Similarities and Differences: Future World Lines of Embodied AI (Part II)

Hardware always takes longer to materialize than software. What story are general-purpose robots actually telling? Where do embodied intelligence applications have real potential? How difficult is it to truly understand scenario-specific needs? For industrial applications, does this wave of AI favor newcomers or incumbents? How does the public market view all this? Is there an "elegant" solution? ... Picking up where we left off, let's continue listening to our guests draw from their firsthand experience to analyze the present and imagine the future. Enjoy.

Guest Introductions

*Listed alphabetically by surname

What We're Discussing

• Where's the intersection between future application scenarios and strictly controlled robot costs?
• What's the real potential for robots in elder care? In the near term, will robot technology, elder care needs, and costs align well?
• Is the hesitation around industrial scenarios really about cost competition?
• Why is there so much conviction around service scenarios, especially home environments?
• What value does AI bring to robots in food service settings?
• For industrial applications, does this wave of AI favor new companies or established ones?
• From a public market perspective, how is the government approaching embodied intelligence policy? How are listed companies thinking about this?
• Can Chinese teams build embodied intelligent robots matching international levels within five years?
• What are your expectations for embodied intelligence over the next 3-5 years?

Key Insights

• General-purpose robots aren't a unit-price story — they're a unit-price-times-volume story
• In Japan, 68% of elderly don't save money for their children; in China, roughly 20% of seniors are already willing or able to spend — they just don't know where to procure the age-appropriate services they need
• The goal of robots is to replace humans, but more precisely to solve the most critical needs: delegate the heavy, dirty work to machines while leaving the warm, physically nurturing care to people
• The most pressing demand is in personal care: among nearly 300 million elderly, at least 40-50 million need long-term assistance with lifting, supporting, and turning; add people with disabilities needing similar care, and the total exceeds 100 million
• The conviction around service scenarios, especially home environments, comes down to two things: massive scale, and not being bound by pure ROI logic in selling the product
• A robotic arm is like a server; a food delivery robot is like a good memory — give it ChatGPT and it still won't run. What matters is having a hand
• All our past automation paradigms didn't complete tasks the way humans do; they restructured tasks to fit automation. What needs re-examining: even when a task must be automated, does it still need to be solved the way a human would?
• The crucial question is the boundary between humans and automated equipment, or between humanoid robots and automated equipment
• Embodied technology's biggest advantage over previous tech is scene understanding; this wave of AI solves the leap from motion-level intelligence to task-level or scenario-level intelligence
• The technological leap this time is substantial, bringing potential across the entire industry — industry has only shown a fraction of it
• From a public market perspective: a product with smartphone-scale volumes multiplied by car-level unit prices has never existed in human technological history
• Back when industrial robots were being developed, the government subsidized for 3 years and the entire industrial robot sector expanded rapidly from 200 to 50,000 companies (including the full supply chain)
• Right now nobody knows what's good, or even what kind of large model humanoid robots actually need, or how the various models differ, or where NVIDIA and Tesla fit in
• Past experience shows that when technological advances were already validated, deployed, and proven feasible in developed countries, heavy government subsidies could rapidly catalyze an industry. But with embodied intelligence, nobody knows its potential or ceiling — because no one has reached it yet
• Finding a precisely defined application scenario is the key to this wave of entrepreneurship
• Current AI progress likely offers no elegant solution. By elegant, I mean a single theory that universally solves nearly everything
• When cutting-edge concepts truly reach deployment, China will be at the forefront

Oasis Capital: Let's talk about how embodied intelligence connects with real-world scenarios. What have you all been thinking and researching lately?

Meng Pengfei: The government and society have poured enormous capital and resources into this field, and most practitioners are highly educated elites. There's definitely an expectation that the robots produced will deliver greater value. Yet if you strictly control robot costs, most deployable scenarios end up being pretty basic. So we need to think: where exactly is the right point of deployment?

Gao Yang: There are scenarios where costs are relatively low in China but remain high abroad due to expensive labor. My postdoc advisor started a company called Covariant, focused on pick-and-place — assembly line work where you grab an object and move it to another position. Originally annual human labor cost was $100,000; now with robot leasing, it's $50,000 per year. I think this is one of the more promising scenarios going forward — starting with European and American markets to replace relatively high-wage basic labor.

