The hardest part of a humanoid robot isn't walking, it's threading a needle and peeling a shrimp
Bipedal walking and autonomous navigation can both inherit from Tesla's self-driving stack; precise force control and touch cannot — that requires simulating human skin, which driving never needs. The first- and second-generation robots are probably only at 20% to 30% of a human.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
It can inherit navigation, but not the sense of touch
To judge what this Tesla robot can do, you first have to separate which parts can be carried over from self-driving. Bipedal standing and walking and gait control rest on decades of research from Boston Dynamics and Waseda University; real-time modeling in complex environments, autonomous navigation, and not losing your own position are ‘very close to Tesla's self-driving capability’, and the underlying Dojo, D1 chips and vision algorithms can all be reused. Two things can't be carried over: first, precise force control — getting the robot to wrap an egg, wrap a crawfish, or install something with elasticity — which requires simulating human skin and touch, and which every company today does only at a very basic level; second, fine control tied to vision, since Tesla's self-driving camera resolution is still very low, and reaching the granularity of threading a needle requires iterating the hardware from scratch.
— Cao WeiThe first-generation robot can't embroider
How fine can fine control actually get? Cao Wei uses a human as the yardstick: 5%, 10%, or 50% of a person. His estimate is that the first and second generations reach only about 20% to 30% of a human — that is, very basic box-carrying, large-scale operations that don't need precise alignment. The most-performed Boston Dynamics trick is also box-carrying; nobody has ever performed watermelon-slicing. So his expectations for Tesla's September 30 robot are very specific: body shape and foot size have to match the mechanical structure, don't be ‘sparse on top and one big foot below’; being able to walk quickly around the venue without losing itself and do basic box-carrying would already be very impressive. Threading a needle, embroidering, cleaning up a bathroom — he says all of those are impossible.
— Cao WeiSaying humanoid robots will upend everything is amateur talk
Some articles say that the moment Tesla ships a humanoid robot, traditional robot companies have no room to survive. Cao Wei's judgment is that ‘a lot of people don't really understand, they're just fantasizing’. His framework: every robot is designed around the pain points and difficulties of its own scenario, so it is itself the ultimate solution for that scenario. The four-layer cart that moves wafers in a semiconductor fab can carry many wafers at once and has extremely high vibration-control precision requirements — a humanoid robot doing that is not cost-effective; the flatbed forklift robot in a factory can pull three TVs plus two big boxes at once, and ‘you can't have 500 robots in one cargo hold, running around on two legs, each carrying just one box’. Heavy industry, welding, high-payload industrial robots all have nothing to do with humanoid form. He has an even more abstract line: in the end the robot's intelligence is all in the cloud, the form doesn't matter, and humanoid is the optimal solution only when the task can only be completed in human form.
— Cao WeiTesla's big flat-panel face sits in the uncanny valley
The uncanny valley is a curve: no matter how extreme you make a large robot, it still doesn't look human, and people instead find it a beautiful, perfect machine and like it a lot; but once it's ‘both like and not like’ — nose and eyes, weird movements, creepy gaze — people's scalps crawl, and this is ‘a very big innovation trap’ for humanoid robots. Cao Wei thinks we are in that stretch right now: Tesla's robot has no face yet, just a big flat panel, and at night a glowing black shadow pacing around the house every day would be unpleasant. To skip past this stretch, experts say 10 years, some say 20, or even longer. Privacy he isn't worried about: an iToF camera gives depth information but not texture information, and only after the owner authorizes it does it overlay ordinary RGB — the robot knows this is a box, but not that it says cash on it.
— Cao WeiA delivery robot's value equals the length of the aisle
What BlueRun first invested in back in 2017 was cleaning robots; it looked at delivery but was very hesitant. Cao Wei breaks the value of a delivery cart down cleanly: getting the food out requires a waiter to put the dish on the cart, serving it requires a waiter to carry it from the cart to the guest, and the cart contributes only the stretch of corridor from point A to point B. ‘That is, the longer your aisle, the greater the cart's value; the shorter your aisle, the smaller the cart's value’ — and restaurants in Beijing, Shanghai, Guangzhou and Shenzhen are small with narrow aisles, yet waiters there are the most expensive and the busiest, so the place that needs it most doesn't match; instead it's stores with especially long aisles, or huge spaces like Vegas casinos, where the cart makes sense. He mentions that the companies doing well in this area now are Keenon and Pudu. The cleaning scenario got invested in because the service process is unrelated to user experience: in the morning you just see the floor is clean, and nobody cares how long the robot dawdled the night before.
— Cao WeiThe next wave of form is soft, not a lump of iron
The soft robotics company BlueRun invested in has ‘not just a soft hand, the whole arm is soft’, like an elephant's trunk, able to wipe a table and twist off a gas cap when refueling. Why this is a direction: past industrial robotic arms were designed around standardized assembly lines, able to lift two or three kilograms while weighing perhaps 5 kilograms themselves, all lumps of iron; it's hard to imagine that the service robots everywhere in future life will all be like that — a child takes a tumble and cracks their head open. Soft robots have no motors inside, and the cost structure is completely different — a robotic arm has to sell for 20,000, 30,000, 40,000, while these might be 5,000 to 10,000, and a very small soft robot costs only a few hundred yuan. Cao Wei says they will invest in such companies when the product has just come out of the lab and there is still no revenue.
