53. When AI Gets a Body: The Opportunities of Physical AI
Available now
-
Available now -
There is a moment in every technological revolution where the thing stops living on a screen and enters the room. We are at that exact moment with AI. Not the chat window kind. Not the one that writes your emails or summarises your meetings. I am talking about physical AI, artificial intelligence that has a body, moves through the world, and learns to do things that until very recently only humans could do.
That is the conversation I had with Vitaly Bulatov, founder of UFB Ultimate Fighting Bots. And trust me, this episode is not about robots punching each other. Well, it is also about that. But mostly it is about something much bigger.
What Physical AI Actually Means
For the past few years, AI has mostly lived behind glass. Your phone screen. Your laptop. A chat interface. Vitaly puts it simply: we are now entering the era where AI gets a body.
The humanoid robots being developed today, machines with 29 electric motors, depth cameras, LIDAR sensors, and AI brains trained through something called reinforcement learning, are not science fiction. They exist. They fight in cages. And the same technology that makes them fight will make them walk your streets, build your homes, and work in your hospitals.
"The fact that we'll have way more robots in the world will actually make us more human," Vitaly told me. His idea is that today, our phones are this weird middle layer between us and the world. We are always looking down, always half-present. In the future, your digital life might be embedded in a robot that sits next to you in the physical world. You stay present. The robot handles the interface.
I find that beautiful, actually.
The Formula One of Robotics
One of the best ways to understand UFB comes from an analogy Vitaly uses: UFB is the Formula One of humanoid robotics.
In F1, every team races on the same track with the same rules. What wins is what is inside the car, the engineering, the software, the intelligence. UFB works the same way. The cage is just the track. What actually competes is the AI brain each team has built and trained inside their robot.
"I really like the Formula One analogy because that's exactly what Formula One is," Vitaly said. "Throughout Formula One, cars became so much safer, new materials have been tested. But also, it's a cultural moment."
That is the dual function UFB is going for: a testing ground for real technology, and a cultural moment for robotics. The two things feed each other.
How You Train a Robot to Fight (And Why It Matters)
Here is where it gets fascinating. Training a humanoid robot is not like programming a dishwasher. You cannot just tell it what to do. You have to show it, let it fail a hundred thousand times in a simulated world, and reward the versions that get closer to what you want.
That is reinforcement learning. And the reward is not a cookie. "Mathematically it's not really a cookie or anything," Vitaly explained.
The robot learns in simulation, not physically, because if it tried every variation in the real world, you would destroy every robot you own. Once it has learned in the simulated world, you transfer that knowledge to the physical machine. But there is always a gap between the two. Vitaly calls this the sim-to-real gap. The physical world has dust, humidity, electromagnetic interference, and wet floors. The simulation does not.
Closing that gap is one of the central challenges of physical AI right now. And the progress is accelerating. NVIDIA just released a model called Sonic that has been pre-trained on thousands of human motions. You give the robot a new movement to learn, and it can already do most of it, because it already knows what movement feels like.
The Data Nobody Thought to Collect
Here is something that surprised me in our conversation, and it might surprise you too.
When training language models like ChatGPT, researchers used the internet. Billions of words, conversations, books, articles, all there, already collected. But when you want to train a robot to clean a toilet or take dishes out of a dishwasher, that data does not exist anywhere. Nobody filmed themselves doing those things in first-person. Why would they?
"It turns out that's exactly the data we need to show the robots how to clean toilets and train them," Vitaly said.
So now an entire new industry has emerged. Companies are paying people to film themselves doing boring household tasks from a first-person perspective, what Vitaly calls egocentric data. Google DeepMind, Physical Intelligence, and dozens of startups are all buying this data to train robotic AI models. The race is on.
This connects directly to something Elio Challita discussed on a previous episode about micro-robots learning from biological systems. The challenge is always the same: how do you teach a machine to navigate the chaos of the real world? For Elio, the answer was looking at insects. For Vitaly, it is looking at humans doing dishes.
Moravec's Paradox: The Hard Thing Is the Easy Thing
One of the most mind-bending ideas in this episode is something called Moravec's paradox. And once you hear it, you cannot unhear it. The things that are easy for humans are incredibly hard for robots. And the things that are hard for humans are incredibly easy for robots.
A backflip? Easy for a robot. You simulate it 10,000 times, you get a perfect backflip. But picking up a small bolt that just fell on the floor in a tight space? No robot on earth can reliably do that. Not even the most expensive ones.
"The real world still presents a lot of challenges for robots," Vitaly admitted. "But the pace is accelerating a lot and robots are learning how to generalize the tasks now, which is impressive."
This paradox explains why physical AI is such a hard problem, and also why solving it is such a big opportunity.
The City That Comes Alive
I want to paint a picture here. Imagine a city in 2073.
Traffic lights that read the real-time flow of cars and people and adjust without a timer. Bridges that monitor their own structural stress and send alerts before any human notices a problem. Delivery robots negotiating with each other for sidewalk space. Trash bins that signal when they are full. Buildings that open their doors based on crowd density.
This city is not a fixed thing. It is a living organism.
Vitaly is confident that autonomous driving will be the first wave, it is already happening in San Francisco. The second wave is service robots: delivery rovers, hospitality robots, machines doing the dull, dirty, and dangerous jobs that humans either do not want to do or should not have to do.
"It's not necessarily that robots will be doing jobs or replacing jobs," Vitaly said. "They're making the jobs better, making people more productive. People can focus on creative output, applying their creativity and brain to the problems that's what people are good at."
This idea connects to a conversation I had with James Glattfelder about consciousness and the future of the physical world. James argued that as we offload more cognitive tasks to systems around us, what we are really doing is freeing human attention for what matters most, meaning, connection, creativity. Physical AI might be the most powerful tool we have ever built for doing exactly that.
Gaming Gets a Body Too
Here is where things get a little trippy. And I love it.
Right now, gaming lives on a screen. You sit, you hold a controller, you move pixels. But UFB has already demonstrated something different: people anywhere in the world can log in and remotely control a physical robot in a real gym, fighting a real robot controlled by someone else on the other side of the planet.
