soft. A controller watching the current knows the moment a foot touches down, and how hard. It can then order the motor to give way by just the right amount, and then push back. The joint behaves like a spring or a shock absorber, one whose stiffness is set in code. Make it stiff for a sprint and soft for a rough landing. Change the setting between one stride and the next, all in software, with no physical spring or rubber anywhere in the leg. The softness lives in the control loop, not the hardware, which means it can be tuned, copied, and improved like any other program. And the decision that gives the leg its sense of touch, refusing the heavy gearing and letting the motor stay transparent, is the same decision that makes the leg cheap. Strength, speed, and touch, the three jobs of the muscle, come out of one mass-produced part and a few lines of code. The low price follows from there.

A Million Falls

A robot’s motors, no matter how cheap or how forgiving, do nothing on their own. Something has to decide, many times a second, exactly how hard each one should push. That decision is the hard part. And it is hard because of a problem that sounds trivial but isn’t: staying upright.

A standing robot is top-heavy. Most of its weight, the torso, the battery, the computer, rides high above a small patch of ground covered by its feet. Physics calls this an inverted pendulum. Anyone who has tried to balance a broom upright on an open palm knows how it behaves. The instant it tips a little off vertical, gravity does not nudge it back but pulls it further. And the further it leans, the faster it goes. Left alone, it accelerates to the floor. The only way to keep it up is to keep moving the base. You slide your hand back under the broom’s center of mass again and again, without stopping. A walking robot does exactly this at every step. Each footfall is the hand sliding under the broom. The robot begins a fall on purpose, then catches it by planting a foot in exactly the right place at exactly the right moment.

The catch has to happen fast. If a person had to consciously watch their body sway and think “I am tipping left, I should step left,” they would be on the floor before the thought finished. It takes a human about a fifth of a second to see something and consciously respond. A falling body does not wait that long. Balance has to be a reflex, not a decision, a correction made before there is time to reason about it. A legged robot works the same way. Its low-level balance control runs about a thousand times a second. It senses the machine’s tilt and corrects the force at each joint in a millisecond, far faster than the robot could think about anything. So the whole problem of making a machine walk boils down to producing the right one of those thousand-a-second corrections, and producing it in time.

For thirty years there was one serious way to produce them. Engineers wrote the robot down as math. They captured its masses, its limb lengths, and the physics of how a push at the ankle travels up through the body. All of it became a system of equations, which they then solved in real time to decide what to do. This is model-based control. Given where the robot is and how it is tipping, the controller works out where to plant the next foot and how hard each joint should push. It does this dozens of times a second, to stay balanced and go where it was told. The machines behind the most famous demos ran on versions of this idea. MIT’s Cheetah quadrupeds and Boston Dynamics’ Atlas both computed their motion from an internal model of their own dynamics. Each one solved a fresh optimization problem several times a second. That is how such a controller can chain a run into a jump into a backflip. It can see the physics of the next move coming, and set the body up for it.

Model-based control is powerful, and when it works, exact. But it has one fatal weakness. It is only ever as good as the model. The equations describe a particular robot on a particular kind of ground. The moment reality departs from the math, the whole thing wobbles. And the world keeps producing things the equations never included. A patch of ice the model thought was dry, gravel that shifts underfoot, a load of cargo that moves the center of mass, a kick from the side that shows up in no equation. Any one of these, and the carefully computed footstep is now wrong. Engineers spent years making their models handle more surprises, hand-coding responses to slips and shoves and slopes. But every new situation had to be seen in advance and written in by a person. The method was brittle at exactly the edges where the real world is messiest.

The alternative turns the problem inside out. Instead of telling the robot how to walk, you let it find out for itself. You build a video-game version of the machine inside a physics simulator. It is a digital copy with the same weights and joints, and the same gravity acting on it. Then you set it loose with no instructions. At first it is hopeless. It twitches, buckles, falls on its face. But every attempt is scored. Stay upright and move forward, and the score rises. Fall over, and it drops. The robot’s controller is a program that maps what its sensors report to what its joints do. After each try, it gets nudged toward whatever raised the score. Repeat this millions of times. Little by little the program moves from random flailing to a stumble, to a stagger, and finally to a clean, efficient walk. And no one ever wrote down what walking should look like. This is reinforcement learning: learning a skill from reward and failure, rather than from instruction. The trained result is what roboticists call a policy, a fixed and very fast reflex that turns sensation straight into action.

A robot learning this way falls a huge number of times before it succeeds. No real machine could survive the beating. What makes the method practical is that none of the falling has to happen in the real world. And it does not have to happen one robot at a time. A modern graphics chip, the same kind that renders video games, can run thousands of copies of the robot at once. Each one lives in its own little world, and they all practice in parallel. In 2021 a group at the Swiss university ETH Zurich, led by Marco Hutter, a roboticist whose lab builds four-legged machines, showed how far this goes. They ran more than four thousand simulated robots side by side on a single graphics card, in a simulator called Isaac Gym. With that, they trained a quadruped to walk on flat ground in under four minutes, and over rough terrain in about twenty. Thousands of copies each get the equivalent of weeks of practice. Add them up, and in an afternoon of real time you have more walking experience than one machine could gather in years. Millions of falls, and not a single bent bolt.

But there is a trap in all this. For years it was the thing that broke learned robots the moment they left the lab. A policy trained in a clean simulation learns to exploit that simulation. The simulator’s physics are simplified, smooth and predictable in ways the real world never is. A robot practicing millions of times will happily find a gait that leans on the sim’s exact quirks: a floor with perfectly even grip, a motor that responds with zero delay, a body whose weight is known to the gram. Now drop that policy into reality, where the friction varies, the motors lag, and the weight is a little off. The flawless simulated gait falls flat on the real ground. The distance between the tidy virtual world and the messy real one is called the sim-to-real gap. Closing it is where most of the craft now lives.

