Latest / Automated with Brian Heater / Deepu Talla: Building the Future of Robotics at NVIDIA

Transcript

Deepu Talla: Building the Future of Robotics at NVIDIA · Automated with Brian Heater

• 0:00
  SPEAKER_01: The best in the world are gravitating towards it to solve the number one problem that we need to solve, which is to create a general purpose brain.
• 0:08
  SPEAKER_01: And testing the general purpose brain in a humanoid form factor makes logical sense because in the last 200 years we built the whole world to be around human form factors.
• 0:19
  SPEAKER_01: But job number one right now is we need to create a chat GPT equivalent for a physical robot.
• 0:25
  SPEAKER_01: And that does not exist, and that's where the whole world is at, which is to create it.
• 0:44
  SPEAKER_00: Obviously, Jensen's out there at GTC with his leather jacket swag, but it was Deepu's talk at our events on synthetic data that had an overflow crowd in the largest ballroom that we had.
• 0:57
  SPEAKER_00: I didn't get a chance to speak to him at Detroit, so I started a podcast and we did it virtually instead.
• 1:03
  SPEAKER_00: We talk about the recent release of Jets and Thor, training robots at scale, and why NVIDIA seems so committed to humanoids.
• 1:13
  SPEAKER_00: You started at NVIDIA back in in 2013 on the Tiger business, but you were only there for it looks like one or two months, and then boom, you move over your VP and GM of robotics and edge AI.
• 1:28
  SPEAKER_00: That's a really short amount of time.
• 1:30
  SPEAKER_00: That's a that's a quick shift.
• 1:32
  SPEAKER_00: Why did you make that jump so quickly after joining NVIDIA?
• 1:36
  SPEAKER_01: So what happened was when I first joined Nvidia, we were in the mobile business.
• 1:40
  SPEAKER_01: So we had a Tegra line of processors that was targeted at uh smartphones and tablets so I started in business development.
• 1:48
  SPEAKER_01: Uh and within two months, uh Jensen decided to promote me to report into him and run that business.
• 1:54
  SPEAKER_01: And as we were building that business, uh in 2015, we decided to uh wind it down and move into more into artificial intelligence and portable gaming and autonomous uh driving, which essentially led to robotics in the future.
• 2:09
  SPEAKER_01: So that wasn't an abrupt transition as it might look on this, but that was uh planned out and uh you know executed over a couple of years.
• 2:17
  SPEAKER_00: So a little advice.
• 2:18
  SPEAKER_00: I mean, obviously you you had a background, you had been working in the industry for some time, but what can one do to get promoted within two months at a company?
• 2:27
  SPEAKER_01: Oh wow.
• 2:28
  SPEAKER_01: Um, you know, I uh that's an interesting question.
• 2:31
  SPEAKER_01: Um well I guess uh you know I was uh partly lucky, but but also you know, I was working on the best technologies with the best uh people around always.
• 2:41
  SPEAKER_01: And uh I I I felt at that time Jensen did definitely take a risk and a chance in uh promoting me.
• 2:48
  SPEAKER_01: I I felt maybe I was at least one, if not two years away from that kind of a leadership role.
• 2:53
  SPEAKER_01: But I'm um, you know, I guess I was uh lucky.
• 2:56
  SPEAKER_01: Uh but uh you know the rest is history.
• 2:59
  SPEAKER_00: Yeah, it also a risk just because again, you know, even 10, 12 years ago, getting into robotics and AI wasn't wasn't a surefire bet.
• 3:10
  SPEAKER_00: And certainly not a surefire bet from somebody coming over from the mobile side.
• 3:14
  SPEAKER_00: I mean, you had come from TI, you know, you were working in mobile processors.
• 3:18
  SPEAKER_00: How difficult of a transition was that for you?
• 3:22
  SPEAKER_01: Yeah, it actually was not that difficult.
• 3:24
  SPEAKER_01: Because you know, if you start from first principles, the idea that robotics is going to be very big someday is not a surprising thought, right?
• 3:33
  SPEAKER_01: I mean, growing up, you know, watching Star Wars and Star Trek, it was obvious it's going to happen.
• 3:38
  SPEAKER_01: The question then is when.
• 3:40
  SPEAKER_01: And and basically technology had to catch up.
• 3:44
  SPEAKER_01: And uh, you know, working on mobile phones, we had to create you know extremely powerful super phones, if you will, processors that were extremely capable, but at the same time energy efficient, low power.
• 3:58
  SPEAKER_01: So it was actually you know perfect uh you know segue into the future world of robotics, future world of autonomous machines, which still need to be based on an SOC type of architecture, because you're talking about robotics having you know multiple uh sensors, whether it's cameras and uh LIDARs, radars, ultrasonic audio, and so on and so forth.
• 4:21
  SPEAKER_01: And now as we go into the world of humanoids in the future, there's got to be a lot more actuation, right?
• 4:27
  SPEAKER_01: Fine-grained uh dexterous manipulation type of tasks, which will require even more uh real-time uh processing and actuation at uh hundreds, if not thousands of hertz, right?
• 4:38
  SPEAKER_01: Uh operations per uh you know, uh movements per second that needs to happen.
• 4:42
  SPEAKER_01: So from my perspective, you know, if you're in the technology world and uh you know working on the latest and greatest technologies, the the evolution from one application to the next is just uh on first principles, is just uh fairly linear.
• 4:56
  SPEAKER_00: Yeah, it it's funny because you know, I granted my arc has taken uh quite a bit longer to make that transition from you know writing about smartphones and into writing about robotics, but I've I got there eventually.
• 5:09
  SPEAKER_00: And something that I realized pretty early on when I started covering some of these technologies when I went to research facilities is how much the success of mobile technology of smartphones has really enabled uh robotics in a big way because of how important and the role that manufacturing and scale has played in the success of robotics.
• 5:32
  SPEAKER_01: Yeah, absolutely.
• 5:32
  SPEAKER_01: I mean, if you if you go back to the um early 2000s, the the best in the world were working on in the mobile industry, right?
• 5:40
  SPEAKER_01: Smartphones.
• 5:41
  SPEAKER_01: That's when you know iPhone came about in 2007, Android came right after that, and people are working on that was the best technology.
• 5:47
  SPEAKER_01: And as the world transitioned from mobile phones becoming fairly mature in the early 2010s, the best in the world started working on cloud computing, right?
• 5:56
  SPEAKER_01: And you can see that, you know, uh how much of you know data center cloud computing happened.
• 6:00
  SPEAKER_01: And now today you can clearly see whether it is uh the top researchers in the world, and whether you can see where the VC, you know, venture capital funding is going.
• 6:11
  SPEAKER_01: You look at some of the large AI native companies, how much they're moving back into robotics.
• 6:16
  SPEAKER_01: Um, you know, you can see that the best in the world now are starting to work on physical AI and robotics.
• 6:22
  SPEAKER_01: And in fact, and the reason why now is two main pieces of technology have reached a tipping point, a sufficient level of maturity, which was needed for us to really double down in physical AI and robotics.
• 6:35
  SPEAKER_01: Number one is what happened with the chat GPT and large language models in November 2022, that extending it to vision language models and action models and agentic AI, that technology can definitely be applied to physical AI and robotics problems, where the accuracy requirements, of course, are far higher compared to in the digital world, as you know.
• 6:58
  SPEAKER_01: So that technology has recently matured, and that's why we see that everybody is kind of gravitating towards that.
