Discussion keeps the world turning.
This is Roundtable.
What if AI could step out of the screen and into the real world?
The new buzzword isn't just about smart or chat, it's embodied intelligence.
Have you heard of it?
We're talking robots that don't just calculate a chess move, but physically make that move themselves.
So what does it mean for AI to finally have a body?
Get ready for the leap from digital theorist to physical doer.
We're live from Beijing.
This is Roundtable.
I'm Steve.
Thank you very much for being with us today.
And for the show, I'm with Yushun and Feifei.
First up.
Forget the digital screens and purely abstract thinking of today's common AI assistants.
A new frontier is capturing the tech world's imagination, one that finally grants artificial intelligence a physical presence.
Welcome to the era of embodied intelligence.
We're moving beyond the talkers, like ChatGPT what experts call disembodied intelligence and entering the age of AI that can perceive, learn from and act within the real world.
But does every smart robot qualify as true embodied intelligence?
And is this concept simply a fancy term for an advanced humanoid robot?
We're going to unravel the realities of a technology that promises to move AI out of the cloud and into the tangible spaces of our lives, from your kitchen, for example, to the factory floor, to remote disaster relief zones.
There's a lot to get into today, Fei-Fei and Yuxin.
Good afternoon to you both.
But let's start with what it is.
What exactly is embodied intelligence?
Yeah, this word is absolutely one of the heated discussed words in recent years embodied intelligence or embodied artificial intelligence and it is a frontier domain intersecting artificial intelligence and robotics.
It defines an intelligent agent, such as a robot, a drone or a smart vehicle that achieves the integration of perception, cognition and decision-making and action by interacting with its environment in real time physically, through a physical entity.
And, in simple terms, the embodied intelligence is a smart system that is able to think, able to perceive and able to act.
Not like an abstract calculator.
Embodied intelligence moves into the real world and actively completing tasks by using its eyes to see, its hands to operate and body to move.
Just like we do.
Exactly.
So we can understand that, as AI now has a physical body that we can talk or touch maybe, and also it can think and also respond to our you know, i don't know movements or decisions or speech, just like we do, it reacts, i guess, to the real world, but it is still a little bit abstract.
So what would some good examples of this be?
Yeah um, and to help to first look at it um, you know it's, it's opposite to under the embody intelligence.
Better is The opposite way is disembodied intelligence.
Like chat GPT, like I said.
As you said.
But these things are just.
They're thinking things and they are probably just the brains of these embodied intelligence.
They can read text, understand images, talk to users and generate information, but they just exist purely in the digital world.
They're softwares.
They can think and talk, but they cannot directly touch or change the physical world. while um uh one classic example is alpha go that we kind of know for a decade ago yeah yeah that was alpha go so that was from google i think right yeah and there was a huge story in korea it was 2016. it was called the deep mind challenge they held it in seoul where alpha go played against the Korean Go champion.
The game is called Go, right?
Yeah.
Lee Se-dol, I think was his name, if I'm saying his name correctly.
They played a few matches against each other.
I think AlphaGo won the majority of the matches, like four out of five.
And then Lee Se-dol was able to win one of those matches.
But that, oh, that was huge news.
God, that was like, yeah, almost 10 years ago.
Yeah.
And also the same here in China.
I think AlphaGo also played against the Chinese Go player called Ke Jie.
Yeah.
And the AlphaGo also won that top player, internationally top player as well.
Yeah, this AI program that became really really famous for defeating some of these really top human players in this board game.
And it is just incredibly good at playing chess and it can calculate millions of their moves ahead.
The thing is that AlphaGo, when we were watching the matches that AlphaGo were competing with this, AlphaGo needs a real human to play the chess for it.
So it is just mapping out the moves that it's going to be. but it cannot actually move the piece.
Physically move the piece itself, right.
So it's a computer, essentially.
So then, I'm still a little bit confused about what's the difference between a robot and embodied intelligence.
So we talked on the show about the robot chefs that are cooking in some kitchens now here in China.
Would that be an example of embodied intelligence?
Yes.
Actually, they are quite or very, very vivid examples of embodied intelligence.
But the thing is that if they are just repeating the same motions over and over, then that's not embodied intelligence.
That's just a machine.
That's a robotic cook, right? uh but they if they can see ingredients they can see the tools they can see the cups um with their like computer vision and then they can move they control forces and timing like how hard it grip how long to pour and when to stop and then react when something changes like a like a cup not being in the right place the pot not in the right place then it is embodied intelligence but the thing is that right now the current stage is that they are great at fixed menu and basically they're just programs that humans set for them and then they don't usually learn new recipes on their own so therefore a better example might be autonomous vehicles yeah and how they interact with other traffic on the roads and the environment of course This is like one of those robots or vehicles in hotels that deliver our food or orders to our rooms.