The fix is counterintuitive. Instead of making the simulation more perfect, you make it worse on purpose. And you make it worse in a different way for every one of those thousands of copies. This is domain randomization. One virtual robot trains on ground as slippery as ice, the next on ground like sandpaper. One gets motors that respond a bit slower than commanded, another motors that push a little too hard. Their body weights are scrambled, their limb lengths nudged, and at random moments an invisible hand reaches in and shoves them mid-stride. No two copies practice in the same world. And none of them practice in a world that is accurate. A policy that only worked under one exact set of conditions dies right away under this treatment. What survives is a strategy general enough to keep the robot upright across all of them. And a strategy this general treats the real world as just one more strange version it is already ready for.

The most striking example came not from a leg but from a hand. Researchers at the AI lab OpenAI trained a robotic hand to handle a Rubik’s cube, entirely inside a randomized simulation. They varied friction, sizes, and forces across countless virtual trials. Then they moved the trained hand onto real hardware it had never touched. It kept turning the cube even while the researchers actively worked against it, tying two of its fingers together, or poking it with a stuffed giraffe. None of these disturbances appeared anywhere in its training. It had not learned the cube. It had learned to cope with a world that refused to hold still.

Domain randomization gets a policy close, but never exact. The trained reflex will still misjudge a landing by a centimeter, still push a little too hard, because no simulation ever matched reality to the last detail. What saves it is the body. The learned brain can afford small mistakes only because the joints it commands are physically forgiving. The cheap, back-drivable actuators give when the world pushes on them. So when a foot comes down a moment early, or catches the edge of a rock the simulator never placed there, the joint yields and takes up the shock. It does not pass it on as a jolt that tips the robot over. The policy fires its commands to the joints many times a second. The built-in give of the hardware quietly covers the difference between what the policy expected and what the ground actually did. A stiff, heavily geared industrial joint would do the opposite, passing every millimeter of error straight into the frame as a bang. The forgiving joint and the learned controller are a matched pair. The soft body makes the imperfect brain workable, and an imperfect brain is all the method can ever produce.

That pairing is what turns a lab idea into the clips that travel the internet. When a robot dog takes a running kick from a curious engineer, staggers, plants a foot, and keeps trotting, it is not following a script written for that kick. It is running a reflex trained against millions of random shoves in simulation. It meets a new one and handles it the way it learned to handle all the others. Unitree’s four-legged Go2 rights itself after being flipped over, and stays up when pushed, for the same reason. The two-legged machines do it too. In March 2024 Unitree’s H1 became the first full-size humanoid to land a standing backflip on electric joints rather than hydraulics. The move was trained in simulation, then run on the real body. A year later the smaller G1 was throwing punches and roundhouse kicks in a kung fu routine. It picked up the sequence by watching recorded human motion, then practicing it against reward in simulation before it ever moved a real limb. In a 2025 project from Carnegie Mellon and NVIDIA, a G1 copied Cristiano Ronaldo’s leaping mid-air spin and Kobe Bryant’s fadeaway jump shot. The researchers closed the last of the sim-to-real gap by measuring where reality still disagreed with the simulation, then training a small correction to cancel the difference. None of this is puppetry. Each one is a reflex, practiced inside a computer until it became reliable.

Here the two halves of the machine, the cheap body and the learned brain, meet in a way that reshapes the whole contest. The training pipeline is neither secret nor expensive. The simulators, the learning algorithms, and the academic frameworks that made all of this work are free to download. Unitree openly publishes its own reinforcement-learning code for the Go2, the H1, and the G1. So a buyer can train fresh skills on the same machine. The intelligence that makes a robot move well has become close to a shared commodity. It is within reach of anyone with a graphics card and the patience to run it. But what is not a commodity is an affordable body worth putting that intelligence into. The same free pipeline that teaches a research humanoid worth many times the price will also teach a $16,000 one. And with a good policy, the cheap machine moves like the expensive one. When the brain can be copied for nothing, the advantage swings to whoever can build the body for the least.

From Walking to Working

A robot that runs, climbs stairs, takes a kick and stays on its feet has solved a genuinely hard problem. It has also solved the wrong one. Staying upright is hard, but it is close to a solved problem, and a machine that can balance still cannot do anything useful. What everyone wants from a robot is not balance but work: fetching a box, wiping a counter, opening a drawer, loading a dishwasher, folding a shirt. Doing something with a pair of hands is nowhere near solved, and it is hard for a different reason than walking was.

Walking was solved because it is, at its core, a narrow problem. It has one clear goal, stay up and move forward. So a machine can tell at every moment whether it is doing well or badly. It obeys the same physics every time, a foot pressing on solid ground, and a computer can fake that convincingly. And because the physics can be faked, a robot can rehearse it millions of times inside a simulation, falling harmlessly until it learns. Manipulation has none of these three properties. Losing each one breaks the strategy.

Start with the goal. Upright or toppled is something a machine can read for itself. Folded correctly is not. There is no simple number that tells a robot it folded a shirt well rather than badly. So it cannot practice its way to the answer by chasing a higher score. Then the physics. A stiff foot on flat ground is easy to model in software. A floppy towel is not. Neither is a bag of rice shifting as it lifts, water leaving a pitcher, or a drawer that jams halfway. None of these can be faked faithfully. And a robot that rehearses against a bad imitation of cloth learns to handle cloth that does not exist. Last, the variety. Every drawer, cup, shirt and kitchen differs a little from every other. This is an unbounded space that no engineer can script by hand and no simulation can cover.

The trick that solved walking, letting a machine rehearse millions of times in software, does not transfer to any of this. Teaching a robot to work needs a different teacher. And the obvious teacher is the human who already knows how. The method is called imitation learning, and it is what it sounds like. Rather than discover a skill from scratch, the robot copies a person doing it. Someone performs the task while the machine records two things at once: what its cameras see and what its joints do. Then it learns to reproduce the second from the first. Given this view of the world, make this motion. The demonstrations are gathered by teleoperation. A human operates the robot by remote control, wearing a rig or holding handles that map their movements onto the robot’s arms. For a few minutes the machine becomes a puppet worked by a person.

