He Xiaopeng: General-purpose humanoid robots are 99.99% doomed to die — we're betting on that 20% win rate
XPeng shut down an autonomous driving system that cost it a few billion RMB, because what you get from a software methodology plus an AI toolbox is just an "AI chimera." He puts the odds for general-purpose humanoid robots at two in ten, and says robot startups are 20 to 100 times harder than building cars.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
What you build with an AI toolbox is still software
He Xiaopeng calls the old autonomous driving approach — "AI algorithms plus software rules" — a chimera: it isn't designing the machine with AI driving the whole thing, it's still software-first logic. He offers an analogy: you're just building a more complicated chimera inside software logic, like renovating a house — you'll use more materials, more craft, use AI, but it's still the methodology of renovating an old house, and what comes out is still the original house, just maybe renovated a bit faster. So last year XPeng stopped that old system, and that system cost a few billion RMB.
— He XiaopengThe new route's floor is only 100 points
He Xiaopeng uses a scoring analogy: the old route had a ceiling of 1,000 points and a floor of 900, decent capability; the new VLA route might have a ceiling of 100,000 to a million points, but at that point the floor is also brutal, maybe only 100 points, lower than the floor of the company's other products, and there are a great many engineering problems. He still made the bet, because he judged the old method would "never be infinitely smart." His concrete example: today no autonomous driving company's software can drive smoothly in an underground car park — everything that drives in car parks is memory-assisted driving, it knows your parking spot and roughly the route after one run, and its understanding of the physical world is very low.
— He XiaopengThe physical world makes you fight on three boards at once
In the digital AI market many models are just benchmark versus benchmark, and the core is the long board; the physical world is ridiculous, because you're not only comparing ceilings, you're comparing floors and breadth — quality, cost, materials, details, what policy and regulation allow are all short boards. His exact words: "The long board and the narrow board — the narrow board has to be made wide, the short board has to be made long, and the long board has to be made even longer." This is also why he judges that companies in the physical world either "don't dare to bet, or feel they have so many boards, what do I do." He also offers a value judgment: today hardware accounts for far more than 50% of what the customer needs, and in the next decade software may account for 50%, with users willing to pay half a car's price for software's central capability.
— He XiaopengThe biggest opposition is voting with your feet
He Xiaopeng says there was certainly opposition inside the company when he made the bet: most non-AI executives felt that whether you did A or B you might be wrong, because at that time their understanding of AI wasn't strong enough; most AI-related executives were in a middle state, unsure whether this was a good pace. The biggest opposition, he says, was that "a lot of people voted with their feet" — they didn't believe in this thing, felt it couldn't be done, and left to do other things. His line on the organisation is "be sure not to cut a big tree with a small knife — cut slowly, and once you've thought it through, cut it off." Last year the knife went from the business layer all the way down to the root.
— He XiaopengA 300-person team kept fewer than 60
In 2023 XPeng went from not believing in brains to believing in brains, and the method was to keep fewer than 60 people out of the original 300-person robotics team and dissolve the rest. He Xiaopeng says the judgment then was: a team that understands robots very well and is very familiar with them has no ability to build a good, all-new-generation robot. The new lead he picked understood a bit of AI, a bit of cars, a bit of engineering, a bit of robotics, on the grounds that "his direction line matched my thinking line fairly well." He mentions that ten startup teams came out of the people who left, and most of them got funding.
— He XiaopengRobot startups are 20 to 100 times harder than carmaking
He Xiaopeng gives a very wide-ranging judgment: robot startups are roughly 20 to 100 times harder than car company startups, and he stresses, "I even gave a minimum of 20 times." The reason is that the pits robots fall into aren't ones you can avoid just by understanding them — you only grasp the whole logic after you've actually stepped in them. He also gives a survival-rate judgment for general-purpose humanoid robots: companies taking the general-purpose humanoid route today are 99.99% doomed to die, but differentiated routes may have a great many solutions. XPeng itself chose the hardest one, and he puts the win rate at two in ten, saying that is already the highest he has seen among Chinese companies.