• 7:03
  SPEAKER_01: Uh, the second uh piece of technology that was really needed to mature was simulation.
• 7:09
  SPEAKER_01: We've been working on simulation for gaming for well over two decades, and even for physical AI robotics applications, we started working on omniverse, even though the term omniverse was not coined until you know recently.
• 7:21
  SPEAKER_01: We've been working on it for gee, I mean, at least 10, if not 15 years in our company, trying to advance it.
• 7:27
  SPEAKER_01: Because we need to model the physics accurately, uh, we need to model the visualization as uh accurately, and we need to scale to you know tens, hundreds, if not thousands, of robots uh simulated in parallel.
• 7:40
  SPEAKER_01: So that technology has reasonably matured with Omniverse now.
• 7:44
  SPEAKER_01: And why simulation is fairly obvious, right?
• 7:46
  SPEAKER_01: I mean, it's faster, safer, and cheaper.
• 7:48
  SPEAKER_01: You know, every chip that we do in our company and the whole industry for 30 years, we 100% simulated, emulate it before we tape out.
• 7:56
  SPEAKER_01: But in robotics, that has not been true because the symptorial gap uh is too far.
• 8:03
  SPEAKER_01: I won't say that we solved it, but we are well on a path to making it better and better.
• 8:08
  SPEAKER_01: So I think because of uh generative AI and uh large language models and vision language models and number two simulation, you can see that now we can address or try to solve some of these problems that are needed to be solved in the next five, 10 years for robotics and physical AI.
• 8:23
  SPEAKER_00: Yeah, I I mean, like you're saying earlier, yes, I mean, abstractly, it's very clear that robotics, that automation, that AI are going to be a thing, that are going to be a very important part of our lives eventually.
• 8:38
  SPEAKER_00: But it's the kind of technology that, you know, every five to 10 years we feel like we're on the cusp of something.
• 8:43
  SPEAKER_00: And that's never really necessarily been true.
• 8:47
  SPEAKER_00: You even backing up again, like 10 to 12 years ago, and I understand why Jensen's the kind of guy who like people will follow him to the face of the earth.
• 8:56
  SPEAKER_00: Uh like I totally get that.
• 8:57
  SPEAKER_00: But it must have felt like a pretty big risk for you to make that shift from mobile phones, which were and continue to be like a very, you know, um, I guess, uh steady or stable space into something that still wasn't, you know, to a certain degree, continues to be fairly risky, like robotics and automation.
• 9:18
  SPEAKER_01: When we were in the market for mobile phones, every year we were launching a new uh Tegra processor at the time to address the highest performing smartphones.
• 9:26
  SPEAKER_01: But what we realized at that time in 2014, 2015, was the market was going to come out at us fairly fast.
• 9:32
  SPEAKER_01: And the value contribution that we can bring over a period of time, we can put our resources the rest of the world was doing phenomenally.
• 9:40
  SPEAKER_01: I mean, very well without us.
• 9:43
  SPEAKER_01: So we felt that it's the right time for us to you know try and do something different.
• 9:47
  SPEAKER_01: And in fact, we have a saying inside Nvidia that, you know, um if if the if you're not at least 10 years before a market or a technology takes shape, we are probably late.
• 9:57
  SPEAKER_01: In that spirit, I think it was a perfect timing for us to jump into physical AI, autonomous um, you know, robotics type of applications.
• 10:07
  SPEAKER_01: And then we also had the you know, the the benefit of autonomous vehicles were about to take shape in 2015.
• 10:14
  SPEAKER_01: So it wasn't like you know, we were completely coming out of left field.
• 10:18
  SPEAKER_01: So there were early indicators that it's clear that you know deep learning came about and we should be able to apply this technology.
• 10:25
  SPEAKER_01: So a lot of our resources went towards uh autonomous vehicles.
• 10:29
  SPEAKER_01: And then we were able to leverage that same technology and along with our complete uh CUDA stack, if you will, and uh start building um, you know, our Nvidia Jetson computers, which is where we started first in uh 2014, in fact.
• 10:43
  SPEAKER_01: Uh and and and the proof is in the pudding, right?
• 10:45
  SPEAKER_01: So if you look at it now, our latest generation is Thor, which is now available in August.
• 10:51
  SPEAKER_01: But the version before that, Orin, which we launched in 2022, has really taken off.
• 10:55
  SPEAKER_01: We have over 2.2 million developers in robotics that are working on our platform, over 7,000 companies that have either launched some sort of autonomous or semi-autonomous or automation product based on Jetson, or in the process of uh you know, serious uh development uh that is going on.
• 11:13
  SPEAKER_01: So that's uh really you know, yeah, proof points that even though it feels like we started early, uh, we had to actually start that early in order to bring the market of the technology to a place where you know it can be broadly uh adopted.
• 11:29
  SPEAKER_00: I suppose it's one of those things too, like as an outsider looking at it, it it seems like it's less left field, but the the pieces are all clearly there, right?
• 11:37
  SPEAKER_00: And they they fit together really nicely.
• 11:39
  SPEAKER_00: You know, simulation, GPU.
• 11:42
  SPEAKER_00: You're really just positioned in the exact right place in order to get put all of that together and deliver it to the community.
• 11:50
  SPEAKER_01: Yeah, that's right.
• 11:51
  SPEAKER_01: Um, we are building a platform for computing, right?
• 11:55
  SPEAKER_01: So uh accelerated computing, and likewise we're building a platform for robotics developers and companies.
• 12:00
  SPEAKER_01: We are not building robots.
• 12:02
  SPEAKER_01: So, what we do, uh what we realized is you need three computers in order to solve the robotics problem.
• 12:09
  SPEAKER_01: The first computer is for training the uh AI.
• 12:12
  SPEAKER_01: The second one is for testing the that AI, and simulation is the best place to test it in.
• 12:17
  SPEAKER_01: And the third, of course, is the computer where you deploy the AI, which is the runtime, the robot brain, right?
• 12:23
  SPEAKER_01: And when we actually started the journey, we actually started with the third computer first, because that was the only technology that was uh available at that time.
• 12:32
  SPEAKER_01: And in fact, it turns out what we've learned in this uh you know 10 plus years is that the slowest way to get to the destination, which is to deploy the robot in the end, is actually work, start working on the third computer first.
• 12:45
  SPEAKER_01: But we had no choice 10 years ago because simulation technology did not exist.
• 12:49
  SPEAKER_01: We did not 10 years ago, believe it or not, our first DGX training computer was launched in 2016.
• 12:55
  SPEAKER_01: So when we started in 2014, we we had uh we didn't have the technology.
• 12:59
  SPEAKER_01: So we started building Jetson and we started engaging with the robotics community to start doing some computer vision and initial deep learning sort of uh applications.
• 13:10
  SPEAKER_01: And obviously, as the technology evolved, now we have the three computers.
• 13:14
  SPEAKER_01: And and what we do is we provide the infrastructure, which is the three computers, and there's four uh major steps, right?
• 13:21
  SPEAKER_01: So if I walk, if you walk backwards from the end to the beginning, uh the the final step, of course, the fourth and the final step is the physical deployment.
• 13:31
  SPEAKER_01: And in order to deploy uh AI into a robot, of course, you need a lot of runtime, real-time acceleration libraries, runtime models, the whole tool chain that we provide, right?
• 13:42
  SPEAKER_01: So we do that.
