# Transforming Industry With Physical AI (Presented by Agile Robots)

## Abstract

Agile Robots is driving a shift in industrial automation—from systems that follow predefined rules to machines that perceive, learn, and act in the physical world through AI. This talk explores how Agile Robots builds on deep experience in industrial automation and combines it with advanced AI to realize physical AI as a new paradigm for designing work, manufacturing, and production, where intelligence is embodied in robotic systems, rather than confined to software. The presentation highlights how Agile Robots is developing today’s and the next generation's robotic systems—including humanoid robots—by integrating NVIDIA Isaac open frameworks and libraries to bring AI-driven flexibility, autonomy, and collaboration to the factory floor, and outlines the technological trajectory ahead.

## AI Summary

- Agile Robots, founded in 2018 as a spin-off from the Aerospace Institute DLR, presented their unique approach to physical AI and their extensive portfolio of robotic solutions.
- The company highlighted their development of force-sensitive robots, which became the leading choice for AI research and physical AI in top labs such as DeepMind and Meta.
- Agile Robots discussed their data collection methods, including in-house data gathering, simulations, and the use of production line processes to train their AI models, emphasizing the importance of combining real-world and synthetic data.

## Transcript

**[00:00:05 – 00:00:09]** Hi everyone, how are you doing?

**[00:00:09 – 00:00:10]** Yeah, thank you for coming.

**[00:00:10 – 00:00:12]** My name is Dmitrij Forssilow.

**[00:00:12 – 00:00:15]** I'm head of strategic accounts at Agile Robots and also a

**[00:00:15 – 00:00:20]** co-founder of one of the subsidiary companies, Franka Robots.

**[00:00:20 – 00:00:24]** And today I want to talk about Agile Robots, who we are, and about

**[00:00:24 – 00:00:28]** our unique approach to physical AI.

**[00:00:28 – 00:00:32]** So Agile Robots, you might have not heard about us, we are quite

**[00:00:32 – 00:00:37]** a young company, founded in 2018 in Germany, Munich, as a spin-off

**[00:00:37 – 00:00:41]** from Aerospace Institute, DLR.

**[00:00:41 – 00:00:45]** At DLR we were developing sensor technologies and robots

**[00:00:45 – 00:00:49]** for space operation, teleoperation, and then we applied this technology

**[00:00:49 – 00:00:52]** to industrial robotics.

**[00:00:52 – 00:00:54]** We are quite a young company, but still quite big.

**[00:00:54 – 00:00:59]** We have more than 3,000 employees today, with quite a big presence

**[00:00:59 – 00:01:04]** in Asia, to be very close to our customers for consumer electronics

**[00:01:04 – 00:01:06]** and also car manufacturing.

**[00:01:06 – 00:01:11]** In Germany, we have our R&D powerhouse, so all the development

**[00:01:11 – 00:01:13]** is happening in Germany, in Munich.

**[00:01:13 – 00:01:17]** And one of the newcomers is Palo Alto, so we are opening an

**[00:01:17 – 00:01:24]** office here to be closer to our key customers in the Silicon Valley.

**[00:01:25 – 00:01:29]** We grew quite fast, not just by hiring talent organically,

**[00:01:29 – 00:01:32]** but also by acquiring companies.

**[00:01:32 – 00:01:35]** Let's start with the first one, my favorite, Franka Robotics,

**[00:01:35 – 00:01:37]** which I'm a co-founder of.

**[00:01:37 – 00:01:41]** At Franka we focus on force-sensitive robots,

**[00:01:41 – 00:01:45]** meaning that we can sense the forces of the objects

**[00:01:45 – 00:01:47]** and we produce very safe robots.

**[00:01:47 – 00:01:50]** And today we are number one robot which has been used

**[00:01:50 – 00:01:53]** in AI research and physical AI.

**[00:01:53 – 00:01:57]** All the lead labs like DeepMinds of this world and Metas are

**[00:01:57 – 00:01:59]** using our robot for that.

**[00:01:59 – 00:02:01]** Another one, Idealworks.

**[00:02:01 – 00:02:03]** It's a spin-off from BMW.

**[00:02:03 – 00:02:06]** They focus on mobile platforms.

**[00:02:06 – 00:02:11]** These are basically the working force of in-house logistics at BMW.

**[00:02:11 – 00:02:17]** There are thousands of those shipping products in BMW factories,

**[00:02:17 – 00:02:21]** and they have a very advanced fleet management system.

**[00:02:21 – 00:02:23]** Another one, Bear Automation.

