This National Day, the first batch of robot employees have taken up their posts

Text | Vertical Youth, author｜Jieshi, editor｜Shanghai

During the National Day holiday, a group of employees who do not have to take off work are shining at their jobs. Some of them are mascots, and some are real workers.

Nowadays, whether it is DQ ice cream stores, FamilyMart convenience stores, Bawang Chaji stores, or 4S stores of BYD, Xpeng, and Li Auto, you may encounter robot employees at any time and enjoy the services they provide.

Humanoid robots replace real employees and rely on the three major capabilities of recognition, processing, and execution. However, there is still a gap between robots and real people in terms of work efficiency and flexibility.

To this end, companies such as JD.com and MiFeng Technology are promoting real-person data collection services, while some migrant workers are facing a possible employment crisis and have launched a “non-violent non-cooperation” movement.

From a cost-benefit perspective, it seems too early to worry about the loss of workers; however, GPT-6 Astra’s effective control of robots has brought embodied intelligence manufacturers and large AI model companies on a narrow path. And a Golden Week holiday has become the main battlefield for testing these capabilities.

What kind of work can humanoid robots do?

Liu Hui traveled to Shanghai during the National Day holiday, and a DQ ice cream store on Nanjing West Road gave her the deepest impression of the trip.

This is because this DQ store has only one employee, and the employee is not an ordinary worker, but a humanoid robot.

As the only one in the store, this humanoid robot from Sharpa can make DQ’s signature: blizzard ice cream on site for customers just like other real employees.

Following the store’s operating instructions, Liu Hui placed an order through the DQ applet and asked the robot to make a cup of Oreo Blizzard.

After starting work, the robot first took out a paper cup, then operated the ice cream machine to fill the cup with milk syrup, then added crushed Oreos, stirred the ice cream in the cup, and finally handed the blizzard cup to Liu Hui.

After one operation, the robot has performed a total of 55 steps, and the specific operation process of each step will be displayed on the store’s large screen.

It is worth mentioning that after the robot has finished making a cup of Blizzard ice cream, it will perform a unique trick in front of everyone: turning the entire cup of Blizzard ice cream upside down, showing off its arm strength to steadily pour the cup without spilling it.

At this time, the robot will not only cheer itself up: “Next, it’s time to witness the miracle.” After it is completed, it will also give the customer a comparison of their feelings or six-six-six, which really adds to the emotional value.

However, compared with ordinary human employees, the speed at which the robot makes ice cream seems to be slowing down, which reminds people of the “Lightning” in “Zootopia”.

On the other side, some netizens observed the scene and said: “The robot first spilled a cup, and then knocked over another cup. It took the same time for a human to make about 10 cups.”

Although the robots in DQ stores move slowly, Zhang Hui has broadened her thinking and cognition. She works in Zhongguancun, Haidian, Beijing. She had always thought that if humanoid robots wanted to work, the most they could do was go to convenience stores to deliver things to customers.

Six months ago, in an office building in Zhongguancun, Haidian District, Beijing, a FamilyMart convenience store hired a new employee. This humanoid robot named “Galbot” is a product manufactured by Galaxy General.

After the robot takes office, every time Liu Hui walks into a FamilyMart convenience store, “Galbot” will take the initiative to say hello to her; if she cannot find the location of the product for a while, she can also ask the robot for answers.

In addition, every time Liu Hui needs to pick up cooked food or make coffee, she can also send voice instructions to the robot. The robot will say while working: “Although I only have two fingers, I have to calculate the best angle every time I grab it.”

Of course, like the DQ robot mentioned earlier, FamilyMart’s “Galbot” is also inefficient. Liu Hui once wanted to eat corn after working overtime at night, and had to wait for more than a minute for the robot to get it. She almost fainted in front of the counter from hunger.

Nowadays, more and more humanoid robots are beginning to appear in offline consumption scenarios.

Not long ago, Bawang Chaji cooperated with Honor to launch the “Drink Good Tea Encounter Robot Parade” activity. Consumers found in some Bawang Chaji stores that the clerks wearing aprons were not real people, but men of steel.

On the other hand, automobile manufacturers such as Xiaopeng and BYD have also developed their own humanoid robots and arranged them to work in automobile 4S stores to answer questions for people who come to look at the cars.

Real person data is very important

Why can humanoid robots from various manufacturers replace real employees? From the perspective of implementation, it can be roughly divided into three aspects:

First of all, humanoid robots need to observe and perceive information about the surrounding environment through cameras, lidar and other equipment, which tests its recognition capabilities.

Secondly, the robot must use its “brain” to calculate the corresponding behavior after acquiring surrounding information, which tests its processing capabilities.

Finally, when everything is deduced in the mind, the robot still needs to use its legs, mechanical arms and other components to complete a series of prescribed actions, which tests its execution ability.

Relying on the above three abilities, humanoid robots can replace front-line service personnel in certain scenarios; at the same time, some robots continue to improve their abilities and have the possibility of becoming higher-level workers.

For example, a research team from the University of California, San Diego, used two Yushu G1 humanoid robots to perform surgery and successfully performed live cholecystectomy on two large primates.

