NVIDIA exec on future of robots
Interview with NVIDA's vice president of omniverse and simulation technology.
Category
🗞
NewsTranscript
00:00Is anybody else coming or are we all done?
00:03No, no, you can start.
00:05Let's go.
00:06So you said the next coming is robots.
00:10But you didn't understand it.
00:13You said, when are those robots?
00:16It's happening now.
00:17I know it's happening now, but the future that you are driving is with human-like robots.
00:29When is that going to happen now?
00:30That's a really, really good question.
00:32Look, I didn't have enough time up there to really go into details about this, but the basic thesis is this.
00:41With humanoid robots, the mechanical, physical part of it, we probably could have solved that problem 10, 20 years ago.
00:53With these phones that we have here, the screen problem, building those screens, we didn't have the technology for that a long time ago.
01:00At the moment, there was a use for it.
01:02We figured out how to manufacture them.
01:04It got better and better and cheaper and cheaper.
01:06And now it's like touchscreens that are flat and everywhere are all over the place and whatnot.
01:12The same thing would happen with humanoid robots.
01:15We can figure out the technology, and if there's enough of a market for these things, we'd eventually get to the point where the physical part of the robot, largely the cost of that, will be just the materials that go into the robot.
01:30You can think about an extreme example, and this is a car.
01:33You can buy a very nice car for $30,000, and the margins on a car are notoriously low.
01:40Mostly the cost is the actual materials.
01:44If you look at the relative weight of a humanoid robot to the weight of a car, it's a fraction of it.
01:51We should be able to, if we had enough robots that we're building, and now we have the physical technology to build these physical robots, get robots down to a few thousand dollars.
02:01Now, the question is why did it happen to come to Seattle?
02:04The reason for that is if we had built such a robot and we had them everywhere, they would be useless to us because we couldn't actually have them do useful things.
02:12We don't have a brain for the robot.
02:14That was the missing ingredient for creating a useful, general-purpose robot that has human-like proportions that could operate in spaces built for humans.
02:26But that changed.
02:28We finally have the technology to build a reasonable, general-purpose brain to operate a robot, and the world realized that essentially we have that when ChatGPT came out.
02:41That same technology, the LLM technology that can understand, generally, language, is exactly the same thing you need to generally understand the world and act upon.
02:51So we know we have the tech to do it. We just have to go do that.
02:54The science part of this is largely solved.
02:57Now we're going into engineering, figuring out how to transform the basic technologies we have into a product.
03:04What you're seeing is many startups and even other larger companies all building humanoid robots right now.
03:10At our conference, the GPU Technology Conference, GTC, back in April, Jensen Wong, our CEO, at the end, he ended it with humanoid robots.
03:21We had a whole wall, I think a dozen or so humanoid robots behind him, and he's continued to talk about that.
03:27That's no accident.
03:29We believe now the time has finally happened.
03:31We've been preparing for this.
03:33We've been building all the core technologies to build the AI, the brains, and we've been building the simulation technologies, which is my job with Omniverse,
03:41so that we can train these robot brains and test them to bring them out into the real world.
03:48Now we finally have enough brain capability on site that we're going to start seeing useful humanoid robots.
03:56Once those start getting out there and we start seeing the uses from them, it'll be like the phones.
04:02Creating them is going to get cheaper and cheaper, and we're going to see them everywhere else.
04:05We're at that inflection point.
04:07Can I predict it's going to be in one year versus four years?
04:10I don't know exactly.
04:12That's not how technology and markets and stuff work.
04:14But all the fundamental things in order to get there are in place, finally.
04:20And so we're going all in on this.
04:23Who would be responsible for bigger actions and broader actions?
04:29I mean, if a robot just pushed someone in the street, who would be responsible?
04:37The one who created it?
04:40Honestly, I can't answer that specific question.
04:44First of all, I'm not a lawyer, a politician, a regulator, or whatever, to be able to answer that.
04:54The answer to that is it kind of all depends.
04:56It depends on which country you're in, which city, what the task is, what the robot's doing, who owns the robot, who operates it, and whatnot.
05:03There are some analogies you can already look at for non-intelligent robots today.
05:09Like who's responsible for when you go into a factory and some machine harms you inside that factory?
05:18The machine's not responsible.
05:20It's whoever's operating that.
05:22Actually, even if a human inside there harms you, the person who's responsible is actually an entity.