Liu Qi: I feel like deployable scenarios are everywhere. Spending 100,000 RMB on an all-purpose nanny is viable. I actually think humans are unreliable — most people can't do most things well. Over 90% of people will employ robots in the future. Everyone's debating how long until AGI arrives, but look how much has changed since 2022 alone. Massive numbers of AI professionals have flooded into embodied intelligence, general-purpose robots are advancing rapidly, and the entire industry has been massively propelled forward.

Han Fengtao: I think general-purpose robots aren't about unit price — they're about unit price multiplied by volume. Once you're talking general-purpose, anything that needs to scale won't be particularly expensive, except cars. Cars, beyond their transportation function, are to some degree a mobile private space. Think about it: what do people own that everyone has and is expensive? But selling water can make you China's richest person. Beyond luxury goods, there aren't many things with practical value that are truly expensive. So first, volume — a robot at 30,000–50,000 RMB, sold to elderly people globally, is a trillion-RMB annual product. Second, look at scenarios, but you need to consider timing phases together with this. Securing enough funding during this period matters.

Oasis Capital: Speaking of specific deployment scenarios, many people have been mentioning the potential for robots in eldercare lately. Granny Liu, what's your take?

*Granny Liu: The most immediate need in eldercare right now is age-friendly renovation. Simply put, this means making urban, community, and residential environments more senior-friendly — currently a labor-intensive service. Age-friendly renovation, in simple terms, means going into elderly people's homes and optimizing aspects that aren't senior-friendly. The 2C market is just getting started. Wait another five years or so for today's 50–60-year-olds to age further, and this demographic's willingness and ability to spend will increase substantially. For example, in Japan, 68% of elderly people don't save money for their children. In China, roughly 20% of elderly people are already willing to spend or have the financial means — they just don't know where to purchase needed services.

Some service providers are now partnering with insurance companies, private banks, and family offices. Take one insurance company: they've launched home-based eldercare products ranging from 1 million to 30 million RMB, with beneficiaries being both parents of the policyholder. Services include age-friendly renovation, nannies, cleaning, rehabilitation care, medical visits, healthcare — covering all aspects of daily living.

Going forward, I believe eldercare will definitely involve both machines and humans serving together. Machines can hardly fully replace people; they're mainly for solving certain problems. But machines have no warmth. Elderly people are psychologically fragile and fear loneliness above all. Many say they want to find some scenic, sparsely populated place when old — that's because you're not old yet. When you truly are, you absolutely won't think that way. When your physical functions genuinely decline, you must find a lively place. Even if you can't get a word in, just sitting in a corner listening to others chat feels wonderful.

The purpose of robots is to replace humans, but more precisely to solve the most hardcore needs — let robots handle heavy, dirty work, while humans provide warmth and physical care.

Oasis Capital: How difficult is it for engineers/companies to truly understand scenarios?

Granny Liu: You absolutely must understand real user needs. Real scenarios are completely different from what most engineers imagine. Engineers aren't on the front lines, their physical condition is completely different from users', and mostly they're just guessing. This is an extremely common problem right now. Especially in eldercare — many product developers and researchers have never personally experienced it and simply don't know the real pain points. For example, voice-activated call buttons: able-bodied people might see no need, but users find them incredibly convenient.

My sense is that humanoid robot deployment may not be that far off. Take age-friendly renovation — this scenario includes both hardware and software services. Hardware that can deploy immediately includes age-friendly appliances and furniture, smart fall-prevention devices, and so on. Some intelligent robot companies have started collaborating with us on partial scenario deployment. They lack real user demand data from home scenarios, and this information is very hard to obtain.

I think you must find the most hardcore-need scenarios. For example, we've found that currently the greatest demand is in personal care for those with physical disabilities — services that nannies or caregivers are reluctant to provide, especially long-term handling of excrement and filth. This portion can be handled by robots. Based on extensive research, we've found both wealthy and less wealthy people have some probability of being willing to purchase. At this stage, if it's not truly a hardcore need, wealthy people might buy, less wealthy people definitely won't. As for volume, I think we can broaden our thinking — not just elderly people, but also disabled people. Among our research subjects, there are people in their 20s who also need age-friendly cleaning products. But cost matters here too. Some very dirty, unpleasant jobs aren't completely without takers — for people with little competitiveness, 4,000–5,000 RMB monthly for short-term work is acceptable.