• 13:44
  SPEAKER_01: But before you deploy it, you obviously have to test it, otherwise you're never going to deploy it in the physical world.
• 13:49
  SPEAKER_01: And testing in simulation, that's a big piece.
• 13:52
  SPEAKER_01: Um, you know, uh, as I mentioned earlier.
• 13:55
  SPEAKER_01: In fact, I believe that in the next five, 10 years, for every one physical robot, you probably are going to be running thousand, if not a million times more in simulation just because it's faster, safer, and cheaper.
• 14:08
  SPEAKER_01: So the third step is testing in simulation.
• 14:10
  SPEAKER_01: To test it, the second step is, of course, you know, you you have to train it, you want to create the AI model.
• 14:16
  SPEAKER_01: And Nvidia, you know, we've been pioneering the workflows for training, not just for robotics.
• 14:21
  SPEAKER_01: Of course, we started with uh you know um deep learning and then moved into uh relevant.
• 14:26
  SPEAKER_01: So that's the second step.
• 14:28
  SPEAKER_01: Then the first step, of course, walking backwards, it turns out, is actually the biggest challenge we have in robotics today.
• 14:35
  SPEAKER_01: And if you can guess what it is, you know, for Chat GPT or LLM, the whole internet existed.
• 14:41
  SPEAKER_01: We, the humanity created, you know, the whole internet in the last 25 years, and so all of the data pretty much that we need exists.
• 14:48
  SPEAKER_01: But for robotics, the data does not really exist.
• 14:52
  SPEAKER_01: And human capture data is necessary, but it's barely going to be sufficient.
• 14:57
  SPEAKER_01: It's not even going to be one person the total data.
• 14:59
  SPEAKER_01: Like the data is out there, right?
• 15:00
  SPEAKER_01: The world is out there, we just don't have a database to train on.
• 15:03
  SPEAKER_01: That's right.
• 15:04
  SPEAKER_01: The world is it's it's out there and and it's a lot of it is not purposeful enough in order to do these, especially the problems that we're trying to solve for labor shortages or dangerous jobs, where you know, humanity, for us, it's so easy to do dexterous uh manipulation.
• 15:22
  SPEAKER_01: Uh, but but in the case of uh AI robotics, that technology doesn't even exist.
• 15:26
  SPEAKER_01: In fact, it's in research today.
• 15:28
  SPEAKER_01: So we are you know creating workflows for synthetic data generation, augmenting real-world capture.
• 15:35
  SPEAKER_01: So data generation first, training second, testing and simulation third, and deployment fourth.
• 15:41
  SPEAKER_01: And if you look at where much of the effort in the industry right now is, especially as it relates to solving the future problems of dextress manipulation, and it turns out a lot of dexterous manipulation can be solved with bimanual versus just a single arm.
• 15:55
  SPEAKER_01: If it's a fairly big, rigid body and you're doing pick and place, the existing cobalt arms and existing industrial arms with suction grippers or two-fingers, kind of it works reasonably well as AI is advancing.
• 16:09
  SPEAKER_01: But many of the problems that we're trying to solve are not exact are sort of deformables in a way, right?
• 16:15
  SPEAKER_01: Even if it's not extremely fluid or uh, you know, but but still deformables, wires and so on.
• 16:20
  SPEAKER_01: So data generation is much of the effort in the industry right now is being spent on technologies for making data generation more efficient so that you can then train the model.
• 16:30
  SPEAKER_01: And then once you train the model, testing it in simulation is where we are seeing rapid uh explosion of uh ecosystem partners that are jumping on it.
• 16:40
  SPEAKER_00: I was at, I mean, obviously I was there because it was an A3 event, but I was watching the talk that you gave at Automate, and I think that the term that you were using was synthetic data that really jumped out at me because it's basically this like almost this infinite loop of training and simulate and simulation where it's constantly just training itself on data.
• 16:59
  SPEAKER_00: And I'm wondering if that's a way to, I know this word is used a lot in the context of LLMs, but to start introducing like hallucinations and to start, you know, like classically, if you're making Xeroxes, the more copies you make of something, the further away it gets from the original.
• 17:18
  SPEAKER_00: So the more you train something in synthetic data, the more opportunity you have to introduce false data into the process.
• 17:26
  SPEAKER_01: Yeah, absolutely.
• 17:27
  SPEAKER_01: But there's two two main ways we can address it.
• 17:30
  SPEAKER_01: Number one, of course, there's always back it up with uh human data collection, right?
• 17:36
  SPEAKER_01: So so that much of the synthetic data that you're generating could be based on real-world data, but then you're changing the trajectories, changing the background, changing the environment, changing scenarios.
• 17:49
  SPEAKER_01: The uh, you know, the odds of elucidating from that would be fairly low.
• 17:52
  SPEAKER_01: And second, even if it's purely synthetic data generation, you should have reinforcement learning, right?
• 17:58
  SPEAKER_01: So you can always have, you know, what the same thing was true with even back in the day with LLMs a couple of years ago.
• 18:04
  SPEAKER_01: The first check for all that was done with humans, but you know, what's called reinforcement learning human uh based on human feedback.
• 18:12
  SPEAKER_01: But increasingly, you know, of course you can use human feedback, but we can also build AIs that are very purpose-built to make sure that the the model, the the data is not elucidating and it's exactly in the domain and and conditions that you want it to operate in.
• 18:27
  SPEAKER_01: So you can use other AIs to uh basically curate it and fine-tune it.
• 18:32
  SPEAKER_00: Yeah, I I think one of the things that that it sounds like you've discovered is that it's not is that training robots is it's gonna be a combination of everything, right?
• 18:41
  SPEAKER_00: It's gonna be video training, it's gonna be simulation, it's even gonna be.
• 18:44
  SPEAKER_00: I was talking to the founders of SKILD recently, and you know, they're they're talking about pre-training, post-training, that even like teleop has a role to play in that process.
• 18:53
  SPEAKER_01: That's right.
• 18:54
  SPEAKER_01: So there's multiple uh things that need to be done, right?
• 18:58
  SPEAKER_01: So first is you know, real-world data collection.
• 19:01
  SPEAKER_01: Real world data connection can be done through teleoperation, is a good scenario.
• 19:06
  SPEAKER_01: Uh, there's also imitation learning where you can just do the, you know, show the robot how you're doing without teleoperation.
• 19:12
  SPEAKER_01: So those are the two most popular.
• 19:14
  SPEAKER_01: And then, of course, third is uh showing videos, right?
• 19:17
  SPEAKER_01: It's the same as you know, essentially it's imitation learning.
• 19:19
  SPEAKER_01: So those are the three biggest ways.
• 19:21
  SPEAKER_01: Uh uh, you know, the whole industry is crafty towards.
• 19:24
  SPEAKER_01: And we think that all of those are good to have and necessary.
• 19:28
  SPEAKER_01: Uh the question then is are they sufficient?
• 19:30
  SPEAKER_01: And the answer is they're not sufficient.
• 19:32
  SPEAKER_01: Do as much as you can.
• 19:34
  SPEAKER_01: If we can collect more uh real-world data, we should.
• 19:37
  SPEAKER_01: But but the belief is that that's not sufficient.
• 19:40
  SPEAKER_01: So why not just use both Gen AI, but also use traditional heuristics to uh to compute more trajectories.
• 19:49
  SPEAKER_01: In order to compute more trajectories, for example, let's say I'm picking up a glass.
