Let a robot peel a potato and you can be an academician — that's harder than going to Mars
Humanity already has most of the technology needed to go to Mars; it's just an economic problem. But how to make a robot work in a kitchen is something humanity does not yet know — has not even got an approach. Whoever can make a robot peel a potato can be an academician.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The humanoid robot is like the moon landing: useless in itself
The guest compares humanoids to the moon landing: the landing itself produced no serious value — you just haul back a bit of soil from the moon — but it generated a lot of technology. Ready meals were first developed for spaceflight, because on the ground you don't need ready meals, but once you go up there you must have them. The same goes for humanoid robots: it may not immediately produce enormous value in itself, but it can pull along a whole chain of technologies and attract talent. His order of magnitude: for it to be truly mass-producible and usable, maybe 10 years, maybe 20 — it will be very far off. So he understands it as a kind of market behaviour — you need a goal big enough that everyone can see it.
— Shao TianlanMusk talks about Mars, but actually launches satellites
The same rocket: described as "we may need hundreds of thousands of people, we need to transport at least hundreds of thousands of people to Mars," it is a grand dream. Stripped down, what he actually does is launch low-Earth-orbit satellites and ferry cargo to the space station — things China, the US, Russia and Europe can all do; the remarkable part is only making a rocket land upright and be reusable. The guest says if it were up to him to tell the story, he would only talk about more cost-effective rocket launches — "deadly vulgar, you're just the Pinduoduo of rockets." The grand narrative buys three things: more talent, more capital, and greater outside tolerance for failure. Musk ran this playbook once on electric cars, and on robots it may be another run.
— Shao TianlanMake a robot peel a potato and you can be an academician
The gap between the robot the public feels from videos and a real professional robot is on the order of decades. He cites the DARPA Robotic Challenge: that video is played at 18x speed, and you still find it slow. Getting a robot to take the washed clothes out of the washing machine, shake them out and hang them up can make you an academician; getting a robot to pick up a potato in its left hand and a peeler in its right hand and peel the potato can also make you an academician. Three extremely difficult things hide in there: in-hand manipulation, two-handed coordination, and force control. He says he is absolutely not exaggerating: the demands robots place on sensing, perception, planning and execution far exceed today's human技术水平.
— Shao TianlanA robot working in a kitchen is harder than going to Mars
He runs a comparison between autonomous driving and robotics along the dimensions of sensing, perception, planning, decision-making and execution. Autonomous driving's sensing has to see 300 metres, but very coarsely, and on perception it only needs to recognise two classes of thing: obstacles and lane lines. A robot has to see the pits on a potato, and has to tell laundry detergent from conditioner among a pile of bottles — "it even has to read." So the difficulty of a humanoid robot is no lower than, even higher than, autonomous driving. Compared with going to Mars it is another dimension again: most of the technology needed to go to Mars already exists — it is only an economic problem; but how to make a robot work in a kitchen is something humanity does not yet know, has not even got an approach to.
— Shao TianlanRobots today are computers in 1980
He benchmarks the current stage to computers in 1980: big American companies were already using them, small and medium firms and individuals had not yet. The reference points are the adoption curves of computers and smartphones — in 1997 using a computer for office work was a remarkable thing, by 2002 it was nothing; in 2006-07 using a smartphone was very advanced, by 2011 it was nothing. Robots are walking the same road: just a few short years ago they were the exclusive preserve of big companies like car plants, and today they are migrating to medium-sized and even small firms. He admits customers like Geely, Sany and Toyota are still big companies, but his example is a small factory that sprays glue on seat foam — even that kind of factory is using robots now.
— Shao TianlanA robotics startup usually gets only one shot
Internet companies rarely die for technical reasons; robotics startups do. The product involves optics, mechanics, electronics, software, algorithms — "a very long noodle" — and that one shot may cost you three years. His company was founded at the end of 2016, and only by 2019 was the product actually sellable. "If the product I shipped then hadn't worked, I had no resources to fire a second shot." He estimates 80% of companies die at the first hurdle: if the first product does not preliminarily prove itself (not required to sell in volume, only that someone buys it and it shows promise), you cannot keep going. Musk can try a second time, a third time; a startup does not have that option.
