# Building Europe’s Autonomous Future With Open Models and Physical AI

## Abstract

Key takeaways Discover how Europe&#x27;s leaders are building for the AI era with NVIDIA’s Physical AI platform, including open frontier models, libraries and frameworks. Learn how simulation and AI agents are accelerating their development and deployment of safe physical AI and autonomous systems at scale. Explore real world examples of how they are reshaping their industries and reimagining industrial engineering, mobility, manufacturing, and cities.

## AI Summary

- The presentation highlighted how advanced AI and simulation technologies are being integrated into European industrial processes to create digital twins and super agents, enabling the design and optimization of physical products and systems.
- It covered the introduction of AI physics and learned simulators, which can generate synthetic data to train AI models, significantly reducing the need for real-world data and manual expertise.
- The presentation showcased partnerships with companies like Dassault Systèmes, Schneider Electric, and Audi, demonstrating how these technologies are used to enhance factory efficiency, product design, and autonomous vehicle development.
- It emphasized the importance of safety in autonomous vehicle (AV) deployment, with the introduction of the NVIDIA HaloSafe platform, which ensures continuous safety through frequent software updates and extensive simulation testing.
- The presentation also discussed the democratization of robot brain development through the use of agents and tools like Astra and NVIDIA Isaac, making it easier for developers to train and deploy robotic systems.
- The speaker concluded by inviting attendees to explore additional sessions and resources to learn more about applying these technologies to their specific industrial challenges.