• 19:52
  SPEAKER_01: You know, I that's an example, I'm just going to pick up a glass.
• 19:55
  SPEAKER_01: But the reality is you don't need to just pick it up like this.
• 19:58
  SPEAKER_01: You could probably go like this, you could probably go like that, and experiment with multiple trajectories.
• 20:03
  SPEAKER_01: And if you make that as part of the training data set, it's highly likely that the model that you're training is going to be more resilient and more general purpose.
• 20:14
  SPEAKER_01: Right?
• 20:14
  SPEAKER_01: That's that's fundamentally the objective.
• 20:16
  SPEAKER_01: What is everybody trying to do in the end?
• 20:18
  SPEAKER_01: We're trying to create a reasonable general purpose brain or to put into the robot.
• 20:24
  SPEAKER_01: And then once you build a reasonable general purpose brain, surely we're going to be post-training it for each and every domain that we want it to excel in, right?
• 20:34
  SPEAKER_01: Very similar to uh if you look at Chat GPT, uh LLMs.
• 20:38
  SPEAKER_01: So pre-LLM, uh from call it 2014, 2015 to 2022, there were AIs that have been deployed in a lot of uh applications, especially in the cloud, right?
• 20:48
  SPEAKER_01: For consumer type of applications.
• 20:50
  SPEAKER_01: You could call them so-called specialist models, uh, very specialized doing certain things.
• 20:56
  SPEAKER_01: And they were doing reasonably okay.
• 20:58
  SPEAKER_01: But but then ChatGPT came about it was a very good generalist model.
• 21:03
  SPEAKER_01: It could do as good or better than almost all those specialists and do more general purpose tasks, right?
• 21:10
  SPEAKER_01: So is that good?
• 21:11
  SPEAKER_01: Of course it's good.
• 21:12
  SPEAKER_01: But then you realize, okay, that's a good generalist, but for many of my enterprise tasks, I want to specialize that generalist.
• 21:20
  SPEAKER_01: So now the third step's happening, which is you take these fairly good general purpose models and then derive generalist specialists from that.
• 21:28
  SPEAKER_01: It's kind of like this.
• 21:29
  SPEAKER_01: Uh you know, a generalist is a good, imagine like for a company, you're hiring an undergrad student or a master's student, PhD student, right?
• 21:37
  SPEAKER_01: The so-called specialist before that was like a middle school student.
• 21:41
  SPEAKER_01: Can they do tasks?
• 21:42
  SPEAKER_01: Yes, but it's very narrow.
• 21:44
  SPEAKER_01: But so but we don't hire them, right?
• 21:45
  SPEAKER_01: So you hire a generalist, like a 20 year old, 25 year old.
• 21:48
  SPEAKER_00: A generalist is in baseball, they call it a utility player, right?
• 21:52
  SPEAKER_00: Somebody who maybe isn't like great at any one position, but can play each of them competently well.
• 21:57
  SPEAKER_01: Aaron Powell That's right.
• 21:58
  SPEAKER_01: And it's a good, and they're a good athlete.
• 22:00
  SPEAKER_01: They understand the game.
• 22:01
  SPEAKER_01: They they they can, you know, absolutely right.
• 22:04
  SPEAKER_01: So it's but but you need that generalist in order to derive a specialist, otherwise you won't get a good baseball player, right?
• 22:10
  SPEAKER_01: If you're not a good athlete, uh well, maybe I mean not in today's world.
• 22:13
  SPEAKER_01: Maybe 50 years ago you can get away without being a good athlete in baseball, but but not in today's world, right?
• 22:18
  SPEAKER_00: Aaron Powell One of the conversations I keep coming back to, and and and I think that this is kind of this this bifurcates people quite a bit, is ultimately how useful using LLMs, um, using generative AI as a basis will be for robotics.
• 22:33
  SPEAKER_00: And you and you're getting into this, uh you alluded to this a little bit before.
• 22:37
  SPEAKER_00: Um, you know, obviously it doesn't translate directly, language doesn't translate directly to the real world.
• 22:43
  SPEAKER_00: And when I ask a lot of folks in the robotics space, the ways in which LLMs, um these transformers will be useful, it's uh you know, communication, right?
• 22:55
  SPEAKER_00: Uh natural language.
• 22:56
  SPEAKER_00: This is something that I know is a big part of this new um the Thor update.
• 23:02
  SPEAKER_00: Uh but beyond that, that maybe you have to think entirely differently when it comes to physical AI.
• 23:09
  SPEAKER_00: I mean, how how difficult of a transition are we talking from this world of of you know LLMs into like actual embodied AI?
• 23:18
  SPEAKER_01: Where the industry mostly is looking at right now is fundamentally using uh multimodal reasoning.
• 23:24
  SPEAKER_01: So this would be you you could say it's LLM, VLM, reasonable size because you got to run it at the edge inside the robot brain.
• 23:33
  SPEAKER_01: Uh so today, the best uh, you know, you mentioned uh Thor, uh, you know, that's our latest uh generation uh over two uh pedaflops of uh compute that's available in a you know in a robot brain.
• 23:45
  SPEAKER_01: That's amazing.
• 23:46
  SPEAKER_01: Compared to when we started Jetson 10 years ago, uh it's 7,000 times more compute that we're providing uh, you know, in the same uh same uh form factor.
• 23:57
  SPEAKER_01: So so it's typically, you know, what we're seeing right now is you know you end up using uh several tens of billions of uh parameter models.
• 24:06
  SPEAKER_01: That's where the technology is today.
• 24:07
  SPEAKER_01: You'd use it for reasoning because you need to understand.
• 24:10
  SPEAKER_01: And you're typically connecting to multiple sensors at the same time.
• 24:14
  SPEAKER_01: And then once you do the reasoning, you have to do the action after doing the reasoning.
• 24:18
  SPEAKER_01: And that's where a vision language action model comes in.
• 24:21
  SPEAKER_01: So what we're seeing most of uh most of the research, including you know our project group, typically what we do is we you know, we have a reasoning, we use a reasoning model as the front end, and then we have a vision language action model.
• 24:33
  SPEAKER_01: Um, and that's kind of where most of the, you know, most of the industry is currently you know exploring.
• 24:39
  SPEAKER_01: And vision language or action models, VLAs as they're called, uh typically you know, you know, in the in the two to five billion parameter model.
• 24:47
  SPEAKER_01: So that's kind of where the state of the art today is.
• 24:49
  SPEAKER_01: But as we and and the job number one right now is accuracy, right?
• 24:55
  SPEAKER_01: If you don't get it fairly accurate to where you want to get to, if you just run it super fast, it might not be that useful.
• 25:02
  SPEAKER_01: So what we are working on is on both the fronts, which is the accuracy needs to be sufficiently good for whatever task you're trying to do, and then and then optimize it to get reasonable performance.
• 25:13
  SPEAKER_01: And then we know that you know, based on all the work that we're doing in in terms of performance and perf per watt and perf per watt per dollar, you know, this that trend, you know, Nvidia with accelerated computing, we are aggressively on a roadmap to whether it's data center or whether it's in a you know workstation or in a in a Jetson, we are aggressively, you know, on that path to keep improving generation over generation.
• 25:35
  SPEAKER_01: So that's kind of where we are today.
• 25:37
  SPEAKER_01: So so leveraging multimodal models for uh reasoning and then and then creating a action model for doing the physical uh actuation.