The physical world becomes the game.
I asked Vitaly: when the world becomes the game, does simulation theory stop being a theory?
He laughed and said he does not think about that much anymore. He is just building.
But I think about it. Because we talked about this exact blurring of digital and physical with Roman Axelrod, who is building contact lenses that overlay digital information onto the real world. The direction is clear: the line between the digital and the physical is dissolving. Physical AI is one of the biggest forces pushing that dissolution forward.
The Arms Race Nobody Is Talking About
Let me end with this thought.
There is a new arms race happening. Not nuclear. Not satellites. The question is: which country trains the smartest robots?
The electric vehicle boom improved battery technology and electric motors, the exact same components that power humanoid robots. The AI boom produced transformer models that are now being adapted for robotics. The data collection industry is emerging to feed those models. The foundational models are being released by NVIDIA, Google, and Physical Intelligence.
Everything is converging. And whoever trains the best robots will have an enormous advantage, economically, militarily, culturally.
Vitaly has been in this space for over a decade. He has seen the slow years and the fast years. "Right now," he told me, "the most I've ever been" enthusiastic. Because the progress has just accelerated so much over the past few years.
That is the sound of a man who was right early, and is watching the world finally catch up.
_____
This article was produced with the assistance of AI tools. The Mizter Rad Show is hosted by Mizter Rad.
Guest: Vitaly Bulatov, Founder of UFB Ultimate Fighting Bots — ufb.gg
Follow UFB on LinkedIn.
Listen to the full conversation with Vitaly on the Mizter Rad Show.
Stay curious, question everything, and maybe just maybe… the most sophisticated machine ever built is not the one in the cage, it's the one reading this sentence right now.
Mizter Rad
-
Mizter Rad (00:00)
Hello, beautiful humans. Today's episode kept me up last night, but in a good way. I can tell you that. Before I bring in my guest, as always, I wanna go first to the future. Let's go 25, 30, 50 years from now.(NEWS FROM THE FUTURE)
Mizter Rad (02:11.36)
Hi, thank you so much. That was Kai and that's exactly why we're here today because that man that Kai talked about, the one who saw this coming, is sitting right here with me today. But before we go to this man and talk to him, let me ask you all something. What if the next arms race isn't about nuclear warheads or satellites but about
Which country trains the smartest robots? My guest today has been quietly building the arena where that race is already happening. He spent over a decade deploying robots in the real world. Rovers, industrial machines, humanoids. When everyone else was still arguing about whether it was even possible. Then he did something no one saw coming.
He built a fighting league for humanoid robots and he called it ultimate fighting bots. But here's where it gets interesting. UFB isn't really about fighting. It's about what fighting forces robots to do. Adapt, learn, survive under pressure and repeat. Think Formula One. Every team races on the same track.
and they're the same rules and they're the same conditions. But what wins isn't really the car. It's what's inside the car, the engineering, the intelligence. UFB works the same way. The cage is just the track. What actually competes is the software brain each team has built and trained inside this robot. This matter the brain, the better the robot fights. And the implications
think about this, the implications reach far beyond a fighting cage. This is not about putting a cage and putting robots to punch each other on the face. That's entertaining and that also happens. But I believe we are at that exact moment where AI stops living on the screen, chat GPT, Claude, et cetera, and gets a buddy.
Mizter Rad (04:26.749)
The robots learning to fight today are the same robots that we will see walking our cities, building our homes and making decisions in the physical world tomorrow. Whoever trains them best matters most than people realize. Please welcome Vitaly Bulatov, founder of UFB Ultimate Fighting Bots. Vitaly Dobropochalovitch
How are you?
Vitaly (04:56.462)
I'm doing good, I'm doing good. Thank you so much for having me.
Mizter Rad (05:02.917)
Excuse my my my Russian is very rusty. That's all I know actually
Vitaly (05:06.638)
It was very good. Thank you.
Mizter Rad (05:10.559)
Okay, Vitaly, look, before we go into robots fighting in cages, because I want to get into that, of course, I want to take you all the way back. And I would like you to tell us who were you before all this? Like, what were you studying? What were you obsessed with as a kid, as a young man? Take me to that beginning.
Vitaly (05:35.084)
Well, if we are talking about the kid, actually, I did really like to play Legos when I was a kid. So I think that's where it all started. But then I was always interested in automating things and kind of that somehow was interesting to me.
like industrial settings and industrial processes are interesting for me. However, that's not what I went to study. My first degree was in economics and I just studied some math and statistics but not really engineering directly.
So, but then when I graduated, I worked a little bit as a management consultant and realized that very practical and useful, but not very interesting long term, it seems to me. So that's where my journey of building with robots started. And I started my first company doing drones.
Mizter Rad (06:47.369)
Thank
Mizter Rad (06:59.859)
Drones. Okay.
Vitaly (07:00.982)
Yeah, so the first company we did was called Drone Employee and that one was a software that helped you automate missions for autonomous drone flights and coordinated flights within the same airspace. Our first use case was a paintball park where you could come and
call a drone to different locations and take photos and you know we've then explored zipline parks and different outdoor facilities ski resorts and outdoor videography but then regulatory
Mizter Rad (07:49.087)
Hmm. Hmm.
Vitaly (07:53.167)
framework only requires you to always have a pilot in command when you fly a drone, even if it's flying like autonomously completely. So that kind of limits the scaling of some business models. So that's when I started working with different types of robots like rovers, boats and industrial integration. And that's what led me to Humanoids.
Mizter Rad (08:21.861)
Okay, interesting. but okay, let's go back a bit. You studied economics. What really, what was that trigger? What was that specific moment where you said, okay, that's it, robots. Like I'm gonna, you know, this is what I want to do with my life. Was there a single moment or was it more like a slow pull into the topic?
Vitaly (08:48.472)
Slow pull into the topic for sure. And then I was interested in reading up on that, going to conferences. And then in one of the events I met my first partners in a startup that we did 10 years ago. And that all started from there to shape up to be a business.