05:28Usually it's the company that owns that, and you can go sue them and whatnot.
05:32So I think we're going to have to start figuring that out.
05:34If we think of these autonomous agents that are operating out in the real world as sort of something that's kind of human-like,
05:41or a place we're going to figure out a special place for them inside our rule structure.
05:47And you're starting to see that play out already.
05:49The first robots that are going into wide use in a meaningful way are self-driving cars.
05:56And so the same questions exist there.
05:58If a self-driving car harms you, who's responsible?
06:02And so it kind of depends.
06:04It depends on the city.
06:05It depends on who's operating it, and so on and so forth.
06:09Will they replace the company?
06:12Just two more questions.
06:14Will they replace people who are working?
06:18Well, you know, I think nobody can say, of course they're not.
06:24There's going to be some things that are going to replace.
06:26That's already happening to some extent in the knowledge space.
06:31We're just surprised, like, what people didn't expect that AI was going to take knowledge work jobs away first or whatever.
06:41The truth is it's not really going to take away any jobs, in my opinion.
06:44And it's not doing it in the knowledge space either.
06:48It's more like we're creating new jobs.
06:52Today we have this ‑‑ there's actually a crisis moment right now for these kinds of jobs that we want to employ these robots for.
07:00If you go talk to any retailer, they don't have enough people to hire to go stock shelves.
07:05There just isn't.
07:07I was in Germany a few months back.
07:09We worked closely with all the big auto manufacturers.
07:12I was talking to BMW.
07:14And they have this problem.
07:16The most expert people who operate their factories who work in them are retiring.
07:23And they don't have enough people to replace them.
07:25You go talk to people who are shipping and logistics.
07:28There's not enough truck drivers in the world.
07:30And that number is shrinking.
07:32The total number of actual people are shrinking.
07:34And the young people don't want to do those things.
07:37And so we have to get those things done somehow.
07:40Well, how are we going to do that?
07:42The only real solution to that is to go create these autonomous systems that can actually do that.
07:48The other problem that we have is that essentially throughout human history, we've been ‑‑ we've created this pyramid scheme.
07:56Every generation depends on there being more people in the next generation to make productivity output to pay for them in their old age.
08:05This is the problems we're seeing everywhere in every country.
08:09And so depopulation is a serious issue everywhere we go.
08:13And elderly people need more care.
08:16They need more people to support them.
08:19So how are we going to solve that problem?
08:21So I think instead of worrying about it taking away jobs, we should be worrying about this crisis that's happening.
08:26This pyramid is inverting, which doesn't work for us as a society.
08:32All the assumptions we made about the economy is growing, productivity is growing, that ends if the populations get smaller.
08:40Less than that.
08:42So there's plenty of jobs for robots.
08:44We're going to ‑‑ there's already unfilled jobs that we need to put the robots on.
08:49We don't have to worry so much about them taking away other jobs.
08:53Sorry, I'm going to bang on about the robot side of things.
08:58Again, the thought of them being humanoid, humanoid robots actually being used, the tech companies like making humanoid robots because they're a great demo.
09:11But if you look at the application of robots like autonomous cars or in factory automation, they're not humanoids, are they?
09:20Is it a gimmick?
09:22That's what I'm saying.
09:23That's what I touched on in my talk.
09:28I first defined robotics as being extremely general.
09:31There's all kinds of robots.
09:33Spaces, buildings, cities can also be robots.
09:38Anything that perceives, decides and plans and acts is a robot.
09:43And the right robot for a given task is, well, you know, often it's going to be something very different than what a human looks like.
09:52However, for a general purpose robot, a robot that can do many different kinds of tasks for different things and do that in spaces that are built for humans, the best form factor is one that's human-like.
10:08If we wanted to stock shelves but also go and maybe help take out the trash, the same robot to do both of those things, it would be hard to take one robot specialized for this one purpose and have it go do the other thing.
10:27But we know humans can already do those things.
10:30We've designed the spaces and all the products and stuff around us.
10:34So I think we'll start from the basis of a humanoid robot and we'll deviate from that.
10:38Maybe it doesn't need to have legs.
10:40It could just go around on wheels.
10:42That's fine in some cases.
10:44If there's stairs in this space, you maybe need the legs.
10:47Maybe it can have three arms instead of two.
10:50It might be more useful.
