Skip to content
Just as the 1960s saw a space race between the U.S. and the former Soviet Union to see who would reach the Moon first, a technological battle is now underway to see who will launch the first humanoid capable of being sold on a large scale without the need for programming.
Currently, the main challenge in humanoid robotics is obtaining the data necessary for physical AI to successfully perform the tasks at hand. In other words, for a robot to perform an action that is simple for a human—such as folding a garment or sweeping—it needs, on the one hand, precise data on how to carry it out and, on the other, to have successfully completed it a minimum number of times.
To illustrate this point: the number of iterations required to learn how to fold a single cloth can reach one million attempts. There are several ways to streamline this process: you can use simulators, watch videos, or utilize data from other models performing the same task. Finally, it is essential to conduct iterations with the physical hardware itself to evaluate its behavior in a real-world environment.
In the past, getting a humanoid robot to fold a cloth was an impossible task. You had to program the first fold line by line and then iterate until it was useful—and all that for just one type of cloth. If you wanted to add different types of garments, the process required years of work, resulting in extreme inflexibility and zero profitability.
All of this has changed radically with the introduction of physical AI. Now there’s no need to write line after line of code; the robot has the ability to learn and perform the tasks we assign it if it has processed them beforehand. However, although this will eventually become a reality, it is not yet possible today. Manufacturers are clear on the reason: a lack of data.
How to obtain that information is, in fact, the very point on which the two major technology blocs cannot agree. This has led them to adopt virtually opposing strategies, which is fascinating.
The American version
If we look at U.S. companies in the field of humanoid robotics, the main focus is on Figure AI, Tesla, and Boston Dynamics. It should be noted that, although Boston Dynamics is a U.S. company, it is currently majority-owned by the South Korean corporation Hyundai.
These companies do not yet market their product on a large scale or offer a closed-loop model. They do not do so because they still need to program the robot to make it truly “useful.” For now, all of them have focused on designing humanoids for the automotive manufacturing sector, and the first units are already operating in real-world tests: Figure AI is working with BMW, Tesla is using them in its own vehicle plants, and Boston Dynamics is deploying them in Hyundai’s factories (as well as in its spectacular testing center built specifically for them).
Although American brands prioritize the automotive industry, the real revolution will come when robots can live alongside us and perform household chores. These companies’ target price has always been around €30,000–40,000, a figure their executives have publicly mentioned on more than one occasion. If they aren’t in the mass market yet, it’s because they still can’t meet those production costs. But is this the best strategy?
The mechanical aspects are considered to be virtually resolved, and although improvements in load capacity may arise, the real bottleneck is data collection. Figure AI has invested more than 35 million euros in its Index program, capturing data from real people performing everyday tasks to extrapolate it to its physical AI. This approach presents certain challenges, but its engineers have determined that the value of the information justifies the investment.
Visitors to Figure AI's offices can see the robots operating on the street almost every day. Their ease of movement and fluidity in uncontrolled environments are remarkable. However, the reality is that their data remains limited, since these robots only go outside when escorted by company staff.
In short: U.S. tech companies have technically very mature products, but since their current production capacity is limited, they focus on refining the robot’s technical feasibility rather than prioritizing the massive collection of data through third parties.
The Chinese version
At the other end of the spectrum is China, whose companies manufacture and supply the vast majority of the world’s humanoids. According to official data from 2026, Chinese firms account for more than 90% of the global market for humanoid robots.
Agibot and Unitree are the two leading Asian manufacturers with models currently on the market. Between them, these two companies alone shipped more than 20,000 units in 2026, and exponential growth is projected in the coming years. In fact, both companies are building several megafactories dedicated exclusively to the production of humanoids.
From a software perspective, these robots lag behind the physical AI developed by companies like Figure AI. However, they have a massive competitive advantage: a vast amount of data collected directly from real-world scenarios. What once began as “toys” with clunky movements has become the perfect vehicle for extracting high-quality data through sensors in real-world work environments.
This vast database, combined with its production capacity and the fact that the Chinese government is one of the main shareholders in these companies, creates a combination that is extremely difficult to beat. Furthermore, due to the unique nature of the Chinese state system, both the data collected and the physical AI systems belong in part to the government, which allows it to standardize processes and make strategic decisions about with whom to share the information.
Conclusion
Regardless of the nationality of the expert you consult, the conclusion is unanimous: humanoids are going to change the world, and we are witnessing a truly disruptive era that will make generative AI seem like child's play.
The ideal product has already been defined, and the industry agrees on the specifications: a cost of between €20,000 and €40,000, full autonomy, and capabilities similar to those of an average human.
The next step is the one that divides opinion: Is it better to launch a fully finished and polished product, or is it preferable to release a less mature model so it can reach technical maturity based on data collected in real-world environments? We are faced with two opposing approaches to selling the product that will transform our society. Who will win this race?