Mizter Rad (09:14.015)
Hmm. Interesting. I understand that you've been doing this for over 10 years and like you said, rovers and drones and now humanoids and we get into the humanoids topics because I think that's the more interesting part. 10 years is a long time. It's a long road. And I can imagine as an entrepreneur as well, I know that it's not.
always a straight line. So I want you to be honest with me. Were there moments of real doubt, like moments where you thought this thing with the robots is too hard, no one is really paying attention? Not talking about right now, because I think right now a lot of people are paying attention, but some years ago, this is not working. Maybe I should just stop. What was the feel like from the inside and what kept you moving forward?
Vitaly (10:12.43)
Oh yeah, mean for sure definitely had ups and downs and not all of the ideas worked out. in robotics you are, you you're always limited by the physical world. So every new deployment that you do...
is unique. every new... like even though it's the same kind of system, the same like hardware used, you still need to, you know, build another one and do another integration. Even if it's similar, it's never completely the same. Like one line...
different incline, have different parameters, have different something. It's in a different room with different humidity. Like there's so many variables in the real world that make scaling hardware products obviously way harder than anything in software. So yeah, I would say what kept me
Going is the... mean, just... I don't know. I am interested in up toation since I was a kid and it's just curious to me. despite the fact that it is hard, I guess the good part about robotics is also once it works, it's pretty steady as well.
Especially like historically robotics was mostly warehouses and industrial and you know once the robot is set up in a factory they're probably going to be there for like five plus years or something like that. So it's a
Mizter Rad (11:54.311)
Hmm.
Vitaly (12:00.59)
It has its pros and cons, I guess. But the major pro is that, like, you know, we live in a physical world. I was always confident that we would need more automation and this industry will continue to evolve no matter what. The question was just how fast and like for the past 10 years, was especially like for the first five years, I would say it was quite slow and...
Mizter Rad (12:15.647)
Hmm. Hmm.
Vitaly (12:28.46)
just like for 30 years robots were just industrial doing pre-planned things. Like the mission that just repeated the same motion over and over again for like 10,000 times a month. Or even like a day rather. But...
Now we start seeing robots that can reason and decide and be around people more, not just in a caged environment, in an industrial site or warehouse.
Mizter Rad (13:02.271)
So you would say you're more positive, enthusiastic about the whole thing right now?
Vitaly (13:09.13)
I mean, yes, definitely. I mean, I was always enthusiastic. That's why I was kept doing it. But right now, of course, the most I've ever been because the progress has just accelerated so much over the past few years.
Mizter Rad (13:14.653)
Right.
Mizter Rad (13:26.269)
Why do you think that progress accelerated so much in the last years?
Vitaly (13:31.264)
I mean, it's kind of a bit of a perfect storm from both hardware becoming available, just electric motors becoming way better, electric vehicle market exploded starting 2020. So...
Mizter Rad (13:47.411)
does that help actually your sector? For example, the electric vehicles boomed.
Vitaly (13:51.502)
I'm...
I mean, generally the supply chain is very similar for electric, like it's an electric motor. You can build the electric motors for cars or those could be smaller electric motors for humanoid robots. That's why majority of automotive companies are also really interested already have a humanoid robot program. Like the...
Mizter Rad (14:16.158)
I see
Vitaly (14:21.398)
skillset and the supply chain is very similar. So, yes, so that became just more robust, more available batteries became better as well and more affordable. The price per kilowatt hours dropped over the past 10 years for battery and...
Mizter Rad (14:24.952)
Hmm, interesting.
Vitaly (14:46.484)
Obviously, like we all know, the boom of AI happened and really the transformer models that first got so popular with child GPT got later use in robotics because it was so great for text because we had so much text on the internet to train on.
And then when we started to think of how we use it in robotics, it turns out there's not that many examples of boring tasks that robots need to do, that we can train on. So we didn't think that it's important to record first-person videos of how we clean toilets. know, not that many people were doing that. So that internet, the data is not available on the internet because nobody was recording it. Because why?
Mizter Rad (15:33.544)
Hmm
Yeah.
Mizter Rad (15:41.212)
Right. Yeah.
Vitaly (15:42.275)
But it turns out that's exactly the data we need to show the robots how to clean toilets and train them. So now for the past couple years, there's just an explosion of companies that do different forms of data collection and incentivizing data collection for robot training.
Mizter Rad (15:48.062)
Interesting
Vitaly (16:05.55)
all the different ways of how they can record data with phones, different special cameras, specialized hardware, VR headsets.
Mizter Rad (16:14.595)
Tell me more about this, because I think this is super. I never thought about it, of course. mean, data collection for normal AI, LLM tools like Chai GPT and so on that that happened for many years consciously, I guess, purposely in the Internet online in your computer. So every click, everything was collected and that trained the models for the AI tools that we use. But when you talk about physical
machines actually performing a task automatically, the kind of data that they require to do that is completely different. You put the example of cleaning a toilet. So and then you said that their company is actually doing the data collection now specifically for these humanoid robots. Tell me more about this because I think it's fascinating. So what kind of companies are collecting what kind of data, for example?
Vitaly (17:16.686)
I mean, ultimately, it's the companies that train their own models that need the data. So ultimately, they are the biggest buyers of it. So all the big AI labs or labs that build humanoid robots are buying data for training. It's not a very public market because nobody wants to signal which types of data they're buying.
Mizter Rad (17:43.856)
Mmm.
Vitaly (17:46.167)
what exactly it's working and what's not working. But yeah, all the big AI companies are definitely buying data for training robots like Google DeepMind and for example, definitely, and have also internal data collection problems, of course.
So as well as startups, there is a number of startups that do data collection, some incentivize layers of networks for collecting data, some just do it directly.
Mizter Rad (18:27.003)
But what kind of activities are they collecting data on? Like, for example, cleaning toilets. Is that a real example?
Vitaly (18:33.708)
Yes, exactly. Yes, it is a real example. But also, like, like, taking dishes out of the dishwasher and doing the scraping of a dish inside the sink and then, you know, putting dishes on shelf and, you know, picking up cups, dusting, all the different things we need to do. We need to show the robots how to do.
Mizter Rad (18:37.693)
Okay.