Oh, yes.
They know how to interact, how to avoid people on the road, and they know when to wait for the lift to come.
Yeah.
They're a bit rude in the elevators though, because they stand right in the middle, no matter kind of where you are.
And I always have to back into the corner to stay out of the way.
On this topic, and this is kind of why we're discussing this today.
Yuxuan, you had a pretty exciting visit recently.
Where did you go?
Exactly.
A few days ago, I traveled all the way to Yizhuang in southeastern Beijing and visited this Beijing Innovation Center of Humanoid Robotics.
And there I saw some quite cutting-edge robotics and embodied intelligence robots as well that can work collaboratively in the logistic industry.
They coordinate with each other to handle movement and sorting and also packaging, no human control at all.
Yeah, so they are like one robot is telling another robot now you can start sorting.
Oh a robot boss, a robot, a robot.
That's how they.
They call it like embodied intelligence, because they can interact with each other.
Um, but i, i was also still thinking about because we're discussing it right, these things, some of these things can be achieved through these uh machineries, through these automated machines, for decades already, actually.
They've been there for so long.
They can do it so good.
Why are we still needing these kind of embodied intelligence?
And that's why I asked Tang Jian, the CTO Chief Technology Officer of this Beijing Innovation Center of Humanoid Robotics, and he explained how embodied intelligence differs from mechanical and traditional automation and why it is needed.
If you look at industrial settings today, traditional industrial robots are already extremely good at what they do.
For example, tasks like tightening screws or installing car doors.
These robots can do them with incredible efficiency and precision.
Humanoid robots are not meant to replace these systems, and there's really no need to.
Traditional robots have already reached near perfection in those specific tasks and it's very hard to beat them.
The real value of embodied intelligence lies in flexible production lines.
Imagine this scenario.
You buy one robot, today you want it to move boxes and tomorrow you want the same robot to tighten screws.
With embodied intelligence approaches, you can reconfigure the robot very quickly and easily, allowing it to adapt to different environments and different tasks.
In other words, it's a robot that can do many different kinds of work rather than just one fixed job.
With traditional automation, reconfiguring the system could take a long time and a lot of engineering efforts.
So that's why embodied intelligence is important, because it fills the gap between rigid automation and human labor.
So if we think about on a car production line, you know you'll have a robot that would tighten the screws or attach a door.
That's the only thing that that robot does. and it can't do anything else.
And if you wanted it to do something else, you'd have to completely reprogram that to do something else.
But with embodied intelligence you can very easily have them complete more than one task or a different task on a different day.
And for a lot of scenarios in these settings, as he said.
I think he was also saying that for many circumstances, people and human needs to like bend over or squat.
That is the actions that may like.
If you do that for a long time, that may harm your body.
And that's when these embodied intelligence come into the situation, because they can do different tasks through your different orders and they can maybe change the tasks for on a daily basis and because they have these different operators and sensors, they can do just different, Like in many, many different, maybe some dangerous and like very hard scenarios that they can solve the problem.
Where humans wouldn't want to do that.
Yeah.
But my question would be like, how far are we in these embodied intelligence?
I mean...
The picture is something that we want in our daily life and also on production lines, for example, for workers.
But how well can this embodied intelligence perform in real life?
Yeah, where do we stand now in terms of the development?
Yeah, first of all, I was thinking that because we also talked about humanoid robots, right?
And the current stage that we are at right now is, I think it is quite um, initial or at a very early stage.
And but the thing is that, why we do that, why we need embodied intelligence, why we need human or robot, it's because maybe in the future, now we are setting a very, very grand goal in the future and now it is only the early stage.
And now then, maybe in the future, when we are having these kind of general intelligence, our generator, artificial intelligence, or general application of these uh, embodied intelligence, then we're having the future of they can actually do things by themselves.
But it is part of the national strategy now, correct?
Yeah.
It's been formally incorporated into the national strategic planning as a very crucial component of future industries.
Exactly.
Recently, for example, China holds the Central Economic Work Conference that basically set out plans for next year.
And it also mentioned the efforts will be made to advance the AI artificial intelligence plus initiative, improve AI governance and foster innovation in science and technology finance.
And also, I think, according to another report called China Development Report 2025, released by the Development Research Center of the State Council, China's embodied intelligence industry is currently, as Yusheng said, in its early stage of development.
But the country has a solid foundation in research of these models and also in product manufacturing.
The market size would reach 400 billion yuan, which is about 50 billion US dollars by 2030.
So we can see that people have a lot of expectations for this.
I would say cutting edge technology now.
That's going to apply to things like transportation and the logistics and commercial services and industrial manufacturing.
It really is on the radar for sure.
And that report also noted that China has really strong capabilities in terms of developing a lot of different types of models.
They call them multimodal models that include vision and language and action.
And it's also able to manufacture key components for embodied agents.