10:52There's still going to be plenty of uses for other specialized robots.
10:55And as we start building other spaces, especially factories and warehouses and stuff, where there's no humans, greenfield ones that we're building from scratch, then we can hyperoptimize the robots for those spaces.
11:08But we're going to have spaces for humans for a long time.
11:12Most of the world, most of the factories, most of the warehouses, most of the hospitals, most of everything is designed for us.
11:19So this form factor is really important.
11:22And as the first robot that's going to take off and become ubiquitous, we're betting it's going to be humanoids because of that fact.
11:30Well, that just absolutely answers my second question.
11:34But as a third one, if you look at – you say that AI – sorry, this is being devil's advocate type question.
11:41AI is the next – physical AI is going to be the next wave.
11:46Robotics is going to be the next wave.
11:49Is that a little bit sensationalist, perhaps, when really the real heavy lifting that AI is doing is in medical applications or satellite data or agriculture or things like that?
12:01Again, those are really, really broad things.
12:04Medical applications are also robotics.
12:07There's many of them in there.
12:09It's about the physical world.
12:12So one place is robotic surgery.
12:16This is something that's taking off right now.
12:18There's surgical procedures you can't do with human hands that robots are doing currently.
12:24We're doing telesurgery and telehealth care and all of that.
12:31There's also drug creation.
12:35That's physical AI as well, understanding the world of chemicals and atoms and whatnot.
12:43All the advances.
12:44Nubar was talking about a lot of things that they're doing that's really amazing.
12:47Then once you actually design that drug, you have to actually go manufacture it and create it.
12:52And that's going to require all kinds of robotics.
12:55So almost anything you talk about, any area of interest, sure, there's stuff that is largely knowledge or information-based.
13:04And that's where computers and the existing AIs that we have already play that naturally because those things are in digital form.
13:11But if you want to start applying the AI to all of the other parts of our economy, the IT information part at maximum is about $5 trillion a year.
13:23Everything else is $100 trillion.
13:25How are we going to go take AI into that area?
13:29Well, that's robotics.
13:31And now you can say, well, robotics, you know, that just makes for nice demos or whatever.
13:38But if you could actually create a robot that could address those $100 trillion over there, that's enough demand.
13:44The market conditions are right to go create that innovation.
13:49We've been trying.
13:50People have been trying to do robotics for a long time.
13:52I was at the BMW factories.
13:54They've had robotics for 20, 30 years that helped them build their cars.
13:58But the robots that we've had have been essentially blind and not very smart.
14:04They do the same thing over and over again.
14:06It takes a team of engineers programming it to do this one thing over and over again.
14:10And any deviation from that and it fails.
14:13And it's sort of useless if you have a kind of factory or production where things are changing.
14:18You're doing different products, small numbers of things that require a lot of dexterous manipulation.
14:24But because AI gives us the ability to have these generally intelligent things that might be able to do many tasks,
14:32now we have the possibility of taking the world of compute and bringing it to the areas of our economy that have been largely untouched for years.
14:44Go into a car factory today and they're different than what Ford did when he created the first factories.
14:51But they're not that different.
14:53That's the most advanced robotics that we have in the world is inside these automotive factories
15:00because they create these high-value items for many years, the same thing over and over again.
15:04They can afford to invest.
15:06And even there, there's barely any robots.
15:08Is it because they're task-specific robots?
15:11Yes.
15:12They're blind and they're dumb.
15:17You program them to do the same thing over and over again.
15:21They're not seeing anything.
15:23And so they have to be caged off from the humans because they're dangerous.
15:27They're not going to react to anybody around them.
15:30You can't give them something new or if there's something wrong with the input coming in, it won't adapt to the situation like a human would.
15:37So there's lots of things that only humans can do there.
15:40But now we have the possibility of a brain.
15:43And if we can create a good robot brain, that unlocks all these opportunities and the application of these robots to a much, much larger market.
15:53And when you have those conditions, what we've seen is the democratization of those technologies, just like these phones.
16:04Sorry.
16:05Let me go again.
16:07If the robotics takes off, what would someone's home look like?
16:11Home?
16:12Yeah, like home and their commute.
16:14Yeah.
16:15Will it be your...
16:16It's a good question.
16:18I definitely believe that these robotics are going to get into the home and consumers' life and world and whatnot.
