I went to see 300 robots at work on National Day, but they were actually taking exams

In the past few days of National Day, a group of special "employees" have been added to the Changlong Spacecraft Park in Hengqin, Zhuhai.

According to the information released before Zhiyuan opened the park, on September 24,more than 300 robots worked collectively here.More than 100 robot interaction points were set up in the park, covering seven types of scenes: performances, popular science studies, guided tours, retail services, escorts, hotels, and sports competitions.

The report did not mention that there was an additional charge for the robot project, and the tickets were still the original price.

When most people see this news, their first reaction is: Are robots entering the park to perform and attract customers?

This is certainly part of it. But if you just look at it as a performance, you will take this matter lightly.

For robot companies, this is more like an examination room.

!

1. What did the official say?

This judgment is not my guess, it is in the official wording.

When the two sides signed a cooperation agreement in August, the official said they would jointly build a joint laboratory to"take root in multiple parks in Chimelong to carry out technical research and scenario verification." On the day of the opening of the park, a specific joint research institute for intelligent culture and tourism was unveiled simultaneously.

There is another saying to put it more bluntly: the rich scenes and huge passenger flow make Hengqin Changlong an ideal scene for robots to enter the real world and achieve large-scale application.

The four words "scene verification" are translated into plain Chinese: Take it into a real environment to take a test.

Why do I have to take the exam in a place like this? This has to start with what the robot industry lacks most.

2. What robots lack most is real-world data

The big model can write articles and chat, relying on mountains of words on the Internet, with a scale of trillions.

Robots do not have this benefit. What it has to learn is to do it in the real physical world: grab a cup, avoid a person, find its way in crowded places. This kind of data cannot be extracted from the Internet and can only be created in a real environment time and time again.

According to the industry, high-quality robot real operation data is currently on the order of millions of hours worldwide. Compared with text data, it is several orders of magnitude worse.

There are three methods to collect data, each of which has its own flaws:

!

One is to simulate it in a computer. It is cheap and large in volume, but there is always a gap between the simulated physical world and the real world. It is called the "virtual and real gap" in the industry.

The second is to film how people work and let robots learn. However, human hands and robot hands look different and may not be useful when learned. This is called the "configuration gap."

Third, people remotely control the real robot to do it and record the process. The highest quality, but the most expensive. According to the domestic market price in the report, this data costs500 to 1000 yuan per hour.

So the entire industry is stuck in a circle:

!

Without large-scale commercial use, there will be no high-quality real data; without data, the robot's generalization ability will not improve; if the generalization ability will not improve, there will be fewer scenarios that can be entered; if there are few scenarios, there will be no large-scale commercial use.

To pry open this circle, there must be a place that is willing to let a large number of robots in, operate continuously every day, and face a large number of real people.

3. What is in the park that cannot be built in the laboratory

There are already many specialized data collection bases, so why do you still need to enter the park?

Because most of what is done in the collection base is structured tasks: place things on the table, grab, place, and fold them over and over again. The environment is controllable.

Paradise is the opposite.

!

The light is changing. The indoor exhibition area, light show, and daylight by the window change several times a day. The flow of people is changing, sparse when opening the door, and crowded with people on holiday afternoons. The interaction is random. No one talks to it according to the script. Children will suddenly rush to touch it and ask strange questions. Moreover, you have to run continuously, from opening the door to closing the door, running every day. The five-minute demonstration does not count.

Some media comments have said a similar meaning: this kind of open scene with dense crowds, variable lighting, and random interactions will continue to generate a large amount of real-world data and can feed back model iteration.

I want to make a boundary clear here: the official did not say that the data from the park would be used to train models. There is an article on the Internet that is very specific, saying that every time a jam occurs, it will send back and drive algorithm iteration in real time. However, this does not appear in the official materials of Zhiyuan and Changlong, so it can only be regarded as speculation.

What can be determined is the direction. A batch of data sets that Zhiyuan opened source in June, and the official statement was100% based on real-life scenarios. We can see how much this company values real scenes.

4. My judgment: In this business, tickets are not the focus

Putting the above several things together, my understanding is as follows:

For a robot company, the most valuable part of this kind of cooperation is a real scene where there are a large number of real people every day and hundreds of robots can run continuously. Of course, ticket revenue and publicity effects are there, but that is the park's account.

The scale is also worth a look. According to public reports, Zhiyuan announced in December last year that the 5, 000th universal embodied robot was delivered offline. More than 300 units are concentrated and operated continuously in one park, which is a considerable centralized deployment for a company.

A robot in the laboratory performs well, which means that it can do it; only when hundreds of robots run down day by day in the crowd of people on holidays shows that it can withstand it.

The distance between these two things is the step the entire industry is taking now.

5. If you go to see it on National Day, you can see these three "subjects"

If you happen to be there in the past few days, in addition to taking photos, you may wish to pay attention to three things. They happen to be what this exam is testing:

!

First, how to avoid it when there are many people? Should we flexibly circumvent it, or should we just stop where we are and wait for people to leave? Is the performance the same during peak hours and when there are few people?

Second, if you don't understand it or do it wrong, how will it end? Any robot will make mistakes. The difference lies in the aftermath of the mistake: whether it gets stuck or can round the scene back on its own.

Third, whether there is any staff following it? Someone is following, either a security officer or helping it remotely. This ratio explains to some extent how far it is now from truly "independent employment".

Last

In the past few years, the impression that humanoid robots have given to most people is a video from a press conference: somersaults, dancing, and carrying boxes.

This time is different. They were put in a place where thousands of ordinary people come and go every day, and they had to stay there for a long time.

Being able to somersault is one thing, but being able to stand in the crowd all day on National Day is another. The latter ability can only be tested in such a place.

few sentences boundary

Statements such as the opening time, the number of robots, the number of interaction points, seven types of scenarios, the unveiling of the research institute, and "technical research and scenario verification" come from the information released by Zhiyuan officials before the opening of the park and reprinted by the media; I have not yet seen it at the time of writing. Official operating data after the opening of the park.

The statement that "real-time data transmission drives algorithm iteration" has not been found in official materials, and it has been clearly stated in the article that it is not a fact.

The scale of real operating data (million-hours), the shortcomings of the three collection methods, and the market price of 500 to 1000 yuan per hour are the industry reporting standards.

"Tickets are not the focus" and "This is an exam" are my judgments and do not mean that the company has expressed its position. This article does not constitute any play recommendations or investment advice.