## Transcript

**[00:00:05 – 00:00:11]** Welcome to our second session of Physical AI in the GTC;

**[00:00:12 – 00:00:13]** hope you're enjoying your time.

**[00:00:13 – 00:00:18]** For those of you who've been to the Mercedes session and seen how you

**[00:00:18 – 00:00:22]** take the edge of technologies and do a responsible job for bringing

**[00:00:22 – 00:00:28]** it into production, now we're going to go and I'm going to invite

**[00:00:28 – 00:00:35]** Rev Lebaredian, and he's going to talk how The leaders of Europe

**[00:00:35 – 00:00:45]** are taking agentic AI, simulation and twin models and using it to

**[00:00:45 – 00:00:50]** bring physical AI into production.

**[00:00:50 – 00:00:53]** Please welcome Rev Lebaredian.

**[00:00:59 – 00:01:00]** Thank you, everyone, for coming.

**[00:01:00 – 00:01:03]** It seems like there's a lot of interest in this session.

**[00:01:03 – 00:01:08]** I think that's probably not a fluke.

**[00:01:08 – 00:01:13]** This is Europe, where there's an extraordinary amount of engineering

**[00:01:13 – 00:01:18]** knowledge embedded inside all of the countries of Europe.

**[00:01:18 – 00:01:20]** This is where the Industrial Revolution started.

**[00:01:20 – 00:01:24]** And AI is now giving us the opportunity to scale this

**[00:01:24 – 00:01:31]** knowledge to a place that was unimaginable before.

**[00:01:31 – 00:01:34]** We can now build things that we have yet to imagine.

**[00:01:35 – 00:01:40]** But this is built on top of an industrial foundation that has

**[00:01:40 – 00:01:43]** has been here for over 100 years.

**[00:01:43 – 00:01:46]** The Industrial Revolution started in Europe.

**[00:01:46 – 00:01:52]** Europe is where electrification started.

**[00:01:52 – 00:01:55]** But we are now in the midst of a new Industrial Revolution.

**[00:01:55 – 00:01:59]** This new Industrial Revolution takes as its input the output

**[00:01:59 – 00:02:03]** of that first one, electricity.

**[00:02:03 – 00:02:10]** Europe is in an amazing position, given the install base of

**[00:02:11 – 00:02:14]** all of the vehicles; we have over 400 million vehicles,

**[00:02:14 – 00:02:18]** we have hundreds of thousands of industrial robots already installed

**[00:02:18 – 00:02:22]** here, we have countless factories, and we have hundreds of billions

**[00:02:22 – 00:02:26]** of dollars being invested in R&D.

**[00:02:26 – 00:02:30]** Europe has fallen behind in recent decades in the software

**[00:02:30 – 00:02:34]** evolution, but it turns out in this new revolution that's

**[00:02:34 – 00:02:36]** no longer a problem.

**[00:02:36 – 00:02:41]** The software is something that everybody can do.

**[00:02:41 – 00:02:46]** What really does matter is the knowledge about the domain

**[00:02:46 – 00:02:51]** that you are applying to, and the biggest value domain that we can

**[00:02:51 – 00:02:54]** apply AI to is the physical world.

**[00:02:54 – 00:03:01]** So this is where Europe gets to enjoy a reset and potentially

**[00:03:01 – 00:03:06]** leapfrog the rest of the world, as long as you choose to do so.

**[00:03:09 – 00:03:15]** Now, in order to build physical AI, you need many things,

**[00:03:15 – 00:03:20]** but one of the key aspects of building any AI is feeding

**[00:03:20 – 00:03:26]** data into the computer that's going to create the intelligence.

**[00:03:26 – 00:03:30]** Data about the digital world has been freely available to

**[00:03:30 – 00:03:38]** us because information by its very nature is digital or digitizable,

**[00:03:38 – 00:03:42]** and we had lots and lots of that laying around on the internet.

**[00:03:42 – 00:03:45]** But in the physical world, we're not so lucky.

**[00:03:45 – 00:03:49]** We have to actually go collect this information somehow.

**[00:03:49 – 00:03:55]** It's too time-consuming, too dangerous, too expensive,

**[00:03:55 – 00:03:57]** and in many cases just impossible to get accurate

**[00:03:57 – 00:03:59]** information from the world.

**[00:03:59 – 00:04:02]** The only solution to this is to use simulation.

**[00:04:02 – 00:04:06]** If we can take the rules of the physical world, put them

**[00:04:06 – 00:04:11]** into a computer, that computer with these rules of the physical

**[00:04:11 – 00:04:14]** world become a generator of data about the physical world

**[00:04:14 – 00:04:16]** we can use to train it.

**[00:04:16 – 00:04:22]** But we've had this big problem over the past few decades.

**[00:04:22 – 00:04:26]** We've understood that representing the physical world in the computer

**[00:04:26 – 00:04:29]** as a digital twin is a superpower.

**[00:04:30 – 00:04:31]** Essentially, it's like a time machine.

**[00:04:31 – 00:04:36]** If you can reconstruct the world and predict the future

**[00:04:36 – 00:04:39]** with it, you are essentially creating a time machine that lets

**[00:04:39 – 00:04:41]** you predict many possible futures.

**[00:04:41 – 00:04:46]** But putting together these simulators and feeding these

**[00:04:46 – 00:04:52]** simulators with the appropriate information, the initial conditions

**[00:04:52 – 00:04:59]** to the worlds that are necessary to drive the simulations, there's only

**[00:04:59 – 00:05:02]** been a handful of humans around.

**[00:05:02 – 00:05:04]** We have a scarcity of humans who know how to do that, and

**[00:05:05 – 00:05:07]** it's very time consuming.

**[00:05:07 – 00:05:13]** Fortunately, AI in recent months has crossed the threshold,

**[00:05:13 – 00:05:16]** and now this problem is solved.

**[00:05:16 – 00:05:21]** What we were lacking before was the ability to stitch

**[00:05:21 – 00:05:25]** together all of the technologies we already had for simulating

**[00:05:25 – 00:05:29]** many aspects of the world and physics into combined simulators

**[00:05:29 – 00:05:32]** that are specific for each purpose.

**[00:05:32 – 00:05:36]** Six months ago, with the advances in coding agents, that problem

**[00:05:36 – 00:05:37]** was effectively solved.

**[00:05:37 – 00:05:38]** We can now do that.

**[00:05:38 – 00:05:42]** And in the past month, the new frontier models that have

**[00:05:42 – 00:05:45]** come about have solved the second problem, which is,

**[00:05:45 – 00:05:50]** how do you construct the virtual worlds that match the physical

**[00:05:50 – 00:05:53]** world accurately enough so that we can simulate them?

