Robots in the limelight are actually still a bit far from "going to work."

As we all know, humanoid robots have been a super - popular technology this year. However, I've noticed that the more popular they become, the more controversial they get.

Since the robots danced the yangge during the Spring Festival Gala, the development of this industry has been incredibly rapid. They were just participating in a running event in the robot marathon in April, and by August, they were able to hold a sports meeting.

There have been quite a few robot exhibitions this year. Even WAIC, an exhibition that used to focus on large - scale models, has dedicated a large area to robots this year.

But don't be fooled by the lively scene. Amid the hustle and bustle, another voice has emerged.

At the beginning of this year, Zhu Xiaohu from GSR Ventures bluntly stated, "Humanoid robots are just a bubble."

Many netizens seem to share this view. The reason is quite obvious. In the videos, these robots seem to be capable of extraordinary feats, and people even wish they could become Terminators tomorrow. However, in reality, these steel boys can't even open a door...

Let me first state my stance. I believe that embodied intelligence will definitely achieve great things in the future; it's a sure - win bet. However, to figure out why there are so many controversies, I've dug into the current situation of the industry.

Unexpectedly, I found that some of the problems are not entirely baseless.

Because in the field of embodied intelligence, there are indeed some difficulties. Not only netizens but also many industry insiders haven't reached a consensus on these industry - related issues.

For example, the industry hasn't even unified its technical routes yet. Will reinforcement learning or world models be more powerful in the future? Should we focus more on data or models?... There is no definite answer to these questions, so everyone is going their own way and can't form a united force.

Friends who understand the industry might say at this point that at least the goals of these different routes are the same.

To be honest, that's true. If we only look at the goal, the ultimate value of the entire robot industry is to participate in labor and improve productivity. Just like us humans, "Labor is the most glorious."

However, even if we put aside these differences, there is a more direct and crucial problem facing the industry: lack of data...

I'm not kidding. The entire industry is currently facing the problem of "waiting for rice to cook."

To enable a large - scale model to exhibit intelligence, it requires at least 10 billion to 1 trillion tokens of data, which is more than 10 times the number of model parameters. But currently, most research projects only have a data volume of a few hundred million, and the largest public dataset is only around 1 billion.

As the saying goes, repetition is the father of learning. If the training volume is insufficient, there's no way to improve skills.

As a result, the current robots can perform very few types of tasks, and their generalization ability is extremely poor. In short, without enough data for training, especially data from real - world scenarios, robots are like "captive" babies in the laboratory and are at a loss in the real world.

The data bottleneck has firmly blocked the way for robots to move from the laboratory to factories and homes. At the Bund Summit this month, Wang Xingxing, the founder of Unitree, also pointed out that one of the multiple challenges facing the current development of embodied intelligence is the data problem.

Take the VLA model as an example. The data for interacting with the real world is currently insufficient.

However, I think it's a bit hasty to deny the industry just because of these difficulties. We'd better see if there are any corresponding solutions within the industry.

We found a friend who works on embodied intelligence at Huawei Cloud, and they said, "It's true that there are industry problems, but you can't just focus on this aspect."

What does this mean? To solve the new problems in these new technologies, we definitely need new methodologies and platforms.

For example, using the cloud to systematically solve various problems in the robot industry.

Let's take the most pressing data problem we mentioned earlier.

Since data collection and training in the real world are extremely difficult, can we move these tasks to the cloud? It turns out we can.

In fact, this is also an industry trend. For example, NVIDIA's recently developed Cosmos basic model uses cloud - generated synthetic data to train physical AI.

In China, Huawei Cloud also has a CloudRobo Embodied Intelligence Platform. It can create a digital world in the cloud that is exactly the same as the real world and then generate data for training in this digital world.

This is like opening a training mode in "The Matrix" for robots. They can master all kinds of skills in the virtual space and then go back to the real world to work.

So how do they actually achieve this? It's not that complicated and mainly involves two steps.

The first step is to solve the data problem, that is, to get the "rice."

Behind Huawei Cloud's CloudRobo is a self - developed engine called MetaEngine, which is used for data reconstruction. It can create a digital twin of a real physical scene in the cloud automatically with minimal human intervention.