Mizter Rad (18:42.377)
Hmm
Mizter Rad (18:52.062)
Right.
Mizter Rad (18:58.185)
And how do they put like...
Vitaly (19:03.438)
.
Mizter Rad (19:03.549)
And do they put cameras in restaurants and collect this kind of data? Is that that's how it works?
Vitaly (19:07.854)
I mean, that's where it gets a little bit to the trade secret zone, I guess, because everybody has their own methods and right now it's cutting-edge research. major labs just buy everything and experiment. Some startups claim to have unique insights of what works. I mean, it could be either... It's called usually...
Mizter Rad (19:13.245)
Right.
Mizter Rad (19:30.099)
Hmm
Vitaly (19:35.403)
the useful data, like the term for it is called egocentric data. the data that comes from the, you know, from your kind of eyes and you see both hands in the video. like, you know, via camera kind of with the hands visible doing the task.
Mizter Rad (19:44.831)
I see we like a POV camera
Vitaly (19:59.843)
But then there is a lot of detail exactly of like which camera, which conditions, like do you wear gloves to track or not. right now everybody experiments with all the different ways of how you can collect data with VR tracking hands, with specialized gloves, with just an iPhone.
Mizter Rad (20:00.191)
Mmm.
Mizter Rad (20:06.654)
Hmm.
Mizter Rad (20:11.103)
Hmm.
Mizter Rad (20:14.591)
Hmm.
Mizter Rad (20:25.885)
Hmm, interesting. Yeah, yeah, yeah.
Vitaly (20:29.527)
or N.
Mizter Rad (20:31.835)
Interesting. Let's get into the machine. Literally, I already used the F1 formula one analogy. I said same track, same rules for F1. And in my opinion, and it's not only my opinion, this is research, this is known publicly. What wins is the intelligence inside the car, the engineering inside the car. At UFB, and you correct me if I'm wrong, the cage is sort of the track.
Vitaly (20:54.979)
you
Mizter Rad (21:00.923)
the software brain in the robot is sort of the engine in the car. Is that a fair way to describe it or am I oversimplifying something important?
Vitaly (21:13.549)
I I would say the comparison is more even direct than that, because the engines are the motors of the robot. The robot itself also has electric motors, right? And the car in Formula One has software, so the robot also has software. So it's more even direct than that. Literally, the components are almost the same. Just, you know, a humanoid robot has more motors.
Mizter Rad (21:35.487)
Mm.
Vitaly (21:41.39)
In particular, the model we use the most in our boxing shows is Uni3G1 and that one has 29 electric motors.
Vitaly (21:55.821)
But more or less they're kind of the same electric motor.
Mizter Rad (22:01.895)
Okay, and so this robot, Vitali, just for me to understand, this robot, okay, they have 29, you said 29 motors, more or less, and they learn to fight, to balance, to stand up through something called reinforcement learning, as far as I understand. Explain me, like I'm 10 years old, what is reinforcement learning? And when you say you reward the robot,
because I know there's some reward mechanism there. What does that actually mean? Does the robot get like a gold star, a cookie, a good review? What does the robot get when it does something right?
Vitaly (22:42.208)
Yeah, reinforcement learning means that we have a... I can give you an example. example, we want to train a particular style of a punch, like cloning, example, like Muhammad Ali or some famous kind of motion. So we have a video of that.
Mizter Rad (22:56.244)
Yeah.
Vitaly (23:11.284)
Then we run one process to extract the pose from the video, like a stick figure skeleton kind of a thing, to just estimate mathematically of where the elbows are, where the wrists are, and have that pose estimation. And then we...
Mizter Rad (23:29.459)
Mm-hmm.
Vitaly (23:33.227)
retarget that pose to the body of the robot because the body of the robot is different from the body of a human even though it's a humanoid but it turns out that humanoid we use for example have seven elbows so it has seven joints in the arm so you know it can bend differently than a human like it kind of can bend the same as human hand can but also in way more ways than a human can as well
Mizter Rad (23:50.469)
Mizter Rad (24:01.309)
Hmm
Vitaly (24:02.254)
So we then do the process called retargeting to kind of compare the body of the human skeleton to the robot skeleton. And then that's when we come to the reinforcement learning part. So reinforcement learning essentially is now we have this motion on the robot that we want to repeat. But...
In reality this robot doesn't know how to repeat it and keep stability like not to fall down So if we just play it out, it will just roll over because it wouldn't keep balance. It doesn't know the physics of this motion
Mizter Rad (24:43.327)
Hmm.
Vitaly (24:43.39)
and to understand the physics of this motion. That's when we use reinforcement learning training. What that actually means is we create a simulated environment, like a simulated world, and put this robot in the simulated world, show him the example, and say, this is what you need to do.
Mizter Rad (25:05.363)
Mm-hmm. Mm-hmm.
Vitaly (25:06.094)
And until you learn how to do this motion, you are going to be in this environment repeating this motion for like 10,000 times or 100,000 times. But that happens because we now have big data centers and all that can happen over like half an hour or an hour for like 10,000. So essentially we run 10,000 robots at the same time to train. then, and they over time...
Mizter Rad (25:16.479)
Okay.
Vitaly (25:35.511)
they get closer and closer to repeating that motion and the ones that do get closer we say yeah that's we reward that that's mathematically it's not really a cookie or anything it's mathematical reward saying like we see that you got closer to the original motion and then we see that you you're actually pretty far up
Mizter Rad (25:53.727)
Mm.
Vitaly (25:57.837)
you're not getting closer. So we can reward the robots that are getting closer to their original motion. And over so many hours or like iterations, we get the robots only, we filter out only for the robots that can actually do this motion perfectly. And after this reinforcement learning training,
Mizter Rad (25:59.507)
Mm.
Vitaly (26:20.748)
It's kind of like, you know, in any sport your trainer keeps telling you like, repeat it again, repeat it again, until you do it perfectly. Yeah, kind of like that.
Mizter Rad (26:25.915)
Right. Right. Right. OK, I get it. So if I understand correctly, when you say that they do these repetitions 100 times over and over until they get it, they do it in a simulation, meaning they do it on the screen. They don't do it physically. Correct? OK.
