Zuckerberg's mistake: metaverses are more necessary for robots than for humans

In early August, Chinese startup Unitree unveiled its new robot - A2 Stellar Hunter. The "robot dog" demonstrates incredible tricks, from dancing on one leg to confidently traversing difficult terrain off-road. Almost simultaneously, Google's Deepmind lab announced the development of Genie 3, its new AI capable of generating not just videos, but entire three-dimensional worlds that can be freely explored. Why, out of all the huge stream of news, do these two seem the most interesting to me? They demonstrate two vectors of technology development that, when combined, are capable of turning the world upside down.

"In 2025, there are agents capable of doing real cognitive work; writing computer code will never be the same again. In 2026, there will probably be systems capable of finding new ideas. In 2027, there will probably be robots that can perform tasks in the real world... If we have to create the first million humanoid robots the old-fashioned way, but then they can manage the entire supply chain - mining and processing minerals, driving trucks, running factories, etc... - to create more robots that can build more chip factories, data centers, then the pace of progress will obviously be very different," wrote OpenAI CEO Sam Altman two months ago.

This is not just a question of economics. There are researchers who believe that the full potential of AI development to the level of AGI (general AI capable of performing all tasks as well as or better than humans) is basically unattainable unless artificial intelligence has a physical embodiment, thanks to which it can explore the real world and interact with people, just as a human child does, writes Mark Lee, Emeritus Professor of Computer Science at Aberystwyth University (UK) and author of the book "How to Grow a Robot".

The professor is a theorist by position, and Sam Altman is a bit of a dreamer and visionary, as befits the founder of a technology company that wants to change the world.

But here is what the most practical and down-to-earth people in the world - investment bankers from Morgan Stanley - write. According to their estimation, by 2050 together with people the Earth will be inhabited by at least 1 billion humanoid robots, their production will create a market worth $5 trillion. They will be mainly used for industrial and commercial purposes, but as the scale of production increases, the price of a robot will decrease from the current $200 thousand to $15 thousand and gradually they will appear as assistants in homes, predict analysts of the investment bank.

"The forecast for robot use in households is much more conservative, with only 80 million humanoid beings in homes by 2050. We won't see a robot in every home overnight," believes Adam Jonas, head of global automotive and shared mobility research at Ma Stanley. Nevertheless, he estimates that by 2050, about a third of wealthy U.S. households with incomes over $200,000 a year will have a robot servant, and maybe even "a whole staff of humanoid butlers".

And the study's third conclusion, less palatable to the U.S.: according to the investment bank's analysts, China is now winning in robot development. "In our view, China's leadership in AI robotics may need to strengthen before competitors, including the US, pay closer attention," the report quotes Sheng Zhong, head of industrial research at Morgan Stanley, as saying.

Unitree founder Wang Xinxing is doing his best to live up to those predictions.

This is what Forbes called Wang Xingxin. The entrepreneur is only 35 years old, yet his startup is already a "unicorn". According to the magazine, in June this year Unitree received a round of venture capital funding at a total valuation of 12 billion yuan ($1.7 billion).

The investors were Chinese automaker Geely Automobile, fintech giant Ant Group and investment firm HongShan Capital Group (formerly known as Sequoia Capital China).

Perhaps it was the fact that Wang Xingxing was the youngest technology company CEO invited to meet Chinese President Xi Jingping in February this year, with a front-row seat of honor, according to the South China Morning Post. Or maybe it was the dance of his 16 H1 robots, which performed with live dancers at the 2025 Spring Festival - a show on China's national TV watched by more than 1 billion people.

Most likely, it all worked together. Unitree's successes are inspiring not only bosses, the public and venture capitalists, but also retail investors. "Because Unitree remains a private company, investors are instead going after its suppliers. The share price of Zhejiang Changsheng Sliding Bearings in Shenzhen has jumped 62% in the past five trading days and more than 600% in the past 12 months. Shares of Ningbo Shuanglin Auto Parts jumped 575% year-on-year," the South China Morning Post wrote in February.

And a demonstration of the A2 Stellar Hunter sci-fi robot dog led to a 5 percent rise on Aug. 6 in the Solactive China Humanoid Robotics index, which includes companies that manufacture and develop robots, Bloomberg writes.

