Jiang Zheyuan: Make Robots Cheap Enough and Let Them Actually Do Work

21st Century Economic Report Journalist Kong Haili

Jang Geuyon speaks at a rapid pace, full of vigor. As the founder of a popular embodied intelligence company, this 28-year-old entrepreneur has a distinct personal style. During conversations, he often says, 'I'm about to say a bold statement,' followed immediately by truths that no one in the industry is willing to speak about.

For example, he believes that "the entire market currently has no more than five companies that can deliver high-quality, product-level ontologies in bulk"; about why robots work slowly, he said "I think this is purely a matter of capability"; talking about industry landscape, his judgment is "the final outcome will definitely be winner takes all". As for when robots will be able to go home and do housework, he gave a relatively optimistic outlook: "Maybe in a few years, it will be possible to complete specific tasks in generalized scenarios." He then added a sentence: "This is not a conclusion, just my personal guess."

Recently, 21st Century Economic Report and 21Tech reporters interviewed Jiang Zheyuan, founder and chairman of Songyan Power. This young man, born in 1998, left school to start his own business while pursuing a PhD at Tsinghua University, has already taken his company to the CCTV Spring Festival Gala, making the robot body sufficiently cheap and reliable. This year, he is leading his team to heavily invest in the robot's brain.

In 2026, the narrative of the humanoid robot industry has undergone a noticeable change. Almost all mainstream players need to prove whether robots can actually do tasks. Zhiyuan, Yushu, and Songyan Power are all increasing their investments in the robot's brain.

In September, Songyan Power announced three new developments in a short period, all targeting industry challenges: the HERON-World Model allows robots to understand and learn about the world, the HERON-CRA model enables robots to remember the past and learn from mistakes, and the HERON-Spatial Intelligence spatial intelligence model builds an interactive and trainable world for robots. The ultimate goal is to give the robot a different environment, a different object, or even a different body, and it can still accomplish tasks effectively.

When asked about whether Songyan Power has changed its route, why they decided to develop a brain, and why they started now, Jiang Zheyuan's answers were very candid: 'Our decisions have always been to develop our own brain, but in the past, it wasn't the right time.'

In 2023, more urgent matters were to first get the body and motion control right. At that time, the technical direction for brain development was still quite scattered. Now, according to his judgment, there is already some consensus on the technical direction, and the body is also ready. In terms of cost and reliability, 'Songyan Power may be slightly ahead of the industry as a whole in some aspects. Now, we have basically completed all the preparations we can do,' Jiang Zheyuan said.

The public's understanding of Songyan Power often starts with a robot that has excellent mobility. In the 2025 Beijing Yizhuang Half Marathon for humanoid robots, the N2 'Xiaowantong' won second place. By 2026, the company's robots appeared in the CCTV Spring Festival Gala sketch 'The Grandmother's Favorite,' performing alongside Cai Ming. Mobility, stage performance, and price became the most memorable tags for it.

In October 2025, Songyan Power launched a small-sized humanoid robot 'Xiaobumi' with a limited-time pre-sale price of 9,998 yuan. This price brought the robot closer to the position of a household consumer decision. It is still not a household helper, and the company's positioning is mainly on children's programming education and companionship, but buyers have started to expand from schools and research institutions to ordinary households.

Jiang Zheyuan's entrepreneurial journey started at Tsinghua University. He studied undergraduate in the Department of Electronic Engineering at Tsinghua University and then entered the Cross Information Institute to pursue a doctorate. In 2023, during his third year of the doctorate, he left campus to start a business, and Songyan Power was founded in September of that year.

His expression carries a strong personal tone. When discussing the future of robots replacing jobs, he said, 'If one day Xiaobumi can sit here to replace me in a meeting, 'I would be happy, and I could just drive to Tibet for a trip.'

Jang Cheol-youn refers to the next hurdle as generalization. In simple terms, it means that even if the robot changes its body, environment, objects, and mode of operation, it can still perform tasks with high success rates. He doesn't believe that a few impressive demonstrations are enough to cross this threshold. In his view, the real competition will come down to data: whether the robot can continuously obtain the experiences it needs, whether it can organize, annotate, and use these experiences for training, and then use new models to generate new experiences.

In Jang Cheol-youn's vision, Soonyeon Dynamics' approach is to first make the robot body cheap, reliable, and capable of mass production and delivery, allowing for large-scale deployment in real-world scenarios, so that the robot can accumulate a large amount of real-world operational data and experience through actual work; then use these scaled data to train the brain, achieving true intelligence.