On September 10, 2026, Yushu also open sourced a basic humanoid robot model, which can support the robot to complete 64 tasks, including folding towels, storing dinner plates, packaging mobile phones, taking out garbage, etc.

Although the technology tree of humanoid robots is becoming more and more leafy, the work performance of many robots is still much inferior to that of real people.

In order to make humanoid robots as flexible and coordinated as real people at work, some companies have begun to collect real-person movement data to guide humanoid robots how to work.

JD.com announced in March 2026 that it will build an embodied intelligent data collection center in Suqian, planning to collect 10 million hours of high-quality data within two years, and plans to mobilize 100,000 employees of the company and 500,000 external people to participate.

Recently, front-line employees such as cleaning, gardeners, and maintenance staff at JD.com’s campus have begun to wear JD.com’s self-developed motion capture equipment to record work movements, tool usage and other information from a first-person perspective.

On the other side, on August 25, 2026, the robot company Figure launched the Index project to train humanoid robots by uploading videos of actions related to work and life by ordinary people.

At present, the project has more than 1 million registered users, a total of more than 16 million videos have been uploaded, and more than 15 million US dollars have been distributed in rewards. The company also proposed that in the next 12 months, it will strive to expand the collection scale by approximately 100 times.

On September 23, MiFeng Technology released the “Chinese version of Index”: MiFeng Pai. In one month of internal testing, it has gained more than 20,000 registered users and completed 13,000 valid task submissions.

Some migrant workers take the initiative to upload their own movement data in order to obtain rewards; some employees realize that humanoid robots may replace themselves at work, and therefore say “no” to data collection.

Previously, Tesla required employees at some factories in the United States to wear motion capture suits to participate in assembly line work, collecting hand grasping, tool operation, parts handling and other movements for training its humanoid robot: Optimus.

In the factory, some workers refused to wear motion capture suits equipped with sensors and used this behavior similar to “non-violent non-cooperation” to express their attitude towards humanoid robots.

Not only that, South Korea’s Hyundai Motor Union also stated that the introduction of robots into production is a robbery of employment, and made it clear that the company is not allowed to introduce automation technology, including robots, without the permission of the labor agreement.

Who will control the robot brain?

In fact, workers’ worries about the possibility of humanoid robots replacing themselves are, to some extent, unfounded.

At present, if humanoid robots want to replace real people, in addition to the constraints of efficiency factors, there is also an economic account to be settled.

For example, in FamilyMart convenience stores, the basic version of Galbot G1, Galaxy’s universal robot, costs 699,000 yuan; while the minimum salary for employees at FamilyMart convenience stores in Beijing is 4,500 yuan a month.

It can be calculated from this that the cost of purchasing a Galbot G1 is equivalent to 155 months’ salary of a real employee, excluding electricity bills, losses and other expenses.

According to the staffing configuration of a convenience store working three shifts, the cost of a Galaxy General Galbot G1 is equivalent to the salary of three real employees working for 51 months.

Although humanoid robots still have a long way to go in replacing real front-line employees, the sales growth of humanoid robots on the market is very gratifying.

According to the “Global Humanoid Robot Tracking Report” released by IDC, in the first half of 2026, global humanoid robot shipments were close to 25,000 units, a year-on-year increase of 432%; the humanoid robot market size exceeded US$740 million, a year-on-year increase of 323%.

At the same time, more and more offline consumer brands are beginning to use humanoid robot employees as marketing tools to promote themselves. Even if the robots do not do any work, they can still attract consumers to stop by them.

The rapid growth of the incremental market has given humanoid robot manufacturers a promising future, but a recent technological advancement in the field of AI has put pressure on these manufacturers.

On September 3, 2026, OpenAI released GPT-6 Astra; the next day, after Robocurve connected it to a humanoid robot, it was found that it performed outstandingly in controlling the robot.

In RoboDojo’s official evaluation, GPT-6 Astra achieved an average success rate of 22.48% in 42 robot tasks and 2100 tests, and ranked first on the list with a score of 28.97 points.

In addition to OpenAI’s GPT-6 Astra, some other large AI model manufacturers are also designing brains for robots.

To this end, Qwen-Robot developed by Alibaba Qianwen allows the robot to have hands, feet and a brain; Bytedance learns physical laws through massive videos and studies its own model against Google Genie.

At the same time, large AI model companies such as Zhipu, DeepSeek, and MiniMax are also developing their own VLA (visual, language, and action) models and supporting robot manufacturers to access them.

Big AI model companies are trying to control robot brains, while embodied intelligence manufacturers have their own ideas.

The Genie Reasoner released by Zhiyuan allows the brain, which is responsible for the robot’s IQ, to coordinate with the cerebellum, which is responsible for the manual ability.

In June 2026, Zhiping released the embodied intelligence system: NeuroVLA, which replicates the three-layer architecture of the human brain’s “cortex-cerebellum-spinal cord”.

It can be seen that regarding who should control the brain of humanoid robots, large AI model manufacturers and embodied intelligence companies are divided into two groups, each of which wants to dominate the industry.

Maybe the next time Liu Hui walks into another offline store, the humanoid robot will recognize her in a second because they share the same brain; or it may always treat her as a stranger because their brains are incompatible.