In early 2024 a Stanford team built a cheap version of this, a system called Mobile ALOHA. The operator steers both arms and a wheeled base through a body tether. With around fifty demonstrations of a chore, their robot learned to stir-fry shrimp in a pan or push in a chair. It succeeded about ninety percent of the time.

But here is the catch, and it reshapes the whole industry. Every demonstration is one piece of training data, and one piece only. A robot that learned to fold a towel from fifty examples knows nothing about folding a fitted sheet, a sock, or a shirt with buttons. Each of those needs its own pile of demonstrations. And every demonstration requires a human to sit down and puppet the machine through it in real time. The world holds a near-infinite number of small, distinct tasks. Covering it means collecting an astronomical amount of human demonstration. And that collection, not the motors and not the math, is the true bottleneck.

The way out of teaching every task from scratch is to change the kind of brain the robot runs. The models that power chatbots work by predicting the next word. Fed a stretch of text, they output what should come next. And because they have read much of the internet, they carry a rough working knowledge of how the world is described. A vision-language-action model borrows that machinery and points it at movement. Feed it the robot’s camera images plus a plain instruction spoken or typed in ordinary language, like “pick up the red cup.” Instead of the next word, it outputs the next motion. It treats a command to the arm as one more kind of word to predict. One model, trained across many tasks, replaces the old setup of a separate hand-written program for each task. It is a single general-purpose brain pointed at the whole messy category of physical work.

This approach took off in 2023. Google DeepMind, the AI division of Google, released a system called RT-2 that did exactly this. It encoded each motion of the arm as a string of text, so one network could produce words and actions at once. Because the model had taken in huge amounts of internet images and text before it ever saw a robot, it came knowing things no robot demonstration could teach it. Told to pick up “the object about to fall off the table,” or to move something onto “the sum of two plus one,” it acted sensibly. These objects and phrases appeared in none of its robot training. It imported common sense from the web the way a chatbot does. That borrowed knowledge is what lets one brain generalize past the exact demonstrations it was fed.

A wave of these foundation brains followed, each trained on a larger and more varied set of demonstrations. One landmark effort in late 2023 pooled more than a million robot recordings from twenty-two different machines across dozens of labs into one shared dataset. It found that a brain trained on the mixed pile did better than one trained on any single robot’s data. More data, from more bodies, made a better brain. Physical Intelligence, a San Francisco startup founded in 2024 by veterans of academic robot learning, built a general model it calls π0, pronounced pi-zero. It folds laundry, makes coffee, and assembles cardboard boxes. Within two years investors valued the company at more than five billion dollars, because of that model. The direction was clear. Whoever could feed one of these models the most real-world experience would own the best robot brain.

This is the loop that makes cheap hardware look less like a discount and more like a strategy. Picture it as a flywheel. Put robots into the world by the thousand, in warehouses, labs, factories, and eventually homes. Each of them generates fresh experience, whether it is doing its work or being puppeted through a new chore. That experience trains a better brain. A better brain makes the robots more capable, which sells more of them, which puts still more data-gathering machines into the world. The wheel turns faster the more bodies are spinning it. It is the same logic that let the company with the most cars on the road collect the most driving data. And the key feature of this loop is that it rewards sheer volume of hardware above almost everything else.

Unitree sits in an unusually good spot on that wheel. It already ships more humanoids and more four-legged robots than any competitor. So more of its bodies are out in the world producing data than anyone else’s. It has also built the tool for turning them into data collectors. This is a five-fingered hand it calls the Dex5, with twenty joints and ninety-four pressure-sensing contact points. It is back-drivable like the rest of its machines, so a human teleoperator can feel through the robot’s fingers while puppeting it. And it has begun closing the loop in public. In March 2026 it released what it called the largest open collection of humanoid teleoperation data: thousands of real-world demonstrations recorded on its own G1 robots and updated continuously. This came weeks after it open-sourced its own vision-language-action brain, built to run on the same machine. The body-first bet is a business plan: sell the bodies, harvest the data, give away the brain.

That last move, giving away the brain, looks strange until you see the logic. And everyone is doing it. Physical Intelligence open-sourced its laundry-folding model. NVIDIA, the chip company whose processors train nearly all of this, released a free foundation brain for humanoids in 2025. It hands the brain to any developer who will buy the chips to run it. Unitree has open-sourced its own. A brain, once trained, is a file. It can be copied, licensed, leaked, or open-sourced across the world overnight. And its advantage disappears the moment it does. A fleet of cheap, capable bodies is not a file. No one can download ten thousand robots. Whoever makes the hardware sits physically closest to the one thing a better brain cannot be built without, a growing stream of real-world data. And that position cannot be copied at the speed a model can.

There is a large and honest hole in all of this, and it sits right at the center of the argument. The brain may be turning into a commodity and the bodies may be cheap, but the hands do not yet work well enough. Reliable manipulation remains unsolved. The polished clips are real but curated. Figure, a California humanoid startup, showed its robots folding towels on their own in 2025. The demonstrations are impressive. But a folded towel on a clean table under good light is one thing, and a stranger’s cluttered kitchen is another. Machines that impress in a two-minute video still fumble, drop things, and freeze when the world stops cooperating. Rodney Brooks, the MIT roboticist who co-founded iRobot, the company behind the Roomba vacuum, argued in 2025 that the current approach will not produce true dexterity for decades. He noted that a human hand carries roughly seventeen thousand touch sensors, and that no robot comes close to feeling what it holds. A machine that walks beautifully but cannot reliably work is a remarkable toy, not an economic revolution. The whole body-first case rests on a bet that the hands, and the data to train them, arrive in time. If they do not, the flywheel spins without ever catching.

Why Two Legs?

The cheap motors and the learned balance answer a narrow question, how to make two legs walk. They dodge a bigger one. Why build a machine shaped like a person? A wheel rolls farther on less power than any leg walks. A single arm bolted to a rolling cart can stack boxes all day and never tip over. Two legs are the most expensive, most dangerous, hardest to control way to move a robot across a flat floor. So why does much of the industry insist on them?