The robot will eventually need to learn to enter a home it has never seen before, be able to store unfamiliar clothes, and when someone moves a cup, know what to do next.

To accomplish this ultimate mission, the company has redirected a significant portion of its investment toward brain research this year. Jang Cheol-youn believes that although the funding is good, money should be spent carefully; at the same time, the body business can also take on the role of generating revenue, allowing today's sold robots to support the potential intelligence that may come tomorrow.

He has a very far-reaching vision for the future of robots, yet his assessment of the present is very clear. When asked how far the robot industry's marathon has come, he still believes it's not even ten meters.

The following is the edited and compiled interview:

One: Now 'going all out' to develop the brain, everything is ready

Question: Soonyeon Dynamics used to be better at the body, but now decided to develop the brain themselves. What was the decision point?

Jang Cheol-youn: Our decision has always been to develop the brain ourselves. However, we did not previously seriously commit to developing the brain with full force.

Why did we decide to go all out this year? It's because we saw the overall technical direction, and there was already some consensus on the big picture. It was just in some details that each company had its own different approach. At this point, we felt it was possible to invest in this area.

Additionally, we also felt that our body was already ready. Our body has already achieved extremely low costs, high reliability, and reached a level that may have never been seen before. In this context, I believe we need a major investment in the brain, and we are fully prepared.

Question: Is developing a brain for robots a bit late?

Jang Cheolsu: There are too many brains in the world, but currently, in the entire market, the number of companies that can mass-produce and scale reliable, usable, product-level embodied robots is absolutely no more than five.

In essence, a brain cannot exist independently of the robot's body. We are in an era driven by data. Data needs scale and good diversity. I believe a significant portion of data cannot exist independently of the body.

Companies with the ability to mass-produce and low-cost manufacture embodied robots will have an unparalleled advantage in data scaling. I don't think the timing is late; I think this timing is exactly right.

Question: What are the pitfalls of developing a brain?

Jang Cheolsu: The main pitfalls are four: one is reliance on real machine data; two is being tied to a single body, and the robot won't work if you switch to another robot; three is only doing work in simulation; four is only doing demonstrations without building data infrastructure.

Question: Does this year's investment in the brain have a high proportion?

Jang Cheolsu: It's still quite high. A large portion of our money this year has been spent on the brain, investing a lot of money. Specifically, we cannot disclose the exact figure, as our financial reports are not public.

II. The robot industry doesn't need that many companies; it will be a 'winner-takes-all' market

Question: What has been the biggest change in the industry from last year to this year? Ordinary viewers didn't see significant changes at WRC and WAIC.

Jang Cheolsu: The changes you can see at WRC and WAIC are limited. This year, there are a few different aspects.

The hardware embodied motion capabilities are gradually becoming more equal. Last year at WRC, few companies showed very high dynamic capabilities, and we were one of the few. This year, many booths have shown high dynamic capabilities.

On the brain side, very few companies showed real operational scenarios last year, and this year, we also saw that. Outside of exhibitions, the cost, reliability, and popularity of robot bodies are gradually improving. These are all positive signals of the industry's rapid development.

Question: Has the focus of competition changed this year?

Jang Cheolsu: Competition occurs at different levels and layers in each stage. Everyone is ultimately heading towards a universal embodied brain robot era. Before this era arrives, the competitive landscape will be different every year.

Last year, the main competition was on motion capabilities, shipment volume, profitability, and how impressive the body was. This year, motion capabilities are not the core competition anymore, as everyone has already reached a very high level. This year, the competition is more about scaling, brains, full-body motion control, and perception-motion control.

Question: What will be the long-term industry landscape?

Jang Cheolsu: This industry indeed doesn't need that many companies. After the capital market becomes relatively calm, some companies will definitely fall behind.

In the end, it will be a 'winner-takes-all' market. Because the winner can obtain the best data loop. Once a flywheel effect is formed, it gets better and better, and it's hard for later entrants to break in. It's difficult to push, but once you enter this stage, it just keeps spinning faster and better.

Question: Who can become that winner?

Jang Cheolsu: I think it's about choosing the right path, selecting the correct data collection route, and investing along this direction to collect the truly needed data and train a model with sufficient generalization ability.

In the field of robots, I think the key is who can start selling at home first and who can sell explosively at home.

Question: Large model or world model companies can also buy body data. What's your advantage?

Jiang Zheyuan: First, our cost is low, we don't necessarily sell cheap, but there will definitely be a funding gap.