Vitaly (26:45.814)
Yes, exactly. I mean, so you can absolutely do it physically, it's just you'll break a lot of robots while you are getting to the result. So it's practically impossible. So that's why...
Mizter Rad (26:56.337)
I can't imagine,
Right. Right. It doesn't make sense. mean, you can simulate it on the screen on the computer. it wouldn't. Or is there a difference when you simulate it? Yeah.
Vitaly (27:09.464)
There is a difference and that's where there is another jargon term in the industry comes up often that's called sim to real gap. And there's always, of course, a gap between simulation and reality because we cannot simulate everything in the physical world, like every little dust, every little atom. Like we cannot simulate the whole real world in its completeness.
Mizter Rad (27:19.87)
Mmm.
Mizter Rad (27:32.265)
Mm-hmm.
Mizter Rad (27:37.317)
I see the wind, everything, every little detail, and maybe even every electromagnetic wave that you don't even see or realize.
Vitaly (27:39.532)
Yeah. So.
Vitaly (27:46.063)
Yeah, so essentially we simulated really close enough to like some forces, but there's still always some things we're missing that in the real world exist. And that's when your robot works in simulation doesn't mean that it will work on your real robot.
Mizter Rad (28:07.337)
Mm.
Vitaly (28:09.142)
and that's like you would still need to spend some time to then polish your move when you transfer from simulation to real robot usually doesn't work from scratch but just this March Nvidia released a pre-trained AI model or policy for the robot
Mizter Rad (28:21.791)
from the beginning.
Vitaly (28:39.15)
called Sonic and that was already trained on a huge data set of motions of all the different motions that like people could do just the motions. So now without the need to do reinforcement learning like they've done it before on so many different things
you can just give it the motion, like just a retargeted motion, and it will most likely repeat it, and we are seeing really good results with this model. So now in this year, we are moving towards those foundational models from big labs, which...
Mizter Rad (29:11.262)
Hmm.
Vitaly (29:20.33)
start to generalize and can like you can give them more intentions and they kind of already know a lot of things because they've been trained on so many things. So we are starting to see that happening.
Mizter Rad (29:36.329)
So these foundational models that NVIDIA and these big labs are releasing, do you see this happening more and more? Is this like a trend or labs like this? Yeah.
Vitaly (29:46.033)
yes. Definitely. So, NVIDIA released this model called Sonic. They also have a model called Groot for hand manipulation and like they released a lot in general. There is a company called Physical Intelligence that is really strong in training models for robotics. And...
Mizter Rad (30:12.671)
Hmm.
Vitaly (30:16.056)
Google, for example, have Gemini Robotics, which is their version of the foundation model for robotics skills. That's closed source and you would probably be using it through the API.
Mizter Rad (30:35.059)
Vitaly, you talk about policy, robots policy, what do mean there?
Vitaly (30:40.686)
I mean, I use it somewhat interchangeably with essentially AI model. And I think in the context of this conversation, that's pretty much, yeah, it's essentially the pre-trained AI model that we give to the robot. And that's what's the software, the software brain that controls the robot.
Mizter Rad (31:05.562)
OK. OK, the intelligence, let's say. OK, I see. When you were talking about how you reinforce the robots, reinforcement or learning by reinforcement with your.
Vitaly (31:17.218)
Yep. So as a result of that process, you get the policy that you can then put on the robot that will be controlling that robot to do that motion. But if you do that a lot, a lot, you can get some impressive results.
Mizter Rad (31:27.729)
OK, OK. Yeah, it's interesting.
Mizter Rad (31:37.171)
Very interesting what you're saying because I interviewed not so long ago Anna Maria Jakcic. She's a scientist in Switzerland who makes flies actually smarter by rewarding certain behaviors. So she selects the flies that performs best, let them sort of reproduce. And then after enough iterations and generations, she only ends up with the smartest.
Vitaly (31:59.193)
Hmm.
Mizter Rad (32:05.427)
flies, so the smartest flies survive, basically. And it's evolution on fast forward. So basically, you're doing the same. Or it sounds to me like it's kind of the same principle that you're using here.
Vitaly (32:09.804)
Yeah, very similar.
Vitaly (32:16.908)
Yeah, very nice. It's just we can simulate this kind of, well, not really exactly evolution, but the training process of, yeah.
Mizter Rad (32:25.993)
Mm.
That's a controversial, whether it's evolution or not. That's good question, actually.
Vitaly (32:34.551)
Yeah, yeah, yeah, I don't think it is this is yeah
Mizter Rad (32:40.605)
Not necessarily. So tell me something just to close on this topic. So who is the one deciding what gets a reward? Is it the system itself?
Vitaly (32:53.23)
A researcher. like you when you when you started training process you define what the reward would be and you define it like an example I gave the reward is this original motion we want to repeat but then like
Do really interesting experiments of what you can give as a reward. For example, Disney has a really interesting research I like around the robot named Olaf, a snowman, they've trained it to make less noise when it walks.
Mizter Rad (33:22.335)
Mmm.
Vitaly (33:30.238)
because its goal is to be in a theme park around kids. Usually robots are pretty noisy, they step, and they're pretty loud and robotic, you know? And so their goal was to optimize for stepping more softly and not scare kids around by stomping and being scary. So you can be pretty creative of what you can be optimizing for as a reward.
Mizter Rad (33:30.943)
Thank
Mizter Rad (33:42.227)
Mm.
Mizter Rad (33:51.821)
Mmm.
Mizter Rad (33:57.108)
Hmm, interesting, So going back to the policy topic, the intelligence of the robot. The intelligence is the policy is the intelligence. I guess the robot has some sort of heart that pumps it pumps the blood, the energy throughout the body, which is the heart, which is the blood and where does the actual energy come from?
Vitaly (34:17.326)
I mean electrical system.
Vitaly (34:23.79)
The battery? Yeah, it's battery powered.
holds battery for about two hours usually for now.
Mizter Rad (34:35.227)
Okay, okay, okay. And can the robot actually sense the world around it? Can it feel heat, for example? Can it hear?
