A robot company's biggest enemy isn't technology — it's teleoperation and storytelling
Wang He says this generation of embodied AI companies has only two dead ends: drifting along telling stories, or rebuilding from scratch in every new scenario. He's betting on synthetic data — real data is just 1% of training volume.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
This generation of robot companies has only two dead ends
Wang He sorts the failure modes of this generation of embodied AI companies into two categories. The first is long-term drifting: telling humanoid robot stories for decades, with no commercial application cases, all skills essentially choreographed for filming. The second is the math not working out: you solve every skill in one scenario, but move to the next scenario and everything is redone from scratch, with no reduction in editing cost — this is common in industrial vision, where different vendors' materials require re-modeling and new templates, and every project consumes fixed manpower, R&D, and delivery cycles. His judgment is that falling into either one means the ceiling cannot exceed that of the previous generation of companies, because you cannot, as in the US, have capital support a company for decades without profit and still keep it alive.
— Wang HeHiring people to teleoperate and collect data burns hundreds of millions a month
Wang He did the math: a full-size humanoid robot costs at least 100,000 to manufacture, so 10,000 units means a billion spent on manufacturing. Once built, they still need teleoperators — at least two shifts per unit, or even two shifts with two people each, four people; the monthly cost of teleoperators per robot is tens of thousands, plus annotation and quality inspection. All told, maintaining 10,000 robots costs hundreds of millions to a billion per month. So his conclusion is that nobody in the world can do this, and those US companies don't work either — they can only film demos into videos for you, they cannot invite you to watch live. Cars are different: you sell the car and still collect money, and data is collected at negative cost.
— Wang HeReal data is only 1% of training volume
Wang He says Galbot dares to livestream a VLA doing shelves at the BAAI conference, and dares to demonstrate continuously in the exhibition hall, because it does not rely entirely on real data — real data is about 1% of training data, or even less, and the self-developed synthetic data pipeline carries most of the work. He started researching synthetic data back in 2017 with the NOX work, and by this year, using synthetic data to build models and solve sim2real problems, it has been eight years. He directly names the motivation of opponents: many people say simulation is not enough, there is a sim2real gap, it cannot solve the problem, you must use real data — and how do you use real data? First you buy my machine, then you teleoperate. It is a closed-loop business model that makes sense.
— Wang HeThose who say sim2real doesn't work must answer why walking works
Wang He refutes the uselessness of synthetic data directly: today all humanoid robot walking, jumping, and running skills come through sim2real, all obtained through large-scale reinforcement learning in simulators. So those who say it doesn't work must answer why leg sim2real works. They will say it's because legs have no vision, and adding a visual modality means sim2real cannot work. Wang He thinks this rebuttal is also wrong: VLA is based on VLM, and VLM training data already includes a large amount of animation and movie data, a mix of real and fake — a VLM that can understand the real world can also understand the plot when watching Donald Duck and Mickey Mouse. The gap between physically rendered grasping and placing and the real world is always smaller than the gap between Mickey Mouse and the real world.
— Wang HeThe biggest chaos is not being able to tell who is creating productivity
Wang He says the biggest problem now is: who can generate productivity, and who cannot and is just telling stories — this is the messiest part, and it originates from the US — the US has a stronger tolerance for innovation, it's fine if you don't make money, fine if you have no product, as long as you can still sell. He takes Figure as an example, saying there are two logics for its valuation: the positive logic is that if humanoid robots can really work on production lines, the long-term value should be in the trillions of dollars, and today's 30-40 billion valuation still has a hundredfold growth; the negative logic is whether it actually has working capability, having shipped only 10 or 20 units, and the work done is not done by the method it claims. The domestic problem is seeing this talk abroad and also exaggerating one's own hardware and software capabilities.
— Wang HePublic demonstrations must not allow teleoperation — that's the first gate
Wang He proposes two verification standards. First is public demonstration: at the BAAI conference, tens of thousands of viewers surround you and watch on site, and teleoperation is not allowed — he says the US has normalized teleoperation, they tell you 'I am teleoperating, just with the person hidden behind'; some people in China are very bold now, not telling others they are teleoperating, but actually teleoperating. Second is that you really enter the venue, how much work is done there every day, and whether there are long-term reports that investors and the public can understand, like their store doing hundreds of orders a day, all certified by the platform and operating normally. In contrast, what is said in videos, or signing strategic agreements claiming to have done something, is increasingly unconvincing.
— Wang HeIf 10,000 units can't be done in five years, the industry is falsified
Wang He gives a clear time criterion: within five years, peers in China and the US must have applications of over 10,000 units, and the leaders must have at least 10,000 units of autonomous robot applications in that year. If it cannot be done in five years, the industry will likely be like industrial vision — initially telling a story of tens of billions, finally finding only a few hundred million in revenue, the time cycle will be greatly extended, and everyone may lose enthusiasm. He repeatedly emphasizes that Galbot must have productivity now, because if 10,000-unit-scale productivity cannot be formed in five years, this field is falsified, the bubble is all bubble. His reason is China's aging and low birthrate: in ten or twenty years, the labor force may be less than half of today's.