Then, data augmentation is carried out in this virtual scene. In fact, it means simulating various types of robots in this digital world and generating a large amount of first - person data, such as RGB images, depth data, and time - series data. You name it, and it comes with automatic annotation.

It is said that in the future, for robot training in certain scenarios, by adjusting the ratio of real data to synthetic data, the training efficiency can be improved, and the problem of "lack of rice to cook" can basically be solved.

Previously, Wang He, the founder of Galaxy Universal, even said that synthetic data will account for the majority of the training data, and it's not something ordinary people can do. It requires long - term accumulation and core technical know - how from manufacturers.

The second step is to solve the training and operation problems, that is, to let the robots learn to work.

CloudRobo's training platform allows robots to conduct countless "virtual labor" through imitation learning in this virtual world, which can significantly reduce the cost of trial and error and accelerate skill learning.

This combination is actually a very cutting - edge idea.

In the past, to train a robot model, people had to input data about how the robot should move. Some data collection even required motion capture, just like in "Real Steel." You move, and it records the data, and then the robot learns repeatedly.

But if all these are moved to the virtual world, it will be very convenient because cloud - based training is completely determined by computing power and electricity. With the support of current cloud providers, while you're struggling to set up the environment for a whole day outside, they may have already trained the model for two and a half years inside, just like the Lakers at 4 am every day.

Moreover, after the robots learn in this virtual world, its operation platform can seamlessly connect to physical robots and directly transmit the learned knowledge to the robot's "brain." Once powered on, the robot can sing, dance, and work. This is why many large companies are exploring this direction.

Actions speak louder than words. Previously, Huawei Cloud demonstrated a dual - arm robot trained by CloudRobo performing high - precision operations in a small optical splitter box with a success rate of over 90%. It can also enable Effort's industrial spraying arm to quickly learn to spray new parts and let Leju's humanoid robot handle material handling and feeding on the automobile production line.

So I think there is great potential in solving data and training problems on the cloud.

In addition to data simulation training on the cloud platform, there are many complex problems in the robot industry, and now there are also cloud - based solutions.

For example, another problem in the industry is the chaos of industry standards. Similar to early mobile phone manufacturers like Nokia, Motorola, and Ericsson, robot manufacturers have different systems and charging ports. In this situation, it's impossible to achieve large - scale multi - machine collaboration like iOS, Android, and HarmonyOS.

Therefore, a unified protocol is needed to enable them to understand each other. The guy from Huawei Cloud said they have a solution called the R2C (Robot to Cloud) Protocol. This is like the "Type - C" interface in the robot world. Its main purpose is to integrate the ecosystem and promote industry standardization.

As long as partners pre - install the R2C interface, they can achieve "plug - and - play." It's like when you buy a new mouse. Whether it's a Windows or Mac computer, you can use it as soon as you plug it into the USB port without having to look for the driver disk everywhere.

Leading players in various fields, such as the National and Local Joint Humanoid Robot Innovation Center, Topsun, and YouiBot, have all joined the R2C Protocol. It's like when someone shouts and all the sects respond. Everyone has started to get on board.

To be honest, I'm not afraid of offending anyone. Although cloud - based solutions are good, they are definitely not a panacea and can't be used in all situations.

Think about scenarios with extremely high requirements for real - time performance and security. They may still prefer local computing, and we should understand this.

In fact, although we often talk about moving to the cloud, the real value of cloud computing lies in more complex scenarios. Tasks such as complex scenario recognition, task planning, and model invocation that consume the most computing power are likely to be handed over to the cloud in the future. The robot itself will focus more on execution, becoming lighter and cheaper.

From another perspective, you don't want your robot to carry a big computer on its back, right? It's just inconvenient. So in this way, it provides a possible path for robots to move from the laboratory to factories and into families. This is what is called cloud - based robot ontology.

Previously, at the end of his speech at the WRC, Wang Xingxing also mentioned that it's impossible to directly deploy a large - scale computing power on the humanoid robot's body. So in the future, this problem will definitely have to be solved by distributed cluster computing power, which actually uses cloud - based computing power.

Just these days, Huawei has also made new arrangements in AI computing power. It has released the latest super - node products, Atlas 950 SuperPoD and Atlas 9