• 25:47
  SPEAKER_00: It's interesting to hear you caveat like that's where we are today, which seems to imply to me that that's probably not where things are gonna be tomorrow, that things are like still constantly evolving.
• 25:56
  SPEAKER_00: Is that a safe assumption?
• 25:57
  SPEAKER_01: Yeah, I mean, that's the best we have today, but of course, you know, you know, the progress, you know, there could be a new technology that comes in, and maybe there's a you know, there's a new version of a LLM, VLM combination that would suddenly, you know, the thing that I I've at least realized in the last 10 years working on AI and all of this is improvements don't come in uh, you know, they come in step functions.
• 26:19
  SPEAKER_01: Yeah, it's suddenly when ChatGPT came out in 2022, it wasn't like it's like 50% better or 100%.
• 26:25
  SPEAKER_01: It's like suddenly, you know, you're you're going like this, and then suddenly the curve turns like this, right?
• 26:30
  SPEAKER_01: So uh maybe there's a step function or two coming in the next five years.
• 26:34
  SPEAKER_01: And when that happens, um, you know, that's why I'm I'm saying, you know, I'm saying just looking at everything that will work.
• 26:40
  SPEAKER_00: Well, yeah, I mean that's something I was curious about.
• 26:42
  SPEAKER_00: You know, when you look at something like Chat GPT from the outside, from, you know, and and and and chat GPT is a really good example because that's something that like my parents who don't know tech at all, like, you know, they're probably familiar with, they probably try it out with.
• 26:56
  SPEAKER_00: It you know, the and the the classic Arthur C.
• 27:00
  SPEAKER_00: Clark quote of like anything that any technology that's sufficiently advanced, you know, is is is indistinguishable from magic.
• 27:07
  SPEAKER_00: Like it feels like magic, right?
• 27:08
  SPEAKER_00: You know, you type in and you can write a book, you can draw, you can draw a picture, and that feels like just immediate like overnight.
• 27:15
  SPEAKER_00: Um, but I suspect that people working on it on the inside, you know, maybe it feels a little bit more gradual.
• 27:20
  SPEAKER_00: But hearing you talk about it, somebody in the industry, it does feel like things just kind of overnight to a certain extent took off.
• 27:28
  SPEAKER_01: Yeah, so so there's actually two dimensions, right?
• 27:31
  SPEAKER_01: So when we think of the the impact of uh you know, impact of the step function improvements that that happened.
• 27:38
  SPEAKER_01: One, of course, is the the technology itself in the case, accuracy, what it can do, right?
• 27:43
  SPEAKER_01: So uh whether it's the speed at which you can run or whether it's the you know the accuracy of solving problems, that's fundamentally, of course, important.
• 27:52
  SPEAKER_01: There's a second thing that a lot of people do not look at or underestimate, which is actually equally valuable, is the ease of use.
• 28:00
  SPEAKER_01: Right?
• 28:00
  SPEAKER_01: The reason Chat GPT, you know, you you you mentioned it exactly, like you know, your parents use it, is because it's virtually indispensable because it's so easy to use, right?
• 28:10
  SPEAKER_01: So one of the biggest challenges in in robotics and physical AI is even if you had a good technology, how do you integrate that into existing processes and workflows?
• 28:20
  SPEAKER_01: It's not easy at all, especially these are industries that have been around for tens of years, if not uh, you know, a hundred plus years, right?
• 28:28
  SPEAKER_01: And and just deploying technology is not at all seamless.
• 28:32
  SPEAKER_01: So what's happening with with with LLMs and Gen AI in in general is the integration points, we are hopeful and we are seeing early indicators that the integration is becoming easier and easier, just like it happened in the consumer world.
• 28:47
  SPEAKER_01: So so if a step function improvement comes next with both technology and ease of use, uh, I don't think it's going to be linear adoption.
• 28:55
  SPEAKER_01: Suddenly you'll you'll find like you know, there's a sudden massive uh you know spike.
• 29:01
  SPEAKER_00: Yeah, that's interesting because when I'm thinking about workflows and integration, I'm thinking about it from two angles.
• 29:07
  SPEAKER_00: And and I think the one that you're getting at right now, which is you know the obvious one and and for good reason, is people in existing industries.
• 29:15
  SPEAKER_00: Let's take obviously we're talking about manufacturing a lot and we're talking about fulfillment a lot, but let's take like ag tech, right?
• 29:23
  SPEAKER_00: That this is obviously a really big one, and you know, this is that's as ancient as an industry gets.
• 29:28
  SPEAKER_00: And because of that, like it can be difficult to convince people who have been doing things the same way for centuries, if not millennium, to update, um, to lower the barrier of entry.
• 29:39
  SPEAKER_00: Um but but then the other side of it, and I think this is something that you're working on as well, is to make it easier to actually build the technology.
• 29:47
  SPEAKER_00: That's a big piece of it, too, that you're working on.
• 29:50
  SPEAKER_01: Because we're a platform company, you know, uh we we are providing the infrastructure, we are providing the acceleration libraries and reference models and tools and workflows for companies, whether they're building robot brains, whether they're building the physical robots themselves, or whether they're building the simulation environment, or a combination of any of those.
• 30:09
  SPEAKER_01: So that's where we basically spend 99% of our time is provide those three computers, like I talked about.
• 30:15
  SPEAKER_01: And then on each of those computers, we provide acceleration libraries and the different tools and workflows to make it easier and easier uh to develop uh these uh applications.
• 30:26
  SPEAKER_01: And then that's a testament that we're actually proud of.
• 30:30
  SPEAKER_01: Seeing the fact that, you know, from 2022 to 2025, you know, from going from zero to seven thousand plus companies launching products or in the case of deploying products, surely, you know, we had to do something reasonably well to enable so many companies, many of which, by the way, we do not directly uh majority of which we do not directly know who they are and engage with.
• 30:52
  SPEAKER_01: So it's all the power of the platform and the ease of use, not just about performance and features, but easy to use, whether it's from documentation for reference examples, or even a partner ecosystem that we have built, which enables uh, you know, uh enables all of these companies to leverage many different layers in the ecosystem to build these products.
• 31:15
  SPEAKER_00: It it seems like uh you and and and Jetson and and NVIDIA are also really involved in uh the open sourcing side of things as well, like especially with hugging face and the robot entering the picture.
• 31:28
  SPEAKER_00: Um and I and I think you've had an involvement with uh with Ross as well.
• 31:32
  SPEAKER_00: Like why is that an important aspect in all of this?
• 31:35
  SPEAKER_01: Yeah, I mean, we we are an open platform and you know we we we open source our source open, you know, both both scenarios are true.
• 31:44
  SPEAKER_01: You know, we open source as much as we can.
• 31:46
  SPEAKER_01: Where we cannot open source, we try to source open.
• 31:48
  SPEAKER_01: The difference being that you know we give access to source code, right, to to many of these companies.
• 31:53
  SPEAKER_01: And and that's super important because at the end of the day, if you think about it from first principles, right?
• 31:58
  SPEAKER_01: Imagine you are imagine you're an Agtech company because you you mentioned that.
• 32:02
  SPEAKER_01: Uh you have a sizable business to run, and and you're going to be building an autonomous machine.
• 32:07
  SPEAKER_01: Imagine it's a tractor, or imagine it is a fruit picker, for example, right?
• 32:12
  SPEAKER_01: With with multiple arms and all doing all sorts of AI in the future.
