Gasgoo Munich-August saw the hype around embodied intelligence hit a fever pitch, fueled by two major industry gatherings.
Inside the National Speed Skating Oval, known as the "Ice Ribbon," the second World Humanoid Robot Games is in full swing.
Over the past few days, humanoid robots have shattered human records across multiple events—including the 100-meter, 400-meter, and 1,500-meter races, plus the high jump. In the 100-meter semi-finals, Beijing Humanoid’s Tiangong robot clocked 8.86 seconds. That’s not just a sharp improvement over its 9.32-second heat time; it far surpasses Usain Bolt’s 9.58-second world record, which stood for 17 years.
Just days earlier at the 2026 World Robot Conference (WRC), the focus on the show floor had shifted. Gone were the "talent shows"—dancing and backflips. In their place: sorting packages on conveyor belts, tightening screws at workstations, and tidying up homes.
Taken together, it feels like the "singularity" for embodied intelligence is just around the corner.
Yet at the WRC 2026 main forum, Unitree founder and Chairman Wang Xingxing offered a sobering reality check: the "ChatGPT moment" for embodied intelligence is still at least two to three years away—and could take as long as a decade.
Humanoid robot performances are getting flashier, and real-world deployment seems closer than ever, yet mass adoption remains elusive. So, where exactly is the bottleneck?
Faster Than Humans, Slower at Work
At last year’s games, simply finishing the race was an achievement for many teams. This year, not only are more teams crossing the line reliably, but the top performers are making massive leaps in speed and agility.
Image Source: Beijing Humanoid
Beyond the 100-meter dash, Tiangong claimed victory in the 400-meter large-group final with a time of 38.15 seconds, leaving the human world record of 43.03 seconds behind. In the 1,500-meter final, the Tianzhuo team used the Tiangong Ultra to win in 2 minutes 21.63 seconds—beating the human record by 1 minute 5 seconds.
Field event results were equally impressive. Tiangong cleared 3.4 meters in the standing high jump—up from 0.95 meters last year. Its long jump of 7.97 meters is already brushing up against elite human performance.
These results demonstrate that the hardware performance and motion control of current humanoid robots have reached a new level.
But even more telling than the scores is a subtle shift in how the industry judges success: it’s no longer about how good the demo looks, but how well the robot works in the real world.
At last year’s WRC, dancing and backflips were standard. This year, the crowds gathered around demonstrations of robots performing sustained, real-world tasks.
Consider UBTECH’s Cruzr series handling car body parts and logistics sorting, or Wujie Power’s K15 executing long chains of tasks—serving, cleaning, bussing tables—in a coffee shop setting. Then there’s Yujiang’s "one-brain, multi-body" platform, enabling humanoids, quadrupeds, and collaborative robots to work together on production lines.
The logic is straightforward: technology isn't for show—it’s for use. No matter how smooth the demonstration, it ultimately has to prove its worth on the production line and stand up to cost scrutiny.
Yet while robots are bustling at the Games and WRC, the reality on the ground looks quite different.
During Unitree’s IPO roadshow, Wang was blunt: real global demand for humanoid robots is currently concentrated in research, education, and commercial consumption. Large-scale deployment in industrial settings will take time. In the short to medium term, applications will remain focused on research, education, cultural performances, intelligent services, and high-risk emergency scenarios.
The data tells a clearer story.
According to the latest stats from Counterpoint Research, global humanoid robot shipments surpassed 22,000 units in the first half of 2026—a nearly 300% year-on-year jump. But break down the usage, and the picture shifts: entertainment and performance account for 33.6%, while data production and research make up 27%. Together, those two categories represent over 60% of the total. Service and guidance take about 19%, smart manufacturing 12.8%, and warehousing and logistics just 4.9%.
In other words, robots actually entering smart manufacturing and logistics make up less than 20% of the total. The home sector is barely a blip.
And even those that have been deployed often underperform.
Take Unitree. It began collaborating with auto factories on production line deployments back in 2024, but Wang reveals that these robots are currently limited to simple assembly. Not only is their efficiency lower than human labor, but new tasks require retraining, making deployment inefficient. As a result, large-scale promotion hasn't happened yet.
If Unitree is struggling, other manufacturers that haven't undergone sufficient production line validation are likely in even worse shape.
Notably, this struggle to gain industrial traction is already echoing in the capital markets. The sharp volatility in Unitree’s stock since its debut is the most direct reflection of this sentiment.
Image Source: Screenshot from Eastmoney
On August 19, Unitree officially listed on the STAR Market, becoming the first "humanoid robot stock" on China’s A-share market. The IPO was priced at 150.80 yuan per share, valuing the company at about 61 billion yuan. But the stock opened at 1,100 yuan—a 629% surge—at one point pushing its market cap past 440 billion yuan during trading.
Yet in the trading sessions that followed, the stock continued to retreat. By the close on August 25, it had fallen to 602.8 yuan, wiping out nearly 200 billion yuan in market value and leaving the company worth 243.8 billion yuan.
One core reason for this pullback is the market’s growing realization that mass commercialization of humanoid robots will indeed take time.