The answer its champions give is that the world itself is the reason. Every staircase, door handle, light switch, and countertop was built to the measurements of the human body. Tools have grips sized for a human palm. Cars and forklifts have pedals spaced for human legs and seats shaped for a human frame, so a person-shaped machine can climb into the vehicle a factory already owns and drive it. Change the shape of the worker and none of it fits. You would have to rebuild the building. Keep the shape and the machine drops into the sockets that already exist. Brett Adcock, the founder and CEO of Figure, frames it as a plug. “You can plug a humanoid into the world, and it can just do everything a human can,” he has said, and the appeal is that no renovation is needed. Elon Musk makes the same pitch for Tesla’s Optimus, a two-legged machine he says will someday do factory work, mow lawns, and look after children. One body, any job, in the world that already exists.

The critics answer that this is a story, and often a story that hides a weak business case. For almost any specific job, a wheeled base with one arm is cheaper, faster, safer, and more reliable than a walking humanoid, and a machine built for that single job beats both. Wheels waste no energy fighting to stay upright. They roll about three times faster than legs walk across a flat factory, they cannot fall over, and their failures are predictable, unlike those of a machine balancing on two legs. A rolling cart with a gripper already handles something like 80 to 90 percent of ordinary factory work: moving material, tending machines, repeating an assembly step. Legs add cost, weight, and the constant risk that a machine of 60 kilograms or more topples onto a person nearby. Guy Hoffman, an associate professor at Cornell University who studies how people interact with robots, has asked plainly why the field is so obsessed with building a copy of ourselves, when a differently shaped machine would do the work with less trouble.

Buyers are voting, and many are voting for wheels. In early 2026 the carmaker BMW put a wheeled humanoid, built by the company Hexagon, to work at its plant in Leipzig, Germany, a torso and two arms riding a wheeled base that crosses the floor faster than any biped and never has to catch its balance. A German parts supplier signed a deal the same year to deploy as many as 2,000 wheeled humanoids across its factories by 2032. Hospital delivery robots that already carry supplies down corridors ride on wheels, not feet. By the middle of 2026 the trade-off had become an open argument inside the industry itself. The pragmatic reading is that legs solve a rare problem, a flight of stairs with no ramp, a ladder, a vehicle built for a human driver, and everywhere else they are an expensive show.

Unitree’s response to this argument is to refuse to pick a side. It sells both shapes and lets one pay for the other. The four-legged machines are not a bet on the future, but a business in the present. Robot dogs go to work now. They walk inspection rounds through oil refineries and chemical plants, patrol power stations, and carry thermal cameras and gas detectors into spaces too cramped or too poisonous for a person. They sit on the benches of university labs and relay what they find in real time. The customers pay, and the quadruped line is the cash engine. The humanoids are the speculative wager stacked on top, a gamble on a market that may not exist for a decade. Most of Unitree’s humanoid rivals in the West build nothing but humanoids and must raise money against a future that has not arrived. Unitree funds the gamble out of a business that already works. It does not have to be right about humanoids soon. It only has to survive long enough to be right eventually, and the dogs buy that time.

There is an irony buried in the question of why two legs. The legs turned out to be close to the easy part. Balance got solved. The hands did not, and they are the reason a humanoid is hard. The single least-finished piece of a person-shaped robot is not how it walks but how it grips, because the hand it is copying is a machine nothing built has matched. Roughly two dozen independently controllable joints fold into one palm, worked by a thumb that swings to meet every finger. The same hand crushes a walnut and threads a needle, lifts a crate and turns a key, wrapped in skin that feels every point of contact. The best research hands have fewer joints than a real one and a far coarser sense of touch, and Adcock, making the case for the humanoid, gives away the difficulty himself. “Nothing is better than the hand,” he says, which is exactly the problem. The entire argument for a person-shaped machine depends on copying the one human part that is hardest to copy. This is why the wheeled humanoid exists. It keeps the hands, the necessary part, and drops the legs, the expensive part. Until the hand works, a robot that walks flawlessly still cannot reliably do the labor that would pay for it. The hand, not balance, is the gate, and it has barely begun to open.

Which sharpens the skeptics’ question to its finest point. If the hands do not yet work, the legs are overkill, and a wheeled arm does today’s jobs more cheaply, why does the money pour into two-legged machines? Because of the size of the prize. A specialized robot does one task in one narrow market. A humanoid, if it ever works, can do anything a human worker does, which means its market is the entire labor economy. Human labor is worth around $30 trillion a year, about a third of everything the world produces. The forecasts for capturing it are wildly far apart, a sign of how unsettled the bet is. Goldman Sachs, the investment bank, sees a modest $38 billion in humanoid sales by 2035. Morgan Stanley sees $5 trillion a year by 2050, with more than a billion machines at work. Musk has told investors that Optimus will one day be worth more to Tesla than everything else the company does combined. When the target is all of human work rather than a slice of it, the math flips. A small chance at a $30 trillion market is still worth an enormous wager. That expected value, a low probability times an astronomical payoff, is what drives capital toward the humanoid, while the wheeled arm quietly handles the work that exists now. The two-legged machine is a lottery ticket on the largest market there has ever been, and the legs, in the end, are almost beside the point. The bet is on the hands.

The Cost Machine

A cheap joint, repeated across the body, is the single biggest reason these robots cost what they do. But it’s not the only reason. A $16,000 humanoid is the product of a whole system built to squeeze out cost at every step. And that system looks less like the way the West builds robots, and more like the way China already learned to build smartphones, drones, and electric cars. The idea behind it fits in one sentence. Build a robot the way a company builds a phone, a product to be made by the million, not the way it builds a spacecraft, a custom instrument assembled by the dozen.

