A robot handed me an ice cream at the World Robotics Conference in Beijing. Clean cone, no drips. Behind me it made popcorn for someone else.
That is the show. I went for the thing behind the counter.
Every humanoid company I visit in China hits the same wall, and it is not the hardware. Ask a founder what is holding him back and you get one word, three times. Data.
The expensive way to teach a robot
The standard method is people. You put humans in motion capture rigs and let them repeat a single task for thousands of hours until the model has seen enough of it.
It works. It is also why so many humanoids can do exactly one job, in exactly one corner of one factory. New customer, new setup, and the clock starts again from zero.
DexForce comes at it from the other end. I have been to their Shenzhen office several times. Founded in 2021, around 300 people, roughly 60 percent of them in R and D. That ratio is the whole company in one number.
Type the scene, let the engine build it
Their system is a generative simulation environment. You describe a scene in plain language, the engine builds the 3D assets around it, and the robot learns the task on synthetic data before it touches a real object.
Collecting the data and training a skill took months. Now we train on synthetic data from our own simulation engine and deploy in hours.
On the screen next to me the instruction was as plain as it gets: put the cans upright. The simulated robot worked it out. A minute later the one on the floor did it with real cans.
Four stations, four objections
The booth was built to answer the things a buyer actually asks, in order.
Can it deal with a human customer? That was the ice cream and the popcorn. Simple machines, built for human hands, operated by a robot in front of a crowd that never stops moving.
Can it run all day? They sat a robot in front of a Chinese chess board and let visitors play against it. On a trade show floor that is the closest thing to an eight hour shift: long task, constant visual change, no reset.
Does it fall apart when the world moves? The retail station was a full checkout. The robot scanned barcodes instead of working from a stored product list, handled bottles of different shapes with both arms, and picked items back up after visitors knocked them over. DexForce calls this zero shot generalization. In a shop it means nobody has to teach the machine every new product.
Does it do real work? The last station was a production line with two robots packing, unpacking, assembling and inspecting.
Why a humanoid and not a cobot
This is where European engineers push back hardest, and the honest answer has nothing to do with the robot looking human.
Put a cobot on a station and you design the station around the robot. Move the job and you rebuild the cell. A mobile two armed machine walks to the next bench, packs this month and inspects the next, and the line stays as it is.
In most plants I walk through, that flexibility is worth more than cycle time. It is also why the deployments that survive are the unglamorous ones: the dangerous station, the night shift nobody wants.
What I still want to see
A trade show is everyone's best day, mine included. The number that decides this is how many hours the machine runs in a real store or plant with nobody standing behind it.
That is the question I put to every manufacturer we take clients to, and the answers are getting better every quarter.
If you are deciding where automation money goes next year, watch the training side rather than the robot. That curve is bending fast, and it is bending fastest in Shenzhen.
