# The AI Flywheel for Robotics: From Simulation to Physical AI (Presented By Agile Robots)

## Abstract

Key takeaways Explore how Agile Robots leverages the AI flywheel to make physical AI a reality. Learn about the latest NVIDIA technologies driving the future of physical AI. Discover the benefits of simulation and synthetic data in enhancing robot deployment. Gain insights into the strategic importance of simulation and AI for executive decision-making.

## AI Summary

- Presented the AI flywheel for robotics, focusing on the journey from simulation to physical AI and the deployment of robotic applications in industrial environments.
- Highlighted the company's broad product portfolio, including robotic arms, AMRs, AGVs, semi-humanoid, and humanoid robots, with over 20,000 applications deployed at customer sites.
- Emphasized the importance of one brain for different robotic embodiments to ensure scalability and flexibility, allowing the same AI to serve various types of robots.
- Showed the use of real-world data, simulation data, and synthetic data generated with NVIDIA Cosmos to train and validate robotic algorithms, optimizing their performance and adaptability.
- Demonstrated a proof of concept with Telecom, using AI RAN to split intelligence between infrastructure and onboard compute, improving decision-making speed and efficiency.
- Introduced the Agile One SN, a 1.4-meter tall, 43-kilogram humanoid robot equipped with NVIDIA Thor and Isaac GR00T, designed for research and development to bridge the gap between academia and industry.