**[00:05:53 – 00:06:00]** This required very, very specific expertise in design, in using

**[00:06:00 – 00:06:05]** 3D tools, spatial understanding, understanding of physics;

**[00:06:05 – 00:06:08]** very few people can do it, and even those that could,

**[00:06:08 – 00:06:09]** it's very time-consuming.

**[00:06:09 – 00:06:11]** And now these new models can actually do that.

**[00:06:11 – 00:06:14]** We have agents that can take the real world and represent

**[00:06:14 – 00:06:16]** it in this form for us.

**[00:06:16 – 00:06:18]** So we are in a place now where we have an abundance

**[00:06:18 – 00:06:20]** of these two things.

**[00:06:20 – 00:06:24]** This problem is solved, from world generation to feeding

**[00:06:24 – 00:06:31]** those worlds into the simulator, learning from those simulations,

**[00:06:31 – 00:06:37]** evaluating those simulations; all of that could be automated with AI.

**[00:06:38 – 00:06:45]** To build physical AI, you need a very complex and deep stack.

**[00:06:45 – 00:06:49]** NVIDIA is essential in this domain.

**[00:06:49 – 00:06:53]** We have been working towards this for decades.

**[00:06:53 – 00:06:58]** At the very base of this stack is the computing.

**[00:06:58 – 00:07:02]** To build physical AI, to deploy physical AI, you need three

**[00:07:02 – 00:07:03]** types of computers.

**[00:07:03 – 00:07:05]** You need the training computer.

**[00:07:06 – 00:07:07]** This is the one we are most well known for.

**[00:07:07 – 00:07:10]** This is the AI computer.

**[00:07:10 – 00:07:14]** You need a computer that can simulate a world, which is

**[00:07:15 – 00:07:17]** the data generator that feeds that training computer.

**[00:07:17 – 00:07:21]** And then you finally need the computer that AI is going to be

**[00:07:21 – 00:07:24]** deployed into the real world, the one that's in your robot or in the

**[00:07:24 – 00:07:27]** system that is acting autonomously.

**[00:07:27 – 00:07:30]** We build all three of those, and on top of that we have

**[00:07:30 – 00:07:34]** a rich set of software that we've been developing for decades.

**[00:07:35 – 00:07:43]** Software and, in more recent years, models, AI models that we provide

**[00:07:44 – 00:07:51]** freely with our computers so that You can make our computer sing.

**[00:07:51 – 00:07:56]** By themselves, these libraries, these technologies

**[00:07:56 – 00:07:58]** are not that useful.

**[00:07:58 – 00:08:02]** They become useful when they are integrated into tools,

**[00:08:02 – 00:08:07]** into software, into applications, into solutions that our developers

**[00:08:07 – 00:08:09]** in our ecosystem build.

**[00:08:09 – 00:08:15]** And here in Europe, we have many of our best industrial software

**[00:08:15 – 00:08:20]** partners that we've been working with for over 20 years to do this.

**[00:08:20 – 00:08:25]** One of the earliest applications or class of applications that

**[00:08:25 – 00:08:32]** we have accelerated in this physical space are

**[00:08:33 – 00:08:37]** All of the libraries and math and algorithms that we need to

**[00:08:37 – 00:08:42]** accelerate simulation, to simulate accurate simulation of the physical

**[00:08:42 – 00:08:45]** world so we can build and design things in the physical world.

**[00:08:45 – 00:08:51]** These are integrated into tools from our great partners.

**[00:08:51 – 00:08:56]** And in recent years, as we've introduced AI technologies,

**[00:08:56 – 00:09:03]** these principled simulators have now become a data source

**[00:09:03 – 00:09:07]** to train a new kind of physical simulation, what we call AI

**[00:09:07 – 00:09:12]** physics, or learned simulators, and we provide tools and frameworks

**[00:09:12 – 00:09:14]** such as PhysicsNeMo and Warp.

**[00:09:14 – 00:09:23]** These new types of simulators allow us to do much more simulation,

**[00:09:23 – 00:09:26]** much faster, and in some cases even more accurately when

**[00:09:26 – 00:09:33]** we combine the synthetic data that we train with real-world empirical

**[00:09:33 – 00:09:36]** data that we've collected.

**[00:09:37 – 00:09:43]** So our partners are taking all of these tools and providing

**[00:09:43 – 00:09:49]** it to their agents, integrating it into their tools and providing it

**[00:09:49 – 00:09:53]** to their agents, combining it with their knowledge, with their tools,

**[00:09:53 – 00:09:58]** in order to create very specialized agents, what we call super agents.

**[00:09:58 – 00:10:00]** Here are some examples of some of our partners doing

**[00:10:01 – 00:10:01]** some great things.

**[00:10:02 – 00:10:07]** We have Autodesk, which has been using PhysicsNeMo and Warp to

**[00:10:07 – 00:10:13]** integrate into NavPack and Alias.

**[00:10:13 – 00:10:19]** We recently released cuDSS, our direct sparse solver,

**[00:10:19 – 00:10:28]** as GA, and direct sparse solvers are critical for many types

**[00:10:28 – 00:10:31]** of physics simulation and optimization problems.

**[00:10:31 – 00:10:37]** Synopsys has integrated it into their EDA tools, making them an

**[00:10:37 – 00:10:39]** order of magnitude or more faster.

**[00:10:39 – 00:10:45]** And DSOC STEM has also used cuDSS to accelerate structural

**[00:10:45 – 00:10:53]** simulations, also giving us a 10x increase or more.

**[00:10:54 – 00:10:56]** There are countless examples of this.

**[00:10:56 – 00:11:00]** Unfortunately, I know I'm going to be offending Almost

**[00:11:00 – 00:11:03]** everyone because I can't talk about everyone on every slide.

**[00:11:03 – 00:11:05]** They won't give me enough time to do that.

**[00:11:05 – 00:11:07]** So please forgive me, but I'm going to try to mention

**[00:11:07 – 00:11:10]** everyone as we go through this.

**[00:11:10 – 00:11:12]** But let's go a little bit deeper into what Dassault

**[00:11:12 – 00:11:15]** Systèmes has been doing.

**[00:11:16 – 00:11:22]** With the advent of AI, Dassault Systèmes has been on the

**[00:11:22 – 00:11:26]** forefront of integrating these AI technologies into their suite.

**[00:11:26 – 00:11:31]** In the past, the industrial software companies were

**[00:11:31 – 00:11:34]** developing tools that were primarily for humans.