A Unitree spokesperson told Forbes that the company controls more than two-thirds of the global robotic dog market and is the world's leading seller of humanoid robots, but declined to give specific numbers.

Yes, while Tesla is struggling to launch its Optimus humanoids, Unitree is already selling robot dogs starting at $2800 and R1 humanoid robots starting at $5900 (excluding shipping). The robot, which is about 1.2 Ma tall and weighs 25 kg, can interact with its surroundings using built-in microphones, a speaker and cameras. It can also walk, run, walk on its hands, somersault, fight (or at least mimic the movements of a kung fu fighter), lie down, stand up, and more.

However, let's not exaggerate - so far it is more of an expensive toy than something really useful in the household. "The Unitree R1 is unlikely to replace your housekeeper, it has no "dexterous hands" and little Xi because of the limited torque. The company advertises it as a 'smart companion'...It could be useful for education and research, as it is much more affordable than other humanoid robots of this size," CNX Software wrote in a review .

Why is creating a truly useful humanoid robot such a daunting task? I wrote about it in detail last year. The problem was formulated back in the 1980s by Canadian engineer Hans Moravec, and in his honor it was called the Moravec paradox. Its essence is that complex tasks - calculations, logical reasoning - require relatively little computing power from a computer, while elementary human actions like movement, orientation in space, and image recognition require huge resources.

Moravec believed that the explanation for the paradox lay in the realm of biology. Basic skills of the human body and mind, such as the ability to recognize prey or predators in time, to navigate a complex situation, to manipulate one's body and objects accurately, are polished by millions of years of natural selection, so they seem simple to us, but are incredibly complex for machines lacking such evolutionary "baggage".

And what to do now, too, wait millions of years for robots to train the necessary dexterity and ingenuity?

No, there's another option. Mankind has already become adept at creating virtual worlds or meta-universes. And although the famous, large-scale and expensive meta-universe project Horizon Worlds by Meta CEO Mark Zuckerberg seems to have failed, bringing him about $70 billion in losses and legal problems, in general this idea can be extremely useful from a completely unexpected side.

Metaviews (such as the game environments of Fortnite, for example, or Roblox) are of course enjoyed by people, mostly as a means of entertainment.

However, if you can create a digital meta-universe that accurately reflects the physical parameters of the real world, you can launch various AIs (from robotaxi "minds" to humanoid robots) into it so that they can quickly, safely, and relatively inexpensively learn to interact with the real world in it.

A kind of "evolution on steroids", the process is officially called "learning from synthetic data", generated by special AI models to teach other AIs with it.

This is the approach that Nvidia CEO Jensen Huang, for example, has proposed for training robots and robotaxis - his company already offers a special Cosmos World Foundation training meta-universe.

Amazon, which according to its data already employs over 1 million autonomous robots (though not humanoid ones) also uses synthetic data to train them, simulating various scenarios that robots might encounter, especially on peak days like Cyber Monday.

In this light, the arrival of Google DeepMind's new Genie 3 interactive world generator from Google takes on special significance. It's not very powerful yet - it's capable of generating several minutes of interactive 3D environments at 720p resolution and 24 frames per second - but it's a significant improvement over the 10-20 seconds Genie 2 was capable of.

While Genie 3 has value for educational experiences, gaming, or prototyping creative concepts, its true potential will manifest itself in training AI agents to perform universal tasks, and is critical to achieving AGI, Google believes.

"We believe that world models play a key role on the path to AGI, especially for embodied (i.e., acting in the real world, like robots, for example) agents, where modeling real-world scenarios is particularly challenging," TechCrunch quoted Jack Parker-Holder, a researcher with DeepMind's openness group, as saying.

The model represents a step forward in teaching agents to go beyond simply responding to input, allowing them to potentially plan, explore, identify uncertainties, and improve through trial and error - the kind of autonomous, embodied learning that many believe is the key to achieving AGI, TechCrunch summarizes.

Perhaps in the future we will remember the failure of the Ma meta-universe not as a business drama, but as a harbinger of an era in which digital worlds will be needed primarily by machines. And then the question "why build meta universes?" will get an unexpected answer: to teach a billion robots to live among us.

This article was AI-translated and verified by a human editor