16:25I think it will particularly be interesting in healthcare, in hospitals, in taking care of our elderly and whatnot.
16:31But I think it's going to be uneven how it's deployed and where.
16:36Some cultures are just more amenable to it.
16:39I think in Japan they'll be just fine with it.
16:41They consider robots just culturally as a good thing.
16:47In the West, particularly the U.S., our vision of a humanoid robot is Terminator.
16:52There's cultural elements at play here.
16:55So we'll see how consumers react to this and how they adapt and whatnot, where it will come in unevenly.
17:01But I think that's actually...
17:03I don't know how this pans out, but our bet is that we're going to first apply it in industry, where we have this great demand.
17:14And these are great big industries with lots of money in the industry.
17:19There's a lot of need for it.
17:20There's really no other solution anyway.
17:22So that's where I personally am hyper-focused.
17:26I think they're going to show up in factories, in warehouses.
17:32And you can see Amazon, the most advanced company in the world as far as logistics and shipping things around.
17:41They've been investing in...
17:42They're the most advanced in robotics because of what they need to do.
17:47So we're going to see it happen there first.
17:49Just a second.
17:50No, you go first.
17:52In a market like India, how do companies create an acceptance for robotics?
17:57Because there is already unemployment.
17:59If there is one job, there are 20, 80, 30, many people trying to apply for that.
18:06So how do enterprises penetrate in that market with their robotics products?
18:11And how do companies buying those robots create sort of an acceptance?
18:15Yeah, I don't know if I have all the answers to every problem in every region and whatnot.
18:19But I think the way I would do it, if I were strategizing for this for India, is instead of trying to go figure out how to put robots into jobs that humans there are already doing,
18:30go find all of the jobs that humans are not doing today that only robots can do.
18:36You want to put robots in the kinds of jobs where it would be too dangerous for a human.
18:42We're already seeing it, say, in mining.
18:45There are very dangerous situations.
18:47No human should have to take on those kinds of risks.
18:50Let's put a robot in that situation.
18:52There are some cases where we need skills that are superhuman.
18:58And so we don't build those things today.
19:00And there are going to be robots that can do things.
19:05I mentioned surgery earlier.
19:07There's kinds of surgery where a human hand is just not good enough.
19:11It's not precise enough, and it's not reliable enough to be able to do something.
19:17It's too dangerous to do that.
19:18Let the robot do that.
19:19So it won't be a robot taking a job away from a human.
19:22It's a new job that we're creating that didn't exist before.
19:28Sorry.
19:29And just another one.
19:30If someone's got a child in primary elementary education and thinking of career paths for them,
19:36what would you recommend to steer them towards as a future proof?
19:41For their child?
19:42Why, you have a child?
19:43No, no, no.
19:44Just for the future.
19:46Well, I do.
19:47And my 18-year-old son just started college.
19:51And I had to have this discussion with him a few years ago.
19:55You know, in the past, people would ask me, what should my kid do?
20:00And it was for most of my life, the answer has been, well, whatever they do,
20:06they should also learn how to code.
20:08Because if you can combine coding with whatever is your domain expertise,
20:12then you'll be really valuable.
20:14You bridge these worlds together.
20:16That may not be so true now.
20:21Because we now have programs, we have software that can write software.
20:28It can actually interpret your intentions from natural language and create software.
20:33They're not great at it yet, but they're actually better than a lot of real coders,
20:37human coders already.
20:39And we have every reason to believe that they're going to get much better very faster.
20:43It's happening as we speak.
20:46And so then somebody asked me yesterday, I did a thing at AUA,
20:50what will happen to all these coders?
20:55Do we not need coders anymore?
20:58Because this used to be a thing.
21:00Actually, I think the way to look at it is, essentially,
21:03because we have this new technology that can go from natural language into software,
21:10it's almost like we've added a new layer of our compiler stack.
21:15The history of computing, we started with programming them just by flipping switches physically,
21:22and then we would type assembly, or a hexadecimal, and then assembly,
21:27and then we had languages like BASIC and stuff that were more English-like, and so on and so forth.
21:31Now we've taken it all the way to the end, to natural language.
21:34You can program a computer just by talking in your language, whatever it happens to be.
21:39And so in some sense, it's not that there's no job for coders, it's that everybody is a coder.
21:45Every person on earth becomes a coder.
21:47You can write your own software.
21:49And so what are the implications of that?
