As we all know, humanoid robots are the technology that has been popular this year. But I found that the more popular this thing is, the more controversial it is.

Because since the Spring Festival Gala robot dances yangko, this industry has developed really fast. The robot marathon is running in April, and the sports meeting will be held in August.

There are many various robot exhibitions this year, and even WAIC, an exhibition that has held large models in previous years, has also vacated a lot of land to put robots this year.

But don’t look at the scene being lively, another sound will appear under the noise.

Zhu Xiaohu of Jinshajiang Venture Capital said bluntly: "Humanoid robots are bubbles."

Many netizens actually have this meaning, and the reason is obvious. The robots on the video seem to be able to go to heaven, and they are eager to become the terminator tomorrow. However, in reality, these steel boys can't even open the door. . .

Let me talk about my position first. I think that embodied intelligence will definitely make great achievements in the future and win. But in order to figure out why there are so many controversies here, I also went to dig out the current situation of the industry.

Unexpectedly, I found that some problems were not entirely nonsense.

Because there are indeed some difficulties in the embodied intelligence industry. Let alone netizens, many industry insiders have not reached a consensus on these industry issues.

For example, this industry has not even unified its technical route yet. Will reinforcement learning be awesome in the future or the world model be awesome in the future? Should we pay more attention to data or models? ...These issues are not concluded, and everyone can only do their own things and cannot form a joint force.

Friends who are familiar with this may say it when they see this. No matter what, the purpose of these routes is at least the same.

To be honest, indeed. If we only look at the purpose, then the ultimate value of the entire robot industry is to participate in labor and improve productivity. Like us humans, "labor is the most glorious."

But the problem is that even if we put aside these differences, there is another more direct and fatal thing in front of the industry: there is no data. . .

This is not nonsense to my buddy. Now the entire industry is facing the problem of "waiting for rice to be put into the pot".

Because if you want a big model to emerge intelligently, you need at least 10 billion to 1 trillion token data, which is almost 10 times the model parameters. But what about now? Most of the research studies have only a few hundred million data volumes, and the largest public data set is only about 1 billion.

As the saying goes, repetition is to learn from his father. If you don’t have enough training, you will definitely not be able to increase your skills.

So this leads to the fact that today's robots have pitifully fewer tasks and their generalization is ridiculously poor. To put it bluntly, there is not enough data to train, especially data in real scenes, which leads to the robots being fat babies "carried" in the laboratory, and they will be blinded when they come to reality.

The data bottleneck has stuck the way the robot walked from the laboratory to the factory and the house. At this month's Bund Conference, Yushu founder Wang Xingxing also said this. Among the multiple challenges facing the development of embodied intelligence, one of them is data issues.

Take the VLA model for example, the data that currently interacts with the real world is not enough.

However, I feel a little anxious to deny the industry because of difficulties. We should check if there are any corresponding solutions in the industry.

We found a friend who was working on embodied intelligence at Huawei Cloud, and they said, "The industry problem is correct, but you can't just see this layer."

What does it mean? To solve new problems in these new technologies, there must be a new methodology and platform.

For example, to go to the cloud, use cloud-based methods to systematically solve various problems in the robot industry.

Take the data problem that we mentioned earlier, which is the first to bear.

Since data collection and training are very difficult in reality, can we move these things to the cloud? It's really OK.

In fact, this is also an industry trend. For example, the basic model that Nvidia recently developed is to use the cloud to generate synthetic data to train physical AI.

Including in China, Huawei Cloud also has an embodied intelligent platform. It can also create a digital world in the cloud, exactly the same as reality, and then generate data in it for training.

This is like opening a training mode for a robot in "The Matrix", proficiency in all eighteen martial arts in the virtual space, and then returning to the real world to work.

So how do they do it specifically? It's not complicated, it's mainly divided into two steps.

The first step is to solve the data problem, that is, you must have rice first.

For example, Huawei Cloud relies on a self-developed engine to carry out data reconstruction and replicate a real physical scene in the cloud. The entire process is low in manual and automatically completed.

Then, data augmentation is carried out in this virtual scene. In fact, it is in this digital world that simulates robots of various forms to generate massive first-view data, such as RGB images, depth, time series data, and everything they need, and they are also marked with automatic labeling.

It is said that in the future, robot training in certain scenarios, the ratio of real data and synthetic data can be adjusted to improve training efficiency, which will basically solve the problem of "no rice to be put into the pot".

Wang He, the founder of Galaxy General Motors, even said that synthetic data will account for the vast majority of training data, and ordinary people cannot do it. Manufacturers need to have long-term accumulation and core technology know-how.

Then the second step is to solve the training and operation problems, that is, let the robot learn to do work.