**[00:11:34 – 00:11:40]** Now these same tools are being adapted for use by agents.

**[00:11:40 – 00:11:44]** And what we are providing to the humans are agents that are

**[00:11:44 – 00:11:47]** acting as tools for those humans.

**[00:11:47 – 00:11:52]** Much more scalable, we can automate things that previously

**[00:11:52 – 00:11:57]** were extremely laborious and manual to do, and you can

**[00:11:57 – 00:12:03]** see some examples here of what Dassault Systèmes is doing with it.

**[00:12:03 – 00:12:07]** They provide virtual companions or these super agents

**[00:12:07 – 00:12:09]** for different areas.

**[00:12:09 – 00:12:14]** I think they have cool names like Leo, Aura, and Marie;

**[00:12:14 – 00:12:17]** I think is Marie Curie.

**[00:12:17 – 00:12:23]** Let's take a look at this in the video.

**[00:12:23 – 00:12:26]** Isn't there supposed to be a video?

**[00:12:31 – 00:12:32]** There was supposed to be a video.

**[00:12:33 – 00:12:35]** Oh, it's after this one.

**[00:12:37 – 00:12:40]** Oh, this is a new slide that was not in the deck you guys

**[00:12:40 – 00:12:43]** gave me; I'm sorry.

**[00:12:45 – 00:12:48]** With the Dassault Systems platform, an engineer needs

**[00:12:48 – 00:12:52]** to adapt a Volkswagen Golf door design for the Tiguan.

**[00:12:52 – 00:12:56]** Aura and Leo use NVIDIA Nematron to reason over

**[00:12:56 – 00:12:57]** engineering requirements.

**[00:12:57 – 00:13:01]** The Harness coordinates the specialized sub-agents and

**[00:13:01 – 00:13:05]** preserves context across planning, design, and simulation.

**[00:13:05 – 00:13:08]** Aura uses Inovia to plan the work around the vehicle's

**[00:13:08 – 00:13:13]** dimensions, weight limits, and other engineering requirements.

**[00:13:13 – 00:13:17]** Leo Design Agent selects a compatible golf door design

**[00:13:17 – 00:13:20]** and uses CATIA to reshape it for the Tiguan.

**[00:13:20 – 00:13:23]** Leo's simulation agent uses Simulia, accelerated by

**[00:13:23 – 00:13:27]** NVIDIA CUDAx libraries, to test the variants, revealing

**[00:13:27 – 00:13:29]** a vibration problem.

**[00:13:29 – 00:13:32]** The results return to the design agent for refinement.

**[00:13:33 – 00:13:38]** The engineer reviews the refined designs and selects the final one.

**[00:13:44 – 00:13:49]** Every industrial company is going to have two factories,

**[00:13:49 – 00:13:51]** two types of factories.

**[00:13:51 – 00:13:55]** There's the traditional factories that produce the physical products,

**[00:13:55 – 00:14:01]** but you also need an AI factory that produces the intelligence that

**[00:14:01 – 00:14:07]** underpins every aspect of what you are doing in the industrial space.

**[00:14:07 – 00:14:11]** Both of these factories need digital twins.

**[00:14:11 – 00:14:13]** You need it in the design phase, you need it in the

**[00:14:13 – 00:14:18]** construction phase, and you need it in the operations phase.

**[00:14:20 – 00:14:25]** We'll talk a little bit here about the digital twin of the AI factory.

**[00:14:25 – 00:14:28]** We have our DSX ecosystem.

**[00:14:28 – 00:14:32]** This is our reference.

**[00:14:33 – 00:14:45]** We provide the ability to simulate every aspect of these

**[00:14:45 – 00:14:49]** data centers before you build them by bringing together all of the

**[00:14:49 – 00:14:54]** amazing tools and simulators that are already in the ecosystem, many

**[00:14:54 – 00:15:02]** of which are industrial software developers right here in Europe.

**[00:15:02 – 00:15:09]** This, the ability to simulate an AI factory, isn't just a nice

**[00:15:09 – 00:15:14]** to have; it's actually the only way we're going to be able to design

**[00:15:14 – 00:15:19]** these factories so that they are truly sustainable, so that we can

**[00:15:19 – 00:15:22]** minimize the usage of resources.

**[00:15:22 – 00:15:26]** We're already reducing the usage of water by using dry

**[00:15:26 – 00:15:36]** chiller techniques by 100%, but most importantly, we need

**[00:15:36 – 00:15:42]** to maximize the output of these AI factories given a unit of

**[00:15:42 – 00:15:45]** input, the input being electricity; for every watt, we need to

**[00:15:45 – 00:15:47]** maximize the amount of compute.

**[00:15:47 – 00:15:51]** To do that, it's a massive systems problem that involves

**[00:15:51 – 00:16:00]** cooling, Airflow, liquid fluid dynamics, simulation of the

**[00:16:00 – 00:16:06]** power systems and all of it co-simulated together to ensure

**[00:16:06 – 00:16:09]** that we can maximize throughput and stand these factories

**[00:16:09 – 00:16:13]** up as quickly as possible because when you're talking about billions

**[00:16:13 – 00:16:18]** of dollars in investment, every day that it's not online, just like any

**[00:16:18 – 00:16:20]** other factory, you're losing money.

**[00:16:20 – 00:16:24]** So we have many great partners that are part of our ecosystem here.

**[00:16:24 – 00:16:29]** You can see many of them listed there.

**[00:16:30 – 00:16:34]** I'm going to go to one particular example.

**[00:16:34 – 00:16:39]** We have a great partnership with Schneider Electric, Aviva.

**[00:16:39 – 00:16:41]** Aviva is a Schneider Electric company.

**[00:16:41 – 00:16:45]** They have been employing

**[00:16:46 – 00:16:52]** The DSX technologies and simulating that to help one of our other

**[00:16:52 – 00:16:58]** great partners, EnScale, build and design their factories

**[00:16:58 – 00:17:03]** well and quickly and stand them up and ensure that they're efficient

**[00:17:03 – 00:17:07]** and working as quickly as possible.

**[00:17:08 – 00:17:13]** With the Aviva Schneider Electric System, we unify all of the

**[00:17:13 – 00:17:16]** equipment data, we can get all of the asset information,

**[00:17:16 – 00:17:20]** all of the operational data, and we can enable simulations

**[00:17:20 – 00:17:27]** of the airflow, power and liquid cooling all together in one place.

**[00:17:32 – 00:17:36]** Now we can go to the other type of factory, the normal kinds, where we

**[00:17:36 – 00:17:39]** actually produce physical products.

**[00:17:40 – 00:17:46]** Every factory today actually has elements of it, a foundation

**[00:17:46 – 00:17:49]** that allows it to become robotic.