Han Fengtao: For families, why would they be willing to pay for a device that reduces a caregiver's workload?

Granny Liu: You need at least 2 people for 24-hour care, or 1 caregiver if there's a spouse. Even when sleeping, elderly people may need turning every 2–3 hours, or have many emergency calls. Long-term care for elderly people — lifting, supporting, and so on — caregivers' bodies can't sustain it. In Japan, there are even many devices specifically designed to protect caregivers from physical strain. Many elderly people themselves have difficult temperaments, caregiver turnover is high, and children can't provide full-time care. Children are very deferential to caregivers. These invisible costs are substantial. This is the very real pain point of tens of millions of households. Among nearly 300 million elderly people, at least 40–50 million are in this category. Adding disabled people who need similar care services, we're talking at least over 100 million people. Compared to Japan's aging population (over 30% above 65), the Chinese market still has enormous growth potential.

Meng Pengfei: This track is extremely difficult. In recent years we've had much cooperation and exploration overseas — companionship, disability assistance, health, all eldercare-related. Most intelligence-related efforts have failed. At one point Japan had a very prominent exoskeleton company, covered by Japanese health insurance and FDA-certified in the US. Initially priced at over 2 million, later dropped to 700,000, then 500,000 — still wouldn't sell.

Granny Liu: We've studied them. Poor usability is one aspect, but prices were genuinely too high. Even with health insurance covering 90%, it wasn't affordable for ordinary people. And Japanese products are indeed too expensive. China has more advantage here — driving prices down.

Meng Pengfei: Business model is something to consider. In aging Japan, one product category doing reasonably well in eldercare and intelligence is electronic pets — can be petted, can converse, and so on. These can be profitable.

Han Fengtao: The C-end even requires market education. Plus robots haven't established ethical standards or moral judgment standards. Smart pets might be a more comfortable entry point, especially for Japan.

Han Fengtao: In the short term, is it difficult for robot technology to well-match eldercare needs and costs?

Leng Zhe: Probably start with some automated equipment, or pick very precise scenarios, and need government or relevant insurance companies to push.

Meng Pengfei: Unless robot R&D is exceptionally good, otherwise if problems arise, there could be serious safety incidents. We feel we need extremely high reliability before doing specific actions involving human contact, because the service recipients are disabled — so requirements for robots are even higher.

Liu Qi: Because it directly contacts human scenarios. Factories are fine, but home scenarios have considerable legal and ethical issues — no accidents allowed.

Granny Liu: I don't think we need to achieve everything at once, using one humanoid to do all tasks. But there are many scenario pain points robots can solve. Including the lifting problem mentioned earlier — this is genuinely a hardcore need. Lifting elderly people, whether in nursing homes or at home, nobody can sustain long-term lifting. Existing hardcore-need solutions are expensive but low-tech. For example, one solution is a track system that hoists elderly people into sitting positions, remotely controlled.

One disabled person destabilizes the entire family — this scenario is truly painful. These needs are rarely researched or produced; nobody's manufacturing them. For example, a thermostatic shower seat can sell for 70,000 RMB, and customers think it's great — they'll buy as soon as it's available. Or smart pill boxes — demand is actually huge, nearly every client we have uses them as standard. Products between 800 and 1,000 RMB, with very low tech content. Last year's popular "bestie screens" are now approaching us for elderly-targeted modifications. Currently very few companies focus on eldercare, products are still quite primitive — elderly-targeted products can be achieved through simple product modifications.

Han Fengtao: Can start with scenarios where people haven't completely lost ability — robots can make their lives simpler. Human function decline is generally gradual, over 3–5 years; robots can improve elderly people's lives during this period.

A new technology, in initial stages, likely won't have many users. From big mobile phones to Redmi — how long did that take? But long-term it's definitely significant. Why do companies still choose industrial scenarios? Because industry has ready-made scenarios. Currently industrial robots have only solved 2% of manufacturing production problems; the remaining 98% unsolved still relies on humans. China has 100 million manufacturing workers — industry definitely has a wave of opportunity.