The training platform allows robots to perform countless "virtual labor" through imitation and learning in this virtual world, which can greatly reduce trial and error costs and accelerate skill learning.

These two are actually a very cutting-edge idea.

In the past, if you wanted to practice a robot model, how the robot moved, it would require people to feed the data. Some data collection had to be replenished. This was like an iron armored steel fist. You recorded the data while moving it, and then learned it repeatedly.

But it would be very convenient to move this into the virtual world, because cloud training is entirely determined by computing power and power. As for the guarantee of these cloud manufacturers, you may have been training outside for two and a half years, and every day is like Los Angeles at 4 a.m.

Moreover, after they learn it here, its operating platform can seamlessly connect to physical robots and directly pass it into the machine's brain. You can sing, dance and work when you turn on the computer, so this is why many major manufacturers are thinking about this direction.

Just talk but not practice fake moves. Previously, Huawei Cloud showed a graduated two-arm robot on the spot, performing high-precision operations in a small spectrometer box, with a success rate of more than 90%. It can also allow Eft's industrial spraying arms to quickly learn to spray new parts; it can allow Leju's humanoid robots to carry and load materials on the automobile production line.

So I think it is actually very promising to solve data and training on the cloud.

But in addition to data simulation training on cloud platforms, there are still many complex problems in the robot industry, and now there are solutions to solve them by relying on the cloud.

For example, there is still a problem in this industry. The industry standards are chaotic. The robot manufacturers are like early mobile phone manufacturers, such as Nokia, Motorola, Ericsson, and the systems and charging ports. Then this situation will definitely not be as large-scale multi-machine collaboration as iOS and Android Hongmeng.

So a unified agreement is needed so that they can understand each other. Huawei Cloud Brother said they have a solution called the R2C (Robot to Cloud) protocol. This is similar to the "Type-C" interface in the robot world. The main purpose is to aggregate everyone's ecology and promote industry standardization.

As long as it is a partner with pre-installed R2C interface, it can achieve "plug and play". It's like buying a new mouse, whether it's a Mac, it can be used by plugging in a USB port, and you don't have to look for the driver disk everywhere.

Leading players in various fields such as the National and Local Co-construction Humanoid Robot Innovation Center, Tosda, and Youai Zhihe have all joined the R2C protocol. I felt like I was raising my arms and shouting, and all the major sects responded one after another, and everyone started to board the boat.

But to be honest, we are not afraid of offending others. Even though it is good to go to the cloud, it is definitely not a cure all diseases and can be used at any time.

If you think about those scenarios that require extremely high real-time and security, people may still want to do local computing, so we have to understand this.

In fact, although we often talk about cloud access, the real value of cloud computing actually lies in those more complex scenarios. In fact, the most computing-intensive scenario recognition, task planning, and model calling are likely to be handed over to the cloud in the future, and the robot body will focus more on execution, becoming lighter and cheaper.

From another perspective, you don’t want your robot to carry a big computer as a whole, right? I’m going to put it bluntly. So in this way, it can provide a possible path for robots to walk from the laboratory to the factory and into the home, which is the so-called cloud ontology.

At the end of his WRC speech, Wang Xingxing also talked about the humanoid robot body. In fact, there is no way to directly deploy large-scale computing power, so this thing will definitely be solved by distributed cluster computing power in the future. This is actually using cloud computing power to solve the problem.

In the past few days, Huawei has also made a new layout in AI computing power: it has released the latest super node products, Atlas 950 and Atlas 960 super nodes, supporting 8192 and 15488 Ascend cards respectively, and is in the lead in key indicators such as card scale, final computing power, memory capacity, and interconnection bandwidth. It will be the world's strongest super node with the strongest computing power in the future years.

Based on super nodes, Huawei also released the world's strongest super node cluster, with the largest computing power scale of one million cards. As the world's strongest computing power cluster, it will provide stable and surging computing power support for innovative breakthroughs in the embodied intelligent industry.

Huawei All Connection Conference 2025: Huawei Vice Chairman and Rotary Chairman Xu Zhijun delivered a keynote speech

After all, whether it is cloud or local, these technical means are here to solve our problems. Let the robots work early and participate in social practice, which is a good means.

In short, those who say that robots are hype are not all wrong, and they do point out the existing difficulties in the industry.

But seeing only difficulties is just like seeing only the tip of the iceberg, and it is easy to draw pessimistic conclusions.

The hustle and bustle of all technology will eventually pass, and these human creations will eventually go from school to society, verifying their value in tightened screws, transported materials, and welding gaps.

In a word, embodied intelligence is too complicated. One or even a few people running in front still need to stand up and not do robots, but build infrastructure first.

In fact, it is true that you are pragmatic. Instead of arguing on the shore about whether you can walk this path, it is better to build the road.

After all, only when the road is wider can everyone run faster and farther.