## Transcript

**[00:00:05 – 00:00:07]** All right. Hello, everybody.

**[00:00:07 – 00:00:11]** Welcome to GTC 2026 in Berlin.

**[00:00:11 – 00:00:17]** My name is Evan Pouyavakili, and I will be your session chair.

**[00:00:17 – 00:00:21]** Today, we're going to be talking about the AI flywheel for robotics,

**[00:00:21 – 00:00:24]** from simulation to physical AI.

**[00:00:24 – 00:00:30]** So today, our speakers are going to be

**[00:00:32 – 00:00:36]** Jimmy Nassif, the Vice President of Innovation and Engineering

**[00:00:36 – 00:00:42]** at Agile Robots and Co-Founder and CTO of IdealWorks, and

**[00:00:42 – 00:00:47]** Anthony Rizk, Head of Innovation and Engineering at Agile Robots

**[00:00:47 – 00:00:50]** and Co-Founder of IdealWorks.

**[00:00:50 – 00:00:54]** Please join me in welcoming our speaker.

**[00:00:59 – 00:01:01]** Good afternoon, everyone.

**[00:01:01 – 00:01:03]** I hope you enjoyed the keynote as much as we did.

**[00:01:03 – 00:01:06]** We had a special keynote for us today because there

**[00:01:06 – 00:01:09]** was a lot of announcements related to our company.

**[00:01:09 – 00:01:11]** So we are a little bit too proud today.

**[00:01:11 – 00:01:15]** But I have to admit, also, Jensen did make my job today very easy.

**[00:01:15 – 00:01:18]** He did it easier and a little bit harder.

**[00:01:18 – 00:01:20]** I would say easier because I don't have to tell you anymore

**[00:01:20 – 00:01:23]** how physical AI is going to change the world.

**[00:01:23 – 00:01:26]** The main question would be when, but harder for me because

**[00:01:26 – 00:01:28]** I wanted to share something with you that has been already

**[00:01:28 – 00:01:33]** shared on the keynote, so we are also very proud about it today.

**[00:01:33 – 00:01:37]** I think the main question today is not anymore when

**[00:01:37 – 00:01:40]** the robots are going to change, sorry, how the robots are

**[00:01:40 – 00:01:42]** going to change the world, but the main question would be

**[00:01:42 – 00:01:48]** when and how to bring the robotic application to the real world,

**[00:01:48 – 00:01:51]** industrial real world at scale.

**[00:01:53 – 00:01:56]** For us at Agile Robots, we've been working on this

**[00:01:56 – 00:01:58]** for almost a decade.

**[00:01:59 – 00:02:04]** We have been founded in Munich, in south of Germany in 2018,

**[00:02:04 – 00:02:07]** out of the DLR, which is the German Aerospace Center.

**[00:02:07 – 00:02:11]** We grew over the last few years a little bit too much, so we

**[00:02:11 – 00:02:19]** are already by 3500 people, 2000 of them are researchers and engineers.

**[00:02:19 – 00:02:22]** So I cannot see it there, I'm still seeing it here.

**[00:02:22 – 00:02:25]** Good, thank you.

**[00:02:25 – 00:02:29]** So, we are a robotics and AI company that has a very

**[00:02:29 – 00:02:32]** clear ambition, how to bring the AI application, the robotic

**[00:02:32 – 00:02:38]** application from a lab to a real industrial environment.

**[00:02:38 – 00:02:41]** In today's presentation, we are going to show you some of these

**[00:02:41 – 00:02:44]** applications that we've already implemented with our customer.

**[00:02:44 – 00:02:49]** And as I already said, we are by now 3,500 people all

**[00:02:49 – 00:02:52]** around the world across three continents, in the US, in Europe,

**[00:02:52 – 00:02:55]** but also in China and in Asia.

**[00:02:56 – 00:02:58]** We have more than 650 patents.

**[00:02:58 – 00:03:01]** As I said, we have more than 2,000 R&D specialists and engineers.

**[00:03:01 – 00:03:04]** But I think the most important number here on this slide

**[00:03:04 – 00:03:07]** is the 20,000 applications that we already deployed at our customers.

**[00:03:07 – 00:03:09]** So that's something we are really proud of.

**[00:03:09 – 00:03:11]** And we're going to show you a glimpse of these

**[00:03:11 – 00:03:13]** applications later on.

**[00:03:13 – 00:03:17]** As I said, we are in 13 different countries.

**[00:03:17 – 00:03:20]** We have 10 plus production sites.

**[00:03:20 – 00:03:23]** We have three production sites only in the south of Germany,

**[00:03:23 – 00:03:26]** around Munich, where we are producing our humanoid robots.

**[00:03:26 – 00:03:31]** You might have the chance to see our new reference platform for...

**[00:03:31 – 00:03:33]** Thank you very much.

**[00:03:47 – 00:03:50]** Agile Robots is not only the Agile Robots company, but

**[00:03:50 – 00:03:54]** we had also a few acquisitions over the past few years.

**[00:03:54 – 00:03:57]** Franca Robots, it's one of our daughter companies.

**[00:03:57 – 00:03:59]** You may be aware of it in the field of research.

**[00:03:59 – 00:04:01]** So I think every researcher has been already working

**[00:04:01 – 00:04:03]** on Franca Robots.

**[00:04:03 – 00:04:06]** Idealworks, which I'm the co-founder and the CTO of

**[00:04:06 – 00:04:10]** the company, which is a spin-off of the BMW Group that got acquired

**[00:04:10 – 00:04:13]** in August 2025 completely by Agile.

**[00:04:13 – 00:04:16]** We are more focusing on AMR, AGV and the orchestration platform

**[00:04:16 – 00:04:20]** that orchestrate any type of robots in the factory and shop floor.

**[00:04:20 – 00:04:24]** Bayer Automation, I think you heard Jensen talk a little

**[00:04:24 – 00:04:27]** bit a lot about SME and Mittelstand in Germany, so Bayer is a

**[00:04:27 – 00:04:30]** 50 years old engineering company specialized in customized

**[00:04:30 – 00:04:33]** solutions for automation,

**[00:04:33 – 00:04:36]** and also across different industries, different verticals.