**[00:17:49 – 00:17:56]** We have sensors and software and computing already embedded

**[00:17:56 – 00:18:01]** inside our factories for many years or even decades now.

**[00:18:01 – 00:18:07]** With AI, and particularly agentic AI, we can now leverage

**[00:18:07 – 00:18:13]** that base, that ability to sense, that ability to actuate and

**[00:18:13 – 00:18:19]** control things through software, by laying AI and agentic AI

**[00:18:19 – 00:18:23]** on top of that, to turn the whole system into essentially a robot,

**[00:18:23 – 00:18:31]** one that can adapt and optimize itself and be responsive to changes

**[00:18:31 – 00:18:34]** in the needs from the market.

**[00:18:34 – 00:18:37]** Here are some examples of some of our partners we're working

**[00:18:37 – 00:18:39]** with, partners and customers.

**[00:18:39 – 00:18:45]** We have Audi, who has almost one million parts of assembly each day.

**[00:18:45 – 00:18:49]** We've been working with Dassault Systèmes and Agile Robots

**[00:18:49 – 00:18:53]** to use digital twins and agents to test how their mobile humanoids

**[00:18:53 – 00:19:00]** can operate and be orchestrated inside these factories.

**[00:19:00 – 00:19:05]** There's another example with Kion, where Kion is at the

**[00:19:05 – 00:19:09]** center at Accor here in Europe for many, many of our logistics

**[00:19:09 – 00:19:13]** and distribution centers.

**[00:19:13 – 00:19:18]** We're completely automating the trucks, the forklifts,

**[00:19:18 – 00:19:22]** and many of the equipment to move things around.

**[00:19:22 – 00:19:28]** Simulating those is essential before you deploy a fleet

**[00:19:28 – 00:19:31]** of robots into a new facility or even an existing one.

**[00:19:31 – 00:19:33]** You have to make sure that it works.

**[00:19:33 – 00:19:37]** And so Kian has been doing great work there.

**[00:19:38 – 00:19:42]** And of course, there's our old friends at Siemens, who I just

**[00:19:42 – 00:19:47]** visited in Erlangen a few days ago.

**[00:19:47 – 00:19:52]** They have an electronics factory in Erlangen that has been

**[00:19:52 – 00:19:57]** operating for 60 years that they They are always updating

**[00:19:57 – 00:20:02]** to use the latest technologies and the latest systems, going back

**[00:20:02 – 00:20:08]** to the 60s with lean techniques and digitalization and so on.

**[00:20:08 – 00:20:12]** And now they are integrating AI technologies and AI agents

**[00:20:12 – 00:20:18]** into their operations there to help in every aspect of

**[00:20:18 – 00:20:20]** the factory production.

**[00:20:20 – 00:20:26]** There are examples I saw there of where they, they have a particular

**[00:20:26 – 00:20:32]** problem in that, unlike many, many factories that have automation,

**[00:20:32 – 00:20:36]** they have a very high mix of products, hundreds and hundreds,

**[00:20:36 – 00:20:42]** I think 700 products that they're manufacturing at any given time.

**[00:20:42 – 00:20:45]** And many of them are low volume or mid volume.

**[00:20:45 – 00:20:51]** And the demand for these products are changing very dynamically.

**[00:20:51 – 00:20:56]** So as they're introducing these products, One of the

**[00:20:56 – 00:21:00]** things that they have to do is ensure that the quality is good.

**[00:21:00 – 00:21:03]** They need inspection systems; to inspect that, you need

**[00:21:03 – 00:21:05]** to train some AIs to inspect it.

**[00:21:05 – 00:21:07]** But you have a chicken or egg problem.

**[00:21:07 – 00:21:11]** Before you create the product, you have the physical product in hand.

**[00:21:11 – 00:21:14]** How do you train the AI to understand which one's

**[00:21:14 – 00:21:16]** defective or not?

**[00:21:17 – 00:21:22]** So they are employing simulation; they're using digital twins,

**[00:21:22 – 00:21:29]** digital twin composer and Omniverse to, along with Cosmos, to generate

**[00:21:29 – 00:21:34]** the data necessary to train an AI to inspect the product before

**[00:21:34 – 00:21:38]** the product even exists, so that the day they start manufacturing

**[00:21:38 – 00:21:43]** it, they can already automate that process, which is incredible.

**[00:21:43 – 00:21:48]** The other use I saw there of AI was the integration

**[00:21:48 – 00:21:55]** of Nemotron and reasoning models to help the factory

**[00:21:55 – 00:21:59]** and the planning of the factory adapt in real time as they have new

**[00:21:59 – 00:22:01]** orders and they change priorities.

**[00:22:01 – 00:22:06]** It can explore a large space of possible Reconfigurations

**[00:22:06 – 00:22:10]** of the factory that were previously impossible.

**[00:22:10 – 00:22:13]** You had to do it manually, by hand, humans just kind

**[00:22:13 – 00:22:17]** of thinking, how are they going to serve everybody's needs with

**[00:22:17 – 00:22:19]** the limited resources they have?

**[00:22:19 – 00:22:21]** And this has reduced their engineering effort already

**[00:22:21 – 00:22:27]** by 85%, which means more products they can bring faster to market.

**[00:22:27 – 00:22:31]** The time to market is down 40%, and cost savings up to

**[00:22:31 – 00:22:38]** 20%, which, as anybody who is in this area knows, 20%

**[00:22:38 – 00:22:45]** increase or savings is incredible relative to your margins.

**[00:22:47 – 00:22:53]** Here we have eight companies, there are more, but eight

**[00:22:53 – 00:22:57]** very exemplary companies in Europe that are building robots

**[00:22:57 – 00:23:01]** and robotic systems today.

**[00:23:02 – 00:23:07]** This is actually a pretty natural thing to have here

**[00:23:07 – 00:23:10]** in Germany and in Europe.

**[00:23:10 – 00:23:17]** This is where we have been manufacturing many, many heavy

**[00:23:17 – 00:23:20]** industrial systems.

**[00:23:20 – 00:23:27]** We've been building cars; we've had this amazing industrial base.

**[00:23:28 – 00:23:32]** But we obviously are going to need to employ these types of

**[00:23:32 – 00:23:37]** robots in the future factories, so it makes sense to build them here.

**[00:23:37 – 00:23:42]** It also is similar to the types of things we've been building before.

**[00:23:42 – 00:23:48]** Building cars, building electric cars, is actually very similar

**[00:23:48 – 00:23:53]** to, in many ways, has many of the same sorts of properties

**[00:23:53 – 00:23:58]** as a humanoid robot or many of these different kinds of robots.