The agents refers to the vehicles, if you will, themselves.
Different types of systems, different types of sensors.
And it has a clear cost advantage too, which is kind of surprising.
I would have thought that this would be, especially in the very beginning stages, extremely expensive.
But I guess that's why we're seeing such robust investment in this as well.
The China Academy of Information and Communication Technology, they were talking about this as well.
What did they have to say?
Experts from China, the CAICT, also added that since the beginning of this year, large models have shown significant improvements in language and also multi-model understanding capabilities.
And now 27 data collection facilities have been established nationwide, providing some high-value data for training these embodied intelligence models.
The data collection is kind of a challenge, isn't it?
I've heard scientists talk about this, getting the data from various sources, I suppose, and human beings can be one of those sources.
Yeah.
The expert that I just interviewed also mentioned that one approach is to Send robots into real world environments, just like these data collection facilities, and to collect data, and they can get the real time, physical and very high value data that they can have and maybe can be used in the future.
And then, but it can be very expensive and also time consuming because you need to build up these facilities for them.
And another approach is to collect data from humans and then transfer that knowledge to the robots.
And it could be faster.
Well yeah, it can be faster, but it could also take a lot of time, because you need the actual humans to do that.
Very exciting news.
China is clearly placing a very high importance on the development of embodied intelligence and the current pace of progress.
It's also really impressive.
So, while you were there and you were chatting with that gentleman, what else were you able to find out?
That must have been such a cool trip, by the way, to be there and seeing robots interacting with other robots and being right on the front line, so to speak.
Exactly.
Some of these quite vivid robots that are actually showcasing how embodied intelligence can be.
It is not a full body robot, but it's just a robotic hand.
That can play rock, paper, scissors with me.
And it wings every time.
Because it can just detect what I'm going to... Or what I am showing it.
And it gives me the reaction of the one that can beat mine.
Oh, that's fun.
Yeah, that's quite fun.
And I think that is a very clear example of embodied intelligence to have the vision, to see the world and have the physical interaction with the real world.
And that's when I also asked Tangjian, the CTO, to tell us what is the current stage that we are at and what are some kind of advantages that China is having in the field of embodied intelligence.
I think China has clear advantages in two main areas.
The first is hardware, and this includes things like motion control and locomotion algorithms, basically how robots move and coordinate their bodies.
In this area, China is somehow already among the global leaders.
And this is evident in the fancy, eye-catching movements that humanoid robots often perform in demonstrations.
A lot of these impressive demos are coming from Chinese teams.
More broadly, China benefits from a very complete industrial supply chain.
From components to manufacturing.
Everything is close by and well integrated, which gives China a real edge when it comes to building and iterating on robot hardware.
The second major advantage is real-world deployment scenarios.
This is actually a huge strength.
China has many large-scale, high-value application scenarios where embodied intelligence can be deployed.
I personally believe that as early as next year we'll start seeing large-scale rollouts in some very valuable use cases.
For example, many robots working together collaboratively in the same environment.
And that kind of skill is something China is particularly good at.
As for software and algorithms such as VLA models vision, language action models, models that connect what a robot sees, understands through language and then turns into actions, I wouldn't say China is far ahead internationally, but we're definitely keeping pace.
The same goes for large language models or LLMs.
China is not lagging behind in that area, and embodied intelligence builds on many of the same foundations.
One important point is that embodied intelligence relies heavily on real-world data collected from physical robots, not just simulation.
In this respect, China actually has an advantage, because we can deploy more robots in real environments and collect large amounts of real physical interaction data.
Yeah, that would be a huge advantage in having the robots out there collecting the data.
But also, yeah, the industrial supply chain.
That's a huge advantage when you have companies interacting with each other, working with each other, that physically are not that far away from each other.
You don't have to be delayed in any questions that need answering or technical support, and you also don't need to worry about delivery delay times either.
I also think, when it comes to real-world deployment scenarios, which is also very impressive, Because I remembered going to a park here in Beijing and I actually see, I think it's a robot, but it's in the shape of a vehicle, of a really small vehicle, sort of patrolling around that park.
I also saw that as well.
Reminding pedestrians of keeping safe.
For example, if you're riding a bike, do not go too fast.
And also just going around, that robot also knows how to avoid pedestrians.
At the same time, that can also be the same embodied intelligence in the real world scenarios.
Well, speaking of which, what can these robots or embodied intelligence actually do for us?
That's one good example, but what are some other good examples of how we can see these helping us in our lives?
They are getting into the stage of or from talking smart to getting things done.
And during the CIIE, Wang Xinxin, who is the founder of Unitree, which is a kind of leading robotics company in China, clearly outlined the practical goals for embodied intelligence and what he calls the two 80 targets.