Vitaly (34:41.184)
Yeah. I mean, it depends on which types of sensors you attach it to. So by default, the robots we have have a depth camera in the real sense. So it knows the distances to objects and it also has a LIDAR. So it kind of laser points around and like knows that and maps the 3D environment around it.
But yes, you can definitely have a temperature sensor, then will know heat. I did an experiment with a company that we've collaborated with before for a while, Carbon Origins. They are doing automation of construction and heavy equipment.
they were exploring to use a robot to do dust sensing because they are monitoring dust in construction yards so the robot could walk around different places and monitor dust levels by having a dust sensor attached to it. So yes, for example,
Robot dogs are used heavily in nuclear sites to be monitoring for potential radioactive leaks. So just walking around with a Geiger sensor and making sure the safety processes are in place without endangering humans.
Yeah.
Mizter Rad (36:20.703)
Can a robot actually know when the floor is wet without having pre-plant that as the researcher or the manufacturer? I'm a manufacturer, I build robots for a certain specific task, but then in the real world, what you were talking about, the sim to real gap is not zero. So the floor is wet, my robots fall, they all fall.
Vitaly (36:40.642)
Nice.
Mizter Rad (36:49.019)
Is that what's happening in reality or are we closing that gap the faster than we think?
Vitaly (36:49.42)
Yeah?
Vitaly (36:59.054)
I mean, yes and no. Depends on what you think, I guess. So, yes, we are. But this is a really good example, actually, of wet floor being pretty tough. You would probably use computer vision for that. And we can tell with some degree of certainty, but not guaranteed. So still, the real world presents so many unstructured challenges that the robots don't really know how to deal with.
Mizter Rad (37:01.907)
Hmm. Right.
Vitaly (37:28.91)
and in robotics it's actually it's called Moravec's paradox and what it
means is basically the robot. What it's easy for us to do is hard for robots and what is hard for us to do is easy for robots. Example of it like backflips for example and all the different tricks are kind of easy because for robots because they are all pre-planned motions. Like we can do all the crazy trick with rotation however you want because we can just simulate run it for 10 000 hours in simulator and then just repeat
Mizter Rad (37:47.743)
Mmm.
Vitaly (38:07.418)
this motion perfectly. But responding to something dynamic like dropping a small bolt and then like you need to pick up the bolt from the ground from like some tight space, that's impossible. Like no robot, no like most expensive billion-dollar robots cannot do it. No major lab cannot do that. It's an impossible task in robotics.
Mizter Rad (38:26.999)
Hmm
Mmm.
Vitaly (38:37.058)
yeah, the real world still presents a lot of challenges for robots, but, the pace is accelerating a lot and robots are learning how to generalize the tasks now, which is impressive.
Mizter Rad (38:47.761)
Hmm, interesting.
Mizter Rad (38:55.487)
So I'm trying to understand something. The robots, when you say robots are generally learning, are they learning real time? If I'm a manufacturer or a researcher working on this topic, am I pushing updates to that physical robot often?
Vitaly (39:20.042)
Often, but not real time. you do train the way when you use ChatGPT, guess, you have different versions that get released. So they train every version.
So similar way, like kind of you collect more data, you train a better version and then kind of upgrade the brain to a newer version in stages like that.
Mizter Rad (39:50.396)
I see. I see. I see. I want to take you off track here because I want to zoom out a bit. Actually a lot. Because the technology you're building doesn't just live in a fighting cage. I've watched some of the videos that you guys publish on YouTube. If you guys are listening to this conversation, go on now.
on YouTube and look for ultimate fighting bots. These guys have amazing content of the robots actually fighting in the cage.
The technology you're building is not just, like I said before, living in those cages that we see.
Mizter Rad (40:40.615)
in the videos. It lives in the world. And what I mean with that is the following. When I think about the city of the future, and maybe you give me one minute to paint a picture here. If you think about the city of the future of Italy and everyone else listening.
You think about a system, right? Traffic lights that don't just follow a timer, but that they read the flow of the people and cars in real time and adjust or buildings that decide when to open their doors, maybe based on crowd density, delivery robots negotiating with each other for sidewalk space. If you pay more for your delivery, then maybe the robot will have a faster journey because they got better space than the others that paid less.
or bridges that monitor their own stress and call for maintenance, the engineer would get an alert to maintain the bridge before any other human notices the problem, or even trash bins that signal when they're full. That city stops in the future. That future city is not longer a fixed thing.
as we have cities today. They become a living organism. In the next, and this is my question to you, in the next 10 to 20 years, what kind of autonomous systems do you actually think we will see operating in cities?
Vitaly (42:20.846)
I mean, autonomous driving, I'm pretty sure, is getting widespread. So in San Francisco, we can see how quickly it just overtook. And now I think a significant portion of the all taxi rides in San Francisco autonomous drives already. I mean, even looking at Tesla autopilot, it's so impressive now that it almost drive itself most of the time.
So I think autonomous driving would be something common. Then as we were talking about robots, I think 10 to 20 years horizon is enough already to say with confidence and we definitely will see more service robots.
delivery rovers on the streets, hospitality robots everywhere, servers and cleaning and all the back office jobs and like boring, dull and dangerous jobs that people don't really want to do already, where labor shortage already exists.
Mizter Rad (43:37.341)
You mean like construction or something?
Vitaly (43:39.843)
I mean in construction different ones but even within construction there are jobs that are or even not necessarily whole jobs but parts of the jobs that are people don't want to do like just you know carrying tools around or moving things around something repetitive something that people don't want to do there people are not
good at doing even. I will see robots doing more and more of those repetitive things around. And yes, I think it's important to say that like it's not necessarily that robots will be doing jobs or like replacing jobs. They're making the jobs better, making people more productive. People can focus on creative output of whatever the job they do, like applying their creativity and brain.
Mizter Rad (44:08.516)
Right.
Vitaly (44:36.796)
to the problems that's what people are good at. While robots can carry tools around and you know break their backs. Yeah exactly.
Mizter Rad (44:44.051)
Great.