Start with what Unitree refuses to buy. The powered joints are the costliest parts of a legged machine, somewhere between half and 70% of a humanoid’s total parts bill. Every joint bought from an outside supplier arrives with that supplier’s profit already priced in. Build the joint yourself and you keep the money. Unitree makes its own motors, its own gearboxes, and its own joint controllers. And it applies the same logic to the hardware around the joints: the spinning laser scanner, the depth cameras, the battery packs. More than 90% of what goes into the machine is sourced inside China, much of it off Unitree’s own benches. The industry’s term for the full parts list and the cost of each item is the bill of materials. Analysts who have taken a G1 apart estimate its bill of materials near $9,000, a figure they call world-leading. The sticker starts around $16,000 and rises with options. As Unitree scaled its four-legged line, the gross margin on those robots, the gap between what a unit costs to build and what it sells for, climbed from about 42% to 55%. A hardware maker isn’t supposed to earn margins a software company would envy. Unitree does, because it owns the expensive parts instead of renting them. Its own filings say the buying power that comes with scale gives it a cost lead that rivals can’t easily match.

That buying power exists because of where the company sits. The parts a robot needs, precise electric motors, gearboxes, battery cells, laser and camera sensors, machined metal, are the same parts China spent two decades learning to make cheaply for two earlier industries. It became the world’s factory for consumer drones and then for electric cars, and both left behind a dense web of suppliers that now feeds robotics. A robot’s high-torque motors and battery modules come from supply chains built for electric vehicles. Many of its sensors were first built for self-driving. China makes an estimated 70% of the world’s laser scanners and refines most of the rare-earth metal that goes into the powerful magnets inside every robot motor. In the industrial corridor between Shanghai and Hangzhou, where Unitree is based, a company can reach nearly every supplier it needs within a two-hour drive. A custom part that a Western engineer would wait 12 weeks to prototype and receive can be quoted, sampled, and revised in 10 to 14 days. The machine shop that makes it is one town over, and the sample arrives the next morning. The template is DJI, the Shenzhen drone maker. As it grew, the price of a satellite-navigation chip in its ecosystem fell from around $800 to under $15, and a flight controller from $2,000 to $400. The same collapse is now running through robot parts.

Designing for that supply chain is its own discipline, and it’s a phone company’s discipline. When a product is meant to ship in the millions, the first question about any part is not how good it can be, but whether it can be made quickly, cheaply, and identically ten thousand times over. That question steers a hundred small choices toward components that standard machines already produce. Gearing a motor lightly, rather than using the precise, costly harmonic drive that Western arms favor, is as much a manufacturing decision as a mechanical one. A lightly geared reducer can be cut on ordinary gear-cutting equipment in any competent shop, while the high-precision alternative is a specialty item with a specialty price and a lead time measured in months. The Western tradition ran the other way. For 30 years robots were research instruments. They were built a handful at a time, priced to recover their engineering, and packed with the best components regardless of cost. The buyer was a laboratory or a defense program, not a consumer. That habit leads to great machines in tiny production runs, and tiny production runs are expensive per unit no matter how you count. Unitree designed for volume before it had volume, and volume rewarded the bet.

The gap this opens is clearest joint by joint. A Western humanoid built to the usual recipe puts a motor, a harmonic drive, and a separate force sensor into each rotating joint. For each straight-line joint, it uses a motor driving a planetary roller screw, a precise threaded rod that turns spin into a strong straight-line push. Every one of those pieces is a precision part with a precision price. A single harmonic drive sized for a robot joint starts near $3,000 by itself. A single roller screw runs between roughly $1,000 and $3,000, and Tesla, after working hard to drive the number down, still reportedly pays about $800 for the ones in Optimus. The very highest-precision pieces, ground screws and specialty bearings, still come from a short list of German and Japanese firms that own that narrow trade. Now multiply. A humanoid carries dozens of these joints, 28 in Optimus and more in some designs, so the actuator package alone can run from $15,000 to $40,000 at the low volumes these machines are built in. Add the imported precision parts, the labor of hand assembly in a high-wage country, and the overhead of a production run counted in hundreds, and the reported hardware cost of a Western humanoid lands near $55,000. Unitree builds a comparable machine, joint for joint, for a fraction of that. That’s where the tenfold and twentyfold gaps at the component level come from.

Behind the company stands a state that has decided robots matter. Xi Jinping has put humanoid robotics under one of his signature economic phrases, “new productive forces,” his label for the advanced-technology industries meant to power China’s next stage of growth. In March 2025, “embodied intelligence,” the official term for AI that acts through a physical body, entered the premier’s annual work report for the first time. It was a signal to every province and state bank that the sector had Beijing’s backing. Money followed the words. The National Development and Reform Commission, China’s top economic planning agency, set up a state-backed investment fund meant to steer around a trillion yuan, about $140B, into hard technology including robotics. Shenzhen alone set up a robotics and AI fund worth around $1.5B. Beijing offered individual firms subsidies of up to 24 million yuan for qualifying work and built a shared data center to help train their robots. None of this lands in Unitree’s account directly, but it changes the risk math. Capital is cheaper and more patient when the state has blessed a sector, and a company that stumbles has a wide net under it. A Western robotics startup raises from venture funds that demand a return and can vanish in a downturn. It carries the full weight of failure alone.

All of which raises the question every incumbent asks about a cheap challenger. Is the cheap thing actually good enough? Often, the honest answer is not yet. Independent comparisons put Unitree’s machines at roughly 70% to 85% of the capability of the Western best for a small fraction of the price, and the gap is real. The G1 is built for a laboratory bench, not for abuse, and it wouldn’t survive the kicks and falls that a Boston Dynamics machine takes on video. The lightly geared joints that make the robot cheap and sensitive also make it less precise. A harmonic-drive arm can hit the same point to a fraction of a millimeter ten thousand times in a row. And the motors run hot. Early G1 units could hold only a couple of kilograms at arm’s length for a few seconds before they overheated and needed to rest. Unitree has pushed back that limit with active cooling and denser windings, but hasn’t erased it. Cheaper buys something less rugged and less exact. Whether that matters is the open question. The history of manufacturing is a history of good-enough-and-far-cheaper overtaking excellent-and-expensive, in steel, in cars, in solar panels, in the very drones Unitree’s own supply chain grew up making. Each time the incumbents trusted that quality would protect them, and each time it didn’t, because most buyers don’t need the best. They need something that works at a price they can pay. Whether robots follow that same road, or whether the last bit of ruggedness and precision turns out to be the part that counts, is the bet the whole industry is now making.