**[00:04:36 – 00:04:39]** And last but not least, our latest acquisition, Krause

**[00:04:39 – 00:04:42]** Automation, which is like the automation integration company of

**[00:04:42 – 00:04:46]** ThyssenKrupp that we just acquired to help us and support us integrate

**[00:04:46 – 00:04:48]** our solution at our customers.

**[00:04:48 – 00:04:50]** With this product, with this company's portfolio, we are

**[00:04:50 – 00:04:55]** able to serve so many customers all around the globe, and therefore,

**[00:04:55 – 00:04:59]** as I already said, we had around 20,000 plus applications already

**[00:04:59 – 00:05:02]** deployed at our customer sites, productively running on a different

**[00:05:02 – 00:05:06]** product portfolio that I will go on also to share with you later on.

**[00:05:06 – 00:05:09]** Just to give you a glimpse of what we've been doing in

**[00:05:09 – 00:05:11]** the last few years.

**[00:05:11 – 00:05:14]** So this is a small collection, as I said, a glimpse of many

**[00:05:14 – 00:05:19]** of our use cases, our customer onsite, in different industries,

**[00:05:19 – 00:05:24]** different verticals that we've been deploying over the past few years.

**[00:05:49 – 00:05:52]** So, as I said, the main question is how to bring robotics and

**[00:05:52 – 00:05:54]** AI in industrial application.

**[00:05:54 – 00:05:58]** We know that today it's a fact that none of our customers

**[00:05:58 – 00:06:00]** will just deploy robots for fun.

**[00:06:00 – 00:06:02]** At the end of the day, there has to be an ROI.

**[00:06:02 – 00:06:05]** There should be an added value for the customer to deploy

**[00:06:05 – 00:06:06]** a robotic solution.

**[00:06:06 – 00:06:09]** We know for a fact that automation is a must, but

**[00:06:09 – 00:06:12]** it's not always automating use cases with a humanoid.

**[00:06:12 – 00:06:14]** We know that humanoid is a huge hype.

**[00:06:14 – 00:06:16]** We are very proud about our own humanoid, but nevertheless,

**[00:06:16 – 00:06:19]** there are also some application and some use cases at the

**[00:06:19 – 00:06:22]** customer side that only could be automated with just one arm.

**[00:06:22 – 00:06:25]** So that's also one of the reasons why we have such a broad product

**[00:06:25 – 00:06:31]** portfolio that goes across arms, AMRs, AGV, humanoid, semi-humanoid.

**[00:06:31 – 00:06:34]** We are the humanoid, so basically, in order to cover all the

**[00:06:34 – 00:06:37]** type of use cases, all the type of needs that our customers

**[00:06:37 – 00:06:40]** are looking for, because at the end of the day, it's all about ROI.

**[00:06:40 – 00:06:45]** It's all about how we can support our customer to make them have

**[00:06:45 – 00:06:51]** an efficient, but also an optimized production line or even processes

**[00:06:51 – 00:06:54]** on their production or shop floor.

**[00:06:54 – 00:06:56]** For us at Agile, the most important three pillars in

**[00:06:56 – 00:07:01]** order to help us and support us to get to this point are our

**[00:07:01 – 00:07:06]** robotic brain that we're working on in-house, our real physical data,

**[00:07:06 – 00:07:09]** but also our different embodiments.

**[00:07:09 – 00:07:11]** And how does it look for us?

**[00:07:11 – 00:07:13]** So basically, as we already said, so we start from the

**[00:07:13 – 00:07:17]** data, and we had more than 20,000 applications already

**[00:07:17 – 00:07:19]** installed at our customer sites.

**[00:07:19 – 00:07:22]** For sure, the customer data is always owned by the by

**[00:07:22 – 00:07:23]** the customer themselves.

**[00:07:23 – 00:07:26]** But nevertheless, a lot and few of few and a lot of our customers

**[00:07:26 – 00:07:29]** are willing to share their data with us because they know that the

**[00:07:29 – 00:07:32]** added value is coming back to them in order to optimize and make their

**[00:07:32 – 00:07:34]** processes even more efficient.

**[00:07:34 – 00:07:36]** Therefore, we use a lot of different type of data, from an

**[00:07:36 – 00:07:40]** industrial data, which are the real data on the shop floor, but also

**[00:07:40 – 00:07:41]** we have our own data factories.

**[00:07:41 – 00:07:43]** So we have few data factories in Germany.

**[00:07:43 – 00:07:48]** We have also few data factories in the US and few also around Asia,

**[00:07:48 – 00:07:52]** where we are collecting egocentric data that also supporting us like

**[00:07:52 – 00:07:55]** a real world data that's supporting us to train our algorithms.

**[00:07:55 – 00:07:59]** And last but not least, we are also having a synthetic data generation.

**[00:07:59 – 00:08:02]** So basically, we are also going to see it later on.

**[00:08:02 – 00:08:05]** We are building for most of our customers digital twins

**[00:08:05 – 00:08:08]** where we can use these digital twins to create synthetic data,

**[00:08:08 – 00:08:12]** to use the synthetic data to train our algorithms so we can optimize

**[00:08:12 – 00:08:15]** our algorithms, and then we can use them also in the production.

**[00:08:15 – 00:08:17]** This is a part of the data.

**[00:08:17 – 00:08:20]** The second part is, as I said, the brain.

**[00:08:20 – 00:08:23]** Since we have such a broad product portfolio, it's very important for

**[00:08:23 – 00:08:27]** us, in order to keep up, to have one brain on different embodiments.

**[00:08:27 – 00:08:29]** And it doesn't make sense for us to train different

**[00:08:29 – 00:08:32]** brains on different embodiments, so therefore we are focusing

**[00:08:32 – 00:08:35]** on having one brain that can serve the different embodiments.

