Interesting article. A couple of thoughts. Would you know in which order Chinese car companies entered robotics? I.e. which one first, Xpeng, NIO (even just by investing in robotics companies) or Li Auto?
Car companies that are advanced in autonomous driving have very large data to interfere and apply to humanoid robots, especially those companies that shifted early on their lidar approach to intelligent “vision”.
That means a robot can move within any environment/settings (including stairs) and generate in real time a digital twin of its context.
Also, physical AI, approach means humanoid robots teach themselves through virtual or real world interactions. That itself is huge.
Pure play humanoid robot companies lack that vastly amount of data and will take them a long time to catch up.
Finally the gala robots show with martial art was indeed impressive but also very much rehearsed like a choreography.
There is no doubt in my mind that China will lead technological advancement in humanoid robots too as it has done with Solar Power, EVs, etc
You’re right that vision-based autonomous driving data is a real asset. XPeng’s argument for software reuse is partly built on exactly this: shared perception pipelines, sensor fusion, and the data flywheel from millions of kilometers driven.
Where I’d push back slightly: the data that makes a car see well on roads is optimized for forward motion, lane boundaries, and traffic patterns. The manipulation tasks that define useful humanoid robots — gripping irregular objects, applying variable force, working within arm’s reach of people — require a different kind of embodied data that road driving doesn’t generate. That’s the gap I think is underappreciated.
On your question about sequencing: XPeng has been the longest in direct robot development. NIO’s involvement has been primarily through investment rather than internal building. Li Auto is the most recent entrant, and notably the most aggressive in terms of internal restructuring.
On China leading the way: the scale of parallel experimentation is genuinely significant. 15 companies running the same hypothesis under real market pressure will produce data faster than any single player. Whether that translates to leadership depends on which problems turn out to be hardest. That’s what I’m watching.
Thank you Poe for your detailed response. I agree completely on the remaining and decisive challenges, I am quite optimistic on Chinese scientists and the speed of advancement in physical AI.
Interesting article. A couple of thoughts. Would you know in which order Chinese car companies entered robotics? I.e. which one first, Xpeng, NIO (even just by investing in robotics companies) or Li Auto?
Car companies that are advanced in autonomous driving have very large data to interfere and apply to humanoid robots, especially those companies that shifted early on their lidar approach to intelligent “vision”.
That means a robot can move within any environment/settings (including stairs) and generate in real time a digital twin of its context.
Also, physical AI, approach means humanoid robots teach themselves through virtual or real world interactions. That itself is huge.
Pure play humanoid robot companies lack that vastly amount of data and will take them a long time to catch up.
Finally the gala robots show with martial art was indeed impressive but also very much rehearsed like a choreography.
There is no doubt in my mind that China will lead technological advancement in humanoid robots too as it has done with Solar Power, EVs, etc
Thanks for the thoughtful pushback, Luca.
You’re right that vision-based autonomous driving data is a real asset. XPeng’s argument for software reuse is partly built on exactly this: shared perception pipelines, sensor fusion, and the data flywheel from millions of kilometers driven.
Where I’d push back slightly: the data that makes a car see well on roads is optimized for forward motion, lane boundaries, and traffic patterns. The manipulation tasks that define useful humanoid robots — gripping irregular objects, applying variable force, working within arm’s reach of people — require a different kind of embodied data that road driving doesn’t generate. That’s the gap I think is underappreciated.
On your question about sequencing: XPeng has been the longest in direct robot development. NIO’s involvement has been primarily through investment rather than internal building. Li Auto is the most recent entrant, and notably the most aggressive in terms of internal restructuring.
On China leading the way: the scale of parallel experimentation is genuinely significant. 15 companies running the same hypothesis under real market pressure will produce data faster than any single player. Whether that translates to leadership depends on which problems turn out to be hardest. That’s what I’m watching.
Thank you Poe for your detailed response. I agree completely on the remaining and decisive challenges, I am quite optimistic on Chinese scientists and the speed of advancement in physical AI.