**[00:23:58 – 00:24:02]** But no one company can build all of the types of robots that we're

**[00:24:02 – 00:24:08]** going to need, so it's great to see all of these robotics companies.

**[00:24:09 – 00:24:15]** The space of robots and the quantity of robots that we

**[00:24:15 – 00:24:19]** are going to need in the coming years is much greater than

**[00:24:19 – 00:24:24]** the capacity we currently have, and congratulations to all of them

**[00:24:24 – 00:24:29]** for making this shift or appearing.

**[00:24:29 – 00:24:34]** But in order to build these robots and to deploy them,

**[00:24:34 – 00:24:37]** it's not sufficient to just build the body, the physical part of it.

**[00:24:38 – 00:24:41]** We have to build a brain, and the brain of the robots

**[00:24:41 – 00:24:49]** has traditionally been our biggest bottleneck, or the thing that has

**[00:24:49 – 00:24:53]** stopped us from actually bothering to try to build advanced robots.

**[00:24:53 – 00:24:57]** That's changed now with the advanced AIs that we have,

**[00:24:57 – 00:25:01]** but we still have a problem in that we don't have enough data.

**[00:25:01 – 00:25:06]** Especially domain-specific data for the tasks, for the

**[00:25:06 – 00:25:09]** types of things that we need to do.

**[00:25:10 – 00:25:13]** As I mentioned earlier, the only solution to this problem,

**[00:25:13 – 00:25:18]** to get the data that we need, is to use simulation.

**[00:25:19 – 00:25:23]** What's been emerging is that now we have many options for

**[00:25:23 – 00:25:25]** the types of simulation we can do.

**[00:25:25 – 00:25:30]** We've always had, or for basically since the start of computing,

**[00:25:30 – 00:25:32]** have had principled simulation.

**[00:25:32 – 00:25:36]** This is simulators that are essentially algorithms that

**[00:25:37 – 00:25:41]** are based off of our understanding of the equations that govern

**[00:25:41 – 00:25:42]** the laws of physics.

**[00:25:43 – 00:25:46]** We've been doing that forever, and that will continue to

**[00:25:46 – 00:25:50]** be a big part of this mix.

**[00:25:50 – 00:25:54]** But recently, we've also introduced a new kind of

**[00:25:54 – 00:25:58]** physics simulation, the AI physics.

**[00:25:58 – 00:26:01]** Another way to think about it is this is learned physics.

**[00:26:01 – 00:26:05]** We can use neural networks to learn the equations, learn

**[00:26:05 – 00:26:09]** the patterns from the physical world by feeding it information

**[00:26:09 – 00:26:11]** about the physical world.

**[00:26:11 – 00:26:15]** Interestingly, generally most of the data that we feed into

**[00:26:15 – 00:26:18]** it comes from these principled physics simulators.

**[00:26:18 – 00:26:28]** We can generate information about fluid flow or finite elements,

**[00:26:28 – 00:26:33]** et cetera, feed it into a neural network, train a neural network.

**[00:26:33 – 00:26:37]** And that AI physics, that learned physics, has properties

**[00:26:37 – 00:26:38]** that are very useful for us.

**[00:26:38 – 00:26:44]** It can run much faster, even if it's not necessarily as accurate.

**[00:26:44 – 00:26:49]** We can also blend in real empirical data, so we can learn properties

**[00:26:49 – 00:26:53]** from the real world to make it more accurate in some cases, and

**[00:26:53 – 00:26:56]** we can use it for inverse problems.

**[00:26:56 – 00:27:01]** And then we have a third class of AI and simulators, these

**[00:27:01 – 00:27:03]** world models that have been

**[00:27:04 – 00:27:08]** showing up all over the place, one of which is ours, Cosmos, and

**[00:27:08 – 00:27:13]** these models have a more holistic understanding of the world and

**[00:27:13 – 00:27:16]** can be used in many different ways.

**[00:27:16 – 00:27:19]** But one of the most interesting ways to use them is to have

**[00:27:19 – 00:27:24]** them understand the output of a simulator or what it observes in

**[00:27:24 – 00:27:27]** the real world to reason about it.

**[00:27:27 – 00:27:30]** And so you can use these world models to drive the principled

**[00:27:30 – 00:27:32]** simulators and the AI physics.

**[00:27:32 – 00:27:37]** All working together, we can finally generate the data

**[00:27:37 – 00:27:43]** necessary for us to build robotic systems that are truly robust

**[00:27:43 – 00:27:47]** and work for all the different applications that we need.

**[00:27:47 – 00:27:51]** We can combine this now with the new types of agents that we have,

**[00:27:51 – 00:27:55]** these super, super intelligent agents in these domains.

**[00:27:55 – 00:28:02]** And all of the manual human work that required very specialized

**[00:28:02 – 00:28:05]** talent is no longer our bottleneck.

**[00:28:05 – 00:28:07]** We can produce the data that we need.

**[00:28:07 – 00:28:11]** This is very, very new.

**[00:28:11 – 00:28:15]** It's basically only within the past month that we actually

**[00:28:15 – 00:28:17]** have all of the pieces together.

**[00:28:17 – 00:28:23]** And so this is our moment to take advantage of all of

**[00:28:23 – 00:28:26]** these capabilities that just arrived, and for Europe to

**[00:28:26 – 00:28:31]** take that and take all of the knowledge that is embedded

**[00:28:31 – 00:28:38]** deep inside the humans that are in industry here, in the processes

**[00:28:38 – 00:28:41]** that are in the industrial companies, in the factories,

**[00:28:41 – 00:28:46]** and combine it with these techniques to build something great

**[00:28:46 – 00:28:50]** and leap ahead of everyone else.

**[00:28:51 – 00:29:00]** All of these data tools and the simulations that we are

**[00:29:00 – 00:29:05]** using to create the data are made available by us through our suite

**[00:29:05 – 00:29:11]** of Isaac tools, which are now all

**[00:29:12 – 00:29:17]** Being redesigned or reworked so that they are agent-friendly.

**[00:29:17 – 00:29:21]** Agents can use the same tools that we used to use as humans, but use

**[00:29:21 – 00:29:27]** them in the form that they want, so that they can be most efficient.