**[00:02:24 – 00:02:28]** They are focusing on heavy-duty mobile platforms for example used

**[00:02:28 – 00:02:34]** by Audi and one of the last ones Krause Automation, this company we

**[00:02:34 – 00:02:40]** acquired from ThyssenKrupp and they focus on industrial automation for

**[00:02:40 – 00:02:44]** car manufacturers based in Detroit.

**[00:02:45 – 00:02:48]** We have quite a big portfolio of hardware.

**[00:02:48 – 00:02:50]** Let's start with the robot arms.

**[00:02:50 – 00:02:54]** We have all different sizes, cobots, small robots, middle

**[00:02:54 – 00:02:59]** size, even big robots for heavy-duty operation.

**[00:02:59 – 00:03:02]** Robotic hands, in fact, the robotic hand that you see

**[00:03:02 – 00:03:07]** here, it was a PhD thesis of our CEO and founder, Xiaoping Cheng.

**[00:03:07 – 00:03:10]** He was working on this robotic hand, and we are still putting

**[00:03:10 – 00:03:13]** a lot of focus on the development of these hands, also using

**[00:03:13 – 00:03:15]** them in our humanoid robots.

**[00:03:16 – 00:03:19]** The AGVs and AMRs, we have quite a big portfolio of mobile robots,

**[00:03:20 – 00:03:23]** which allows us to combine them

**[00:03:24 – 00:03:26]** to apply to different use cases.

**[00:03:26 – 00:03:30]** You see, like, putting single arms on the mobile

**[00:03:30 – 00:03:32]** robots, even dual arms.

**[00:03:32 – 00:03:35]** And these systems, they have a very high potential.

**[00:03:35 – 00:03:38]** They can cover so many use cases.

**[00:03:38 – 00:03:41]** We are heavily developing the systems today.

**[00:03:41 – 00:03:45]** Of course, today, humanoids, it's quite a strong trend,

**[00:03:45 – 00:03:46]** we are not missing that.

**[00:03:46 – 00:03:49]** In fact, we have two humanoids here, feel free to come and

**[00:03:49 – 00:03:53]** to touch and to play around with them, very advanced.

**[00:03:54 – 00:03:58]** Intelligent operating system, AgileCore, it's our own layer,

**[00:03:58 – 00:04:02]** how to synchronize and how to program the robots, which we

**[00:04:02 – 00:04:05]** use in our industrial automation.

**[00:04:05 – 00:04:08]** In sensing technologies, we try to own the whole stack

**[00:04:08 – 00:04:12]** of the technology, meaning that we develop our own sensors,

**[00:04:12 – 00:04:16]** own gears, own motors, which we are using in our own robots.

**[00:04:16 – 00:04:20]** And last but not least, the foundation model, we are developing

**[00:04:20 – 00:04:24]** our own VLA, Vision Language Action Model, and I will talk

**[00:04:24 – 00:04:27]** a bit more about that later.

**[00:04:28 – 00:04:35]** Our customers in the AI space, as I mentioned, DeepMinds,

**[00:04:35 – 00:04:40]** Metas of this world, are using our robot for AI research, basically

**[00:04:40 – 00:04:43]** using our robots to collect data, to train their models,

**[00:04:43 – 00:04:45]** and to apply in the real world.

**[00:04:45 – 00:04:49]** Consumer electronics, smartphone manufacturing, laptops, all

**[00:04:49 – 00:04:54]** kinds of electronics, we are using our robots to automate that.

**[00:04:54 – 00:04:59]** Automotive, for example, polishing application, very tough.

**[00:04:59 – 00:05:01]** You need very advanced force control to do that.

**[00:05:01 – 00:05:05]** So that's what we are doing with the car manufacturers.

**[00:05:05 – 00:05:09]** Medical healthcare, what you see here on the picture, it's

**[00:05:09 – 00:05:13]** a client, Escape, with an A. And they're using our robots

**[00:05:13 – 00:05:15]** to do a robotic massage.

**[00:05:15 – 00:05:16]** It's very, very good.

**[00:05:16 – 00:05:17]** They have a couple of offices here.

**[00:05:17 – 00:05:18]** You can give it a try.

**[00:05:18 – 00:05:20]** I highly recommend that.

**[00:05:20 – 00:05:25]** Logistics, we are combining our solutions with mobile platforms,

**[00:05:25 – 00:05:30]** robotic arms to solve the logistics problems of the industry.

**[00:05:30 – 00:05:33]** And other applications like power management, automation,

**[00:05:33 – 00:05:38]** where we also combine different kinds of arms, AGVs, to

**[00:05:38 – 00:05:41]** automate those problems.

**[00:05:41 – 00:05:46]** I will give a short video of some of the examples.

**[00:05:48 – 00:05:53]** There should be a sound, there should be a sound.

**[00:05:54 – 00:05:56]** Okay.

**[00:06:21 – 00:06:22]** Okay, let's talk about AI.

**[00:06:22 – 00:06:24]** We are at GTC.

**[00:06:24 – 00:06:27]** We think AI has three main pillars.

**[00:06:27 – 00:06:31]** We start with the robot body, meaning that the physical

**[00:06:31 – 00:06:35]** embodiment of the AI.