**[00:08:35 – 00:08:37]** And last but not least is the different embodiments

**[00:08:37 – 00:08:38]** I already spoke about.

**[00:08:38 – 00:08:41]** But as I said, we cover all types of embodiments from

**[00:08:41 – 00:08:47]** the hand, from the cobot, from the arm, from humanoid, semi-humanoid,

**[00:08:47 – 00:08:49]** wheel-based humanoid.

**[00:08:49 – 00:08:51]** Because, as I said before, each and every use case of

**[00:08:51 – 00:08:55]** our customers need a special embodiment to be covered.

**[00:08:55 – 00:08:57]** And it's not always about how Humanoid can be supporting

**[00:08:57 – 00:09:00]** the production line, but it's more about what's the right use case

**[00:09:01 – 00:09:03]** and what's the right technology for the right use case to get

**[00:09:03 – 00:09:06]** the maximum ROI for our customers.

**[00:09:07 – 00:09:09]** How does this all look like?

**[00:09:09 – 00:09:13]** My colleague Anthony will give you a more deep dive into this topic.

**[00:09:13 – 00:09:16]** Anthony, the floor is yours.

**[00:09:16 – 00:09:18]** Thanks, dude.

**[00:09:23 – 00:09:26]** All right, so since we're at GTC, it so happens that

**[00:09:26 – 00:09:29]** the architecture that Jimmy mentioned fits perfectly in

**[00:09:29 – 00:09:32]** the three-computer story of NVIDIA.

**[00:09:32 – 00:09:37]** And so the brain is built on the GB300s, the data is

**[00:09:37 – 00:09:40]** coming from Omniverse and Isaac, and then whenever we want

**[00:09:40 – 00:09:44]** to deploy intelligence on edge, we deploy on the Jetson family.

**[00:09:44 – 00:09:47]** And Jimmy mentioned a very important point.

**[00:09:47 – 00:09:52]** So we have 22,000 robots deployed plus, which means that these

**[00:09:52 – 00:09:57]** robots are making a hundred million decisions and actions every week.

**[00:09:57 – 00:10:02]** And these actions should fit into what we call the AI flywheel.

**[00:10:02 – 00:10:04]** It all starts with the data.

**[00:10:04 – 00:10:08]** And at the basis of this data, there is the human skills,

**[00:10:08 – 00:10:13]** and the idea here is that how we can take the human skills from the

**[00:10:13 – 00:10:17]** human that actually knows how to do it intuitively or learns it and

**[00:10:17 – 00:10:19]** bring it into robots capabilities.

**[00:10:19 – 00:10:22]** And for us to be able to bridge this gap between humans and

**[00:10:22 – 00:10:26]** robots, we found that there are four types of data that we need.

**[00:10:27 – 00:10:29]** First, real-world robotics data.

**[00:10:29 – 00:10:32]** So this data is coming from the industry, and it's coming from our

**[00:10:32 – 00:10:36]** robots academy, where we collect data, whether it is tele-operated

**[00:10:37 – 00:10:39]** or agents that are working with it.

**[00:10:39 – 00:10:42]** Second, there is the simulation, and this is where we heavily use

**[00:10:43 – 00:10:47]** NVIDIA Omniverse for photorealism, to be able to generate data

**[00:10:47 – 00:10:52]** for training, and to be able to validate as well our fleets.

**[00:10:52 – 00:10:55]** Third, this is something we like to call data in the wild,

**[00:10:55 – 00:10:58]** so it is ecocentric data.

**[00:10:58 – 00:11:01]** This data is very important for manipulation because this

**[00:11:01 – 00:11:05]** is the kind of data that is hard to operate and to collect.

**[00:11:05 – 00:11:08]** And this is data that you can get from videos from humans

**[00:11:08 – 00:11:12]** doing complex motions and that you can transfer to your robots.

**[00:11:13 – 00:11:16]** And finally, it's internet-scale visual data.

**[00:11:16 – 00:11:20]** So whenever you want to know general knowledge about visual

**[00:11:20 – 00:11:22]** aspects of things that you would see, like warehouses,

**[00:11:23 – 00:11:26]** pallets, stage, things that you can find on the internet, this is

**[00:11:26 – 00:11:31]** also something that we can use for the foundation of our visual data.

**[00:11:31 – 00:11:35]** And now, let's say we collected our data, we did training,

**[00:11:35 – 00:11:38]** or we did our software development.

**[00:11:38 – 00:11:42]** Then we need to validate it on a big fleet, so we need

**[00:11:42 – 00:11:44]** to validate on scale.

**[00:11:44 – 00:11:47]** And this is where Omniverse comes into play.

**[00:11:47 – 00:11:50]** So firstly, how do we do digital twin creation?

**[00:11:50 – 00:11:54]** So you can start with a scan or something much worse than

**[00:11:54 – 00:11:55]** what you're seeing here.

**[00:11:55 – 00:11:58]** There are automations that can build the pillars of how

**[00:11:58 – 00:12:01]** your digital twin would look like, and then step by step you

**[00:12:01 – 00:12:06]** go from something that resembles to actually nothing to something that

**[00:12:06 – 00:12:08]** really looks like the actual world.

**[00:12:08 – 00:12:12]** And you have AI agents helping you fill this digital twin

**[00:12:12 – 00:12:14]** because you don't want to do it manually whenever you're

**[00:12:14 – 00:12:18]** working at like 100,000 square meters, 120,000 square meters.

**[00:12:19 – 00:12:23]** And at the end, this kind of digital twin allows us to test our

**[00:12:23 – 00:12:28]** fleets, to test it in production sites, how would it look like,

**[00:12:28 – 00:12:34]** to be able to actually see, in a safe manner, whatever change we do

**[00:12:34 – 00:12:38]** on a fleet level, on an individual robot level, the aspect of it.