— Wang HeJensen Huang invited him to dinner because Nvidia is betting on synthetic data
Wang He explains why Jensen Huang invited him: Nvidia is the major company besides Galbot that values the synthetic data technical route, and their reasoning is simpler — if you have GPUs, they should solve all problems; if you need something else besides GPUs, it won't advance as fast. Synthetic data is using GPUs for rendering and simulation, then using higher-level GPUs for computation, so Nvidia strongly agrees that if synthetic data can be made to work, Nvidia alone can support half the sky of embodied AI. Before this, Nvidia's robotics vice president and specialists visited Galbot multiple times, and only after seeing it with their own eyes did they call him to dinner with Jensen Huang, seated next to him as arranged by Nvidia.
— Wang HeIn their own words · checked verbatim
I can't do the daily math. In this scenario, I solve every skill you want, one by one, using various methods — data-driven, or trajectory viewpoint, drag and replay, anyway I get it done. But the solution for this one scenario, when it goes to the next scenario, is all redone from scratch.
我算不过来日账 就是我现在在这个场景里 我全部你要的所有技能 我挨个都给你这个用 这个各种方式 不管数据驱动 还是用这个就是轨迹视角 拖拽 重放 总之我都给你搞定了 但就这一个场景的解决方案 它到下一个场景 全从头重做
Wang He1:46:28
Everyone now can make a demo, but they all film it into a video for you to watch. They cannot invite you to watch live. They cannot do public demonstrations.
大家现在呢 能做一个demo 都是拍成视频给你看 无法邀请你现场看 无法做公开展示
Wang He1:57:36
It's a closed loop that can be told: I don't believe in real data, I tell you real data is useless, you shouldn't believe it either, you just buy my machine, then you go teleoperate. Once you've teleoperated enough, the skill naturally emerges. That kind of business model.
它是一个能够讲的闭环 就是我不信核数数据 我告诉你核数数据没用 你呢不要信 你呢就给我买我的机器 然后呢你就去摇 摇够了自然这个技能就出来了 这种商业模式
Wang He1:59:37
A VLM that can understand the real world — when it watches Donald Duck and Mickey Mouse, can it not understand the plot? It can.
一个能看懂真实世界的VLM 他看唐老鸭米老鼠 他就看不懂那个剧情吗 看得懂
Wang He2:05:46
What is productivity? It's that in a unit of time, the work you do must be comparable to a human. If you are much slower than a human, or you don't work as long as a human, you are not quality productivity.
什么叫生产力呢 就是你在单位时间 你干的活 必须得跟人相当 如果你比人慢太多 或者你干的不够人久 你就不是优质生产力
Wang He2:12:52
If in five years we cannot form 10,000-unit-scale productivity, we here are falsified again, the bubble is all bubble.
五年如果我们都不能形成万台级的规模化生产力 我们这里又被证美了 泡沫全是泡沫
Wang He2:19:53
Don't do things that smash our industry's signboard. For example, promising others that if they step on it, it can be trained; if you build a factory, you can have skills; I sell robots and you step on them, you step on them and train them, tomorrow it will be your employee. These models are terrifying. These models are smashing the rice bowl of this industry.
不要去搞一些砸我们行业招牌的事情了 比如承诺别人 你踩了就能训出来 你建厂你就能够有技能 我卖机器人你来踩 你踩你来训 明天他就是你的员工 这些模式是很可怕的 这些模式是在砸这个行业的饭碗
Wang He2:22:54
Figures
| Wang He's age | 33 years old | 3:02 |
| Galbot valuation | over $1 billion | 3:02 |
| Manufacturing cost of one full-size humanoid robot | at least 100,000 | 1:55:35 |
| Monthly cost of maintaining 10,000 robots | hundreds of millions to a billion | 1:56:35 |
| Share of real data in Galbot's training data | 1% or even less | 1:58:36 |
| Figure valuation | 30-40 billion USD | 2:14:52 |
| Figure shipments | 10 or 20 units | 2:14:52 |
| Last year's global industrial robotic arm output value | 100 billion RMB | 1:34:20 |
| Galbot's mass production scale this year | thousand-unit level | 1:39:23 |
Glossary
- sim2real
- Transferring policies trained in a simulator to the real world so they still work.
- VLA / vision-language-action model
- A large model that maps visual and language inputs directly to robot actions.
- perception action loop
- The cycle where perception determines action, and action changes the environment, which then updates perception.
- object goal navigation
- A navigation task of finding a specified object in an unknown environment.
- six dof pose
- An object's position plus rotational orientation in 3D space.
How to listen
Founders and investors in robotics and embodied AI, especially teams agonizing over whether to build their own real-data collection pipeline, or stuck in the sim2real debate.
The first 40 minutes of academic marginal history and the rapid-fire Q&A can be fast-forwarded; the judgment starts after 1:44.