The Humanoid Race

Unitree is not running alone. The contest to build a humanoid robot has pulled in more money and more companies than at any point in the field’s history. By the middle of 2026 the runners had sorted into two camps that barely resemble each other. On one side are the Western moonshots, a handful of firms raising billions to build a small number of expensive machines for American factories and, in time, American homes. On the other is a Chinese pack, a crowd of companies iterating fast and racing to ship cheap machines by the thousand. Unitree runs inside the second camp. To see what kind of runner it is, you have to start with the field it is running against.

The loudest of the Western bets belongs to Tesla. Elon Musk has promised that Optimus, the two-legged robot his carmaker has been building since 2021, will one day be produced by the million, and he has set an internal price target near $20,000 a unit at full scale, with a ceiling under $30,000. Those are ambitions, not deliveries. Tesla began building its third-generation Optimus on a converted line at its Fremont plant in early 2026 and put roughly three hundred of the machines to work inside its own factories, what Musk called a learning phase. The Fremont line is designed to make a million robots a year, and Musk has described a second plant in Texas meant for ten million. As of the middle of 2026, not one Optimus had been sold to an outside customer. The scale lives on paper and in speeches. The robots exist in the hundreds. And the $20,000 that Tesla hopes to reach someday is already more than Unitree charges for a humanoid today.

The best-funded of the pure startups is Figure, the California company founded by Brett Adcock. In September 2025 it raised more than a billion dollars at a valuation of $39 billion, roughly fifteen times what it was worth eighteen months earlier. The backers are the biggest names in tech: the chipmaker Nvidia, the software company Salesforce, and Qualcomm, with Intel’s venture arm also on the list. Figure has the most concrete factory record of any Western humanoid. At a BMW plant in Spartanburg, South Carolina, one of its robots worked ten-hour shifts helping build the X3 SUV. In about 1,250 hours it handled parts for more than thirty thousand cars, and moved more than ninety thousand components. A newer Figure model has since taken over sorting parts into the order that assembly workers need them. Unusually for the Western field, Figure broke off a partnership with OpenAI in early 2025, and now builds its own AI in-house rather than borrowing a brain from one of the AI giants.

A different Western bet points not at the factory but at the living room. 1X is a Norwegian-American company founded in 2014 by the roboticist Bernt Børnich, and it builds its machines in both California and Norway. OpenAI’s venture fund was one of its earliest backers. In 2025 it was reported to be raising as much as a billion dollars at a valuation above $10 billion. Its machine is a soft-bodied humanoid called NEO, built for the home. The company opened preorders late in 2025 at $20,000 outright or $499 a month, with deliveries promised through 2026. But there is a big catch. For most household chores as of late 2025, a remote human operator wearing a headset still guided the robot by hand. That put a stranger behind the machine’s cameras inside the customer’s home, and showed how far the home still is from a robot that can take care of itself.

Two more American entrants skipped the home and went to the warehouse. Agility Robotics was spun out of Oregon State University in 2015 by the roboticist Jonathan Hurst. Its bipedal machine, named Digit, has logged more than sixty-five thousand hours moving totes for logistics companies. Its early investors include Amazon, which has run Digit through pilots inside its warehouses. In 2026 Agility agreed to go public by merging with an already-listed shell company at a valuation near $2.5 billion. That is a faster route to the market than a traditional IPO. The other is Apptronik, an Austin company whose Apollo robot works in plants run by Mercedes-Benz and the logistics firms GXO and Jabil. It raised $520 million in early 2026 at a valuation around $5 billion, in a round co-led by Google. Apptronik has tied its robot’s intelligence to Google DeepMind, the search company’s AI division. Its models now run as Apollo’s brain.

A pattern runs through the whole Western camp. The money is enormous. The machines cost tens of thousands of dollars, or aren’t for sale at any price. The target is high-wage American labor. And the intelligence comes from the AI giants, whether OpenAI, Microsoft, Nvidia, or Google, as investors, as partners, or as the source of the software. Even Boston Dynamics is now on the same road. The company’s backflipping machines once defined what an advanced robot looked like. At the start of 2026 it unveiled a fully electric Atlas, and committed its entire first production run to just two customers: the robotics unit of its owner Hyundai, and Google DeepMind. The new Atlas runs on electric actuators from a Hyundai parts division instead of the hydraulics of the old one. Its intelligence is to come from Google DeepMind’s models. And Hyundai has begun building a plant designed to turn out thirty thousand robots a year. A company known for decades for costly, hydraulic engineering is now building cheap electric humanoids at volume, and buying its brain from an outside company. The road Unitree took has become the industry’s road.

China’s pack is deeper and moving faster. The listed incumbent is UBTech, a Shenzhen company traded on the Hong Kong exchange since 2023. Its Walker S2 humanoid began mass production in November 2025, with a first batch of several hundred units and a goal of five thousand a year by the end of 2026. UBTech has put its robots on the assembly lines of Chinese automakers including BYD and Geely. In 2026 it signed a deal with Airbus to test them in aircraft assembly. Its order book for the Walker series had passed eight hundred million yuan, more than a hundred million dollars, by late 2025.

Unitree’s closest domestic challenger did not exist three years ago. AgiBot, also called Zhiyuan, was founded in 2023 by Peng Zhihui. Peng was a star of Huawei’s elite “genius youth” recruitment program, and worked on the company’s AI chips before leaving to build robots. AgiBot has scaled at a striking pace. Its ten-thousandth humanoid rolled off the line in early 2026. Its backers include Tencent, the retailer JD.com, and Sequoia China. It has also found a shortcut onto the stock market by taking a controlling stake in a Shanghai-listed materials company. This lets it reach public investors without a conventional listing. By one industry estimate, Unitree and AgiBot together make up roughly eighty percent of China’s humanoid shipments, and Unitree is the larger of the two.

