Before robots are put into large-scale jobs, a large group of "teachers" are needed first: what reminders does this new occupation bring to the employment sector?

While many people worry that robots will take their jobs, a group of young people have already stood beside the robots.

They put on VR glasses, hold the control handles, and repeatedly demonstrate actions like grabbing bottles, folding clothes, tidying dining tables, and sorting goods. A seemingly simple action may need to be repeated 300 to 1000 times; a 5-minute task will be split into 7000 to 8000 frames of data. Before robots can learn to work, humans need to break down labor into "teaching materials" that robots can understand.

Some of these people are called humanoid robot trainers, and their job titles also include data collectors, model trainers, or artificial intelligence trainers.

New occupations are emerging, but what really deserves ordinary people's attention is not replacing a job with a tech-sounding name, but seeing a clear change: in the future, more and more employment opportunities will appear "between humans and machines".

Teaching Robots to Work Is More of a Physical Labor Than You Might Think

At the Hubei Humanoid Robot Innovation Center, trainers control robots to complete actions such as grabbing and placing. The simple task of tidying the dining table and placing food is broken down into 20 specific steps; after data collection is completed, it has to go through cleaning, labeling and review to become a qualified piece of data.

The center already has nearly 100 trainers, producing about 24,000 pieces of data per day with an annual collection volume of about 6 million pieces. The training scenarios cover supermarkets, coffee shops, production lines, warehouses, apartments and laboratories.

The training base in Beijing has deployed more than 120 robots, which are trained in more than 30 scenarios across six major fields: home, supermarket, office, industry, medical care, and elderly care. The base has a daily data output capacity of over 500 hours, with a data validity rate of over 95% and an annual output of more than 10 million pieces of data.

It sounds like participating in cutting-edge R&D, but the actual work is often repeated over and over: raising hands, rotating wrists, grabbing, moving, and putting down. If the movement is not smooth, the force is not appropriate, or the data is incomplete, you have to start all over again.

This occupation has technical content, but not every position is for algorithm engineers.

Judging from public recruitment information, basic data collection positions usually require a college or bachelor's degree, and the work content includes wearing VR or motion capture equipment, controlling robots to complete actions such as grabbing, transporting, and stacking, while recording, screening and cleaning data. Some positions adopt a six-day work system, with a comprehensive monthly salary of about 4,000 to 7,000 yuan; another recruitment notice in Beijing offers a salary of about 5,000 to 6,000 yuan.

This at least illustrates one thing: don't automatically equate positions with "AI" and "robot" in their names with easy, high-paying and glamorous jobs.

The sense of technology belongs to the industry, and wages are ultimately determined by the scarcity of the position.

Why Do We Need More Humans Even When Machines Are Getting Smarter?

Large language models can learn from massive amounts of text, but robots face a more complicated real world.

Even for the same action of picking up a cup, the material, weight, position, light of the cup and surrounding obstacles may all be different. If you grab it too lightly, it will fall; if you grab it too hard, it may break. If the cup on the table is rotated a little, the originally effective action may also fail.

What robots lack is not only "knowledge", but also physical experience.

Through teleoperation, motion capture and real-scene demonstration, trainers record video images, joint angles, movement trajectories, force and tactile feedback. After these data are cleaned, labeled and trained, robots can gradually learn to transfer one action to different environments.

This also explains why the more the robot industry advances into real scenarios, the greater the demand for data. Industry practices show that the data required to train high-quality embodied large models may reach tens of millions or even hundreds of millions of hours; for a specific industrial action, tens of thousands of pieces of real-scene data may only train the task success rate to 80% to 90%. To further approach stable availability, more data needs to be supplemented.

Commercial problems have emerged accordingly.

If collecting one skill requires a large amount of equipment, venues and personnel, and each robot configuration and each type of application environment are not completely the same, then data may become the most expensive "invisible component" in the embodied intelligence industry.

Therefore, the emergence of trainers is no accident. The industry is shifting from "manufacturing robots" to "making robots work stably", and employment opportunities are also extending from hardware manufacturing to data production, scenario design, testing and evaluation, operation and maintenance.

Essentially, the humanoid robot trainer is the result of this industrial chain starting to make up for its shortcomings.

New Occupations Also Have Their Own "Skill Hierarchy"

When discussing such positions, the most common mistake is to regard all trainers as doing the same job.

In fact, it can be divided into at least three levels.

The first level is the data collection execution post. The staff operates the equipment according to the specified actions to complete data recording, preliminary screening and uploading. The entry threshold for this level is relatively low, but the work is highly repetitive, and it is most easily affected by process standardization and automatic collection technology.

The second level is the data quality inspection and scenario design post. It requires the staff to know what data is valid, why it fails, and how to adjust the position of objects, light, task difficulty and abnormal situations. Humans are no longer just repeating actions, but "setting questions" for robots.

The third level is the model training and deployment post. It requires understanding computer vision, reinforcement learning, control algorithms, sensors and specific industry processes, and being able to connect data, models and real tasks. Such positions have a higher entry threshold, but are closer to the core of industrial value.

The Ministry of Human Resources and Social Security has long issued the National Vocational Skill Standard for Artificial Intelligence Trainers, whose professional competence ranges from junior workers to senior technicians, and includes data collection, data labeling, intelligent system operation and maintenance, and business analysis into the competence system.

Therefore, for people who are preparing to enter this track, what really matters is not to grab the title of "robot trainer", but to judge which level they are at and whether they have the ability to move upward.