Mizter Rad (44:48.477)
The order or their seven joints.
You talked about in other conversations, I watched some of the videos that you guys have on YouTube, but also some of the past videos, maybe from a different lifetime of yourself talking about trustless systems for physical infrastructure actually. And that word trustless, it comes from the crypto and blockchain world as far as I understand.
It doesn't mean nobody trusts anyone, means systems. It means basically the system is designed so that you don't need to trust someone. So the rules are sort of enforced automatically and in theory no single company controls it. No single government can switch it off. Is that a fair understanding of what a trustless system is?
Vitaly (45:47.503)
Yeah, I would say so. It's a bit of you know, jargon term, so like a bit of a... like means a lot of different things, I guess. But yes, I would say that's a fair description and the level of which...
Mizter Rad (46:01.416)
Okay.
Vitaly (46:11.971)
practically we can achieve trustless systems and how idealistic it is. It's, you know, a conversation.
Mizter Rad (46:19.643)
Right, but do you think the cities that we talked about just now, do you think they should be trustless physical infrastructure or not necessarily?
Vitaly (46:27.118)
I personally think they should be. I think that's the ideal scenario. I am a big proponent of both open source software and I do like open hardware. So all the way to the processors.
Mizter Rad (46:31.966)
Mm-hmm.
Vitaly (46:55.192)
for RISC-V architecture being really interesting, especially in like IoT devices and edge infrastructure and RISC-V being the open hardware architecture for processors. I think it is really useful and cool for, yes, spaces to be using more open hardware as well.
Vitaly (47:29.342)
But in reality, it's obviously like availability of hardware, faster speed of execution, know, companies build solutions that work really well, they control them, but you know, they build really good products.
Mizter Rad (47:42.396)
Right.
Right. Yeah, reality is a different thing. mean, when you talk about also trustless physical system in the city, I could imagine that a company like Amazon or Meta would like to have some sort of control of some of the physical infrastructure as well at some point if we go that way.
Vitaly (48:11.362)
Yeah, I mean, it's, It's definitely, like...
I would say case by case discussion as well, because trustless is like, yeah, it's a, it really doesn't mean anything in particular. So when we get to actually details of execution and deployments, that's when it's like all the specific questions matter of like, which specific hardware is being used, which chips in that hardware, what software. And yeah.
Mizter Rad (48:43.551)
Hmm.
Vitaly (48:45.516)
All the details matter in the physical world.
Mizter Rad (48:48.351)
Right, right. I to switch gears a bit and ask you about UFB, your initiative, your venture, Ultimate Fighting Bots in 10 to 20 years. What do you think it's going to look like? Do you think it's still a fighting league or does it become something else entirely, something we don't have a name for yet?
Vitaly (49:13.57)
Definitely not something else entirely. We are the premier competition for humanoid robots. We are the sports league. Boxing was the first sport we've introduced, but we are definitely going to be working with other types of competitions as well. We are, yes, the premier testbed, industry, ecosystem, and sport around humanoid robots.
So depending on how humanoid robotics will keep growing and it seems like it's going pretty well and like there are
serious discussions of humanoid robotics being the size of automotive market and all them all close to every household having a car in the world being similar like that with the humanoid robots. The scale of this industry would be enormous. So our goal is to be the test bed and the premier media competition franchise around humanoid robotics.
Mizter Rad (50:09.844)
Hmm.
Mizter Rad (50:22.399)
And can you do both at the same time? They seem to be very complementary, both sort of focus for the company.
Vitaly (50:33.296)
yes, definitely. I really like your Formula One analogy because that's exactly what Formula One is. Throughout Formula One, cars became so much safer, new materials have been tested, carbon fiber tested in automotive in Formula One. But also, it's a cultural moment, right? Automotive became...
Mizter Rad (50:42.847)
Hmm. Right.
Vitaly (51:00.606)
a cultural moment and Formula One became a cultural moment for the whole industry. So it's both bringing attention towards robotics is our goal, more people building robots, building software and skills for robots, training robots, collecting datasets, participating in our league. We provide resources, we give access to robots.
we want like activity around robotics and robotic growth but also yes being the testing ground for different technologies that are being revealed and launched now in both hardware actuator design motors as well as software new AI models and different things
Mizter Rad (51:48.026)
How do you actually finance it? you like in F1? I know in F1 when you look at the pilots, the uniforms, they're full of brands. They get a lot of sponsorship deals. Do you have the robots also full of brands around their bodies or is this something that you want to explore more or how are you financing it right now?
Vitaly (52:10.062)
Yes, definitely. So we already worked with sponsors and some really strong tech brands that
project we product we are using and they're listed on our website and we are now sourcing more sponsors as well so definitely way more opportunities than Formula One.
Mizter Rad (52:41.471)
Hmm, very interesting. But the way you monetize is by selling tickets or that's not really like the bigger chunk of the revenue. Sponsorship, yeah. Okay, I see, I see. I also understand that and maybe this leads me to my next question because I've been sitting on this idea since I started preparing.
Vitaly (52:53.102)
Sponsorship definitely should be the... yeah.
Mizter Rad (53:10.735)
the video because I understand that gaming today lives on a screen. If I want to play FIFA, I open my laptop or my console and use a remote generally to move the players that move in the screen and the action is all digital and the stakes are digital. But what you're building, and you correct me if I'm wrong, is also remote controlled physical robots, so operated by anyone on earth basically.
And that sort of starts to blur the line of how we know gaming completely at the moment. So I'm imagining a future where gaming is not on a screen anymore, where you control when you log in and control a real robot in a real city instead. So you're on a real quest in a real physical space in Paris while you're sitting in Vladivostok.
And you're not playing on a virtual map, but you're actually playing on the street. And actually, the obstacles are real. And maybe you're competing against someone else sitting in Cairo and the robot is in Paris, like I said. So I guess gaming at that point becomes physical. And the world at that point becomes the game.
And when the world becomes the game, you know what happens, Vitali. Simulation theory stops being a theory. So is this something you think about every now and then?
Vitaly (54:52.629)
not...