**[00:29:27 – 00:29:33]** The process before of building a simulation for robotics purposes

**[00:29:33 – 00:29:36]** required all of these steps, each of which is laborious, from

**[00:29:36 – 00:29:42]** preparing the data and curating it, to setting up the simulations,

**[00:29:42 – 00:29:47]** to actually doing the training, and then evaluating as you're training.

**[00:29:47 – 00:29:50]** Is it learning the right thing?

**[00:29:50 – 00:29:54]** Each one of these things required humans there in the loop,

**[00:29:54 – 00:29:58]** designing the tasks, breaking down the tasks that we want

**[00:29:58 – 00:30:00]** the robot to do into the steps.

**[00:30:00 – 00:30:07]** That was a very laborious process, one that requires expertise.

**[00:30:07 – 00:30:11]** And then deploying it and observing in the real world, did

**[00:30:11 – 00:30:17]** it actually do what you intended it to do, what you had in simulation,

**[00:30:17 – 00:30:19]** to the operations overall.

**[00:30:19 – 00:30:24]** All of these things are now possible to automate

**[00:30:24 – 00:30:26]** through these agents.

**[00:30:28 – 00:30:33]** These tools, when combined with the knowledge and the tools

**[00:30:33 – 00:30:38]** that are already embedded inside the industrial software companies

**[00:30:38 – 00:30:43]** that are here, and the industrial companies, it's just remarkable.

**[00:30:43 – 00:30:46]** So, I'm going to play a little video here that will give

**[00:30:46 – 00:30:51]** you guys kind of a taste of what I've just been talking about.

**[00:30:51 – 00:30:54]** Building robots is hard.

**[00:30:54 – 00:30:58]** Simulation makes training and testing easier, but building these

**[00:30:58 – 00:31:01]** worlds takes time and expertise.

**[00:31:01 – 00:31:06]** GPT-6 Astra makes this easier than ever.

**[00:31:06 – 00:31:11]** Developers describe the assets, environment, and task with

**[00:31:11 – 00:31:14]** a simple prompt and data.

**[00:31:18 – 00:31:21]** Astra builds the training world using NVIDIA Open Libraries

**[00:31:21 – 00:31:25]** and models as tools.

**[00:31:27 – 00:31:31]** It replays real-world motion and tunes the physics to get

**[00:31:31 – 00:31:34]** the simulation ready for training.

**[00:31:37 – 00:31:41]** Next, Astra trains and refines robot skills using Isaac Lab

**[00:31:41 – 00:31:49]** and Newton and evaluates them at scale using Isaac Lab Arena.

**[00:31:50 – 00:31:54]** Every tool is running locally on NVIDIA DGX Station for

**[00:31:54 – 00:31:56]** faster iteration.

**[00:31:56 – 00:31:58]** This is just one skill.

**[00:31:58 – 00:32:02]** Developers are training robots to learn tasks of every kind.

**[00:32:02 – 00:32:09]** At home, in hospitals, stores,

**[00:32:10 – 00:32:14]** Factories and beyond.

**[00:32:14 – 00:32:19]** Then, these learned skills move from simulation to the real world.

**[00:32:20 – 00:32:25]** Introducing Agile One SN, a full-stack reference robot

**[00:32:25 – 00:32:31]** built on NVIDIA Isaac GROOT and powered by Jetson Thor

**[00:32:32 – 00:32:36]** Giving developers a starting point to train and deploy new skills.

**[00:32:36 – 00:32:40]** What once took a team of specialists can be done

**[00:32:40 – 00:32:46]** by any developer with Astra and NVIDIA Isaac tools for agents.

**[00:32:53 – 00:33:06]** So now that we are democratizing the ability to build a robot

**[00:33:06 – 00:33:11]** brain by democratizing the ability to build simulations

**[00:33:11 – 00:33:14]** and train from those simulations, essentially anybody can do

**[00:33:14 – 00:33:20]** that these days, given the right agents, given the right AI.

**[00:33:21 – 00:33:23]** And in order to actually deploy these things, you're going

**[00:33:23 – 00:33:30]** to need an actual robot to test it out on and develop on.

**[00:33:30 – 00:33:36]** And until recently, until this, we've had this problem where

**[00:33:36 – 00:33:39]** anybody who wants to do advanced robotics, particularly humanoid

**[00:33:39 – 00:33:42]** robots, you have to actually become an expert in building

**[00:33:42 – 00:33:44]** and maintaining robots yourself.

**[00:33:44 – 00:33:49]** This is not what most people want to do.

**[00:33:49 – 00:33:52]** Most people want to employ one and have it go solve a

**[00:33:52 – 00:33:59]** particular problem for them, apply it to a particular task, but

**[00:33:59 – 00:34:01]** we just didn't have this option.

**[00:34:01 – 00:34:03]** So this is a really great thing that Jensen

**[00:34:03 – 00:34:05]** announced at the keynote.

**[00:34:05 – 00:34:14]** We are working with Agile robots, building on the Agile 1S platform.

**[00:34:14 – 00:34:20]** We have our compute, our software, our GRUT

**[00:34:22 – 00:34:29]** Open source foundation brain for the humanoid, Along with

**[00:34:29 – 00:34:34]** our simulation, Omniverse, Isaac Lab, and operating system,

**[00:34:34 – 00:34:38]** all working so that you can get one of these things, take it

**[00:34:38 – 00:34:43]** out of the box, and it just works from day one and when you need

**[00:34:43 – 00:34:45]** it maintained, because inevitably you're going to harm your

**[00:34:45 – 00:34:48]** robot, you can have it serviced.

**[00:34:48 – 00:34:50]** You don't have to service it yourself.

**[00:34:50 – 00:34:53]** It's a really great thing

**[00:34:54 – 00:35:02]** Which leads into Automotive, which is our most developed, most

**[00:35:02 – 00:35:05]** widely deployed physical AI domain.

**[00:35:05 – 00:35:13]** We've been working on the Automotive and the AV space

**[00:35:13 – 00:35:16]** now for over a decade.

**[00:35:16 – 00:35:21]** Within this space, we have introduced a reference platform a

**[00:35:21 – 00:35:26]** while ago, a reference architecture called NVIDIA Hyperion, which

**[00:35:26 – 00:35:31]** is a modular open platform for the compute and the sensor

**[00:35:31 – 00:35:37]** suite to make it so that there's a standard form that you can build

**[00:35:37 – 00:35:43]** off of, which is now employed, I believe it's being integrated into

**[00:35:46 – 00:35:50]** Within the next five years, 80% of all of the cars produced

**[00:35:50 – 00:35:52]** in Europe will have this.

**[00:35:52 – 00:35:55]** 43 million vehicles by 2033.

**[00:36:02 – 00:36:06]** Every one of these AV developers who are building on this stack

**[00:36:07 – 00:36:14]** can build robo-taxis that take advantage of our robo-taxi

**[00:36:14 – 00:36:21]** brain, Alpamayo, and at the keynote we announced.