He predicts that within the next year or two, humanoid robots will be able to understand voice or text commands in 80 of unfamiliar daily scenarios and successfully complete roughly 80 of tasks.
And also, I think the same company, Unitree I oversaw it on the news that is also collaborating with the state-grade Hangzhou Electric Power Company, which the company is also headquartered in developing these robots for maintenance work in this grid company.
For example, they are using drones that are equipped with robotic arms and to work on high voltage cables instead of humans.
You know they're really high, very dangerous.
Right and now they are trying to use robots in help and they also use trying testing humanoid robots that are in their transformer substations to learn how to work on the systems which buttons to to, to to push during, when and they're.
They're moving slowly at current stage, but the future goal is to have them incorporated in the human teams.
Yeah, that's one of the biggest goals or the goals that most of these embodied intelligence or robotics companies are aiming for to get these robots into these And they like kind of summarize it into a 3D environment dull, dangerous and dirty environments that the robotics can get into to help or kind of replace human for in that kind of a dangerous environment.
So it's in the beginning stages and always in the beginning stages.
There are a lot of bumps in the road, a lot of challenges along the way.
What are the scientists dealing with or the engineers dealing with in terms of?
Troubles or challenges in terms of embodied intelligence?
The whole setup.
Right now, the whole industry is facing a lot of challenges, as Changjian was saying.
One thing is that also Feifei mentioned the current stage.
That kind of little robot can move only very, very slowly.
That's also what I saw in that showroom.
Speed.
Yeah, the speed.
Because, of course, they are moving, they're sorting, they're packaging, but maybe at a 0.5x speed.
Even though they've been upgrading this whole robotics for another version already and it's already speeded up for double times, but still.
Yeah, but I think that's probably a wise idea.
Do it right before you do it faster.
Mm-hmm.
And also, I think the real world testing is also very important.
And I don't think maybe it can be rolled out very fast, especially in many other countries that you do need government support, industry support and also supports coming from businesses and companies and factories to test, to make this test, to tolerate that the robots at current stage cannot move as fast as human workers.
Sure.
So that can be a challenge.
Yeah.
And from a technical perspective.
I think Tang Jian, the CTO Chief Technology Officer of Beijing Innovation Center of Human and Robotics, have a better explanation on this.
From a technical perspective, embodied AI still faces major challenges.
AI development usually involves three core elements, namely data, algorithms, and hardware.
And hardware here doesn't just mean computing power.
Starting with data.
The amount of data we currently have is still far from enough and there is still a long way to go.
Beyond quantity, data quality is also an issue.
Things like completeness, precision, and coverage across different dimensions all need improvement.
But the biggest problem is lack of diversity, and that's a serious bottleneck.
Collecting real-world robot data is expensive and slow because it requires physical machines operating in real environments.
On the algorithm side, things are far from converged.
There's still no consensus on the overall architecture.
And that includes questions like how many layers should the system have, what operators or modules should each layer use, and how many parameters are needed at each stage.
And none of this has settled yet.
And finally, there are hardware challenges.
Beyond computing power, there are also challenges related to stability reliability, load capacity efficiency, battery life and operating time.
All of these need significant improvement before embodied intelligence can truly be deployed at scale in demanding industrial environments.
I was watching a video of a humanoid robot operating in a home and it was trying to put laundry into a laundry machine.
And the robot.
I guess I don't know what the problem was, but it couldn't recognize the size of the hole to put the laundry into.
So it kept on hitting its hand on the entrance to the washing machine and dropping the laundry and then pick it up and then hit its hand again.
These are the types of... It seems so simple, right?
But these are the types of major challenges that we're looking at before we get humanoid robots embodied intelligence into our homes.
So...
The basic models.
I was reading stories on the internet, maybe as early as next year or the year after, but that's the very entry-level stuff.
If we're looking at having them incorporated into our homes and into our lives, fully integrated, where we're using them on a daily basis.
Maybe 2030s 2040s, something like that.
And also it reminds me of when I visited a data center here in China.
They also mentioned about the quality of the data that they collected.
Right now, it's like water flooding into those centers, bad or good.
And they need extra effort, extra humans, extra facilities to sort things out and how to do that.
It remains a challenge and a question at the current stage.
And apparently the quality of the data is also a challenge for you know, engineers to going ahead with embodied intelligence.
So when you are thinking about data, you are thinking about oh no, you know, companies are collecting my data.
But it's not everything.
It's not all the kinds of data are needed for different technologies.
And, you know, yeah, that could be another question here.
Yuxin, thanks for taking that trip.
It sounds like you had a really worthwhile experience.
I think so.
That was a quite enjoyable experience that I had.
Yeah.
What would that trip look like four years from now?
Right.
You'll see.
Maybe you won't even be speaking to a human anymore.
Maybe you'll be interviewing the robot manager there at a place like that.
Really interesting topic, embodied intelligence.
