NVIDIA Just Changed Robotics Forever With GR00T N1 – See It in Action! (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) Let's go talk about robotics, shall (00:00:02) [Music] (00:00:05) we? Let's talk about (00:00:08) robots. Well, the time has come. The (00:00:11) time have has come for robots. Uh robots (00:00:15) have the benefit the benefit of being (00:00:17) able to interact with the physical world (00:00:18) and do things that otherwise digital (00:00:20) information cannot. Uh we know very (00:00:23) clearly that the world is has severe (00:00:26) shortage of of human labors, human (00:00:28) workers by the end of this decade. The (00:00:31) world is going to be at least 50 million (00:00:34) workers short. We'd be more than (00:00:36) delighted to pay them each $50,000 to (00:00:38) come to work. We're probably going to (00:00:40) have to pay robots $50,000 a year to (00:00:42) come to work. And so this is going to be (00:00:44) a very very large industry. There are (00:00:46) all kinds of robotic systems. Your (00:00:48) infrastructure will be robotic. billions (00:00:50) of cameras and warehouses and factories, (00:00:53) 10, 20 million factories around the (00:00:54) world. Every car is already a robot as I (00:00:57) mentioned earlier. And then now we're (00:00:58) building general robots. Let me show you (00:01:00) how we're doing that. (00:01:04) [Music] (00:01:11) [Music] (00:01:27) Hey (00:01:28) [Music] (00:01:36) hey (00:01:38) a hey a (00:01:44) [Music] (00:01:46) I'm (00:01:55) [Music] (00:01:59) a boy. (00:02:03) [Music] (00:02:24) [Music] (00:02:28) Why? Heat. (00:02:32) [Music] (00:02:40) [Music] (00:02:47) Heat. Hey, (00:02:50) [Music] (00:02:55) Heat. (00:02:56) [Music] (00:03:06) [Music] (00:03:13) [Music] (00:03:18) Heat. Heat. (00:03:33) Heat. (00:03:34) [Music] (00:03:42) Heat. (00:03:43) [Music] (00:03:59) [Music] (00:04:03) Hey, hey, (00:04:07) hey. Heat. Heat. (00:04:14) [Music] (00:04:28) N. Heat. Heat. (00:04:30) [Music] (00:04:54) [Music] (00:05:02) [Music] (00:05:10) [Music] (00:05:18) I'm physical AI and robotics. (00:05:22) are moving so fast. Everybody pay (00:05:25) attention to this space. This could very (00:05:27) well likely be the largest industry of (00:05:29) all. At its core, we have the same (00:05:33) challenges. As I mentioned before, there (00:05:36) are three that we focus on. They are (00:05:39) rather systematic. (00:05:41) One, how do you solve the data (00:05:44) problem? How where do you create the (00:05:48) data necessary to train the AI? Two, (00:05:51) what's the model architecture? And then (00:05:53) three, what's the scaling loss? How can (00:05:57) we scale either the data, the compute or (00:06:01) both so that we can make AIs smarter and (00:06:05) smarter and smarter? How do we scale? (00:06:07) And those two those fundamental problems (00:06:10) exist in robotics as well. In (00:06:12) robotics, we created a system called (00:06:14) Omniverse. It's our operating system for (00:06:17) physical AI. You've heard me talk about (00:06:18) Omniverse for a long time. We added two (00:06:22) technologies to it. Today I'm going to (00:06:24) show you two things. One of them is so (00:06:28) that we could scale AI with generative (00:06:31) capabilities. A generative model that (00:06:34) understand the physical world. We call (00:06:36) it cosmos. Using (00:06:39) Omniverse to condition Cosmos and using (00:06:43) Cosmos to generate an infinite number of (00:06:47) environments allows us to create data (00:06:51) that is grounded grounded controlled by (00:06:55) us and yet be systematically infinite at (00:06:59) the same time. Okay. So you see (00:07:02) Omniverse, we use Candy Colors to give (00:07:04) you an (00:07:05) example of us controlling the robot in (00:07:09) the scenario perfectly and yet O Cosmos (00:07:13) can create all these virtual (00:07:14) environments. The second thing just as (00:07:17) we were talking about earlier, one of (00:07:20) the incredible scaling capabilities of (00:07:22) language models today is reinforcement (00:07:24) learning verifiable rewards. The (00:07:28) question is what's the verifiable (00:07:29) rewards in (00:07:31) robotics? And as we know very well, it's (00:07:34) the laws of physics, verifiable physics (00:07:38) rewards. And so we need an incredible (00:07:41) physics engine. Well, most physics (00:07:44) engines have been designed for a variety (00:07:45) of reasons. It could be designed because (00:07:47) if we want to use it for large (00:07:49) machineries or uh maybe we design it for (00:07:51) uh virtual worlds, video games and such. (00:07:54) But we need a physics engine that is (00:07:57) designed for very fine (00:08:00) grain rigid and soft bodies designed for (00:08:05) being able to train tactile feedback and (00:08:09) fine motor skills and actuator controls. (00:08:13) We needed to be GPU accelerated so that (00:08:16) we these virtual worlds could live in (00:08:18) super linear time, super real time and (00:08:22) train these AI models incredibly fast. (00:08:25) And we needed to be integrated (00:08:27) harmoniously into a framework that is (00:08:31) used by roboticists all over the world, (00:08:33) Mujoko. And so today we're announcing (00:08:36) something really, really special. It is (00:08:39) a partnership of three (00:08:41) companies, Deep (00:08:43) Mind, Disney Research, and Nvidia. And (00:08:47) we call it (00:08:52) Newton. Let's Let's take a look at (00:08:55) Newton. (00:08:59) [Music] (00:09:04) [Music] (00:09:16) [Applause] (00:09:18) [Music] (00:09:21) [Applause] (00:09:24) Thank you. All right, let's start that (00:09:26) over, shall (00:09:28) we? Let's not ruin it for them. Hang on (00:09:31) a second. Somebody talk to me. I need (00:09:33) feedback. What (00:09:36) happened? Who? I just need a human to (00:09:38) talk (00:09:42) to. Come on. That's a good (00:09:45) joke. Give me a human to talk to. (00:09:48) Janine, I know it's not your fault, but (00:09:50) talk to me. We got We just got a two (00:09:52) minutes left. I'm right here. They're (00:09:53) re-rackcking it. They're re-rackcking (00:09:55) it. I don't even know what that means. (00:09:57) Okay. (00:09:59) [Music] (00:10:17) [Applause] (00:10:23) [Music] (00:10:29) What did you do? (00:10:32) [Music] (00:10:42) Tell me that wasn't (00:10:45) amazing. Hey, (00:10:47) Blue. How are you (00:10:49) doing? How do you like How do you like (00:10:52) your new physics engine? You like it, (00:10:54) huh? Yeah, I bet. I know. Tactical (00:10:58) feedback. (00:11:00) Rigid body, soft body (00:11:03) simulation, super real (00:11:05) time. Can you imagine just now what you (00:11:07) were looking at is comp complete real (00:11:09) time (00:11:10) simulation? This is how we're going to (00:11:12) train robots in the (00:11:13) future. Uh just so you know, Blue has uh (00:11:17) two computers, two Nvidia computers (00:11:22) inside. Look how smart you are. (00:11:26) Yes, you're smart. Okay. (00:11:30) All right. Hey, Blue. Listen. How about (00:11:33) let's take them home. Let's finish this (00:11:35) keynote. It's (00:11:37) lunchtime. Are you ready? Let's finish (00:11:40) it up. We have another announcement to (00:11:43) You're good. You're good. Just stand (00:11:44) right here. Stand right here. Stand (00:11:47) right (00:11:49) here. All right. Good. Right (00:11:52) there. That's good. All right. Stand. (00:11:57) [Applause] (00:12:06) Okay, we have another amazing (00:12:09) news. I told you the progress of our (00:12:12) robotics has been making enormous (00:12:14) progress and today we're announcing that (00:12:18) Groot (00:12:19) N1 is open sourced. (00:12:22) [Applause] (00:12:31) I want to thank all of you to come (00:12:35) to let's wrap up. I want to thank all of (00:12:38) you for coming to GTC. We talked about (00:12:39) several things. One, Blackwell is in (00:12:42) full production and the ramp is (00:12:46) incredible. Customer demand is (00:12:48) incredible and for good reason because (00:12:50) there's an inflection point in AI. The (00:12:52) amount of computation we have to do in (00:12:54) AI is so much greater as a result of (00:12:57) reasoning AI and the training of (00:12:59) reasoning AI systems and agent agentic (00:13:02) systems. Second, Blackwell MVLink 72 (00:13:05) with Dynamo is 40 times the performance (00:13:10) AI factory performance of Hopper and (00:13:12) inference is going to be one of the most (00:13:14) important workloads in the next decade (00:13:16) as we scale out AI. Third, we have an (00:13:20) annual annual rhythm of road maps that (00:13:23) has been laid out for you so that you (00:13:26) could plan your AI infrastructure. And (00:13:28) then we have two we have three AI (00:13:30) infrastructures we're building. AI (00:13:31) infrastructure for the cloud, AI (00:13:33) infrastructure for enterprise, and AI (00:13:36) infrastructure for robots. (00:13:43) [Music]