Oasis Capital: Is the hesitation about not choosing industrial scenarios due to cost competition?

Han Fengtao: Once a product starts deploying, cost absolutely matters — Tesla is also competing on cost. Industrial robot penetration in China is already quite high; 60% of industrial robots are sold to China, yet penetration is only 2%, already creating a $100 billion market — globally even lower. If robots benefit from AI and penetration increases one order of magnitude to 20%, that's a $1 trillion market.

Oasis Capital: For this wave of AI in industrial scenarios, is it an opportunity for new companies or old companies?

Han Fengtao: I currently think it's probably a short-term opportunity. After technology goes open-source, existing suppliers will all adopt it. If AI can solve low-efficiency, high-cost production problems in industry, it absolutely can in service scenarios too — advanced technology mostly starts from military, aerospace, and industrial applications, with requirements gradually decreasing.

Oasis Capital: The reason everyone sees such potential in service scenarios, especially home scenarios, is first, huge volume, and second, products aren't sold purely on ROI logic. With robot vacuums, users don't feel cheated if it misses a day. But commercial cleaning — one day of downtime and it's a problem.

Oasis Capital: What does everyone think about back-of-house kitchen scenarios?

Leng Zhe: The biggest problem in commercial back-of-house currently is: without extremely stable recognition rates and success rates, customer willingness to adopt is weak. And there's another problem — robots competing with specialized machines. Take a restaurant like McDonald's: because they use standardized tools processing standardized ingredients, from frying fries to chicken nuggets to making burgers, everything uses standardized tools — they can directly use an automated back-of-house production line. McDonald's would ask: why not use a specialized machine better suited to my scenario? What's the value of flexible robots?

Haidilao was considering full back-of-house automation before 2018 — automatic pot-dispensing machines, robotic ingredient warehouses, and delivery robots were all first scaled by them. Before Haidilao, delivery robot companies had nearly 10 years of history in China, but beyond some restaurants buying them as gimmicks, nobody wanted to use them. Finally Haidilao was willing to use them, setting a demonstration effect, and market penetration improved more quickly. So even when new robots are developed, market education is very difficult. The Haidilao case had considerable randomness — if Haidilao hadn't adopted them, would delivery robot promotion have been delayed 3 years, 5 years, even 10 years?

Many restaurant passages only allow one person to pass forward, or two people passing sideways. In this environment, if you deploy robots, two robots can't pass each other. This means such stores can generally only use 1–2 robots, usually just 1.

Gao Yang: Can't this be solved through algorithms?

Leng Zhe: Space is the constraint — these restaurant passages are largely "dead ends." It's very difficult to satisfy more robots operating simultaneously through scheduling algorithms. Many restaurants would rather spend more space building circular passages (or widening passages) for more robots to operate — or more profitably, add more tables for space efficiency, or reduce area to cut rent costs. So the cost of deploying more robots is substantial.

Oasis Capital: Is there a compromise solution, like designing robots half a person's width?

Leng Zhe: That's a longer-term problem, and there's another consideration — if these robots are too tall and narrow, they might be unstable. Once they hit a slope or tilt sideways, they could topple over, which creates safety issues. On top of that, in many restaurants, human food runners also do lots of other tasks like prep work and cleaning. So customer willingness to buy gets constrained by many factors.

Han Fengtao: Just looking at food delivery robots, a few companies have actually made it work.

Leng Zhe: Yes, there's real demand in this scenario — it's a good business. The reason competition is so fierce right now is simply that there are too many players in the space. China has an oversupply of talent, so when people see a viable direction, several well-credentialed teams all pile in and fight to the death domestically. Then everyone discovers that going overseas actually yields higher returns, because there are so few international competitors. Looking at the global market, Japan seems to have just one mediocre company doing this, and the United States barely has any companies in this space at all.

Oasis Capital: AI has opened up 10X imagination for robots — what's AI's actual value in food service scenarios?

Han Fengtao: It varies by scenario. Objectively speaking, AI may help food delivery robots only so much — first you have the hardware, and a mobile robot is basically a cart. Where do you apply AI? But robotic arm hardware has good boundaries; giving a mobile robot a robotic arm could expand its capabilities significantly. A robotic arm is like a server, while a food delivery robot is like a good memory — give it ChatGPT and it still won't run. What it really needs is a hand. I think the opportunities AI brings aren't that related to mobile robots. As for robotic vacuums, current intelligence levels are already sufficient — obstacle avoidance and mobility tech are relatively mature, so there's limited room for AI capabilities.

