_________

Published on Sep. 04, 2026

On the morning after his company’s stock closed 460 percent north of its issue price, Wang Xingxing had every reason to celebrate. Instead, the founder of Unitree Robotics stood on the main stage of the World Robot Conference in Beijing and explained, patiently and in technical detail, why humanoid robots are still no match for human workers.

His industry will reach its “ChatGPT moment”, Wang said, when a robot can successfully complete around 80 percent of tasks in an unfamiliar environment. In an optimistic scenario, he estimated, this moment could arrive within two to three years. It could also take five to ten, the Chinese business newspaper Jingji Guancha Bao quoted him as saying on August 20, 2026 (in Chinese).

Unitree’s shares fell 18,7 percent that very day. A few days later, the stock at times traded 46 percent below its first-day high. Now, this kind of sharp pullback tends to happen after spectacular stock market debuts in China. Still, founders rarely dampen expectations so bluntly one day after the market has celebrated them. Wang, who is almost never photographed with a smile on his face, seems to be a real no-nonsense type of person.

“When a robot reaches for an object, it looks like it has almost got hold of it — but that last bit of tactile feedback, that final small margin of error, it cannot correct,” Wang said. This is when the success rate of humanoid robots currently collapses.

His own company, Wang said in Beijing, was holding off on a large-scale rollout of humanoids because their efficiency and generalization remain insufficient. He called this insufficient generalization of embodied intelligence the biggest current bottleneck of the robotics industry worldwide. After enough data collection and specialized training, today’s humanoids reach task success rates approaching 100 percent. But that is only in fixed environments, without any surprises.

Once the object or environment changes only ever so slightly, the success rate falls sharply. With this yardstick in mind, a stroll through the robot trade fair in Beijing’s Yizhuang district offered an objective answer to the question of where humanoids stand right now. More than 300 companies showed over 2,000 products from August 19 to 23, 2026, so both the progress and lingering constraints of humanoids were clearly displayed.

A year ago, robots at most booths tried to outdo one another by performing somersaults and boxing matches. This year, quite a few of them sorted medicines in simulated pharmacies, retrieved goods for customers in mock retail stores, or folded clothes and tidied up model homes.

Fewer of the robots were steered by human operators with remote controls, compared to earlier years. This year, they started to walk by themselves. “The most important step for humanoid robots entering real-world scenarios is getting rid of the remote control,” Chen Feng of the robot maker LimX Dynamics told a Chinese reporter at the robot show.

The Unitree founder’s sober remark is spot on, but there was also clear progress compared to just a year ago. Xiaomi’s robot Tieda has been fastening nuts at a fixed station in one of the carmaker’s plants since March. It is an operation of a few seconds, repeated hundreds of times a day.

The robot started with a success rate of 90 percent, Xiang Diyun, general manager of Xiaomi’s robotics division, said during this year’s robot conference. It kept getting better, however. Four months later it stood at 98 percent, only one percentage point below the 99 percent of its human colleagues. This gap should close by the end of this year, the manager thinks.

Right now, in the summer of 2026, a lack of training data and computing power means the robots are still a little bit clumsy. For tasks of three to five seconds, robots reach 70 to 100 percent of human efficiency, depending on the scenario. Beyond one minute, this can drop to around 30 percent.

Wang Xingxing, the Unitree founder, explained the mechanism behind this decay. Every perception-and-control loop introduces a small error. The longer a task runs, the more these deviations accumulate.

Visitors to the World Robot Conference could watch the phenomenon live. At a mock retail booth, a robot took an order via smartphone app. It then picked a bag of soy milk from its shelf and placed it on the counter. At the next order, however, it froze. The crowd of onlookers had confused its obstacle-detection system, a staff member explained.

If a small crowd is enough to paralyze the machine, it will never survive a real store, one visitor commented dryly. At another booth, a robot spent several minutes trying to fold a shirt and failed, the news agency AP was happy to report.

Engineers call such surprises long-tail problems. Each one is rare. Together they come in endless variety. One company currently observing a lot of them is Yunji Technology. It has deployed service robots in more than 40,000 hotels worldwide, it claims. Long-tail problems occur on a daily basis. Long-pile carpets can trap a robot’s wheels. Then the hotel staff has to rescue the machine.

Obviously, this is not how it is supposed to work. “Customers buy robots to solve problems, not to take care of the robots,” Duan Yanbiao, Yunji’s chief product officer, told reporters at the fair. Today’s humanoids resemble interns. The smart type, that is. They may be clumsy in the beginning, but they are learning fast.

To understand the current constraints, you have to look at the robots’ brains. The architectures have recently aligned throughout most of the industry. In 2025, most embodied-intelligence companies adopted vision-language-action models, which process images, language and movement in a single system. This year, most of them added world models on top, which let a robot predict the state of its environment before it acts. What the brains need now is more data and more computing power.

That is exactly why it is so fascinating what is happening in China right now. The retail group JD.com is opening its logistics, retail and healthcare operations to robot makers for joint data collection. It wants to gather tens of millions of hours of high-quality scenario data over the course of the next two years.

Robots are moving from “demonstrating” to “working,” said Zheng Xiaodan, who runs the embodied-intelligence business of JD Retail. Nothing beats data collection for getting better. The cheaper path, training in simulated worlds, has clear limits. Virtual environments can multiply scenarios almost at will. They also tend to simplify physics and miss the crumpled wet tissue that sticks to a robot’s gripper. The industry calls this the sim-to-real gap.

Computing power, the second big constraint, is advancing as well. As single chips deliver more and more TOPS (tera operations per second) with every new generation, it becomes possible to run more and more demanding AI models on the robot itself rather than in the cloud.

American and Chinese companies are trying different solutions, but their common entry point is the robot’s brain. Without a capable brain, after all, a robot is merely a sophisticated piece of hardware.

The late summer of 2026 thus finds the humanoid robot industry on sober middle ground. The stock market, with the Unitree IPO, has priced in the high expectations for the future. Practioners like the Unitree founder dampen exaggerated expectations for the immediate future.

Yet it is also true that hardly anyone in China still argues whether humanoids still have a great future. The argument is about the pace of the learning curve, not the final destination.

At the beginning of this year, you could still compare Chinese humanoids to children who dance and play before the seriousness of life catches up with them. Now, half a year later, school has begun.