00:00It's not just the volume. It's not just the valuation. Some of the strategics that are backing you, NVIDIA, others in consumer electronics, Samsung, LG, and in the software sector as well.
00:11What are they recognizing as your core competence, do you think?
00:15So if you think about robotics, right, it's one of the oldest areas in technology, like in AI. AI was built for robotics.
00:23We have seen amazing demos in robotics for the last 70 years, not four or five years, 70 years.
00:29But where are robots around us? And this is what people often get wrong.
00:34When you think of robots in your mind, you imagine hardware.
00:37Like that comes upon your mind first because that's how Hollywood has primed us.
00:41But the main thing behind not having robots around us today is the brain is missing.
00:46And if you have the brain for robots, you can really enable them around in a variety of areas, variety of applications.
00:53And that's what exactly Scaled is doing.
00:55We are building what we call omni-bodied brain.
00:58Any robot, any task, one brain.
01:00I'm interested in how you're building it.
01:02The limitation that a lot of people are talking about right now is real-world data and training.
01:07There are different approaches.
01:09Digital twins, trying to simulate physics in digital worlds.
01:13This is very much what NVIDIA is focused on, right?
01:15Is that the approach you've taken, simulated or synthetic data?
01:19So I think I will draw analogy to LLMs like ChatGPT-like models.
01:23Like if you see OpenHead ChatGPT.
01:25And soon after, many companies had their own LLM models.
01:28And the reason for that is data exists on the internet for language or vision.
01:33But robotics, there is no internet of robotics today.
01:35So what we do is we look for alternate sources.
01:39One such source is human videos.
01:41We look at how humans operate in their scenarios, in your kitchen, while working in factories, and combine that with training from simulation, which companies like NVIDIA invest a lot on, like Omniverse and all those.
01:55So these two things, watching people and then practicing in simulation, that's how we combine and scale these models to large scale.
02:04Your software can be put within the GPU.
02:08And what's interesting is you're already going from what we understand zero to tens of millions of dollars, Deepak, in revenue in just a few months in 2025.
02:17Where's that revenue coming from?
02:19Who is using your software already?
02:20Yeah.
02:22So the revenue has mostly been from the enterprise applications.
02:27And the idea is, instead of going for, our end goal is a consumer home-like application.
02:36But right now, since it's the beginning of the field, we are taking this model and we are deploying it in a variety of applications.
02:43Like, we are deploying robots in sectors, like point-to-point delivery applications, security applications, data centers, and manufacturing and warehouse.
02:56So the revenue is coming from most of these applications.
02:58You were born out of academia, and you're still Professor Carnegie Mellon.
03:02But what's interesting is you're not without some competition out there.
03:05I think about physical intelligence, for example.
03:07You're all trying to attempt to get robots to become intelligent and situationally aware.
03:12Do you think you can take on?
03:14Are you worried about competition?
03:16Well, competition is always good.
03:18And it turns out, like, the AI community, born from very small circles.
03:24So they're all hard friends and colleagues from earlier.
03:28But what's unique about what we are building is this omnibodied brain.
03:33Like, any robot, any task, one brain.
03:35Now, if you really think deeply about this, it sounds absurd.
03:40Like a humanoid robot with a human-like body, a dog robot, and they share the same brain.
03:46And that's what you see in the results we release.
03:49The difference in approach is fascinating.
03:51It is a crowded field.
03:52You know, Bernt from 1X was on the show yesterday.
03:54They have a world model that essentially allows Neo to carry out a task it's never seen before.
04:02And that, largely, you know, I'm not an engineer, right?
04:05But what you all have in common is VLM.
04:08Those are the inputs.
04:10Distinguish your approach to 1Xs as an example.
04:13Yeah.
04:13So in our case, the robot learns by watching people, right?
04:18Now, imagine, I have to pick up this cup in front of me.
04:21Do I have to exactly imagine?
04:22That sounds scary, by the way.
04:23Watching me?
04:25Well, you learn when you were a kid, you learn by watching your parents.
04:28Right.
04:29You learn by watching others.
04:30Now, if you think about how you operate, every time you pick up a cup, do you imagine in your head a picture of a cup and then putting your finger in the handle?
04:39No.
04:39It's all merged in, like somehow.
04:41It's consciousness inside your body.
04:43And then you imagine in some abstract space and you then do it.
04:46And that is the main idea here.
04:48That's how we do it.
04:49But just like being said, watching alone is not enough because if it was enough, I could play like Federer.
04:55I keep watching his games all day.
04:56And that's where practice comes into play.
04:59Now, practicing in the real world, making mistakes is expensive.
05:03And that's where we combine simulation.
05:05So videos alone is not enough.
05:08Sim alone is not enough.
05:09But put them together and boom, you have a magic recipe to really figure out how to scale robotics.
Comments