— He XiaopengRobot motion control is still in the Model T era
He Xiaopeng thinks most people underestimate robot motion control. Car motion control has been considered solved by many over the past hundred-plus years, so many companies just buy off-the-shelf motion control capability and combine it; robots, by contrast, may be working from some open-source motion control from 2018. His time coordinate: robots are still at roughly the car of the 1930s, the 1920s, maybe still the Model T era. His example: an autonomous car with its left tyre on snow and right tyre on grass, needing to take a 47.5-degree turn — how should it turn? That's hard even for a person; what a robot needs is full-posture, all-AI-combined motion control, probing friction like a human instinct.
— He XiaopengEven if L4 is achieved, it doesn't mean long-term value
He Xiaopeng judges L4 will arrive in roughly 18 to 24 months, but immediately pours cold water: even with more sales, it doesn't mean long-term value. His data: at the end of March last month the first version of the second-generation VLA was released, and in April China's auto sales fell about 20% year-on-year and month-on-month, while XPeng rose roughly 50% to 70%, a considerable part of which was related to the second-generation VLA. But he asks back: if L4 is achieved, whether that capability can be raised by one time, five times or ten times, he finds it hard to judge whether that's good or bad, and views on the impact on society and the industry are not consistent either. His conclusion is that the ultimate success of a car and a company doesn't come from just one capability in AI.
— He XiaopengIn their own words · checked verbatim
Because you're using a software methodology and an AI toolbox, what you produce is a more powerful piece of software — I call it an AI chimera.
因为你是用软件的方法论，使用AI的工具箱，做出来是一个更强力的软件，我认为它叫做AI缝合怪。
He Xiaopeng17:13
Personally I'd say a robot startup is roughly 20 to 100 times harder than a car company startup, and I even gave a minimum of 20 times.
我个人来看机程的创业大概是汽车公司创业难度的20到100倍，我还给了一个最低20倍。
He Xiaopeng57:45
I think today's robot companies, in my own view, if they take the general-purpose humanoid route, 99.99 will die.
我觉得今天的机型公司，我自己觉得，如果走通用能行机型，99.99会死掉。
He Xiaopeng59:45
Have you swum out of the sea of blood today? I think everyone's still swimming. I think — out yet? No. And I don't think any blood has come out either.
你今天从血海里游出来没有，我觉得都在游，我觉得，出来了没，没有，我也不觉得有血出来了。
He Xiaopeng1:24:04
Figures
| Investment in the old autonomous driving system that was shut down | a few billion RMB | 17:13 |
| XPeng's direct rigid annual cost on data | close to a billion RMB or more | 7:04 |
| Target for AI-related investment as a share of the company | 15% to 20% | 11:07 |
| Robotics team headcount retained | 300 people, kept fewer than 60 | 44:36 |
| Number of newly hired PhD graduates | close to 80 in one department | 47:38 |
| Robot startup difficulty relative to carmaking | 20 to 100 times | 57:45 |
| Judgment on the death rate of general-purpose humanoid robot companies | 99.99% will die | 59:45 |
| XPeng's robot win rate | two in ten | 1:07:51 |
| Judgment on when L4 arrives | 18 to 24 months | 1:18:58 |
| Change in China's auto sales in April | down about 20% year-on-year and month-on-month; XPeng up 50% to 70% | 1:19:59 |
Glossary
- VLA / vision-language-action model
- A route that first opens up the ceiling of autonomous driving with a larger foundation model, then converges the floor.
- Chimera / hunhe guai (缝合怪)
- He Xiaopeng's coinage for a more powerful piece of software assembled from a software methodology plus an AI toolbox.
- SOP / start of production
- A key milestone in the car's path from design to mass production; robots also have to reach this step.
- T1 / tier-one supplier
- A supplier that the OEM deals with directly; He Xiaopeng says this setup doesn't work in robotics.
How to listen
Founders and investors watching route choices in humanoid robotics, the autonomous driving technology transition, and organisational change at large hardware companies.
The GX new-car segment starting at 1:09:20 is more product introduction; you can fast-forward.