If after many years, the work content is still just repeating actions according to the script, then the faster the machine learns, the more likely the basic positions will be the first to face wage cuts. On the contrary, people who can find data defects, design training tasks, understand industry processes and solve deployment problems have more opportunities to share the value brought by industrial growth.

In one sentence, being a teacher for robots is divided into two types: some are responsible for copying teaching materials, while others are responsible for designing courses.

Robots Create Positions and Rewrite Existing Positions

China is already the world's largest industrial robot market. Data from the International Federation of Robotics shows that 295,000 industrial robots were installed in China in 2024, accounting for 54% of the global newly installed volume; the number of industrial robots in service exceeds 2 million units.

Such a large installed base of robots means that the impact on employment will not stay in the laboratory.

On the one hand, robots will replace some highly repetitive, dangerous or extremely precision-demanding tasks; on the other hand, they will also bring new jobs such as equipment debugging, data training, on-site operation and maintenance, system integration, safety evaluation and human-machine collaboration. The International Federation of Robotics summarizes relevant studies as that the mechanism of robots affecting employment is not simple replacement, but coexistence of task replacement, productivity improvement and re-emergence of new tasks.

However, there is no simple addition or subtraction that the number of new positions can directly offset the number of old positions between new and old jobs.

The positions lost by assembly line workers may not be turned into robot operation and maintenance positions in the same city or the same company; people who are skilled in assembling products will not automatically acquire the ability to debug sensors and analyze data. New occupations are indeed increasing, but skill mismatch, regional mismatch and age threshold are also real problems.

This is the easily overlooked side of the robot employment story: technology may create positions and generate transformation pressure at the same time, but the two often fall on different groups of people.

Therefore, the most valuable question for workers is not "will robots take my job", but which tasks in their own work are easy to be taken over by machines, and which experience can be transformed into the ability to train, manage and correct machines.

Ordinary People Entering the New Track Should Ask Five Questions First

Faced with a newly booming occupation, you don't have to miss it, but you shouldn't put all your bets on the job title either.

First, see what exactly the position delivers. If the payment is only calculated by working hours, number of actions or quantity of qualified data, it is more like a new type of production line position; if the work requires designing solutions, analyzing failure causes and optimizing models, the growth space is usually larger.

Second, see whether the skills are transferable. The operation methods of VR devices may vary from company to company, but data quality judgment, sensor debugging, basic programming and industry knowledge can be transferred to more enterprises.

Third, see the industry scenarios. The requirements for robots in industry, logistics, medical care, elderly care and home services are completely different. Those who only know robots but do not understand scenarios are likely to stay in the demonstration stage; those who understand both equipment and processes are closer to the problems that enterprises are willing to pay to solve.

Fourth, see the working conditions. Wearing VR devices for a long time and repeating arm movements may cause dizziness, fatigue and occupational injuries; some positions also have night shifts, piece-rate wages and high data qualification rate requirements. You should make clear the working hours, performance calculation, training fees and labor security.

Fifth, see whether the commercialization is real. A well-presented robot does not mean that customers are willing to pay continuously. At present, humanoid robots still have obvious limitations in dexterity, environmental adaptability, cost and stability, and some applications still rely on preset processes and manual intervention.

Only when robots truly generate revenue and save costs in factories, warehouses, supermarkets or elderly care institutions, can the positions formed around them exist for a long time.

The Most Worthwhile Thing to Learn Is Not How to Repeat a Thousand Times

The robot trainer is like a mirror, reflecting an important rule of future employment.

New technologies will not only produce "scientists" and "programmers". Every industrial upgrading will create a large number of intermediate positions between technology and the real world: some people collect data, some design scenarios, some test safety, some maintain equipment, and others are responsible for translating the vague needs of customers into tasks that machines can perform.

These positions open a door for young people and job transferees, but behind the door there is no shortcut to high salary.

The truly long-term valuable ability is to understand how real work happens, know why machines make mistakes, and be able to transform human experience into data, processes and standards. Today humans teach robots to fold clothes, but tomorrow the more valuable ability may be to judge why robots can't fold clothes well, and how to make them fold clothes well in different rooms of thousands of households.

Don't just chase the latest job titles. Strive to stand in the position where the machine needs you to define problems, handle exceptions and take responsibilities.

In the future, humans do not necessarily need to do everything faster than robots, but they must understand better than robots: why this thing needs to be done, how to judge it is done well, and what to do when something goes wrong.

References:

1. People's Daily: Being a Teacher for Humanoid Robots, September 2, 2026

2. Xinhua News Agency: Chief Reporter Talks About the First Quarter Economy · Employment Chapter丨New Business Forms and New Tracks Give Rise to New Occupations and New Positions, April 24, 2026

3. Xinhua News Agency: Data "Feeds" the "Big and Small Brains" of Robots — Visiting the Data and Training Base of Embodied Intelligent Robots at Beijing Humanoid Robot Innovation Center, June 14, 2026

4. CCTV News: Embodied Intelligence Training Ground Has Formed Large-Scale and Clustered Layout, August 19, 2026

5. Ministry of Human Resources and Social Security: National Vocational Skill Standard for Artificial Intelligence Trainers (2021 Edition)

6. International Federation of Robotics: World Robotics 2025

7. Reuters: Investigation on China's Humanoid Robot Industry, August 27, 2026

This article is from the WeChat public account "BT Finance" (ID: btcjv1), author: BT Finance, published with authorization from 36Kr.