**[00:12:39 – 00:12:45]** And in the realm of physical AI and digital twins, we're

**[00:12:45 – 00:12:47]** going to show now a use case that we did with Dassault

**[00:12:47 – 00:12:50]** and Audi for part sequencing.

**[00:12:50 – 00:12:55]** And so mainly we use the virtual twin from Dassault where we

**[00:12:55 – 00:12:58]** can inject the robot, when we can inject the work cell,

**[00:12:58 – 00:13:01]** when we can inject also the whole material flow.

**[00:13:01 – 00:13:05]** Then we use NVIDIA Omniverse libraries to be able to get

**[00:13:05 – 00:13:07]** physics and visuals.

**[00:13:08 – 00:13:11]** We use OpenUSD, which is a universal scene descriptor, to be

**[00:13:11 – 00:13:15]** able to move data between companies and make it in one place, so

**[00:13:15 – 00:13:17]** for everybody to be able to use it.

**[00:13:17 – 00:13:22]** And at the end, we can deploy these robots in this virtual

**[00:13:22 – 00:13:28]** world to be able to see different scalabilities, different KPIs.

**[00:13:29 – 00:13:32]** Whenever we want to change something, how it would look like.

**[00:13:32 – 00:13:35]** And this was the end result of it.

**[00:14:28 – 00:14:32]** So what you saw here is something running kind of in real time.

**[00:14:32 – 00:14:35]** So on the left, you had the agent that was able to generate

**[00:14:35 – 00:14:37]** multiple scenarios.

**[00:14:37 – 00:14:40]** And you can see all the KPIs that we need to kind

**[00:14:40 – 00:14:42]** of optimize against.

**[00:14:42 – 00:14:44]** And this whole thing was done in collaboration with

**[00:14:44 – 00:14:45]** Audi and Dassault.

**[00:14:45 – 00:14:48]** And I would also invite you to check the Dassault booth,

**[00:14:48 – 00:14:51]** our partner, and our booth as well, if you want to know

**[00:14:51 – 00:14:53]** more about this use case.

**[00:14:53 – 00:14:57]** Now that we collected the data, we did the training,

**[00:14:57 – 00:15:01]** we simulated, sometimes simulated data is not realistic enough.

**[00:15:02 – 00:15:05]** And this is where we collaborated with NVIDIA, using Cosmos

**[00:15:05 – 00:15:09]** to have sim-to-real and real-to-real generations.

**[00:15:09 – 00:15:12]** So you see here real data, and usually industrial data

**[00:15:12 – 00:15:15]** such as these are very hard to get.

**[00:15:15 – 00:15:18]** You cannot get diversity, you cannot create edge cases,

**[00:15:18 – 00:15:23]** and this is why we use Cosmos for synthetic data generation.

**[00:15:23 – 00:15:27]** And mainly this use case is a mobility use case, and we

**[00:15:27 – 00:15:29]** have published with NVIDIA a white paper around it.

**[00:15:29 – 00:15:31]** We have published a white paper around how synthetic

**[00:15:31 – 00:15:36]** data can help you with mobility use cases that was online,

**[00:15:36 – 00:15:37]** I think, 10 minutes ago.

**[00:15:37 – 00:15:41]** So you can go check it out on the Idealworks website.

**[00:15:41 – 00:15:47]** Also moving to more stringent and complex use cases, you can have

**[00:15:47 – 00:15:52]** manipulators that you augment, and this is another use case that we do

**[00:15:52 – 00:15:56]** with one of our sister companies, Franca, which is an actual

**[00:15:56 – 00:15:58]** industrial use case for unboxing.

**[00:15:58 – 00:16:01]** And whenever you're doing unboxing, whenever you're

**[00:16:01 – 00:16:04]** collecting real-world data, it's very hard to change the setting.

**[00:16:04 – 00:16:06]** It's very hard to change the lighting conditions.

**[00:16:06 – 00:16:09]** It's very hard to change the background, the floor, if you have

**[00:16:09 – 00:16:11]** people around you, if you don't.

**[00:16:11 – 00:16:14]** So these are things that cost a lot, and with Cosmos you

**[00:16:14 – 00:16:19]** can cut this cost and generate this variety of data.

**[00:16:19 – 00:16:22]** So you might have seen some artifacts in the last video,

**[00:16:22 – 00:16:26]** and this is why we collaborated with NVIDIA to bring this

**[00:16:26 – 00:16:27]** to the next step.

**[00:16:27 – 00:16:33]** So, on the right, it is the actual video coming from the industry.

**[00:16:33 – 00:16:37]** Sorry, on your left, on my right.

**[00:16:37 – 00:16:41]** On this side, we actually generated a close call with a forklift.

**[00:16:41 – 00:16:45]** So, this one never happened.

**[00:16:45 – 00:16:47]** There was never a forklift in the scene, and we were

**[00:16:47 – 00:16:51]** able to generate this kind of close call by adding an object

**[00:16:51 – 00:16:53]** that did not exist into our data.

**[00:16:53 – 00:16:55]** And this is how we can augment data.

**[00:16:55 – 00:16:59]** This is how whenever we capture data like the one on the other

**[00:16:59 – 00:17:03]** side from the industry, where we cannot actually simulate this kind

**[00:17:03 – 00:17:07]** of close calls, this kind of edge cases, collisions, we can add this

**[00:17:07 – 00:17:10]** kind of scenarios using Cosmos.

**[00:17:10 – 00:17:15]** And one additional example, and this is also going live today, it's

**[00:17:15 – 00:17:19]** an agentic AI cookbook by Franca.

**[00:17:19 – 00:17:24]** And this is how we can also do this sim-to-real transfer

**[00:17:24 – 00:17:25]** using agentic AI.

**[00:17:25 – 00:17:29]** So mainly you start by prompting your AI agents.

**[00:17:29 – 00:17:33]** You have multiple agents that can plan, and they are very specific.

**[00:17:33 – 00:17:36]** So you have perception, motion planning, estimation,

**[00:17:36 – 00:17:38]** gripping, et cetera.