21:52Well, before we had a small number of people who could actually write good code,
21:56and it takes a lot of work to go do that.
21:59So they go build these applications.
22:03They try to make them as general as possible so as many users as possible can take advantage of it.
22:09But that tool that they build is not necessarily perfect for each one of those people.
22:15It's even worse if your particular domain is something that's kind of a niche,
22:21a very important one, but a very small number of people really understand it.
22:25It might be the language of molecules, knowing how to put molecules together,
22:31or genes, or I don't know, material science, something in particular.
22:37Now you have to find some computer scientist somewhere who you can explain enough of your problem to
22:45for that person to turn it into a tool that you could use in your field.
22:50And most of the time we just end up not using computing there because we couldn't.
22:55But now that person, that expert in that domain, can become a programmer
23:00and build a tool that's perfect for them for that moment.
23:03And so what I would recommend to any young person is,
23:09I mean if you're really interested in building computers and computer science and stuff,
23:14then this is still a great thing. Go do that.
23:17But if you have an interest in things related to the physical world,
23:21you are interested in how materials work, if you're interested in how cells work,
23:28all the amazing stuff Nubar was talking about, go study that,
23:32and don't worry about the coding part of that.
23:35The computer scientists are going to take care of that for you.
23:37They're going to give you a compiler stack that makes it so that you can write your own programs.
23:42It's the same way you would by telling a really intelligent computer scientist what to do,
23:48but this one will probably listen to you better than that programmer.
23:54I've got another couple of questions.
23:57Are we being fair? Are there any other questions?
24:00I'll take the next one.
24:02Why don't we let him go first, just because I don't know how much time we have left.
24:06I'm going to have to leave, and I want to make sure we're fair.
24:09Well, thank you.
24:11My question was about the infrastructure relating to the application of robotics in the future,
24:16like what 4.0 or industrial revolution, et cetera.
24:21To what extent do you think that the current Internet infrastructure is capable of handling
24:26all the new data transfers that you'd have to expect if you have a more widespread adoption of robotics?
24:31You know, funny enough, I think this new technology has kind of the inverse effect or needs from information.
24:42It's a go.
24:44All right, so I'll finish answering this question.
24:46If you look at classical information theory, you can't reduce the amount of data or information
24:56without really losing something.
25:02And so what we do, say, with video, which was one of the things that drove, like,
25:06a high bandwidth adoption on the Internet, we have an encoder and a decoder.
25:12The encoder takes all of the information, the high resolution and all this information,
25:17encodes it into something smaller, and then the decoder then expands it back out, right?
25:22There's kind of a limit we can get to with what the quality is for that size of that thing.
25:28But if you were to call me on the phone and I described to you the room I'm seeing around here,
25:37that there's a red carpet and there's a wall over here behind me, a row of chairs and whatnot,
25:43in your mind, you start creating an image.
25:46I'm encoding all of this information around me, and because of all the priors you have,
25:52you already know what a chair is, you know what red is, you know what a black wall is,
25:57all of these things, with very little data, because of the priors that you have,
26:02you can expand that into something that's more interesting.
26:05That's essentially what Gen AI is doing.
26:07You're putting a prompt that describes a small thing, and you can get an image out of that,
26:13that has all this information in there.
26:15And it's largely right, kind of matching it.
26:19So really good Gen AI that has good priors and good knowledge of the working world around us
26:27will actually allow us to transmit more information with less bits.
26:33So the stress that it will put on the Internet, in some ways, for at least these kinds of things,
26:38it will have the opposite effect.
26:40It's going to reduce that.
26:43The bigger problem is not bandwidth, it's latency.
26:47Time to the server and back, and now that doesn't matter for all kinds of AI,
26:52but for the things that are interactive, when you're having a discussion.
26:55I mean, that's a problem even just between humans.
26:57If we're on the other side of the Earth, you have one second of latency,
27:00it's very hard to communicate this way.
27:03And so that part of the Internet, I think, we need to deal with more.
27:07Also, like, the on-prem computing, right?
27:10You're saying, like, you have smart cards?
27:14Yes.
27:15And so we're building those.
27:16That's one of the computers that I mentioned.
27:18We have the edge computers, the ones that go in the robots and all that.
27:21And, yes, there's lots of problems to solve.
27:24But those problems I think we can solve.
27:26We know how to build computers.