Second, doing data collection is not just a matter of the body itself, it's a complex solution. How to collect human data, do we need special equipment? After collecting the data, how to store it, how to save it, how to transmit it back, how to annotate? The pipeline is a capability of a complete chain and system.

In addition to the body cost, I think the other parts are replicable, but also there is a cognitive gap and time gap.

We welcome base model companies to buy our body in large quantities, and we are open-minded and willing to build a dataset with everyone. We don't expect to be the only one to achieve general embodied intelligence, but we hope the whole industry can move forward together.

Three, now we are just moving from L0 to L1

Question: What do you mean by L0 to L3?

Jiang Zheyuan: L0 is nothing intelligence, it's just a motion robot. It can only do actions, dance, no operation capability.

L1 is able to complete the landing in some specific scenarios under heavy delivery. What does heavy delivery mean? For each trade, arrange people to go to the site to collect data, do training, optimize, deploy, the whole process, full chain needs a delivery team to complete. This is my own definition, not the industry's definition.

L2 is able to complete specific tasks in generalized scenarios. Home is a typical L2 scenario, buy a robot home, open the corresponding button switch, it can help me make the bed, tidy up clothes, and can achieve complete generalization in several specific tasks.

L3 is truly giving it a SOP, giving it a picture-rich explanation, it starts working on its own, no need for any optimization or delivery actions.

Question: Where is the industry now?

Jiang Zheyuan: I think the industry is now in the stage from L0 to L1. L1 level intelligence, we ourselves are capable of doing, can deliver some scenarios.

The true one with almost infinite ceiling is L2. Can enter the home, all homes can buy, this is completely different ceiling scale.

Question: How long for L2 and L3?

Jiang Zheyuan: Distance L3, I can't answer well, this area is still in the exploration and development stage.

Question: Wang Xingxing's judgment on the AGI moment required years is constantly extended.

Jiang Zheyuan: I like to define the problem clearly, what does AGI moment mean? If it refers to L3, it's definitely a decade or more cycle, because the data scale needed is too big, need to accumulate in various different fields. But for L2, I think it won't take that long, relatively easier to do.

Question: This industry's marathon, is it only ran ten meters now?

Jiang Zheyuan: It's actually about the same. The industry now has not yet any truly large-scale collected body data. I think it's less than ten meters, still just starting.

Our destination is definitely L3, give the robot an SOP, it starts working on its own according to the SOP, no need for any optimization, no need for any delivery actions. This is the ultimate dream. The development process of the whole industry towards this dream, can be understood as the process of collecting data, how much data, this is too far, ten thousand eight thousand li away.

Four, the body price dropped from ten thousand yuan to two thousand yuan, relying on the process

Question: Is the body ready, passed the grade, or already got a high score?

Jiang Zheyuan: A body to get the passing score needs to meet several conditions.

The first, it shouldn't be a consumable, this is the basic baseline for passing. The body itself should last at least one or two years, not break, this is the minimum requirement.

Second, it should have the ability to complete operations. Degrees of freedom should be enough, strength should be big enough, repeat positioning accuracy needs to be high enough, joint clearance, backlash also need to be small enough.

Third, the underlying control needs to support you to complete these operations. We hope it can support doing operation tasks. Once these things are done, it's completed the passing score.

The rest, you say 80 points, 90 points, which is better, we have different evaluation indicators. Some people's answer may be more sensors, such as full-body skin touch; some people's answer may be stronger reliability. This is currently not determined.

Question: The low cost of Songyan Power relies on scale?

Jiang Zheyuan: No, it's relying on process. In the field of cost analysis, there is a saying that is very reasonable, maybe R&D accounts for 80%, supply chain or scale accounts for 20%. The real cost reduction R&D, your design, process should account for more than 80% of the weight.

For example, from 25,000 yuan to 23,000 yuan, 22,000 yuan, you may be able to rely on scale. But if from 100,000 yuan to 20,000 yuan, 25,000 yuan, this is definitely relying on process.

We have used a lot of different processes. Our team is actually helping the whole industry to take the pitfall in this matter.

Question: What specific process changes have you made?

Jiang Zheyuan: I'll give you a few points. The first, our whole machine body is all plastic. All plastic means that the whole machine body structure cost can be much lower than others; if compared with metal mold or metal machining, it may be much lower.

The second point, we made our own gear reducer box, the process is completely different. So the cost is significantly lower than the current supply chain.

Our idea is to definitely achieve the lowest cost under the premise of absolute reliability. This comes from design, process, which are all handled by the R&D team.

Question: You said manufacturing advantage will become data advantage, how to understand it?