**[00:17:38 – 00:17:40]** And then you can validate all of this in Omniverse in

**[00:17:40 – 00:17:45]** the digital twin before executing it and moving into the real world.

**[00:17:45 – 00:17:49]** And this is how the cycle looks like.

**[00:18:17 – 00:18:21]** And so this is one example of how you can have a genetic simulation

**[00:18:21 – 00:18:24]** and real-world all interconnected.

**[00:18:24 – 00:18:27]** So, you collected your data, you did your training, you

**[00:18:27 – 00:18:32]** simulated, you augmented, and now it's time to actually deploy.

**[00:18:32 – 00:18:39]** And this is a proof of concept that we did with Telecom, and

**[00:18:39 – 00:18:41]** I will give a small analogy here.

**[00:18:41 – 00:18:45]** So imagine you have your brain in your backpack, and the smarter you

**[00:18:45 – 00:18:47]** get, the heavier the backpack is.

**[00:18:47 – 00:18:51]** And at some point, you might make a lot of fast decisions, but it will

**[00:18:51 – 00:18:53]** be a bit heavy for you to move.

**[00:18:53 – 00:18:56]** And this is why we partnered with Telecom and this proof

**[00:18:56 – 00:19:01]** of concept to be able to use AI RAN, which combines mobile

**[00:19:01 – 00:19:05]** network infrastructure with AI computing to split this

**[00:19:05 – 00:19:10]** intelligence between the AI RAN infrastructure and the onboard

**[00:19:10 – 00:19:12]** compute that is on the robot.

**[00:19:12 – 00:19:15]** And this is because sometimes it is too heavy to have a lot of big

**[00:19:15 – 00:19:19]** models on the robot itself, so you need this kind of hybrid approach.

**[00:19:19 – 00:19:24]** And this is a comparison for a task for fan insertion

**[00:19:24 – 00:19:26]** for server building.

**[00:19:26 – 00:19:29]** And you can see the difference between the ARN server plus

**[00:19:29 – 00:19:35]** on-device deployment versus the on-device only deployment server.

**[00:20:06 – 00:20:10]** And so the reason why one of the models or one of the

**[00:20:10 – 00:20:13]** one of the task was done faster than the other, it's because

**[00:20:13 – 00:20:16]** when we leverage AI RAN, we were able to dream ahead.

**[00:20:16 – 00:20:19]** So the robot kind of actually knew what the next phase would

**[00:20:19 – 00:20:22]** be, what the next frames would be, and it could compute a little

**[00:20:22 – 00:20:26]** bit faster and more efficiently before taking the actions.

**[00:20:28 – 00:20:33]** And to sum it up, we said we're going to deploy, and

**[00:20:33 – 00:20:36]** as Jimmy mentioned, at Agile, we have multiple embodiments.

**[00:20:36 – 00:20:40]** And the same as the story of the software is kind of

**[00:20:40 – 00:20:44]** a story of evolution, so how the flywheel evolves, the

**[00:20:44 – 00:20:46]** same goes for our hardware.

**[00:20:46 – 00:20:52]** So the hardware for Agile evolved from robotic arms, AMRs, to mobile

**[00:20:53 – 00:20:58]** semi-humanoid robots, and now to full-on bipedal humanoid robots.

**[00:20:58 – 00:21:02]** And we need this kind of family to be able to serve our customers

**[00:21:02 – 00:21:05]** end-to-end because we are serving multiple verticals.

**[00:21:05 – 00:21:10]** And perhaps the last level or the latest stage of this

**[00:21:10 – 00:21:14]** kind of evolution was the Agile One SN that was announced today

**[00:21:14 – 00:21:16]** as a Groot reference platform.

**[00:21:16 – 00:21:22]** So, this robot is 1.4 meters, so not too long, 43 kilograms,

**[00:21:22 – 00:21:24]** also very lightweight.

**[00:21:24 – 00:21:27]** It is meant for research and development, and it

**[00:21:27 – 00:21:31]** is our way to be able to bridge the gap between research

**[00:21:31 – 00:21:34]** and academia to the industry.

**[00:21:35 – 00:21:37]** It is equipped with two dexterous hands.

**[00:21:37 – 00:21:42]** It has NVIDIA Thor integrated inside the robot, so

**[00:21:42 – 00:21:45]** it's not like an external backpack or external GPU.

**[00:21:45 – 00:21:48]** It is equipped with cameras, with a lidar in the chin to

**[00:21:48 – 00:21:52]** be able to have the 3D recognition of the environment.

**[00:21:52 – 00:21:55]** And it is also connectable to WiFi, and it is based on

**[00:21:55 – 00:21:58]** Isaac Groot reference platform.

**[00:21:58 – 00:22:01]** So it comes built-in with the NVIDIA toolchain to be

**[00:22:01 – 00:22:04]** able to kind of apply the flywheel that we discussed

**[00:22:04 – 00:22:07]** right now directly on this robot.

**[00:22:07 – 00:22:12]** This robot will be available in early 2027 for purchase.

**[00:22:14 – 00:22:21]** And so, as we said, we have hundreds of locations, thousands

**[00:22:21 – 00:22:26]** of robots, millions of decisions and actions being made.

**[00:22:26 – 00:22:29]** And this is the power of the flywheel, because these kind

**[00:22:29 – 00:22:33]** of decisions, whenever they fed into the flywheel, every one

**[00:22:33 – 00:22:37]** of them makes the next one smarter.

**[00:22:44 – 00:22:49]** So I think we have time for two or three questions.

**[00:22:54 – 00:22:56]** Thank you.

**[00:23:00 – 00:23:02]** Hi, a very quick question.

**[00:23:02 – 00:23:06]** When do you expect to have an industry-grade humanoid

**[00:23:06 – 00:23:10]** robot to be bought?

**[00:23:10 – 00:23:12]** So, as we already showed, we have already one industrial

**[00:23:12 – 00:23:14]** robot called Agile One.

**[00:23:15 – 00:23:17]** So basically, we have few applications, but they are all

**[00:23:17 – 00:23:20]** still in the POC phase with some of our early adopter customers,

**[00:23:20 – 00:23:22]** where we are deploying it on site.