So, what are the remaining hurdles for humanoid robots to transition from star athletes on the track to reliable workers on the assembly line?
The Scaling Bottleneck: The Brain Is the Core Weakness, But Not the Only One
Wang has a clear benchmark for the "ChatGPT moment" in embodied intelligence: if a robot can complete about 80% of tasks via voice or text commands in 80% of unfamiliar scenarios, that will be the true tipping point for the industry.
Clearly, current embodied intelligence robots are not there yet.
Wang believes the biggest bottleneck preventing general-purpose robots from truly entering our daily lives and homes is their lack of generalization capability.
"Right now, many AI models around the world can achieve a success rate close to 100% in a fixed scenario, provided there’s enough data collection and training," Wang says. But the moment the object being manipulated changes slightly, or the environment shifts, that success rate plummets.
In fact, the limitations of the robot "brain" are widely recognized as the industry’s core bottleneck.
At a forum held during WRC 2026, Zhao Tongyang, founder of ZhongQing Robot, was equally blunt: the single biggest constraint preventing humanoid robots from reaching deliveries of 10,000 units is the "brain’s" insufficient success rate.
"Even a 99% success rate means that if a robot goes out 100 times, it will fail once. That is a very bad thing," he argues. On a production line, a single failure can mean an entire line stops and a batch of parts is scrapped, resulting in massive losses.
Therefore, he believes the robot "brain’s" success rate must climb from 99% to 99.99%—something that isn't possible right now.
Cheng Hao, founder of Accelerated Evolution, agrees that the biggest industry weakness lies in the "brain." Because the technological roadmap hasn't converged yet, the sector hasn't truly entered the stage of engineering implementation.
"In the past, the 'brain,' 'cerebellum,' and body were all lacking. Now, the body is sufficient in many scenarios—for example, when 100 robots march in formation, their steps stay highly consistent. That’s because consistency is quite good, at least in terms of hardware," Cheng explains. "The 'cerebellum' has also basically matured alongside advancements in general motion controller technology."
Still, while the "brain" is the core weakness, it isn't the only one.
Image Source: ZhiShen Tech
According to Liu Yulong, co-founder of ZhiShen Tech, while the current maturity of the "brain" still falls short of the ideal future for embodied intelligence, achieving 10,000-unit deliveries doesn't require perfection. The "brain" just needs to be good enough—meaning it can address specific pain points and solve actual problems within the target application scenario.
The more critical bottleneck, he argues, is whether the hardware can meet various metrics in the target scenario. Can it carry a heavy enough load despite its own weight limit? Can it adapt to high and low temperatures? Does its mean time between failures (MTBF) meet the client’s minimum requirements? And can the manufacturer guarantee consistency in yield, craftsmanship, evaluation, and final delivery across all batches during a 10,000-unit run?
"Right now, the industry lacks clear standards. We need to work with the upstream and downstream supply chains to establish them," Liu says.
Xu Lei, head of JD.com’s intelligent robot business unit, offers a different perspective from the channel side: it comes down to the match between supply and demand—specifically, whether these current embodied intelligence robots can genuinely satisfy user needs.
"If we can match demand well, I think achieving deliveries of over 10,000 units and completing large-scale deployment won't be a problem," Xu believes. The current reality, however, is that the conversion rate for embodied intelligence products is dozens of times lower than for traditional consumer electronics or mobile devices, he reveals.
Thus, the large-scale deployment of humanoid robots is a systems engineering challenge, with weak links in every link. It’s just that the "brain" is the most visible—and the one that most constrains the long-term ceiling.
So, what is holding back the "brain" itself?
The most direct challenge is a severe shortage of data to train the "brain."
"In terms of the 'brain,' the real, effective data we possess—both in quality and quantity—is far from sufficient," Liu states bluntly. Without enough data, there is a discrepancy between virtual simulations and actual hardware performance, which in turn affects the ability to handle long, diverse tasks.
Even more fundamental than the data gap is the model architecture.
Image Source: UBTECH
Jiao Jichao, VP at UBTECH and head of its Embodied Intelligence and Humanoid Robot Research Institute, points out that current embodied large model architectures are basically migrated from natural language models. They aren't necessarily suited to the data types and training paradigms of embodiment. In the future, entirely new architectures dedicated to embodied intelligence are likely to emerge.
After all, language is one-dimensional logic, while embodied intelligence must navigate a three-dimensional physical world. It involves understanding task logic, perceiving spatial position, executing actions in sequence, and fusing multimodal feedback from vision, force, and touch. The problems they solve are on completely different dimensions.
Trying to solve embodied problems using the foundation of large language models is essentially like "trying to build an electric car on the foundation of a gasoline engine." Without changing the underlying architecture, a fundamental breakthrough is hard to achieve.
Insufficient data and mismatched architectures ultimately result in robots that lack generalization capability.
The Endgame Debate: Routes Undefined, Industry Structure Unsettled
There is no doubt that neither embodied large models nor hardware bodies have truly converged yet. And because of that, the debate over the industry's endgame remains fierce. Will the market structure be concentrated or diversified? Will software and hardware be integrated or decoupled? Opinions vary widely.