Jiang Zheyuan: It can be calculated as a very simple primary school math. The unit cost of a robot body, the cost of using it per unit time, can be understood as including labor, the whole robot manufacturing cost, divided by the time it can be amortized, that is, how long the hardware will break, how long it can't be used.

There are two points here. First, you need to be able to make low-cost usable hardware; second, you need to make high-reliability hardware. These two points are the advantages we are hard to compare.

We have seen the situation of C-end users using robots after this year's C-end delivery, and we have done a lot of strengthening accordingly. We also have the ability to mass produce robots.

Question: Doing the brain and reducing costs are two different logics?

Jiang Zheyuan: Doing the brain and low cost are definitely complementary.

You must first reduce the cost, if you don't reduce the cost, there's no talk about scale. Reducing the cost can bring scale, scale can bring scaled data, diversified data can truly bring intelligence.

Cost means you can confidently scale up. When the cost is high, I have no confidence to scale up equipment, fleet. But when the cost is low enough, we can do this painlessly. 

Five, the body needs to make money, the brain still needs to invest

Question: Now the income mainly comes from B-end or C-end?

Jiang Zheyuan: In terms of amount, B-end is more, in terms of volume, C-end is more.

From kindergarten to K12, to universities, vocational schools, colleges, each scenario has different customers and partners. The whole solution doesn't only include robots, but also accompanying courses and services.

In this year's B-end revenue growth, the main part is in education, large education scenarios. Due to past accumulation, this year has shown scaled growth.

In addition, in commercial services, tourism and culture scenarios, many solution partners will do secondary development based on our robots, with very diverse application scenarios. Diverse application scenarios, education sector, and initial exploration in C-end, have built the revenue structure of this year.

Question: Both doing the body and the brain, how to divide the money?

Jiang Zheyuan: I clarify, the body is a business, a business. The real investment cost department is definitely the brain department, and it is a heavy cost department, which needs continuous high-intensity investment.

If we single out the body, I haven't calculated this year, this year may not be. But next year, I think if we single out the body, it may really be profitable, very likely. If we remove the brain part entirely, it's very likely that the whole company is profitable.

Question: Xiao Bu Mi has low gross margin, why still spend effort on it?

Jiang Zheyuan: Brand mindset needs to be accumulated over time. When we first launched Bumi on JD platform, the shop traffic was very low. Later, with the passage of time, we gradually built it up.

In the home end, C-end, brand mindset needs to be cultivated over time. Need to continuously plant seeds, continuously speak out, continuously have delivery, continuously let users have some cognition.

This mindset building takes time, can't wait for the brain to be good before doing it. We must do it, need to continuously invest.

Question: Do investors recognize this route of first doing the body and then the brain?

Jiang Zheyuan: If they have really invested in us with real money, can they not recognize it?

Persuasion is less than choice. Some investors may not want to see companies that do bodies, some investors focus on the brain, some investors only see bodies. We choose more investors who recognize our route.

A company at least needs to have a business that can sustain itself first, then talk about stars and seas. The tricky point of embodied companies is that pure brain companies, in the short term, are difficult to have this income. In my opinion, doing the body is not an option, it's something that must be done.

Question: Is the industry not so focused on making money now?

Jang Cheol-youn: We are still relatively pursuing healthy cash flow, and definitely cannot rely on transfusions all the time. We hope the company can have a clear trend of profitability improvement. We are relatively pursuing this, but indeed the entire industry is not paying much attention to it now.

Question: After fundraising, will the money still be tight?

Jang Cheol-youn: Even though we raised a lot of money, we still feel that money is tight. Survival is definitely okay, but indeed it's tight, and we have to spend it carefully.

If you give me enough money on my account today, I can deploy up to ten thousand robots to collect data, possibly in a short time, at least achieving L2 level things quickly. 

Six, the reason robots work slowly is the company's capability, the hardest part is generalization

Question: Why are robots still so slow?

Jang Cheol-youn: I think this is purely a capability issue. You can look at our demo, which is similar to humans working, and some even faster than humans, which is a capability issue.

Yes. The distance to the ultimate dream can be understood as the process of data collection. How much data you have, right? This is still very far from that goal, like ten thousand eight thousand li away.

Question: You used to be skeptical about simulation data, have you changed your mind?

Jang Cheol-youn: My previous judgment has always been that we cannot rely entirely on simulation data. Simulation data is useful, but cannot be used entirely. This is still my judgment today.

But indeed, a lot of our work is focused on how to establish a large amount of diverse and high-quality assets in the simulator. We have our own data recipe, and the pyramid includes some simulation data.