— Cao WeiChina can build a robot in just seven or eight months
For China's systemic opportunity in this wave, Cao Wei gives several supports: on the demand side, developed countries generally face labor shortages; on the market side, 40% of the world's industrial robots are sold to China, and about 25% of service robots, so China is itself the largest robot consumer; on the community side, over the past 5 years the number of registered robotics-related startups is in the hundreds of thousands, and the robotics field took about 18 billion in investment in 2021 and nearly 100 billion over the past 5 years; on the talent side, more than 400 universities have set up robotics engineering majors, still growing at about 10% a year; on the industrial ecosystem, the Yangtze River Delta, Pearl River Delta and Xi'an all have industry clusters, close to both the supply chain and customers. The result is that a robot can be built in China in 7 to 8 months, versus a year and a half to two years overseas — iteration speed is double that of overseas.
— Cao WeiHold one scenario and even Tesla can't do it
Cao Wei doesn't think robots will be winner-take-all the way the internet was. The dividing line is: if Tesla ultimately gets a humanoid robot down to 20,000 yuan apiece, then it's a highly general, scenario-independent product, and it might be winner-take-all again; but industrial robots and vertical service robots are contextual, with scenario-based functions and design — semiconductor scenarios, 3C scenarios, new-energy photovoltaic scenarios, cleaning scenarios — and ‘whoever holds this scenario may become the king of this scenario in the global market’, with first-mover advantage and accumulated depth as the position. The track risk he gives follows the same logic: the risk is pacing — as long as you survive and persist for 30 years, this thing will certainly have value, and many companies die because they don't manage their pacing well through the fluctuations.
— Cao WeiIn their own words · checked verbatim
You can't expect that he puts out a robot in September and this robot can embroider, right? Sit down there, pick up a needle, and just embroider a flower right out. I don't really expect that.
你不可能上来他9月份发个机器人 这机器人会绣花 对吧 往那一做拿个针 就直接绣个花出来 我不太期待
Cao Wei10:07
You can't have 500 robots in one cargo hold, running around on two legs, each carrying just one box, right? That's not logical.
你不可能一个货舱里边500个机器人 两条腿在那跑 每个人只抱一个箱子 对吧 这不合逻辑
Cao Wei16:11
Why do I say the robot ends up in the cloud? Its form doesn't matter; its form is for completing the key tasks of human society.
为什么我说最后机器人是在云端的 它的形态并不重要 它的形态是为了完成人类社会的关键任务
Cao Wei27:16
That is, the longer your aisle, the greater the cart's value; the shorter your aisle, the smaller the cart's value.
就是说你锅道越长 小车的价值越大 锅道越短 小车的价值越小
Cao Wei32:20
The soft robots we invest in — not just the hand is soft, the whole arm is soft — it can wipe a table, and when refueling, twist off the gas cap.
我们投的软体机器人 就是不光手是软的 整个胳膊也是软的 它就可以擦桌子去加油的时候 拧油箱盖
Cao Wei37:22
In China it takes about seven or eight months to build a robot; overseas it might take a year and a half to two years.
国内大概七个月八个月做出来一个机器人 海外的话可能要一年半到两年
Cao Wei50:29
This is the kind of thing where whoever holds this scenario may become the king of this scenario in the global market.
这种就是谁把这个场景站住了 谁有可能成为这个场景在全球市场的王者
Cao Wei54:30
Figures
| Research period for bipedal gait control | nearly 40 to 50 years | 3:05 |
| Fine manipulation level of first- and second-generation robots | 20% to 30% of a human | 10:07 |
| Expected duration of the uncanny valley | some say 10 years, some say 20 years, or even longer | 21:14 |
| Market-rumored cost of the Tesla robot | 200,000 to 300,000 US dollars | 23:14 |
| China's share of global industrial robot purchases | 40% | 47:28 |
| China's share of global service robot purchases | about 25% | 47:28 |
| Robotics-related startup registrations over the past 5 years | in the hundreds of thousands | 48:28 |
| Robot product iteration cycle | 7 to 8 months in China, a year and a half to two years overseas | 50:29 |
Glossary
- uncanny valley
- The more a robot resembles a human without fully resembling one, the more fearful and repelled people become
- iToF
- Indirect time-of-flight camera: outputs only depth information, not texture information, and can be used for privacy protection
- SLAM
- Simultaneous localization and mapping: the robot builds a map from vision while moving and confirms its own position, replacing guide rails
- force control
- The fineness with which a robot's applied force is controlled; threading a needle, peeling a shrimp and twisting off a gas cap all depend on it
How to listen
Investors and founders watching the robotics track, and hardware product managers who want to know where humanoid robots are actually stuck and how soon they reach the home.
From 44:25, the stretch on ‘service robot companies looking more and more like consumer electronics’ is abstract; you can fast-forward.