**[00:36:21 – 00:36:23]** There are many new European partners that

**[00:36:23 – 00:36:26]** are announcing new lines.

**[00:36:26 – 00:36:32]** So we have Volvo, which has nearly 100 years of safety

**[00:36:33 – 00:36:35]** as their main thing.

**[00:36:35 – 00:36:39]** They trust this enough.

**[00:36:39 – 00:36:45]** Our Hyperion suite is so robust.

**[00:36:45 – 00:36:47]** Our Hyperion platform is so robust that they trust it.

**[00:36:47 – 00:36:50]** We have Volkswagen with the new ID.

**[00:36:51 – 00:36:54]** Bolt, who is using it with the Alpamayo stack for European

**[00:36:54 – 00:37:00]** autonomous ride hailing, starting with 25,000 Lucid vehicles.

**[00:37:00 – 00:37:01]** It's really incredible.

**[00:37:01 – 00:37:04]** I mean, we started on this journey over 10 years ago.

**[00:37:04 – 00:37:06]** It seemed like it was impossible.

**[00:37:06 – 00:37:09]** But pretty much overnight now, it looks like we're going

**[00:37:09 – 00:37:14]** to have robo taxis everywhere.

**[00:37:14 – 00:37:19]** Now, in order to deploy something like this at scale, safety

**[00:37:19 – 00:37:22]** has to be at the heart of it.

**[00:37:22 – 00:37:24]** It can't just be some add-on at the end.

**[00:37:24 – 00:37:30]** We have designed our whole AV stack, starting from the hardware,

**[00:37:30 – 00:37:35]** all the way up to the application layer with the driver, to be safe

**[00:37:35 – 00:37:40]** using our Halos safety platform.

**[00:37:40 – 00:37:44]** This is the first physical AI safety standard

**[00:37:44 – 00:37:49]** platform, I think, ever.

**[00:37:49 – 00:37:52]** It's used by automakers worldwide.

**[00:37:52 – 00:37:58]** 80% of all the cars, all robotaxis that will be developed use it.

**[00:37:58 – 00:38:03]** And along with this safety platform, one of the things

**[00:38:03 – 00:38:09]** that is essential is the ability to do software updates and

**[00:38:09 – 00:38:12]** ensure that it's going to continue to be safe after you update it.

**[00:38:12 – 00:38:16]** The way you do that is by doing a combination of both testing

**[00:38:16 – 00:38:21]** in the real world, just like we normally do, but mostly you have

**[00:38:21 – 00:38:24]** to test in the simulated world.

**[00:38:24 – 00:38:29]** You want to be able to release updates frequently, but you

**[00:38:29 – 00:38:33]** can't afford to go test every possible combination of every

**[00:38:33 – 00:38:36]** scenario for every update inside the physical world.

**[00:38:36 – 00:38:37]** It's just not possible.

**[00:38:38 – 00:38:41]** The only solution to that is to do simulation, and so we provide that

**[00:38:41 – 00:38:45]** as part of our HaloSafe platform.

**[00:38:47 – 00:38:53]** Within the AV domain, we have two big AI platforms.

**[00:38:53 – 00:38:57]** I mentioned Alpamayo, which is the driver, which we open-sourced.

**[00:38:57 – 00:39:01]** But we also have NVIDIA Cosmos.

**[00:39:01 – 00:39:03]** Cosmos is our world foundation model and acts as the

**[00:39:03 – 00:39:07]** backbone for Alpamayo.

**[00:39:09 – 00:39:13]** And Cosmos is actually much more than just the model.

**[00:39:13 – 00:39:19]** With Cosmos, you get the open weights, but we also provide you

**[00:39:19 – 00:39:23]** with everything necessary to create a world foundation model and to

**[00:39:23 – 00:39:25]** recreate the same one that we did.

**[00:39:25 – 00:39:30]** We provide you with the data, we provide you with the pipelines

**[00:39:30 – 00:39:35]** to curate the data, we provide you with all of the scripts

**[00:39:35 – 00:39:39]** and recipes to train this.

**[00:39:39 – 00:39:43]** And then we provide a system for you to customize it, so

**[00:39:43 – 00:39:48]** you can bring in your knowledge, your know-how, your specific data

**[00:39:48 – 00:39:54]** for your domain, for your area, and post-train within the system.

**[00:39:54 – 00:39:58]** And we are now also making it possible to use all of that

**[00:39:58 – 00:40:00]** with other models, with OpenMall.

**[00:40:00 – 00:40:03]** So you don't even, if there's another model that's more

**[00:40:03 – 00:40:07]** appropriate, you could still take advantage of the whole stack.

**[00:40:07 – 00:40:17]** We do this so that everyone, every developer, everyone can have full

**[00:40:17 – 00:40:23]** control and sovereignty over their own knowledge and data and decide

**[00:40:23 – 00:40:26]** what it is that they want to put outside and what they want to keep

**[00:40:26 – 00:40:32]** within their country, within their company, and within their products.

**[00:40:34 – 00:40:39]** With that, I think I'm actually on time; this is kind of amazing, this

**[00:40:39 – 00:40:47]** is, this here is our wall of logos that represents only a small subset

**[00:40:47 – 00:40:53]** of the full physical AI ecosystem that we work with, that we support,

**[00:40:53 – 00:41:02]** that we, we are at the foundation of many of these companies.

**[00:41:02 – 00:41:04]** We build the computing infrastructure and the

**[00:41:04 – 00:41:05]** computers for them.

**[00:41:05 – 00:41:11]** But without them, our technology is meaningless.

**[00:41:11 – 00:41:13]** It's just technology.

**[00:41:13 – 00:41:17]** It's through these partners that we convert our technology

**[00:41:17 – 00:41:22]** into something that actually matters to people, into solutions.

**[00:41:22 – 00:41:25]** So thank you to the whole ecosystem.

**[00:41:26 – 00:41:32]** For all of you here in Europe who are interested and passionate

**[00:41:32 – 00:41:38]** about bringing AI to industry, I recommend you scan this

**[00:41:38 – 00:41:43]** QR code and look, there's many other sessions here that

**[00:41:43 – 00:41:49]** will Teach your teams or educate you on how you can apply all

**[00:41:49 – 00:41:54]** of the amazing technologies coming out at this rapid pace to your

**[00:41:54 – 00:41:58]** particular domain and workloads.

**[00:41:58 – 00:42:03]** And with that, we will have an amazing session following

**[00:42:03 – 00:42:08]** mine here with these esteemed guests who will be talking

**[00:42:08 – 00:42:12]** about the robotaxi movement and what it means for Europe.

**[00:42:12 – 00:42:13]** Thank you very much.