— Shao TianlanMusk and Jobs: whoever imitates them dies
He divides formidable entrepreneurs into two kinds. Lei Jun is "the ceiling of the imitable" — his diligence and his way of doing things can all be learned. Musk and Jobs are extraordinarily gifted, not imitable; whoever imitates them dies. He uses a basketball analogy: Yao Ming's technical moves can be learned, but Allen Iverson's and Kobe's cannot — the talent is too outrageous. "Look at the people imitating Iverson — every single one of them, most of them are idiots." But not imitable does not mean not worth watching: Musk is someone who can combine technology, product and business, "telling the most badass story, then doing the most solid kind of iteration" — electric cars going from luxury cars, to high-end cars, to a 200,000-yuan car, step by step.
— Shao TianlanIndustrial robots served car plants for forty years and nothing else
Industrial robots have essentially taken two steps so far. The first step was meeting the needs of the automotive industry, from 1969 all the way to 2010; over those forty years, the revenue contributed by the automotive industry went from nearly 100% at the start to still over 70% in 2010. All the robot giants came out of cars — Europe's first robot was made by Kuka for Daimler-Benz; the world's robot giants are Germany and Japan, because the giants of the car industry are also Germany and Japan. The second step is the last decade's move into home appliances and other industries, driven by rising intelligence and falling costs.
— Shao TianlanIn their own words · checked verbatim
You take a team, and you can get a robot to take the washed clothes out of the washing machine, shake them out and hang them up — this can make you an academician.
你带一个团队 能让机器人从洗衣机里面 把洗好的衣服拿出来 抖干净 挂起来 这个事情可以让你成为院士
Shao Tianlan9:03
How to make a robot work in a kitchen is something humanity does not yet know, has not even got an approach to.
怎么让机器人在厨房里工作 是人类现在还不知道的 甚至还没有思路
Shao Tianlan18:05
It has no perception, it just executes a fixed program, so we consider it a machine, not a robot.
它没有感知 它就是执行固定的程序 所以我们认为 它是个机器 不是一个机器人
Shao Tianlan22:06
If the product I shipped then hadn't worked, I had no resources to fire a second shot.
如果我当时出卖这个产品不行 我没有资源去开第二枪
Shao Tianlan40:12
Figures
| Chang'e lunar exploration programme budget | 1.4 billion RMB | 19:06 |
| Share of industrial robot revenue from the automotive industry (around 2010) | over 70% | 51:43 |
| Span of time industrial robots served the automotive industry | 1969 to 2010 | 51:43 |
| Death rate of robotics companies whose first product does not preliminarily prove itself (guest's estimate) | 80% | 40:12 |
| Time needed to validate a robotics product | at least two years plus, usually three years | 42:13 |
| Engineer headcount at a leading autonomous driving company | over a thousand engineers | 26:07 |
| Europe's first robot (made by Kuka) | 1971 | 53:14 |
Glossary
- in-hand manipulation
- Changing an object's pose in the hand without putting it down; the guest says this is extremely hard even in academia.
- DARPA Robotic Challenge
- A robotics competition funded by the US Department of Defense; the video shown in the episode is played at 18x speed.
- silver bullet
- A single decisive fix; the guest says no such thing exists in robotics problems.
- product market fit
- The hurdle where a product genuinely has buyers; the guest says this is the most dangerous place in robotics startups.
- DLR
- The German Aerospace Center; it has a robotics institute, the guest's advisor held a post there, and many robotics founders have a spaceflight background.
How to listen
Founders and investors working on robots, hardware or AI applications, and anyone trying to judge which parts of the humanoid robot narrative are real progress and which are just story.
The last two minutes are a Spanish-language robot song; you can skip it.