Question: Looking at it now, what is most underestimated?

Jang Cheol-youn: The most underestimated, I think, is some technologies on the hardware body. Many people think hardware body is no longer something that can have new breakthroughs, but we may prove with a series of future releases that there are indeed breakthroughs on the body.

Question: How far has embodied intelligence development reached, especially with the slowdown in discussions about general AI?

Jang Cheol-youn: It is still very far from being OK, and I think embodied intelligence needs to accelerate. L3 is definitely very far, and we can talk about these things when we reach L3.

Seven, robots need to learn to judge what to do next on their own

Question: What problems are the recent actions on the brain side solving?

Jang Cheol-youn: HERON-World Model is an action-conditioned world model, given the current image and the action the robot is about to perform, it predicts how the world will change next. The actual output of actions is done by HERON-CRA.

A world model can serve as an evaluator for strategies and a data generator, helping to reduce the cost of trial and error with real machines.

You can refer to our published CRA demo. There are two demos, one is folding socks, and the other is making coffee.

Folding the orange socks, at first it found that the gripper wasn't holding the socks well, so it made an adjustment on its own, then reinserted the claw into the sock tube, and continued with the task.

For the coffee one, we intentionally added artificial interference. The robotic arm was supposed to take the cup with frothy milk to mix the coffee, but the staff moved the cup to another location twice. It can be seen that the model paused first, then adjusted the direction of the action to continue executing the task.

In the future, in various service scenarios such as home services and commercial services, we will encounter such issues to some extent. We believe, thanks to the evolution of brain technology capabilities, these issues can be handled and adjusted on the spot.

Question: Which of memory, fine manipulation, and cross-body generalization is the hardest to break through?

Jang Cheol-youn: The truly difficult part, I think, is the generalization capability. We talk about cross-body, cross-scenario, cross-object, cross-operation object, and various different generalization capabilities, which is the hardest, because it is the most data-intensive.

The next step is definitely moving in this direction. How to make robots usable in different scenarios with high success rates, requires enough massive, diverse data, as well as some human-in-the-loop data.

The final generalization is tested by the final data, I think, and it depends on how good the data pipeline and data infrastructure are.

In some tasks, it is indeed necessary to emphasize memory capability. In other tasks, such as simple pick and place, there is no need for memory capability.

Eight, a million hours is just an estimate, what's really important is data we haven't seen before

Question: What data scheme do you adopt? How much data is needed to reach the capability inflection point?

Jang Cheol-youn: We have a data pyramid, and we initially think it's at least a million hours level.

Question: Pursuing data quantity is meaningless, right?

Jang Cheol-youn: Diversity is more important, diversity is the most important. You need to supplement some data you haven't seen before, which is the key.

Question: How should generalization be measured?

Jang Cheol-youn: Cross-body, cross-scenario, cross-object, which is cross-operation object, and cross-operation methods.

For example, putting clothes on a hanger, different hangers have different ways of putting them on, some are clamped, some need to be threaded from below to put on, there are different methods, which is quite troublesome.

Question: How much data can a robot collect? How large a scale should be deployed?

Jang Cheol-youn: We calculated that, if each body is fully operational, a reasonable estimate, about one thousand hours of data can be generated per year. If it's a million hours, then one thousand bodies.

We will initially deploy about hundreds of machines to collect data, gradually increasing to thousands.

Nine, if robots can replace me in meetings, I will go to Tibet to play

Question: Will robots mature enough to cause employment shocks?

Jang Cheol-youn: We believe the truly best general embodied intelligence model should be able to do anything. Give it an SOP, it can start working on the production line; give it a language instruction, it might start working in your home.

When robots achieve this, can we consider that there will no longer be physical labor in the world? We enter an era of greatly increased productivity, where productivity can be infinitely replicated. Because robots can make robots, and the robots made can produce goods for you.

In that era, we can pursue higher-level needs in Maslow's hierarchy of needs, no longer being a workhorse.

When can we pursue poetry and the distant horizon? When can I drive a off-road car on the 318 National Highway, looking at the distant scenery, and the robot brings me a cup of milk tea? This is the era we hope to see.

I hope the robot can come to be the chairman. If one day robot Bumi can sit here, replacing me to hold this meeting, I would be very happy. I will rest at home and go to Tibet to play.

I think I should be a poet. I think everyone here should be a poet or a musician, and let the robots handle the rest of the mundane realities.

(Author: Kong Haili, Editor: Luo Yiqi, Luo Yifan, Designer Liu Yanyan, Video Editor Yang Haokai)