[00:00] Hello. [music] If we want to bring robots into our home, into our hospitals, into our factories, into our daily lives, there are three things that are most important. Data, data, and data. Data is a big bottleneck, but it's still very important. And right now I'm here in Beijing at the World Robotics Conference at the booth of one company that has a really special approach to data. It's called DexForce. They are from Shenzhen. It's a company I have visited a lot of times. And today I want to show you how they tackle this problem. >> [music] >> Here with me is Baiyu. Baiyu, thanks for the invitation. >> Hi. >> Tell us a little bit about when was
[00:47] DexForce founded? >> Uh we founded in 2021 and specialized in body AI humanoid robots. >> Yeah, and this is one of your humanoids. And uh your founder is one of the top 2% scientists in whole China in in this robotics field. So, what is his specialty? >> He specialized in body AI and 3D perception, data collection. And we also have 300 members in our uh company now. About like 60% of them are R&D members. >> Yeah, this is what I saw in Shenzhen. That 60% of their staff are actually engineers, right? And they work in the research and development. That is why they are strong. We will see a lot of robots in the future in the service and industry. And this robot from DexForce
[01:35] is already collecting a lot of data. He can use the ice cream machine and also the popcorn. So, I'm very sure that this robot we will see in the future in a lot of scenarios in real life. >> [music] >> So, here we go. That's a fresh ice cream I got from the robot right here and it looks perfect, right? The way you train it is so different to other companies. Most companies, they will have human beings who repeat one task for 10,000 of hours so they get the data. But you have the matrix which is your training simulation system, right? Can you tell us a little bit about it? >> Yeah, actually we in the greatest software to hardware into our product matrix. And that means we from the
[02:23] robotic brand and to the robot body we can integrate it at all. This demo we already scalable and repeatable landed. So that means you can see our robots is actually in working as Thomas said before. We collect data need training a skill and it need months. Now we use our own self developed a simulation AI engine. That means training our robots with the synthetic data that can deploy in only like hours. >> This is crazy and this is the China speed I'm talking about all the time. is that they will shorten the time for training the robot with a simulation software. >> So you can see here is a generative simulation user interface and it's basically based on the Textor's physics engine developed by Textor's here and you can input your scene description
[03:11] here. The engine will generate 3D assets accordingly. >> Okay, so so here we can see now the task we we gave it, put the cans upright. It's doing it right now which is crazy. So you don't need to train them in real life. With this simulation user interface, you can train it once and then it knows in the future how to put up all the bottles, right? >> Yes. >> As Barry said, it's already deployed in a lot of use case scenarios. You have a lot of customers who use it like in this setting like in shop. But what I'm also very interested in industrial scenarios, right? >> Yes, yes. Let us see. >> Let's go and have a look. What we see here is the robotics version
[04:01] of AlphaGo. We all know this super brain who plays chess with all the world masters and here we have the Dexforce version of this robot who's playing Chinese chess with other people. Whenever you are lonely, he will be here for you to play a game of Chinese chess with you. >> Why we develop this demo is because we want to show our robot's long-term operation capability. >> So actually this is not for just playing. What you can see here is the very strong vision model. As you said the long-term operation, right? So I think for industrial use cases this is really important. It should be fast, it should be reliable and it should be precise. Don't just treat this as a game of Chinese chess. This is actually could be an industrial use case. What we can see here is the demonstration of the
[04:50] brain of the Dexforce robot. This robot actually can work in a retail store so we can see the brain is so smart. have to know all the SKUs in one shop or a factory. It just can take it by themselves, scan the barcode so you don't have to put all the data of the SKU inside. We have different forms of packaging and it can grip the bottles, it can grip the both arms cooperate with each others and cool can do the automatic checkout for 24 hours and they're also very friendly to you. >> Actually for this demo we want to showcase our zero-shot generalization capability. No matter it is two hit or is knocked down, it can always pick it up.
[05:36] >> For the factory we actually also have a use case right here. Let me show it to you. So we have two robots working here on a production line. Can be packed, can be unpacked. >> For this >> demo assembly and inspection product line, FOV is more large and the grasping way is more flexible. >> The flexibility is seen. Today it can do this job, but then you can move it and it can do another job. Why you need humanoids on a production? Because if you just have the cobots, you need to redesign the whole factory and the whole space, but these robots are very flexible and you can easily move them from one place to another. >> Yes, also some dangerous or some very boring stuff can let our humanoid robots do. >> I'm very sure that humanoids will not be here to replace human beings, but to do
[06:24] the work no humans should do. For example, very dangerous places, uh very boring tough repetitive tasks. And this is not the future, this is already reality because the robots of Dexforce, they are already collecting data, they're already working, and they're already doing services and doing real work in real factories, in supermarkets, in shopping malls. So next time you come to China, you can have a look at the shopping mall, at the factory, and also soon we will see them all over Europe and the rest of the world. Thank you so much, Bei Yu. >> Hello. >> [music]
Get These Insights Every Monday
Join 18,000+ professionals reading Asiabits. Free, every Monday, straight from Shanghai.
Subscribe Free →