Below the two leaders sits a crowded field. Fourier Intelligence, which began by building robots for medical rehabilitation, now makes humanoids aimed at eldercare and research. Galbot, founded in 2023, has reached a valuation near $3 billion, with backing that includes a Hong Kong government fund. And Xiaomi, the consumer-electronics giant with far deeper pockets than any startup, has built its own humanoid called CyberOne, and begun testing it on its own production lines. The common thread is a home market where dozens of firms iterate fast, where the entire parts supply chain sits within a short drive, and where capital from corporations and the state is plentiful. It is the most competitive humanoid market on earth, and it is where Unitree fights on price and speed.

Inside that field, Unitree’s standing is easy to state. It is the clear leader on cost and on quality of movement. It has become the default machine for anyone who wants to do robotics research or buy a capable humanoid off the shelf. The clearest sign came in mid-2026, when Nvidia launched an open platform meant to put a standard humanoid in the hands of academic labs. For the body it chose Unitree’s. For the brain it used its own onboard computer. And the labs lining up to use the system were Western and Korean: the Allen Institute for AI, ETH Zurich, and the Stanford Robotics Center among them. That single arrangement captures both Unitree’s strength and its ceiling. The company builds the best cheap body in the world, and it takes the brain from Nvidia. Manipulation, the plain business of using hands to do real work, is a known weak point of its machines. They shine as research platforms and demos more than as factory workers on paid shifts. The better-funded Western firms are further along on the brain, the hands, and the slow work of fitting a robot into a paying customer’s operation. Their robots already move real parts on real assembly lines, and their intelligence is built alongside the AI giants. Unitree leads on the body, and trails on much of what comes after.

The capital markets have priced the body-first bet regardless. A funding round in June 2025 valued Unitree near $1.7 billion. Its investors are the heavyweights of Chinese tech and industry: the internet companies Tencent and Alibaba, the carmaker Geely, the financial group Ant Group, and HongShan, the firm once known as Sequoia Capital China. The food-delivery company Meituan came in from an earlier round. Unitree has turned a profit for two straight years across ten rounds of financing, an unusual record for a hardware startup. It is now heading onto Shanghai’s STAR Market, the mainland’s board for tech firms. Money that could go anywhere is betting that the body is the thing worth owning.

The two camps are betting on different halves of the same machine. That is the asymmetry worth keeping in mind. The West is racing on the brain and on enterprise deals. It wires its robots to OpenAI and Google DeepMind, and places them, a handful at a time, on the floors of BMW, Mercedes, and Amazon’s suppliers. China is racing on volume and price, flooding the world with cheap bodies and collecting the data those bodies send home. The Nvidia platform, an American brain inside a Unitree body, hints at one way the two halves could meet in the middle. Nothing guarantees they do. Whoever masters the hands and the mind first may not be whoever can build the body cheapest. And winning one race does not hand you the other. The contest is running on two tracks at once, and it is far from settled that they finish at the same line.

The Long Bet

Two numbers will settle the argument over the next ten years, and neither of them is a dance routine. The first is the price of a capable humanoid body. The second is the count of those bodies doing real work for money. The bull case is the story of these two lines bending toward each other until they cross.

That case is a chain of ifs, and each link has started to hold. If the powered joints keep getting cheaper, and they are, the parts bill for a useful humanoid has roughly halved between 2024 and 2026. If the general-purpose robot brains keep improving, and they are, then deployments at real scale arrive this decade, in a predictable order. The easy rooms come first. Factory logistics, where the floor is flat, the light is good, and the task repeats: moving totes, sorting packages along a conveyor. Then inspection, the walking patrol through a refinery or a power station that four-legged machines already do for pay. Eldercare and household chores come last, because a stranger’s kitchen is the hardest room in the world for a machine to enter. Governments are already planning for this demand. Japan wants ten million service robots deployed by 2040, many in elder care. China has set a goal of putting robots to work across a hundred real-world settings by the end of 2026. The first machines built for the job are appearing. Fourier’s GR-3, a humanoid shown in 2026, is pitched for elder care. But a robot you can leave alone with a frail person is still years away.

Beneath the chain sits the real argument. Cheap hardware is not just a discount, but a strategy. A better robot brain is built from real-world experience, and that experience comes from robots already out in the world doing things. So the fleet that collects the best dataset is the biggest fleet. And the biggest fleet is the cheapest one to build. This points at one company more than any other. In 2025 Unitree shipped more than 5,500 humanoids, while Tesla, Figure, and Agility each shipped roughly 150. Wang Xingxing, Unitree’s founder, has set a target of 10,000 to 20,000 for 2026. Industry trackers expect the whole market to pass 50,000 humanoids for the year, up more than threefold. If the data flywheel is real, the machine spinning it fastest is the one selling bodies by the thousand at a few thousand dollars each.

Three near-term milestones will show whether the flywheel is catching. The first is money. On July 3, 2026, China’s securities regulator cleared Unitree to list on Shanghai’s STAR Market, and the terms say more than the approval did. The company filed in March and won clearance in 73 days, a fast lane reserved for industries Beijing has decided will matter. It is raising about 4.2 billion yuan, roughly 610 million dollars, at a valuation near 42 billion yuan, about 6 billion dollars. The prospectus names where the money goes: into robot AI models, into new products, and into a new factory. That is the flywheel written as a budget. Sell a slice of the company, and pour the proceeds into building more bodies and training smarter brains.

The second milestone is the price tag, and it is dropping faster than the optimists forecast. The G1 that stunned everyone at 16,000 dollars in 2024 was not the floor. In July 2025 Unitree released the R1, a full-size two-legged robot, at 5,900 dollars, with a stripped-down Air version near 4,900. Whether a capable humanoid could fall below 10,000 dollars, then below 5,000, was answered inside a single year. By April 2026 the R1 was listed for global sale on AliExpress, Alibaba’s cross-border shopping platform, with free shipping to American and European doorsteps. The next thresholds to watch are 3,000 dollars and lower. That’s the range where a walking machine costs less than a refrigerator, and an ordinary household could think about owning one.