Meng Pengfei: I think there's one scenario with real demand. When chefs are cooking, they need someone dedicated to handing them things — different knives, ladles, ingredients — because a chef isn't cooking just one dish at a time. When you have 10 or 20 chefs cooking 30 or 40 dishes simultaneously, these runners are called "da he gong" (kitchen assistants). Nobody wants to do this job — it's exhausting, hectic, you learn no real skills, just menial tasks — so it's extremely hard to hire for, and wages are high. If AI improvements can enhance a robot's movement speed, recognition capabilities, and so on, that could work well.

Han Fengtao: This is exactly what I meant — one or two robotic arms plus mobility could work. But actually having robotic arms cook stir-fry would require much more.

Oasis Capital: From a technical implementation perspective, how difficult is this?

Gao Yang: It depends on how many different things need to be picked up. If the dishes are all in containers, that's fairly simple. If the number of items is limited, or their shapes are limited, it's still doable.

Leng Zhe: Currently in back-of-house scenarios, humans do quite a lot of things. If pre-packaged content has very high standardization, application deployment will be faster. FANUC's cafeteria has a noodle-cooking robot — there's an automatic noodle-making machine next to it that dispenses noodles, the broth is pre-made, and the robotic arm handles boiling the noodles, draining water, and ladling broth. The process is relatively standardized, so robots can do it. What doesn't work well are non-standardized scenarios — say a bundle of scallions arrives, you need to remove the leaves first, then wash them. Can this generation of embodied technology handle this? Yes, but what's the success rate? Since nobody dares guarantee extremely high success rates — what if there's still mud on the scallions? — the vegetables just get sent in.

Gao Yang: I think this kind of thing is genuinely very difficult, way too non-standard. My sense is that washing scallions would be handled by pre-made food factories with automated equipment like spray machines or dedicated machines for specific operations.

Leng Zhe: This brings us back to a paradox. Humanoid robots broadly aim to replace humans, but all our past automation models didn't complete tasks the way humans do — they transformed the task into a form suitable for automation. Take frozen steamed buns: machine-made buns look very similar to hand-made ones, but the actual wrapping process is completely different — they just imprint a few pleats on top to make them look hand-made. This phenomenon needs to be re-examined: even if a task needs to be completed automatically, does it still need to be solved the way a human would do it? There might be another way entirely.

Gao Yang: That's a very good observation. A crucial point for consideration is the boundary between humans and automated equipment, or between humanoid robots and automated equipment.

Leng Zhe: I think embodied technology's biggest advantage over previous technology is scene understanding. The previous generation could recognize what something was, then operate on it. This generation can recognize what scene it's in, figure out what needs to happen next, and complete the task. Whether it's a humanoid robot or one with four arms, that's very flexible.

Han Fengtao: This is actually about layers of AI. How do we define AI? An artificial thing with some degree of automatic capability to complete a function — that's AI. The earliest example: making a motor maintain 100 RPM regardless of interference — to some degree, that's also AI. Previously, robot software was only at the motion intelligence level — you tell it to draw a circle, it draws you a perfect circle. But ask it to observe its surroundings and sketch what it sees, and it can't. So this wave of AI solves the transformation from motion-level intelligence to task-level or scene-level intelligence. What Dr. Leng described as the previous generation of robots was motion-level AI. But motion alone isn't enough — now AI can intelligently decompose tasks, and many tasks robots can complete automatically.

Oasis Capital: Would the first wave be good news for industrial robot companies? At minimum, could it save on system integrator service fulfillment costs?

Leng Zhe: I think everyone's direction is different, because each company has its own application scenarios. Existing industrial robot companies have very clear application scenarios — rather than chasing an unclear scenario with uncertain returns on investment, it's better to make their current scenarios more solid. Higher certainty, higher profits — this is an optimization problem.