**[00:23:22 – 00:23:25]** But before getting to a fully deployed solution

**[00:23:26 – 00:23:29]** in the production, you need a certification for the robot.

**[00:23:29 – 00:23:31]** So the certification is still missing.

**[00:23:31 – 00:23:34]** We are one of few companies supporting on finalizing the

**[00:23:34 – 00:23:40]** certification for the ISO in the US, but also in Europe, in order

**[00:23:40 – 00:23:42]** for us to be able to deploy these robots on the production line.

**[00:23:42 – 00:23:45]** So, in order to answer your question directly, we don't

**[00:23:45 – 00:23:48]** know now the timeline when we will be deploying these

**[00:23:48 – 00:23:50]** robots on the production line, but we have already early adopters,

**[00:23:50 – 00:23:53]** customers that we are working with them, but before solving

**[00:23:53 – 00:23:56]** this issue, we need to solve the issue of the certification.

**[00:23:56 – 00:23:58]** That's also one of the main reasons why, for example,

**[00:23:58 – 00:24:01]** we introduced the semi-humanoid, which is a wheel-humanoid,

**[00:24:01 – 00:24:04]** because it's a different approach when it comes also to safety.

**[00:24:04 – 00:24:07]** And we are able to deploy the robots already now, as we already

**[00:24:07 – 00:24:10]** showed with Audi, for example.

**[00:24:10 – 00:24:12]** And at the end of the day, if the robot is driving on a

**[00:24:12 – 00:24:17]** wheelbase or with bipedal, I would say the only open question back

**[00:24:17 – 00:24:21]** then would be the certification.

**[00:24:29 – 00:24:34]** Um, why only one brain for all the different bodies?

**[00:24:34 – 00:24:37]** So you said you're gonna have one brain for all the different bodies.

**[00:24:37 – 00:24:40]** Why? Because of some flexibility, or?

**[00:24:40 – 00:24:43]** So, as I said before, when I started the whole presentation,

**[00:24:43 – 00:24:44]** it's all about scalability.

**[00:24:44 – 00:24:46]** How are you able to scale?

**[00:24:46 – 00:24:48]** And we believe at Agile, you are only able to scale if you have

**[00:24:48 – 00:24:49]** one brain on different embodiments.

**[00:24:49 – 00:24:52]** So, you are training the same brain to serve different embodiments,

**[00:24:52 – 00:24:54]** because at the end of the day, also, the customers will

**[00:24:54 – 00:24:55]** not only have one embodiment.

**[00:24:55 – 00:24:58]** They will have different embodiments, different use

**[00:24:58 – 00:25:00]** cases, different application, and at the end of the day,

**[00:25:00 – 00:25:03]** if we can serve the same brain, the different embodiments,

**[00:25:03 – 00:25:06]** that's where the power of scaling and being able to scale very

**[00:25:06 – 00:25:09]** fast and being able to be flexible.

**[00:25:21 – 00:25:25]** Yes, about the different, one brain for different on modiment.

**[00:25:25 – 00:25:29]** What is the ratio of real data versus simulation data you need,

**[00:25:29 – 00:25:33]** specifically when you, for example, have real data for robotic arms,

**[00:25:33 – 00:25:39]** but you want to train for HGV or IMR, which is radically different?

**[00:25:39 – 00:25:43]** So, the answer that nobody likes to hear, it depends on the use case.

**[00:25:43 – 00:25:48]** So, for example, in the white paper that came out today, we showcased

**[00:25:48 – 00:25:51]** how, for mobility, you can have a full-on simulated data, and

**[00:25:51 – 00:25:54]** you can deploy it on a real robot, on a Jetson, and it will operate.

**[00:25:54 – 00:25:58]** So, it depends on the complexity of the use case.

**[00:25:58 – 00:26:01]** It depends on the generalization that is needed and the variety.

**[00:26:01 – 00:26:06]** So, there is not one answer that will cover it up.

**[00:26:13 – 00:26:17]** Are you sure many use cases, and are your robots to be

**[00:26:17 – 00:26:24]** able to handle deformable materials like bags, tubes,

**[00:26:24 – 00:26:27]** life science and pharma industry?

**[00:26:27 – 00:26:29]** That's a very good question, and I think, I don't know if

**[00:26:29 – 00:26:31]** you were able to see it in the use case of Audi, but there is also,

**[00:26:31 – 00:26:35]** for example, a plastic cover that need to be unfold in order to get

**[00:26:35 – 00:26:41]** to the second layer of the pillars of the car, so basically, yes.

**[00:26:43 – 00:26:46]** Yes, I have a question regarding the synthetic data that you

**[00:26:46 – 00:26:49]** are generating using Cosmos for NVIDIA, right?

**[00:26:49 – 00:26:53]** Is there a possibility to generate synthetic data more than video,

**[00:26:54 – 00:27:00]** or is it only meant to generate synthetic data based on video?

**[00:27:00 – 00:27:03]** So, what we showed is kind of video-to-video transfer,

**[00:27:03 – 00:27:08]** but we also have in the pipeline a way to generate using prompts.

**[00:27:08 – 00:27:12]** So, you don't have a base, but the way we saw it, it's

**[00:27:12 – 00:27:16]** that you usually need some reference frame, because you're

**[00:27:16 – 00:27:19]** generating for things that are probably specific for you.

**[00:27:19 – 00:27:21]** You want to also, when it comes to economy, you don't

**[00:27:21 – 00:27:24]** want to generate a scale that you wouldn't need, so you would target

**[00:27:24 – 00:27:25]** the use case that you would need.