Not so much anymore to be frank. I'm just keep building. But yes, that's kind of where it's going, especially given that like all we spoke about before the robots right now are not yet ready to do like all the jobs. So similar what we've seen with autonomous vehicles and autonomous taxis. Initially, there was like a remote operator looking after the car all the time.
ready to take control at any moment and then gradually just one operator started looking after three cars and like five cars and ten cars and now they're like almost don't need interventions at all. So in robotics yes clearly we still need
Like for most tasks, to be honest, we need all close to a hundred percent control. Like most like people controlling the all the motions with it's called teleoperation. And for a certain amount of time, this would be how robots will be deployed with a remote person actually controlling them. And that's where we also, technology comes in because we're gamifying this process, but this process is actually needed in the industry for
Any particular talk that Obis would need to do
Mizter Rad (56:14.929)
Which process, sorry, which process do you mean?
Vitaly (56:17.498)
The remote streaming, like the gaming thing that we use for gaming, you can be using it for any other control of the robot remotely doing a task. essentially what...
Mizter Rad (56:28.511)
So people can go now to your website and do that in a robot that is sitting there in San Francisco?
Vitaly (56:34.702)
So it's not live 24x7, but we are organizing events where you can and we are inviting people because right now access is quite limited. only have still so many people wants to work with robots. We have so many still. But yes, the capability is there and people from anywhere in the world can join and we've hosted tournaments remotely where people can come.
Mizter Rad (56:49.447)
Of course. Yeah, yeah.
Vitaly (57:04.494)
to our website, connect to the robots and then fight with another person on the other side of the globe in a physical gym with the robots. But the same way we do those gamified experience, we want to expand and kind of generally just blur the line between physical and digital. You can now be in all the different places, we could use it for so many different things. And that's what I mean when I say testbed, like we are...
showing real useful technologies that are needed by the industry in a fun gamified way, letting people try it out, getting curious and hopefully getting them excited enough to start building skills for robots themselves.
Mizter Rad (57:50.688)
What is one of those examples that are useful for the industry that are tested in how you call it your testbed in the UFC?
Vitaly (58:03.982)
So specifically the remote streaming for example and optimizing the remote streaming part of how people can remote.
Mizter Rad (58:17.875)
You mean teleoperation or what do mean with remote streaming? Yeah, OK.
So for what industry, I mean, I can imagine some industries, but I want to hear it from you. What industry would benefit from an enhanced teleoperation process or technology?
Vitaly (58:40.876)
I mean, even like a factory assembly line, so...
Vitaly (58:48.3)
hotel cleaning robot that would need to get some help from an operator remotely.
for example, things like that. But even the themselves, I think first and foremost, it's the humanoid robots themselves is a new platform. The hardware just launched, it became available, but nobody really know, like it was not deployed at scale to know what breaks, how often, how's the process of maintaining these robots, what needs to happen when something goes wrong.
And we are the place that have ran humanoid robots for many hours in really tough environments.
Mizter Rad (59:33.727)
So you mean that the mechanics of it and the repairments that they require.
Vitaly (59:40.611)
Yes, yes. Purely on the hardware level even, yes. But even on the software as well, like how it degrades over time or like for example how the fact that they're always becomes beaten up and you know not as new. How does it affect the software and like... Yeah.
Mizter Rad (59:58.815)
Yeah, yeah, yeah.
Mizter Rad (01:00:05.247)
Very interesting.
I this question to every single guest of mine, you're not going to be the exception, of course. How do you see the world in 30, 40, 50 years? How do you think?
We will look like we will be more like machine by then machine like or.
as human as we look right.
Vitaly (01:00:34.234)
Definitely as human as we look, think. I don't think we'll be physically changed as much, but the world will change. My hope is that it will... Like, the fact that we'll have way more robots in the world will actually make us more human, because right now we're in this weird moment where our phone is kind of have to be an extension of our life.
because that's our connector to like very important digital life that we need. But maybe in 40, 50 years, that's our robot that has this, that's our gateway to our digital life. So we don't have to, you know, get distracted to look at the screen. It's the screen that now has a form.
Mizter Rad (01:01:21.597)
Mm-hmm.
Vitaly (01:01:21.768)
and it sits in a chair next to us and it's part of our life. It's not that we have to get distracted and get like enclosed in this digital space and not as present in the physical world, but rather we are all present in a physical world and including our digital life kind of is present in our physical space through the robot. So I'm pretty hopeful.
Mizter Rad (01:01:25.704)
Hmm
Mizter Rad (01:01:45.958)
Interesting, interesting. So you're a positive man. Good, that's good to hear. Definitely, 90 % of the people that I interview that are scientists or innovators or entrepreneurs like yourself, they're positive. And that makes me positive as well, because you guys are the ones building the future.
Vitaly (01:01:49.826)
yeah, yeah, I have definitely.
Vitaly (01:02:09.236)
Nice, yeah.
Mizter Rad (01:02:12.041)
For anyone, beautiful humans, for anyone that wanna go deeper, go and check out UFB on YouTube. They have really cool videos, watch a fight. And if you ever wanted to control a robot from your living room, as Vitaly said, there will be an option. It's a limited option for now, but there is that option. And apparently, that is something you can do now. You don't have to wait 10 years to do that.
Vitali, is there anything else that you would like to share with people? Where can people find you, your work, if people want to connect with you, investors that want to know more about your business model, where can they reach out?
Vitaly (01:02:53.742)
Definitely follow us on social media, follow us on X and on LinkedIn and Instagram. Our website uofb.gg has a forum. Please reach out. We're really open and happy to talk about all the collaborations. We're really excited to co-build, work with startups and big companies.
happy to work with good investors. yeah, we are the premier testbed and immediate platform around humanoid robotics.
Mizter Rad (01:03:32.485)
Beautiful, beautiful. Spasiba!
Vitaly (01:03:36.142)
Thank you!
Mizter Rad (01:03:37.823)
Until next time beautiful humans stay curious question everything and maybe just maybe start treating your buddy like the most sophisticated machine ever built because somewhere out there people like Vasily are already training a robot to move better than you do.
Hasta la vista!