**[00:06:35 – 00:06:40]** We have a lot of different robots, different hands, sensor technology.

**[00:06:40 – 00:06:43]** By combining them in a kind of clever way, we built the

**[00:06:43 – 00:06:46]** different types of robots.

**[00:06:46 – 00:06:48]** The next one is robotic brain.

**[00:06:48 – 00:06:52]** The AI model that executes and gives the commands to the body.

**[00:06:52 – 00:06:57]** And last piece, which is very important, is the data.

**[00:06:57 – 00:07:01]** To train a chat GPT model, the large language models,

**[00:07:01 – 00:07:04]** you can basically scan all the internet, all the books, and it's

**[00:07:04 – 00:07:06]** enough data to train the model.

**[00:07:06 – 00:07:09]** For the physical world, this data doesn't really exist.

**[00:07:09 – 00:07:12]** So we need to generate it, we need to collect it, we need to extract

**[00:07:12 – 00:07:17]** it from somewhere in order to train the brain and apply to the body.

**[00:07:17 – 00:07:20]** And this combination of these three pillars is quite unique,

**[00:07:20 – 00:07:24]** and that's why we think we have a very, very strong standing

**[00:07:24 – 00:07:27]** in the physical AI.

**[00:07:27 – 00:07:29]** How do we collect the data?

**[00:07:29 – 00:07:31]** We run data collection in-house.

**[00:07:31 – 00:07:35]** Basically, people collect the data with our robots day and night.

**[00:07:35 – 00:07:38]** We run a lot of simulations, synthetic data.

**[00:07:38 – 00:07:44]** We also equip our workers in the production line to collect data of

**[00:07:44 – 00:07:45]** the processes that they are doing.

**[00:07:45 – 00:07:48]** Basically, when you assemble a gear or electronics, all

**[00:07:48 – 00:07:53]** this data can be recorded and then applied to train the model.

**[00:07:54 – 00:07:57]** Also, you can try to scan all the internet videos.

**[00:07:57 – 00:08:00]** This is one of the next steps, I guess.

**[00:08:00 – 00:08:02]** We are trying that as well.

**[00:08:02 – 00:08:08]** And it's a huge chunk of data that we can use for training as well.

**[00:08:09 – 00:08:11]** This is one of the examples.

**[00:08:11 – 00:08:12]** What you see here is our production line.

**[00:08:12 – 00:08:16]** We use our robots to produce our own robots, and we use

**[00:08:16 – 00:08:19]** a lot of simulation, what you see there, Isaac SIM.

**[00:08:19 – 00:08:24]** So basically it's a digital SIM of the production line,

**[00:08:24 – 00:08:27]** and by using this synthetic data, we can increase the amount

**[00:08:27 – 00:08:29]** of data by 100x or even more.

**[00:08:29 – 00:08:33]** But still, there is a physical copy of the production line with

**[00:08:33 – 00:08:36]** the real-world data where basically the digital twin can always

**[00:08:37 – 00:08:40]** compare and be very close to that.

**[00:08:42 – 00:08:44]** Next step, humanoids.

**[00:08:44 – 00:08:47]** We have quite a history with them, we tried with different

**[00:08:47 – 00:08:51]** embodiments, combining mobile platforms, putting different arms

**[00:08:51 – 00:08:56]** on them, iterating, having like next and next generation, putting

**[00:08:56 – 00:08:58]** more features, more sensors.

**[00:08:58 – 00:09:01]** And then today we are at the biped, basically a walking

**[00:09:02 – 00:09:05]** humanoid, which can execute different kinds of tasks.

**[00:09:05 – 00:09:08]** We are starting to use them at our own factories.

**[00:09:08 – 00:09:11]** We don't try to go to household.

**[00:09:11 – 00:09:14]** We believe it's an industrial machine.

**[00:09:14 – 00:09:17]** We want to use it in our own factories to make it mature,

**[00:09:17 – 00:09:21]** to make it safe, to make it execute the real-world applications and not

**[00:09:21 – 00:09:25]** just dancing or doing backflips.

**[00:09:26 – 00:09:30]** And I will leave you with a video of that.

**[00:09:57 – 00:09:59]** I am Agile-1.

**[00:09:59 – 00:10:01]** How can I assist you today?

**[00:10:01 – 00:10:05]** Agile-1, please help me assemble the F-A3 robot arm.

**[00:10:05 – 00:10:08]** Okay, let's go.

**[00:11:11 – 00:11:14]** I am Agile One.

**[00:11:19 – 00:11:24]** This is our factory in Germany, where we produce all the robots.

**[00:11:26 – 00:11:27]** Thank you very much for your attention.

**[00:11:27 – 00:11:29]** Please come to our booth, you will see humanoids, we

**[00:11:29 – 00:11:32]** have some other robots you can play around, you can touch.

**[00:11:32 – 00:11:33]** Thank you very much.