Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
Perspectives: The rise of AI robotics image

Perspectives: The rise of AI robotics

HSBC Global Viewpoint
Avatar
0 Plays2 seconds ago

Ye Tian, Founder and CEO of RoboScience, joins Zhenyi Tang, Head of Banking, Corporate and Institutional Banking, HSBC China, for a discussion on RoboScience’s start-up journey, the rise of AI-powered robotics, and AI innovation in China. 

This episode was recorded on the sidelines of HSBC’s 12th Annual China Conference in Shenzhen on 2 September 2025.

Disclaimer: Views of external guest speakers do not represent those of HSBC.

Recommended
Transcript

Introduction to HSBC Global Viewpoint

00:00:01
Speaker
Welcome to HSBC Global Viewpoint, the podcast series that brings together business leaders and industry experts to explore the latest global insights, trends, and opportunities.
00:00:13
Speaker
Make sure you're subscribed to stay up to date with new episodes. Thanks for listening, and now onto today's show.

Tian Ye's Career and RoboScience Journey

00:00:21
Speaker
Welcome to perspective podcast series. And I'm Zain Yichang, the head of banking CRV China. And today we have Tian Ye, the founder and the CEO of RoboScience.
00:00:34
Speaker
Hi, Mr. Tian, welcome to HSBC. Thanks for having me. Yeah, and we have known each other for like almost a year. And I have seen you really, you know, start a really brand new career to start your own business, RoboScience.
00:00:50
Speaker
And I also have noticed, and by the way, congratulations on securing a very significant amount of investment into a company. So, you know, of course, I'm sure everybody's is curious about how you have made it made this and how is your journey of starting your own company and by having all this ah advanced technology and they can share with us some of your views and how how you're doing how RoboScience is doing right now.

RoboScience Technology and Demonstrations

00:01:19
Speaker
yeah Thanks for having me here. ah yeah i started my company at the end of last year and before that I spent 10 years abroad abroad in the US.
00:01:31
Speaker
So when I came back with my co-founder Shaolin, it was kind of a like mixed feeling. We stand in the the empty office room just the two of us and wondering what we should do ah in the in the coming days and are we already know the technology roadmap we should have and we are are also had some funding for the start so we were not that nervous and we hired people and we started building our models and started building demos and what what we did was building a robot that can ah so ensemble furniture by reading the manual
00:02:14
Speaker
I saw the video, that was a very amazing one Yeah, so that one, I think that demonstrates the ability of our model and the potential potential of it because it showed how it can understand very abstract information and also complete very complex and precise and long-horizon task. And that demo, I think it's the it's the showcase of our technology and people really love the demo and that's how we that's our first milestone.
00:02:43
Speaker
And I think really, you know this is like seems to me the very, very good and the typical kind of startup story, you know, among our overseas Chinese, you know, excellent students with all the good technology.
00:02:56
Speaker
And when we talk about talking about the technology, can you share with us some of your views on how advanced your technology and then and how it is different from the other for instance, like the normal or ah the 1.0 versions kind of technologies. So and how you are planning to use your technology, not only in your robot science, but also in other scenarios in our industries, what's your plan for that?

Generalization Capabilities of RoboScience

00:03:22
Speaker
Yeah, so we know like robotics has been here been there for 50, 60 years and it has been like widely adopted in all kind of industries, especially in manufacturing. And the ah the property of the ah of Robotics 1.0 was sir precise and efficiency.
00:03:43
Speaker
Basically, people programmed the robot and it will execute the task step by step very precise precisely. and what it lacks is generalization. Basically, if we can have a robot that can follow natural language guidance and complete any task, manipulate anything, that will open up a big opportunity for robotics and also for human kind. Because now we can basically have replicators of over ourselves that can ah be the delegate over over to to to do the tasks.
00:04:19
Speaker
And the technology we are building is especially good at generalization on three axis. The first one is the subject, basically the robot itself. Our model can be applied to different form of robots, whether it is ah like it has ah or it has ah dexterity hands and that will make our model ah applicable to many different domains i see because different domains require different kind of form factor yeah and the second one is the object basically what you are manipulating
00:04:57
Speaker
like we are I can let my robot to grasp a bottle, this is a rigid body object yeah and we we can also ask our robot to not a tie a t-shirt that's a deformable object so our model can also generalize to all different kinds of objects that the robot needs to manipulate And lastly, it's the predicate.

Human Knowledge Integration in Robot Models

00:05:22
Speaker
So basically, ah it can complete different tasks, whether it is a simple pick-and-place or like a complex task like ah ensemble the furniture.
00:05:33
Speaker
So we can generalize ah over subject, object and predicate. That's the generalization capability of our model. The major advantage of our technology is we have a very high data efficiency. ah We summarized the human knowledge about robot manipulation tasks and embedded it into our model. but ah We can let the robot to learn from video. yeah okay Basically we let it to read the to to watch the the videos on the web or we can demonstrate the tasks by ourselves, by human beings.
00:06:11
Speaker
We also built our own and simulation engine. like Our engine will generate as important very vivid physical related data and which is the essential missing data for embodied AI because that's the critical difference between embodied AI and LLM LLM is in digital world but embodied AI is in physical world yes and we can generate physically precise data for our robots I see really, I think the data matters a lot in the whole training process. yes Not only you are, you know, your whole system learning from the real world, but also you are generating a new set of data, right? For the robots to really learn by themselves. so i Really, really add to another new dimension of the whole kind of research. Am I right on that?