Han: This technological leap is substantial, bringing potential for entire industry opportunities, of which industrial applications only show a portion. Industrial opportunities are relatively certain. AI advances make industrial robots more usable and intelligent — intelligent welding, intelligent handling, and so on. But scenarios are still limited to industrial-level production capacity. If robots become more intelligent and can move into commercial, household, and service applications, the volume space would certainly be several orders of magnitude larger than industrial. Take computers as an example: before home computers existed, servers had no monitors. When monitors appeared, they certainly provided a better interface for servers. But monitors' greater significance was that when ordinary people didn't understand complex operations, monitors enabled computers to enter homes.

Leng Zhe: I think embodied technology is more valuable when applied to complex tasks and complex scenarios. For highly structured scenarios, there's no need for this kind of technology.

Oasis Capital: From a secondary market perspective, what's the government's attention and policy on this, and what are listed companies thinking?

Meng Pengfei: First on the policy side, I believe the government will introduce various industrial support policies. Any industry that's genuinely good will definitely have bubbles when it first emerges — an industry without early bubbles is probably not a good industry. Right now, the robot industry's bubble isn't even that large. From a secondary market perspective, this is a product with the volume of phones multiplied by the unit price of cars — in all of human technological history, nothing like this has existed to date. Moreover, this is just the beginning — we believe there's even greater potential ahead.

On the policy side, since the beginning of this year, local governments everywhere have been holding discussions with relevant enterprises in their jurisdictions, and some listed companies or state-owned enterprises are also accelerating exploration of humanoid robot applications. Comparing hardware versus large model software, we have greater advantages on the hardware side — first, hardware is definitely China's strength; second, we have abundant scenarios, many of which Tesla and NVIDIA probably haven't even thought of; and third, Chinese teams, whenever they see opportunity, can rapidly drive costs extremely low and make deployment possible.

In just over a year, companies fully claiming to be making humanoids — I've counted more than 30 now, probably over 50 by next year — this scale still isn't large enough. Just like the industrial robot industry's development back then: the government subsidized for 3 years, and industrial robot numbers rapidly grew from 200 to 50,000 (including the entire industry chain). In China, participating in the first wave of anything is better — capable teams should get involved as early as possible.

Meng Pengfei: But currently nobody knows what's good, or even what kind of large model humanoid robots actually need, what the differences between various models really are, or what NVIDIA's and Tesla's positions are in all this?

Teacher Gao: Academically speaking, there's still considerable uncertainty. It's actually like BERT and GPT — BERT was extremely popular before GPT took off, a very hot natural language processing model. I think embodied large models are currently in an era where BERT and GPT coexist. Judging which technical path will succeed — I think it's still early, but I believe it will eventually be at least something on the scale of GPT, and ultimately there will indeed be one large model applicable to many, many scenarios. If GPT handles speech and embodied large models handle action, I'm more inclined to believe this will rise in tandem with a hardware company, because you need scenarios and hardware to really put this large model to use. It's not like pure internet — it can't exist independently. GPT was an era, autonomous driving was an era, but NVIDIA has consistently played a platform role — all participants can use its cards and get excellent support. Speaking of competitors, I think Tesla may be a major competitor because it has the industry chain, it's seriously building hardware, and has accumulated substantial development experience.

Liu Qi: I think there might be a smaller Embodied AI model specifically for robot interaction — something fine-tuned specifically on robot data — with a large model on top. The vast majority of tasks can be completed through a small model, and when encountering difficult problems, it calls the large model's higher intelligence capabilities. So I think small companies shouldn't try to build a large model.

Leng Zhe: I think hardware-related data is also quite valuable — teleoperation data, super-multimodal datasets. But the data business is a "selling water" business: it depends on how many people are gold-rushing. If there aren't that many gold-rushers, or if they can bring their own water, then this business doesn't work. An independent data business is a small business that might last a few years.

Oasis Capital: Let's wrap up — what are everyone's expectations for the next 3-5 years?

Meng Pengfei: There's one image: in Tesla's factory, robots are working, and humanoid robots can control industrial robots. In the early stages of robot development, costs won't drop quickly, but companies will receive substantial subsidies, including tax benefits and government subsidies.

