00:00NVIDIA wants to make humanoid robots safer around human beings,
00:04allow them to be in closer proximity to human beings.
00:08Deepu Tala, why?
00:11Well, Ed, first of all, thank you for having me.
00:14As you can imagine, physical AI and robotics is quite simply the largest opportunity in front of humanity.
00:19We've been trying to solve this problem for over 50 years with automation.
00:24However, in the world, we need increasingly more intelligent robots.
00:28And as robots become more intelligent and work with other robots and humans beside each other,
00:34safety, of course, is extremely important because in the physical world,
00:38anything that can go wrong is going to be much, you know, the damage can be extremely large.
00:43So what we do at NVIDIA is we don't build robots ourselves or humanoid robots for that matter.
00:48We work with every company and we provide the core technology that will help companies build robots
00:54that are intelligent, safe and reliable.
00:58Deepu, my experience of interacting with humanoids in an industrial setting or in an office setting
01:04is that the default right now to this generation of hardware is for them to kind of stop short,
01:10keep a safe distance, right?
01:12As they perceive the world around them, they play it a bit safe.
01:16What is it that you're trying to accelerate?
01:18What is the kind of literal proximity that you want the humanoid robot to be able to get to the
01:24human on?
01:26Yeah.
01:26So I think there's five things that we actually need for robotics to really proliferate.
01:30The first thing is they need to be intelligent and capable.
01:34And second, we need them to be reliable.
01:37Number three, they need to be safe.
01:39Number four, of course, they need to be economical.
01:41And lastly, I think they need to be not creepy, right?
01:45So right now, if you look at where all the best in the world from research labs to startups to
01:51universities and so on and so forth,
01:54they're all working on trying to make these robots general purpose intelligent, right?
01:59We haven't reached that point yet where they become useful in the physical world.
02:02If you think about what happened in the last three, four years in the digital world with Chad, GPT, and
02:07now Anthropic, and Gemini, and everybody else,
02:10the accuracy has gone higher and higher.
02:13However, there's almost always a human in the loop.
02:16Like, for example, if you're using Cloud to summarize your email or summarize a document,
02:20you probably are going to see the final result and going to tweak it.
02:24But when it comes to physical robots in the real world, there's not going to be a human in the
02:28loop.
02:28So the accuracy requirements are 99 point, how many nines is it?
02:32Nine nines, 10 nines, 15 nines, depending on the application.
02:36So which means we need the general purpose intelligence to become accurate enough.
02:40That's kind of where we are right now.
02:42And along the way, as we are improving the accuracy, safety has to be designed in all the way from,
02:48you know,
02:48from the chip level to this hardware system level, to the operating system level, to the app level.
02:54And ultimately, it needs to be explainable so that third-party, you know, systems or third-parties
03:01can actually validate that you are actually, you know, certified, that the system is built in a safe manner.
03:08Deepu, could we think about this from a sort of localized compute perspective,
03:12inference, you know, at the edge, within the brain, so to speak, of the humanoid?
03:16A humanoid's walking toward me, and it has a very heavy object, and it senses I'm there.
03:22Like, what has been the breakthrough on the compute side that NVIDIA's done for what goes into the humanoid robot?
03:31Yeah, so we've been building a computer for, you know, robotics actually is a three-computer problem.
03:35The computer you're talking about is the third computer, which goes inside the, you know, inside the robot, the brain.
03:41But we also need a computer for training the brain and a computer for testing,
03:45which happens to be in simulation or omniverse platform.
03:47So now coming to the third computer, to your question, what needs to happen?
03:50We started working on this more than a decade ago.
03:53We call this, you know, NVIDIA Jetson.
03:55And number one, of course, it needs to have high amount of compute capability
03:59because models are going to become more larger in order to make them more intelligent to hit the accuracy mark.
04:05Of course, it needs to be real-time.
04:06It needs to be safety.
04:07It needs to be designed from the ground up, right?
04:10Energy efficiency also matters.
04:12And, of course, it also needs to be general-purpose programmable,
04:14and that's kind of what we see with our NVIDIA Jetson computer with over 2.5 million developers
04:19that, you know, are active on the platform, more than 10,000 companies that are building all sorts of robots,
04:25some humanoids, but all sorts of other embodiments as well.
04:29And so that's kind of what, you know, what's needed for robotics, general-purpose robotics to happen.
04:34Right now in the field of humanoid robotics, what is the biggest barrier to progress, software or hardware?
04:42I think there's been tremendous amount of mechatronics miracles that have been done,
04:46except for the hands that are being worked on.
04:48But I think still the number one problem remains.
04:50The brain has to be reasonably, you know, general-purpose and accurate.
04:55And we haven't hit that, you know, what happened with Chad GPT in November 2022,
04:59where you had that moment.
05:00That moment for robotics is right around the corner, but we haven't reached that yet.
05:06Let's bring this back to safety to end.
05:09You say that the HALO system that you guys talked about earlier this week
05:14brings awareness to humanoid robots.
05:19Define what you mean by awareness,
05:21but also what is the state that you want that humanoid to get to from where it is currently today?
05:27Yeah, so I think the safety is a multi-layered, you know, ecosystem.
05:31You've got to start first making the SOC or the chip needs to be functionally safe.
05:36Because NVIDIA, we've been working on autonomous vehicles for over a decade.
05:40We invested over 20,000 engineering years into making, you know,
05:43all sorts of, you know, safety techniques inside the SOC.
05:47That's the first layer.
05:48And now we are basically taking all of that investment and adding it to general-purpose robotics.
05:54And then at the system layer, the hardware layer, you have safety microcontrollers that can, you know,
06:00for redundancy reasons, you can check against, you know,
06:03if the main processor is off for whatever reason, you can do that.
06:06And then the operating system above that, you know, whether it's Linux or safety operating system or QNX,
06:13you know, a combination of all of these,
06:15they all enable you to operate them in a safe and explainable manner.
06:19And then on top of that, you have these algorithms, these AI models, if you will.
06:23You use a combination of traditional computer vision models.
06:26You use a combination of now the latest large language models and vision language models.
06:30And you can use reasoning with those models, which reasoning you can use it for explaining what's,
06:36why the model is doing, what it is doing.
06:38You know, you can kind of see all of that.
06:40And then lastly, on top of that, you bring in all these certification bodies,
06:45which are independent third party that can, you know,
06:48essentially look at the whole stack on a case-by-case basis, use case-by-use case,
06:53deployment to deployment.
06:55They can actually go through the full analysis as to how is, you know,
06:59how safe is this deployment going to be.
07:01And you can actually do all of this evaluation in simulation first,
07:06millions of times over before you bring the robot into the real world.
07:09So I think that's kind of like the breakthrough is, you know,
07:11leveraging all of the full stack, but being able to test it in simulation first.
Comments