• 32:16
  SPEAKER_01: When you're going to deploy this to your customers in your farm, these are going to be in the field for 10 years, if not more, right?
• 32:24
  SPEAKER_01: In some cases, they could be in the field for 20 years.
• 32:27
  SPEAKER_01: And what is the number one expectation with AI in this new world?
• 32:30
  SPEAKER_01: AI is not going to be stagnant.
• 32:32
  SPEAKER_01: You're going to be constantly updating it with new technology, new capabilities, which means you can you're providing new additional value to your customer, which potentially you could monetize, right?
• 32:43
  SPEAKER_01: So the moment you deploy this autonomous machine, your job just starts.
• 32:47
  SPEAKER_01: Even though it took you so long to build that machine and make it good, your job is never complete.
• 32:52
  SPEAKER_01: So you have another 10, 15, 20 years of work to do.
• 32:55
  SPEAKER_01: Now, in order to maintain and update this, you would absolutely like to have 100% control of your own destiny, right?
• 33:03
  SPEAKER_01: Would you, what happens if you find a bug five years after launch and you with a new AI model?
• 33:09
  SPEAKER_01: Uh, would you?
• 33:09
  SPEAKER_01: I mean, obviously they could file a bug on Nvidia, but maybe Nvidia has moved on to something else, and you know, or maybe somebody else that they worked on, right?
• 33:16
  SPEAKER_01: So the the need to have you know deterministic control of your own destiny means that you have to have access to all of the technology, all of the software as much as possible.
• 33:28
  SPEAKER_01: So we've realized this very early on in the process.
• 33:31
  SPEAKER_01: And that's why, you know, yeah, you know, we we make sure that everybody has the ability to control their destiny.
• 33:36
  SPEAKER_01: So open source as much as possible, work with industry consortiums to release, uh, you know, continually update these technologies, right?
• 33:44
  SPEAKER_01: So that's fundamentally the philosophy as to why we keep all of our platforms open.
• 33:48
  SPEAKER_00: I completely understand again why manufacturing and why fulfillment are these like two primary focuses right now.
• 33:58
  SPEAKER_00: But obviously things are expanding quickly into other industries.
• 34:01
  SPEAKER_00: You know, the conversation is being had right now a lot about um uh uh uh home robotics.
• 34:07
  SPEAKER_00: Like ultimately, when you're doing your work, uh whether it's building the hardware or working on simulation, like how much does the use case, how much does the actual end user and how and where they're using and deploying the technology matter to what you're building?
• 34:24
  SPEAKER_01: The technology that we're building is going into multiple use cases.
• 34:28
  SPEAKER_01: The idea of data generation, the idea of training, the idea of testing and simulation and deployment is mostly common.
• 34:36
  SPEAKER_01: But then when it comes to the final deployment, there are variations, right?
• 34:40
  SPEAKER_01: The beautiful thing about robotics is what's the level of accuracy that you need in order for it to be deployed at scale.
• 34:48
  SPEAKER_01: Uh and we, you know, I call it the number of nines of accuracy you need after 99 point, right?
• 34:54
  SPEAKER_01: There are some use cases uh where maybe 99.9 is enough to deploy it.
• 35:02
  SPEAKER_01: There'll be some use cases where you probably need 10 nines, like autonomous vehicles is probably one of those, surgical robots probably will be one of those where you know you absolutely need like you know, uh a trillion, you know, 109s or or famously, you know, self-driving cars, right?
• 35:18
  SPEAKER_00: I mean, the first person that dies on the that's that sets the industry back a decade.
• 35:25
  SPEAKER_01: That's right.
• 35:25
  SPEAKER_01: That's right.
• 35:26
  SPEAKER_01: And then and then you walk to the other side.
• 35:28
  SPEAKER_01: Maybe agriculture is one of those where the number of nines is relatively lower, only because the number of humans that are going to be in the in the farm is lower, right?
• 35:39
  SPEAKER_01: So so that's why you see, like, why do you s why do we see manufacturing and warehousing is where most of the activities today?
• 35:46
  SPEAKER_01: Well, number one, of course, is labor shortage.
• 35:48
  SPEAKER_01: There's the highest need.
• 35:49
  SPEAKER_01: But also it turns out in those spaces, you can kind of fence them reasonably to be structured, right?
• 35:56
  SPEAKER_01: Not not it need not be 100% unstructured.
• 36:00
  SPEAKER_01: If it's in a home, it's probably relatively unstructured.
• 36:03
  SPEAKER_01: But in a factory and warehouse, you can design around, you can probably design a work cell where the robots like relatively, you know, uh compared to humans, right?
• 36:12
  SPEAKER_01: You can, you know, fence them off.
• 36:14
  SPEAKER_01: And although increasingly we are seeing that that there's a need for humans and robots to collaborate, and and this is where the the latest AI is helping us to bridge that gap, right?
• 36:25
  SPEAKER_01: So so that's kind of how we see the technology.
• 36:28
  SPEAKER_01: And of course, it's safety.
• 36:29
  SPEAKER_01: Eventually, I mean, all of these are going to be super safe because of our autonomous vehicle heritage.
• 36:34
  SPEAKER_01: We have invested over 15,000 engineering years into safety already.
• 36:40
  SPEAKER_01: And and this is something that will not be compromised, right?
• 36:43
  SPEAKER_01: When you look at all these folks that are going to be deploying robots, um, you know, so safety is going to be essential.
• 36:49
  SPEAKER_01: And safety comes from in the machine to the surroundings.
• 36:53
  SPEAKER_01: You can deploy safety solutions that are using the surrounding.
• 36:55
  SPEAKER_01: You can put a camera that's not on the robot but in the warehouse that can be monitoring and providing outside in.
• 37:03
  SPEAKER_01: Um so we're spending a lot of time with the ecosystem and building this uh uh safety uh also.
• 37:08
  SPEAKER_01: We we call it Nvidia Halos.
• 37:10
  SPEAKER_01: So we introduced this uh earlier this year.
• 37:12
  SPEAKER_01: So that's going to be a big part of uh bringing robots at uh at scale.
• 37:17
  SPEAKER_00: Well, so the you bring up a really good point.
• 37:19
  SPEAKER_00: And I'm glad that I feel like this is a conversation that people are having more and more in the industry over the last couple of years, especially as it pertains to humanoids, both in the in the factories, and again, as things are increasingly hypothetically moving into the home, you know, age tech, working uh, you know, alongside the most vulnerable members of society, this is going to be really important.
• 37:39
  SPEAKER_00: But you know, you're you're discussing, and when I think of safety, I think about sensors, right?
• 37:44
  SPEAKER_00: I think about the way it's, you know, as you said, externally interacting with people.
• 37:48
  SPEAKER_00: But like what can you do as a company to actually bake that into the actual like processors or the simulation or the things that you're building on your end?
• 38:00
  SPEAKER_01: So safety is uh comes in many layers, right?
• 38:02
  SPEAKER_01: So starting at the very bottom, we are our our processors like Orin and Thor are ACLB certified.
• 38:09
  SPEAKER_01: So we have uh redundancy inside, we have a safety cluster, safety operating system that's running in parallel with the main operating system.
• 38:19
  SPEAKER_01: So we have a lot of capability within the SOC itself.
• 38:22
  SPEAKER_01: And then one layer above that, when you're building the module and the system, the hardware, you have safety microprocessors, safety microcontrollers that you know many of the industry partners like uh Infinian and TI and Renesas, and there's many others that provide these solutions that are typically embedded on the total compute board, if you will, that can also do additional checking.