One detail matters more than the rest, because it’s the tell for everything that follows. The cheapest R1 ships with hands that do not open. Its fists are fixed, fine for cartwheels and stage demonstrations, useless for picking anything up. Part of how the price came down is that the hardest and priciest part, a working hand, was left off. The price curve and the capability curve are not the same curve. The distance between them is where this whole bet is won or lost.

The third milestone is the one that separates a product from a performance: a robot doing paid work all day, not a two-minute clip. That line was crossed in May 2026, and not by Unitree. A fleet of Figure 03 robots, one nicknamed Rose, ran a live warehouse sorting line for 200 hours without stopping. Together they handled 249,560 packages with no hardware failure and no one stepping in to help. When one robot’s battery ran low, another walked over to take its place, and the tired unit shuffled off to a charging pad built into its own feet. Figure now assembles these machines at a plant it calls BotQ, which reached a rate of one robot an hour in the spring of 2026. This is the number to count from here on. Not backflips, not kung fu routines, but shifts worked and packages moved, machine after machine, with no one babysitting.

The bull case has three ways to die, and they share a shape. The body arrives cheap and on schedule, and the thing that was supposed to make it worth owning never shows up.

The first way is that the hands stay broken. Reliable manipulation is still unsolved, and the R1’s fixed fists show it. Machines can walk, run, climb, and dance, and they still cannot be trusted to load a dishwasher they have not seen before. Dexterity, the everyday human skill of handling unfamiliar objects in positions no one scripted, is far behind human ability. There is not yet even an agreed way to measure how far behind. Through 2026 a wave of foundation models aimed straight at this problem. One came from the startup Genesis AI, trained to cook, solve a Rubik’s cube, and play piano. They arrived with impressive demos and unproven generalization. If those models hit a ceiling, if handling the near-infinite variety of real objects needs a leap no one can yet make, then the humanoid stays what it is today, a remarkable and cheap toy for labs. The trillion-dollar labor market it was meant to swallow does not shrink. It just stays out of reach, and the cheap bodies pile up with nothing paying them to hold anything.

The second way is that the brain arrives, but late and unsafe. A robot that works 99 percent of the time sounds finished until you do the math. At a hundred actions an hour, it fails several times an hour. A home or a crowded workplace needs something close to flawless, and the general-purpose brains are not close. In July 2026 researchers at the Technical University of Munich released the first benchmark that tests humanoids while a person actively bumps, shoves, and grabs at them. This is the ordinary friction of sharing space with a machine. Every leading vision-language-action model scored far below the reliability a person would need to feel safe next to it. Even the flawless 200-hour warehouse run was one repeated task in a controlled building, not an open home. The gap between those two settings is the gap the brain still has to cross. The rules for such robots in homes are themselves only half-written. The international safety standard for personal-care robots is under revision for the first time in more than a decade. If reliability and safety take another ten years to get there, the bodies will exist long before there is anything safe for them to do in a room full of people. A warehouse will take them. A living room will not.

The third way is the one Unitree can do least about, because it is not an engineering problem. In March 2026 two senators who agree on almost nothing introduced the American Security Robotics Act. Tom Cotton, the Arkansas Republican, and Chuck Schumer, the New York Democrat who leads his party in the Senate, put it forward together. The bill would bar the federal government from buying or operating ground robots, humanoids included, made by companies tied to China. A companion bill followed in the House. The stated fear is simple. An internet-connected robot from a rival power is a rolling camera and microphone that someone abroad can take over. And the template is out in the open. It’s the drone playbook, run a second time. Chinese-made drones took most of the American market before regulators moved to shut them out. At the end of 2025 they added all foreign-made drones to a federal blocklist. A national-security research group has already urged that Chinese robots join the same list. A split world could follow, a Western robot ecosystem walled off from a Chinese one. In that world Unitree’s prices no longer matter, because its machines are not allowed through the door. Selling the R1 straight to Western buyers on AliExpress, before the wall goes up, is a race against exactly that ending. The irony runs both ways. The same Chinese supply chain that makes the bodies cheap, the motors and gears and magnets, is the chokehold the West is now trying to break, and cannot break quickly.

Set the three failure modes against the bull case, and the honest answer is that no one yet knows which wins. Whether body-first beats brain-first, whether the company that builds the cheapest machine or the one that builds the smartest ends up owning this industry, is genuinely an open question. That uncertainty is the entire point. Unitree is not making a safe bet, but an asymmetric one.

The shape of the bet is what makes it rational. If humanoids stall, Unitree has still built a business out of robot dogs and research machines that has been profitable for two straight years. A real company whichever way the future breaks. If humanoids work, the prize is not a share of a market, but ownership of the low-cost body layer beneath the whole embodied-AI economy. Unitree would be the company that every warehouse, factory, and household buys the physical machine from, while the intelligence is downloaded from the cloud. A modest chance at owning that layer is worth more than a near-certainty of owning almost anything else. That is why the smart money, Tencent, Alibaba, Geely, and the rest of China’s industrial giants, has already placed its bets.

And the one line that has bent reliably, through every one of these years, is the price. A capable walking machine cost roughly a hundred thousand dollars in 2023, sixteen thousand in 2024, under five thousand in 2025, and it is still falling. The industry has voted with its factories on which curve to chase. Boston Dynamics retired its hydraulics and builds cheap electric humanoids at volume. Tesla is tooling a line for a million a year. Figure is turning out a robot an hour. Every serious builder has converged on the method a graduate student once sketched with thirty dollars of scavenged motors: make the body cheap, make it by the million, and let the brain catch up. The bodies are here, cheaper each year and shipping by the tens of thousands. The brains are getting better month by month. The order in which the two finish is the only open question, and that is a question of timing, not direction.

So watch the two numbers, and let the rest go. Not the dance videos, not the backflips, not the kung fu. The price on the machine, and the count of machines doing paid work. Those two lines are moving toward each other, and the year they cross is the year the shape of work changes. On the evidence of the last three years, that year is arriving sooner than the skeptics believe.