According to Cheng Hao of Accelerated Evolution, answering these questions requires looking at the different stages of embodied intelligence development: the body stage, the agent application stage, and the model stage.
In the current body stage, the market is undoubtedly characterized by a diverse range of players. It’s like the early computer era: every company has a different approach, with players like Apple and IBM pursuing distinct paths.
"We think we’re about to enter the agent stage—the application era. At this point, the body systems will slowly converge from many to a few, but the applications on top will become increasingly rich. It will be like the early internet or early mobile internet, with a wide variety of diverse applications emerging."
Once the model stage arrives and embodied large models emerge, they might quickly swallow up the applications. "It’s like how people now just download WeChat or TikTok when they buy a phone—those few platform apps are enough. At that point, applications start to converge again," Cheng says.
Zhao Tongyang of ZhongQing Robot offers a different forecast. The humanoid robot market is massive, theoretically capable of supporting multiple giants worth hundreds of billions or even trillions. But profits will ultimately concentrate among a few top players, giving rise to one or two full-stack ecosystem companies like Apple or Microsoft.
"The current stage of the humanoid robot industry is like the feature phone era. Once the industry enters the smartphone era, I believe two or three highly prominent leading companies will capture 80% of the industry’s profits, leaving the remaining 20% for specialized niche manufacturers," Zhao says.
AI-generated image, Source: Doubao
UBTECH’s Jiao Jichao offers a more nuanced, layered view. Upstream components like motors, reducers, and "brain" chips will form an oligopoly. At the body level, platform giants are likely to emerge. The operations end will be diverse, with third-party developers building scenario-specific applications on top of body platforms—resembling the smartphone ecosystem.
Liu Yulong of ZhiShen Tech holds a more fragmented view.
He believes the market won’t be devoured by a handful of giants. Across a thousand industries, every vertical will have companies that survive by meeting specific needs. This is especially true in the industrial sector, where supply chain divisions, channel relationships, and regional characteristics will allow numerous companies to carve out their own survival models across different upstream and downstream links.
If the debate over market structure is just a disagreement over the final shape, the choice of technological route represents a deeper, more fundamental conflict.
In Cheng’s view, the ultimate winner in embodied intelligence will be model companies. In the long run, software and hardware are bound to be decoupled.
This judgment stands in stark contrast to the positions of Zhao and Jiao.
Zhao’s logic is straightforward: to be Apple, you must integrate software and hardware. "Unless you just want to be a Windows, or a contract manufacturer for hardware OEMs."
Jiao also believes software and hardware cannot be separated. In his view, unlike smartphones where hardware makers build the device and third parties build the apps, humanoid robot hardware and software must be deeply coupled—like the human body and soul, they cannot be torn apart.
"Therefore, I believe future body manufacturers must possess full-stack, in-house capabilities covering both software and hardware," Jiao states.
It is easy to foresee that with neither large models nor hardware bodies having converged, these disagreements will persist for some time to come.
Notably, while the industry hasn't reached a unified conclusion on the future landscape or technological evolution, it has reached a strong consensus on a core principle for deployment: don't wait for the so-called "ChatGPT moment." Instead, prioritize scenarios that are less difficult and where the business case makes sense. However, disagreement remains on whether to push into the business (B-end) or consumer (C-end) markets first.
Jiao points out that while consumer demand is high, current products may not actually meet user needs. He predicts that the industrial sector will see a surge in volume over the next year or two, while opportunities for personal emotional companionship products will arrive later.
Image Source: ZhongQing Robot
Zhao also believes the industry doesn't lack demand; the real bottleneck lies in technology and product capability. "So for someone like me who handles both technology and product, I aim to bring them to a balanced state. We prioritize landing applications that are technologically easy to achieve and realistic for us right now."
Scenarios like patrol, security, property management, and reception can be prioritized, he suggests. Then comes the industrial sector, starting with rough, tiring, and undesirable jobs like transport and sorting. Precision tasks like "threading a needle" remain unrealistic for now.
Additionally, ZhongQing is opening up its underlying capabilities in body and motion control, acting as a "shovel seller" that lets third parties develop various industry applications. "You can't expect to cover every industry. With only a few hundred engineers, most companies can only aim at one or two major fields."
Cheng offers a different take. He believes the business and consumer markets aren't a linear sequence; they can develop in parallel. Each can first validate processes in simple scenarios, then gradually penetrate into more difficult tasks.
But regardless of the chosen order, moving from easy to hard and prioritizing value has become the pragmatic consensus across the industry.
Conclusion
Right now, everyone in the embodied intelligence race is talking about "deploy first, iterate later." Yet few can clearly define what true deployment actually means. Is it a perfect product demo on a show floor? A record-breaking performance on the track? Or is it real productivity that customers will pay for continuously and that can be replicated at scale?
This industry is too good at manufacturing highlight moments, but genuine industrial revolutions are never built on hype alone.
When humanoid robots no longer rely on breaking records to make headlines on social media, but instead prove their worth by securing a steady stream of orders for production lines, that is when the industry will have truly crossed the "singularity."