Leng Zhe: What I'm thinking about is where we are on the timeline. EVs have taken off in recent years, but the last wave of commercially available electric vehicles close to our memory was in 1990, when the United States poured massive subsidies into General Motors to produce a lead-acid battery electric car. Many people back then thought the time for EVs had come — looking back now, it was simply too early. In 2010, people were debating how many years until autonomous driving would deploy at scale — optimists said five years, pessimists said ten. Where are we now? What Professor Meng said earlier makes sense, but the lesson from past experience is that those technological advances had already been validated, implemented, and proven feasible in developed countries. When our country invested heavily to subsidize them, the industry could be rapidly catalyzed. But with embodied intelligence, nobody knows its true potential or ceiling, because no one has hit the ceiling yet. Just like when deep learning first emerged, people thought it could do anything, that its ceiling was infinitely high. After a few years of running with it, we discovered the ceiling was right here.

In three to five years, I think we'll see one or two application scenarios generating hundreds of millions to billions in annual revenue. The key is who can capture that scenario. Because while deep learning ultimately fell short of expectations, facial recognition did eventually land — and that's actually a sizable business. Take electric vehicles: before this explosive growth wave, although the 1990s attempt didn't take off, electric carts used in factory campuses and low-speed elderly vehicles ("laotoule") did have considerable market demand, just not as massive as people expected. The key is still finding a very precise application scenario — I think that's the most critical thing for this wave of entrepreneurship. Focusing solely on building something big, comprehensive, and generalized is pointless.

Liu Qi: In three to five years, I think both hardware and software will see many iterations — motion control, grasping, multimodal control understanding, and so on. I think people will be willing to buy humanoid robots to perform simple operations in living scenarios. After reaching a certain inflection point, the iteration speed of embodied intelligence software and hardware will rise exponentially, and applicable scenarios will multiply rapidly.

Gao Yang: In the next one to two years, we'll see humanoid robots walking around and performing some operations — maybe not particularly stable, maybe not commercially viable yet, but flashier demos will emerge. Actually deploying them in home scenarios remains very difficult. In environments with lower sensitivity requirements and less cost sensitivity, we'll see more deployment within three years, and more ordinary people will have tangible exposure to them.

Han Fengtao: I think within five years, people will see many robots in daily life. The most common type will be toy-oriented. GTC went viral again this time, and last year it sparked a lot of discussion in robotics circles. The Walt Disney Company's robotics division is also quite impressive, though they haven't mass-produced yet. I think as functionality expands and costs continue to drop, more toy robot form factors may emerge. Second, in commercial service settings, we'll definitely see all kinds — with one or two arms, running or walking. I'm not sure about home scenarios. Looking at configurations, I think mobile manipulators with bases will appear in large numbers.

Liu Qi: Can Chinese teams really make it in the next five years? Will they produce embodied intelligent robots approaching foreign levels?

Gao Yang: From a product perspective, if hardware is well-built, reliable, and usable, China can outcompete through sheer effort. And many of our domestic scholars are team members of these top overseas scientists.

Leng Zhe: Looking at it this way, autonomous driving — DARPA held its first Grand Challenge in 2001. Looking at now, if we're talking about autonomous driving for everyday use, I think it's still happening here domestically. I believe current AI progress likely doesn't admit an elegant solution. By elegant, I mean a unified theory, a few formulas, an algorithm that universally solves almost all problems. Autonomous driving is a classic example — initially everyone assumed there was an elegant solution and searched desperately for it, only to discover there wasn't. Once there's no elegant solution, what does it become? A battle of manpower. And that's not something the United States excels at. China's annual engineering graduate output is more than ten times that of the US, exceeding the entire developed world combined. Whether high-IQ talent or general workforce, no country surpasses China — this has become abundantly clear in autonomous driving. There are simply too many edge cases that require human effort to work through; an elegant solution can't resolve them. Real-world AI problems are similar. You can't solve them with a beautifully elegant solution — your information or certain conditions are simply insufficient, or current intelligence is insufficient. While 99% of cases might be handled by elegant AI algorithms, that final 1% still determines whether users ultimately adopt your product. What elegant technology can't solve requires human effort to grind through. Once it becomes a battle of manpower — even PhD-level manpower — the US holds no advantage. Even if the US has many top people, the number of the very best is limited, and they won't spend their time resolving these extreme edge cases. I still believe that when cutting-edge concepts truly reach implementation, China will be the first to land them.

Han Fengtao: History often rhymes, as they say.