**[00:27:25 – 00:27:28]** And so this is why we found this to be the more efficient

**[00:27:28 – 00:27:32]** way to do it, but you have the possibility to do it via,

**[00:27:32 – 00:27:36]** like, prompting or other stuff, or even adding, like, skeletons,

**[00:27:36 – 00:27:41]** depth images, like, it doesn't have to be only RGB as well.

**[00:27:46 – 00:27:49]** So I saw very little about teleop in the data part of

**[00:27:49 – 00:27:50]** the presentation.

**[00:27:50 – 00:27:53]** With simulation becoming easier and easier and frontier models

**[00:27:53 – 00:27:57]** being able to control robots better, do you think teleop

**[00:27:57 – 00:28:00]** will become less relevant in your data mix in the future?

**[00:28:00 – 00:28:02]** I'm sorry, can you repeat the question?

**[00:28:02 – 00:28:05]** Do you think teleops will become less relevant in training

**[00:28:05 – 00:28:08]** data mix in the future, given that simulation is easier

**[00:28:08 – 00:28:12]** and maybe frontier models are able to control robots better today?

**[00:28:12 – 00:28:15]** Yeah, I mean, at some point, it's a question of scale,

**[00:28:15 – 00:28:21]** so scaling teleop versus scaling simulation, it depends also on the

**[00:28:21 – 00:28:25]** resources, like where you want to put your efforts, if where you have

**[00:28:25 – 00:28:28]** your teleop data, your teleop labs, it's something that is reusable

**[00:28:28 – 00:28:30]** for something else as well.

**[00:28:30 – 00:28:33]** So it also depends at some point, but the way it is going

**[00:28:33 – 00:28:37]** right now, also with agents, as you mentioned, and the way where

**[00:28:37 – 00:28:40]** you can have this kind of smartness embedded in the simulation,

**[00:28:40 – 00:28:43]** it could be that you would still need a lot of data, but it would

**[00:28:43 – 00:28:47]** be more complementary as it goes.

**[00:28:54 – 00:28:56]** You mentioned that you only have one brain, but usually

**[00:28:57 – 00:29:00]** customers demand data separation between companies.

**[00:29:00 – 00:29:03]** So do you run different versions of the brain based on different

**[00:29:03 – 00:29:06]** flywheels per customer, or how do you address it?

**[00:29:07 – 00:29:14]** Yes, so basically, the customer data belongs to the customer.

**[00:29:14 – 00:29:17]** When we were talking about also one brain, it's the data that

**[00:29:17 – 00:29:20]** we generate from, for example, our data farm, from some simulations,

**[00:29:20 – 00:29:26]** et cetera, from the internet scale, so because... At some point, let's

**[00:29:26 – 00:29:32]** take an example that you bring a human to do a task in the factory.

**[00:29:32 – 00:29:35]** So you don't teach, for example, the human to open the door.

**[00:29:35 – 00:29:38]** These are skills that the humans, they already know.

**[00:29:38 – 00:29:42]** But you might teach the human a task that is very specific

**[00:29:42 – 00:29:45]** to the thing that you wanted to do at the factory floor.

**[00:29:45 – 00:29:48]** And this is also the whole idea behind it, so for now, because

**[00:29:48 – 00:29:52]** of GDPR, of data governments, for sure, you cannot cross-fertilize

**[00:29:52 – 00:29:54]** the data between customers.

**[00:29:54 – 00:29:59]** You can have one data pool where you can have these general

**[00:29:59 – 00:30:02]** skills, like how a robot can use a screwdriver, how it can pick

**[00:30:02 – 00:30:04]** and place, and all of these things.

**[00:30:04 – 00:30:09]** And later on, you might fine-tune, depending on the customer.

**[00:30:09 – 00:30:13]** And then, yeah, hopefully, at some point, this will be

**[00:30:13 – 00:30:16]** solved for all of us.

**[00:30:21 – 00:30:25]** Yeah, I would like to ask about the skills for the robots and,

**[00:30:25 – 00:30:29]** like, yeah, a bit more low-level.

**[00:30:29 – 00:30:36]** When do you think that it's better to plan, like, plan

**[00:30:36 – 00:30:40]** the motion and all the things, and when to, like, use the policy?

**[00:30:41 – 00:30:46]** And if you have, like, kind of a brain, or brain, like, policy,

**[00:30:46 – 00:30:52]** do you think that it's better to separate, like, between...

**[00:30:57 – 00:31:01]** So, this will also depend on what you're doing.

**[00:31:01 – 00:31:03]** So, I'll give an example on mobility level.

**[00:31:03 – 00:31:07]** So, whenever you have any kind of policy controlling a robot that is

**[00:31:07 – 00:31:10]** driving at 2.2 meters per second,

**[00:31:10 – 00:31:14]** you might need something specific for that, because you cannot

**[00:31:15 – 00:31:18]** take the chance with latency or all of these things, so you need to

**[00:31:18 – 00:31:21]** be real-time, you need to be fast, and you need to account for that.

**[00:31:21 – 00:31:24]** So, for example, when it comes to mobility and driving, you

**[00:31:24 – 00:31:25]** would do it in a way.

**[00:31:25 – 00:31:30]** Whenever maybe you have the luxury, let's say, of time to be able

**[00:31:30 – 00:31:33]** to do tasks where you need to pick and place, and then it is within

**[00:31:33 – 00:31:39]** the range of the workflow and the work cell, then maybe it would

**[00:31:39 – 00:31:42]** be easier for you to have this kind of plan ahead, as you said, like

**[00:31:42 – 00:31:45]** what we also call the dream ahead, to be able to have a smoother

**[00:31:45 – 00:31:47]** functionality of the robot.

**[00:31:47 – 00:31:51]** So at the end of the day, it also depends on the use

**[00:31:51 – 00:31:54]** case itself as well.

**[00:31:58 – 00:31:59]** Thank you, everybody.

**[00:31:59 – 00:32:01]** That's all the time we have, and thank you for

**[00:32:01 – 00:32:03]** turning GDC Berlin 2026.