DeepSeed's Methodology and AI Influence

00:07:01
Speaker
Exactly. Fantastic. So when we talk about all this data, there's one thing that we definitely need to mention, need to and need to communicate with you that, okay, It's really on DeepSeed. Seems to me that DeepSeed has created a kind of new methodology by less input, by doing things like less means more.
00:07:18
Speaker
right Do you think that the methodology that DeepSeed brings to the table, brings to the industry, it helps you as a founder of Robert Sainz to really kind of ah moving things faster in the whole process. Is that that helpful?
00:07:32
Speaker
To me, the extra efforts or the the the success of DeepSeek is they embed human knowledge into this large language model training. they don't just blindly throw data into the system and hope a it will generate some very good new models.
00:07:50
Speaker
They really use their knowledge to to refine the system, making less data, hu ah doing more, like you said. ah That's exactly what we do, like because in body data really need data and physical real data is really hard to obtain. yes If you use teleoperation to acquire data from the physical world,
00:08:12
Speaker
that's ah like you can get 200, 300 per day yeah that's the maximum and if you rely on this blindly brute force way to generate data like then then it's a very long journey for you to achieve generalization what we did was we use our knowledge embed into the model and the data generation process ah so that our system can learn faster comparing to ah those brute force approaches and I think the impact of DeepSeq is it proved this methodology to the general public and to the resource holders like by resource I mean capital like investors yeah and I mean policy makers who can redirect the capital yes
00:08:58
Speaker
That's very important. and I also mean the general public who has probably the most ah important resource which is attention. yeah like Like we know, attention is all you need, right? yeah They can bring all this resource to this ah smart companies not to work hard to embed and their knowledge instead of using brute force. I think the most that that's the most precious ah lesson, DeepSeek, to everybody. And also I think from an investor perspective, it really has pushed us to rethink our kind of methodology for invest.
00:09:35
Speaker
into investing into the new economy sector, right? Because the new method, new model, new dimension can appear. Yeah, I think like investors and i can look deeper into the technology instead of just superficial metrics, like how much compute you have, how much resources you have. The bigger the better. No, that's maybe not the logic. That's the longer the logic. Yes, yes.

AI Development in China vs. the US

00:09:57
Speaker
And also by developing or the new economy, the AI capability in China, do you see it, I mean, along with the development you of DeepSeek yourself, do you see that the difference in technology, for instance, in in AI area, you know but the difference between China and the US, do you see the gap is narrowing?
00:10:20
Speaker
Or someone said, okay, no, the gap is still enlarging. What's your view on that? Yeah, I think that's a really interesting question. Like, you know, I spent 10 years in the US and now I'm ah doing my business in China. china I have observed both sides. You are the perfect one to to speak on that, right? Yeah, I think it depends on how you measure ah innovation.
00:10:42
Speaker
Let's focus on two two sub areas. Like first one is large language models. Second one is embodied AI, which I am working in. So for large language models, the two algorithm pillars the transformer architecture and the GPT architecture they are invent they were invented by u companies ah so that's the history and the physical foundation is the chips GPUs and that are mostly owned by US companies as well so that's the fact that's the the current status
00:11:16
Speaker
US companies have innovated important piece of it yeah to start this new generation of AI which is large language model. ah But if we look at the current status, I think the science part of the larger language models have ah has already like almost converged.
00:11:36
Speaker
Now people are mostly working on scaling and applications. To this, and I think Chinese companies have really caught up or they have big advantages yeah because they have ah the the critical resource for this. Like we have a very vibrant society that are eager to see new technologies. So I think Chinese people... embrace the changes. Yes, we really we have seen all these dramatic changes in the past 30 years and we are used to it. yeah Whenever a new technology comes out, like even my grandma or my... my like like yeah
00:12:11
Speaker
My parents are using over 80s, they are using technologies really, very good. Yeah, they're really embracing this new technology, technology which and like opens up all kind of ah opportunities for application builders, which will further boost the foundation model builders.
00:12:26
Speaker
Positive cycle here. Yeah, that's a really positive cycle. And the complexity of ah the Chinese society and industry really makes it a gigantic experimental bet, or like...
00:12:39
Speaker
like a warm room to to to to improve the ah AI technology because what AI needs is different kind of data yeah and different kind of applications And to this extent, I think the application and adoption of AI in China is really matching, if not beating the US. yeah Even those very traditional industries have started adopting AI and this industry may not even exist in the US.

China-US Collaboration in AI

00:13:09
Speaker
And honestly, I totally agree with you. And I see the marriage of China and the US working together in the AI or the law or the new economy sector in large, because it's really not about the kind of competition between two countries, two markets. yeah It's really about cooperating and by by cooperating with each other and then bring the best of technology to humankind yes on this planet. That's the kind of target we need to achieve.
00:13:35
Speaker
Yeah, exactly. yeah I think if the the people in the two countries can really work together as we did in the past, it's it's good for us, it's good for the US, it's good for the entire world. Yeah, see this is the this the dream ah from a so scientist, from a banker and maybe from all of us, right? yeah We have this kind of common

Future Vision of AI-Powered Robots

00:13:58
Speaker
dream over here. yeah And last question is really in young um on your future plan.
00:14:02
Speaker
And how do you see robot science and yourself and whole industry for instance in the five years time. In China or in the in in a globe? I think my vision is one day AI-powered robotics will become the the coexisting species of mankind.
00:14:25
Speaker
to use fish It's a new species. It will become the friends of people. It will help people work and it will support people in life. That's my long-term vision. In the future, I hope our products, our ah robot systems will become ah ah co-workers and friends and pets and all different kind of family members of people that will empower the productivity and also the life of everybody
00:15:00
Speaker
And hopefully HSBC can help robots and science and the whole ai industry you know to develop fast along the way. That's our dream too. Thank you. Thank you for watching. Thank you.
00:15:11
Speaker
thank you for joining us at hsbc global viewpoint we hope you enjoyed the discussion make sure you're subscribed to stay up to date with new episodes