• 38:44
  SPEAKER_01: And then you the next level is software.
• 38:47
  SPEAKER_01: So we are making an operating system, including our CUDA stack, all the runtime safety certified.
• 38:54
  SPEAKER_01: We started that with autonomous vehicles, and we are taking that same thing, same approach for making it for robotics as well.
• 39:00
  SPEAKER_01: So you have all the different checkpoints, you have all the documentation, and you have the testability points for for the core level operating system.
• 39:07
  SPEAKER_01: And then one level above that, of course, you can make the algorithms themselves to be resilient, right?
• 39:13
  SPEAKER_01: So if you have a certain computer vision, you can you can run your deep learning algorithms, so you can learn Gen AI algorithms, but maybe you can also have your traditional computer vision algorithm to do a similar task and you can cross-check against it, right, at one phase.
• 39:27
  SPEAKER_01: So there is that level of uh safety that comes from algorithmic uh level itself.
• 39:33
  SPEAKER_01: And then all of this, this complete stack uh from all the different layers of software to the hardware, uh, it has to be open for standards, bodies, uh certification bodies to be able to assess.
• 39:46
  SPEAKER_01: So we've partnered with TUV and some of the others to basically have this safety labs where independent third-party companies can can certify this.
• 39:55
  SPEAKER_01: So safety is a long journey.
• 39:57
  SPEAKER_01: We mentioned 15,000 engineering years, that's a lot of investment.
• 40:00
  SPEAKER_01: Uh, but we fundamentally believe that it's not a differentiation, it's going to be a must-have that you would need, especially when you're talking about five, 10, 15 years from now when robots are going to be deployed at scale.
• 40:12
  SPEAKER_01: Can you imagine you you come to the market and say, oh, I'm not safe or it's not certified?
• 40:18
  SPEAKER_01: It just that's not pragmatic.
• 40:21
  SPEAKER_00: Yeah, again, like I think self-driving cars are a perfect example because this is something that like we keep having expected to have come out.
• 40:28
  SPEAKER_00: And and and I I just fundamentally think that the expectations, people have much higher expectations for technology than human drivers, right?
• 40:36
  SPEAKER_00: I mean, like it's built in that a certain number of people every year are gonna die because of some human error, but you know we we we can't ex we won't accept that from technology in the same way.
• 40:50
  SPEAKER_01: That's right.
• 40:51
  SPEAKER_01: Uh the bar is extremely high, and uh I think it should be, it makes sense.
• 40:55
  SPEAKER_01: And and we have technology, you know, one of the things I you know didn't mention is simulation can play a massive part in creating you know not only the accuracy but safety as well, right?
• 41:05
  SPEAKER_01: Because you can really create scenarios that are virtually impossible to encounter in the physical world, or you encounter them in such a such a rare uh way that you're not even comprehending it.
• 41:18
  SPEAKER_01: I think that that ability using simulation, using generative AI to generate lots of scenarios is actually making our going to make our algorithms models more and more resilient.
• 41:30
  SPEAKER_01: And and you can you can you know you can continually add these.
• 41:33
  SPEAKER_01: So I think that's the uh other big step that we think that will help not only robotics, but of course AV as well, to be able to use simulation and gen AI to create all these scenarios.
• 41:45
  SPEAKER_00: Yeah, and you can process so many at once that you know you don't have to do in real time.
• 41:49
  SPEAKER_00: So something like I'm kinda I really want to get your take on this because this is something that I've been thinking about a lot lately, uh, you know, as far as like the job that I do.
• 41:56
  SPEAKER_00: And I think you can relate to this, you know, as somebody again who has sort of made this transition to really writing about consumer electronics to writing about robots.
• 42:04
  SPEAKER_00: I worry that sometimes maybe I cover robotics and automation too much like I've covered consumer technology, and that in doing so, maybe it causes people to underestimate the difficulty of things, you know, how difficult is it is to scale, uh, how important safety is, like, you know, how long it's actually going to take to deploy some of these technologies.
• 42:29
  SPEAKER_00: And I was kind of thinking about this in the context of um like some of these events, right?
• 42:33
  SPEAKER_00: It's and I I I kind of wonder like if I'm I'm a member of the audience and I'm I'm watching like a keynote and you know, I'm seeing all these like cool like you know video game graphics, and then I see humanoid robots on the same stage, it all kind of like, you know, it it takes this sort of same pleasure sensor in my brain.
• 42:51
  SPEAKER_00: But are we setting realistic expectations as far as how long it's gonna take and how difficult it's actually gonna be to deploy these technologies in the real world?
• 43:01
  SPEAKER_01: I think the the answer is both yes and no, right?
• 43:04
  SPEAKER_01: And and if you think about it, you know, robotics ultimately is the largest opportunity in front of humanity.
• 43:10
  SPEAKER_01: The world is, you know, 90% of the world is physical and physical things.
• 43:14
  SPEAKER_01: And the need for robotics has always existed for decades and it's only increasing.
• 43:19
  SPEAKER_01: You're all going to be living a lot longer.
• 43:21
  SPEAKER_01: And the new generation of uh humans do not want to do certain jobs that the previous generations did.
• 43:27
  SPEAKER_01: That's just a fact of life, right?
• 43:29
  SPEAKER_01: So the the reason we have not been able to solve any of those before was technology wasn't good enough.
• 43:35
  SPEAKER_01: Now, as much as it is difficult to solve these problems today, I you know, if you go back 10, 15 years ago when when mobile revolution started, the cloud revolution started, it was equally difficult for humanity to solve those problems with whatever technology was available at that time.
• 43:51
  SPEAKER_01: But the best in the world migrate to that problem and eventually we solved it.
• 43:55
  SPEAKER_01: So suddenly we look at it today.
• 43:57
  SPEAKER_01: The mobile phone looks like, oh yeah, it's a piece of cake.
• 44:00
  SPEAKER_01: Everybody on the planet uses it.
• 44:01
  SPEAKER_01: But what happened in the 1990s?
• 44:03
  SPEAKER_01: Why did we have that capability in the 1990s?
• 44:05
  SPEAKER_01: Right?
• 44:05
  SPEAKER_01: It's just that the technology didn't evolve.
• 44:07
  SPEAKER_01: So the same way, I think robotics is now we have core pieces of technologies that will enable us to solve some problems that were impossible before.
• 44:17
  SPEAKER_01: And surely, as we go in this journey, we'll find that there are still some more problems.
• 44:22
  SPEAKER_01: This technology is not good enough to solve.
• 44:25
  SPEAKER_01: And it's going to be difficult.
• 44:27
  SPEAKER_01: And there'll be more ideas that we want to pursue.
• 44:30
  SPEAKER_01: And so I don't think this is a 10-year journey.
• 44:33
  SPEAKER_01: I think this is a multi-decade journey.
• 44:35
  SPEAKER_01: And we will solve some use cases in the next five, 10 years.
• 44:38
  SPEAKER_01: And there'll be more interesting problems that we need to solve beyond that.
• 44:43
  SPEAKER_00: One of the things, you know, I was going back and watching some of the uh GTC and some of the uh Nvidia keynotes.
• 44:49
  SPEAKER_00: And it's really it was striking to me that the first GTC where Janssen comes out and he's flanked by all the humanoids that you know that they're all these bipedal-legged robots.
