We may not see general-purpose robots in our lifetime
李飞飞收购郑昌熙：有生之年，我们可能看不到通用机器人出现
B · Useful, limited Long interview with Changxi Zheng, a Columbia computer science professor now at World Labs, giving a reasoned practitioner case on industrial robotics, with argument but little data.
Summary based on the full transcript·Original: about 60 minutes to listen
In short
An hour-long Mandarin interview on the channel 课代表立正, published 9 October 2026, with Changxi Zheng, professor of computer science at Columbia University, co-founder of the robotics simulation start-up SceniX and now full-time at Fei-Fei Li's World Labs. He argues that industrial robots will not follow the language-model curve to a single general model. Factories need policies that are robust and faster than people, and economic specialisation favours task-specific systems. He says no embodied-AI company, including his own, is yet used at scale in real production.
Key numbers
- 3 of 10failure rate he calls unacceptable for an industrial robot policy
- 80-90%share of industrial tasks he does not expect a single foundation model to handle at production reliability soon
- GPT-2 stagehis comparison for where embodied AI stands today
As reported by the source.
Chinese·Video·opens youtube.com
What it reports
- From running an incubation lab for Tencent in New York: in business, technology is "not the most critical or useful thing, not even the second".
- On academia: evaluation counts output, so resubmitting a paper to new reviewers is "buying a lottery ticket".
- SceniX builds a real-to-sim-to-real workflow: capture a real task cheaply, rebuild it in simulation, generate data, evaluate, then transfer a policy back.
- Industrial policies must be robust (failing 3 or 4 times in 10 runs is unacceptable) and faster than a person, or the ROI does not work.
- Three axes: generalisation, robustness, throughput. He expects no model like ChatGPT to handle 80-90% of industrial tasks at production level soon.
- Citing Adam Smith on specialisation, he argues mass production will keep needing specialised policies; home robots differ because buyers want one machine.
- Limits beyond data: motor control runs well above a game engine's 60 Hz, hardware moves slowly, and the physical world lacks an agreed 3D representation.
- World model, in his definition: given a state and an action, predict the next state. Video prediction alone is not one, and training one needs "god's-eye" experience, a chicken-and-egg problem.
- He rejects the view that robotics must happen in China: Western labour costs and unions make automation more urgent there.
How they got it
A conversational interview drawing on his research and start-up experience; there is no dataset. The digest uses cleaned auto-captions of the Mandarin audio, which mixes in English terms.
Limits
- Opinion from one researcher whose company sells a simulation-first approach.
- No figures on deployments, costs or failure rates beyond illustrative examples.
- The video title's framing of an "acquisition" by Fei-Fei Li is not explained in the conversation.
Why it matters for companies
Manufacturing and logistics managers face many claims of near-general robots. His test is practical: does the robot finish the task reliably and faster than a person at lower cost? By that standard, near-term value lies in specialised automation for defined tasks, with simulation used to train and test before deployment. Budgets built on a general-purpose robot arriving soon carry high risk.
Key passages in translation
Chinese original
in terms of business(就商业而言) technology(技术)真的真的不是最 关键或者最有用的东西 甚至都不是第二有用的东西
In terms of business, technology really, really is not the most critical or most useful thing; it is not even the second most useful thing.
Chinese original
跑十次可能坏个三次四次对吧 那这无法接受的 然后还有就是需要它足够快
If it runs ten times and fails three or four, that is unacceptable. And it also has to be fast enough.
Chinese original
跟人的速度是一样的 其实这个ROI(投入产出比)都是算不过来的对吧 我不如雇一个人去做这件事情
If it works at the same speed as a person, the ROI simply does not add up; I would rather hire a person to do the job.
Chinese original
很难有这么一个 像ChatGPT一样的一个model(模型)可以 Generalize(泛化)到做各种各样的 尤其是在工业场景当中的这些工作
It is hard to have a model like ChatGPT that can generalise to doing all kinds of work, especially work in industrial settings.
Chinese original
所谓的embodied AI(具身AI)的公司 包括我们自己 就坦率地说 都还没有在真实的生产环境中大规模用起来
The so-called embodied-AI companies, including ourselves, frankly, are not yet being used at scale in real production environments.
Chinese original
如果一定要用ChatGPT来举例的话 我觉得就GPT-2这种阶段吧 GPT-3都还没有到的阶段
If I must use ChatGPT as a comparison, I think we are at about the GPT-2 stage, not yet even at GPT-3.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- AI capability