• 45:00
  SPEAKER_00: And then a subsequent one, I think it happened uh at VivaTech in France over the summer, there were some wheel bases.
• 45:07
  SPEAKER_00: And I feel like it's interesting to watch that this kind of evolution in how people are thinking about these robots that um that that even like what we consider humanoids are going to come in all sorts of shapes and sizes and form factors.
• 45:23
  SPEAKER_01: Yeah, I mean it's it's it's pretty logical to expect that there will be multiple embodiments.
• 45:29
  SPEAKER_01: And in fact, you know, many, many, many people ask me and NVIDIA and for that matter, you know, other experts, is humanoid the right form factor exactly to look like a human?
• 45:37
  SPEAKER_01: And the answer is yes and no.
• 45:39
  SPEAKER_01: There are certain jobs that it makes sense, and there are certain other spaces where mobile manipulators make sense for cost reasons or uh, you know, for other reasons, right?
• 45:48
  SPEAKER_01: Ability.
• 45:49
  SPEAKER_01: But the beautiful thing about what's happening in the industry, irrespective of the embodiment, you know, we think it's got to be many, many different embodiments.
• 45:57
  SPEAKER_01: But the beautiful thing about what's happening with humanoid, even though some might feel like we are pushing it way too much and it's hype, because of because of the opportunity, the best in the world are gravitating towards it to solve the number one problem that we need to solve, which is to create a general purpose brain.
• 46:14
  SPEAKER_01: And putting testing the general purpose brain in a humanoid form factor makes logical sense because in the last 200 years we built the whole world around to be around human form factors.
• 46:27
  SPEAKER_01: It's brownfield.
• 46:28
  SPEAKER_01: So it's it's logical that we can do a lot of testing in that form factor.
• 46:33
  SPEAKER_01: It doesn't mean that our deployments, final embodiments, 100% need to be that.
• 46:37
  SPEAKER_01: They can be many different embodiments.
• 46:39
  SPEAKER_01: But the beautiful thing is if you're able to create the general purpose brain, then we can adapt the general purpose brain to different form factors and different use cases along the way.
• 46:49
  SPEAKER_01: But job number one right now is we need to create a chat GPT equivalent for a physical robot today.
• 46:55
  SPEAKER_01: And that does not exist, and that's where the whole world is, you know, is at the problem right now to create.
• 47:02
  SPEAKER_00: That's a really interesting point.
• 47:03
  SPEAKER_00: And I've never heard it framed like exactly that way.
• 47:06
  SPEAKER_00: And this kind of gets back to what we were talking about at the beginning, as far as the way in which all this money and scaling and manufacturing going into smartphones has informed robotics.
• 47:14
  SPEAKER_00: It's similar into this, that obviously a lot of money and excitement's being pumped in the humanoid form factor.
• 47:19
  SPEAKER_00: So like let's use all of that and let's harness all that and all the innovations that are happening around humanoids can obviously deploy into other any kind of other robotic form factor.
• 47:30
  SPEAKER_01: Absolutely.
• 47:31
  SPEAKER_01: And that's what uh you know is so exciting for me.
• 47:33
  SPEAKER_01: And you know, when we look at our platform, uh you know, I I think it's great that there's so much funding and so many things that are coming in.
• 47:40
  SPEAKER_01: So the moment the best in the world are working on any problem, it's surely going to be solved better than that than not.
• 47:45
  SPEAKER_00: Yeah.
• 47:46
  SPEAKER_00: Uh and and we're at about time, so we can we can close on this, but also case kind of a just a big sweeping question.
• 47:52
  SPEAKER_00: But we've been talking around this quite a bit, this idea of having some orthodoxies challenged, you know, some things that that that we've taken for granted that as new technologies come up, we've had to rethink.
• 48:04
  SPEAKER_00: For example, you know, the the order in uh in which you know you developed the brains that you kind of started from third and have gone back to the first.
• 48:13
  SPEAKER_00: Um when you really think about it, since you started this journey, you know, 10 to 12 years ago, what do you feel like is the sort of biggest orthodoxy that you took from granted for granted that turned out to be just completely wrong?
• 48:30
  SPEAKER_01: That's a deep question.
• 48:31
  SPEAKER_01: Think about there's so many, so, so, so many learnings that we had.
• 48:34
  SPEAKER_01: But but I think the the the biggest thing uh if there's a single big takeaway I have from the last 10 years and what I expect to happen in the next 10 years, I think the role of simulation is underestimated even today.
• 48:49
  SPEAKER_01: And it's not surprising it's underestimated today because I mean if I were to really polarize it, it doesn't work, right?
• 48:55
  SPEAKER_01: I mean, that's kind of why simulation has not been used because the technology wasn't good enough.
• 48:59
  SPEAKER_01: The SIM to real gap wasn't there.
• 49:01
  SPEAKER_01: But as we are solving that Sym to real gap, because we're working with the industry leaders like Google, Deepine, and Disney to create new physics engine for robotics, not physics engine that was used for gaming or something like that, right?
• 49:13
  SPEAKER_01: And as we are working with the industry with you know the ROS consortium to really make ROS for real production, and as we are embarking on these different journeys, in the end, for every physical instantiation, you want to test as much as possible in simulation.
• 49:30
  SPEAKER_01: It just makes sense.
• 49:31
  SPEAKER_01: Like just look at chip design.
• 49:32
  SPEAKER_01: We never ever take a chip to tape out until we 100% know that it's simulated and emulated because the cost of failing in the real world is million, trillion times higher.
• 49:43
  SPEAKER_01: Yeah.
• 49:44
  SPEAKER_01: So the role of simulation.
• 49:45
  SPEAKER_00: And there's virtually no cost in running simulation.
• 49:48
  SPEAKER_00: That's right.
• 49:48
  SPEAKER_01: And it's not only simulation from a single robot perspective, ultimately, when you look at industrial environments, commercial environments, it's gonna be a fleet of robots of different embodiments, including humans, all orchestrated in some way or the other.
• 50:03
  SPEAKER_01: And what's the best way to orchestrate this policy?
• 50:06
  SPEAKER_01: And what's the best AI to coordinate all of this?
• 50:09
  SPEAKER_01: How do you know if this is the best one?
• 50:10
  SPEAKER_01: You've got to run simulations, you've got to run all scenarios, all planning, right?
• 50:14
  SPEAKER_01: So to the point where directionally, 10 years out from now, in terms of the amount of infrastructure spent on, how much do you spend on training, how much do you spend on runtime, how much do you spend on simulation, I think that that one is something that you know I would be surprised if that's not going to be one of the biggest opportunities that's in front of us.
• 50:35
  SPEAKER_01: Deepu, always a pleasure.
• 50:37
  SPEAKER_01: Thank you so much for joining us.
• 50:38
  SPEAKER_01: Thanks, Brian.
• 50:39
  SPEAKER_00: I appreciate the opportunity.
• 50:42
  SPEAKER_00: Thanks again to Deepu and NVIDIA.
• 50:44
  SPEAKER_00: That conversation, of course, was tied to the release of Jets and Thor back in August.
• 50:49
  SPEAKER_00: And we've already seen some incredible work starting to roll out from that project.
• 50:54
  SPEAKER_00: If you've made it this far into the episode, please like and subscribe and check out the automated newsletter over at automated.org.
• 51:02
  SPEAKER_00: Thank you so much for tuning in.
• 51:04
  SPEAKER_00: We'll catch you next time.