# Automation Meets Autonomy: Europe’s Next Era of Robotics

## Abstract

Key takeaways Learn how physical AI and open models are turning fixed automation into intelligent robotic arms, mobile robots, humanoids, and fleets that adapt to changing tasks and environments. Discover how agentic AI, real time digital twins, simulation, and robot agnostic software can accelerate the full robotics lifecycle, from design and learning through validation, deployment, and operations. Gain insights into how Europe can safely scale industrial autonomy through reliable hardware, interoperable software, human robot collaboration, and integration across factories, warehouses, and other industrial environments.

## AI Summary

_No AI summary available._

## Transcript

**[00:00:05 – 00:00:07]** Hello. Oh, great.

**[00:00:07 – 00:00:10]** Hello, everyone. How are you?

**[00:00:10 – 00:00:14]** Excited for the next physical AI session?

**[00:00:15 – 00:00:16]** Yeah, great.

**[00:00:16 – 00:00:19]** So this is our second physical AI session of the day.

**[00:00:19 – 00:00:22]** Whoever was here before, I'm sorry I'm repeating these

**[00:00:22 – 00:00:26]** couple things again, but I'd like you to know that we have

**[00:00:26 – 00:00:29]** an opportunity to review and give feedback for the session.

**[00:00:29 – 00:00:31]** Afterwards, you can go into the app and do it.

**[00:00:31 – 00:00:34]** It's very easy, and it helps us a lot, too.

**[00:00:34 – 00:00:36]** Curate the content for you.

**[00:00:36 – 00:00:40]** The session will be available 72 hours from now, so you can

**[00:00:40 – 00:00:43]** watch it on demand on our webpage.

**[00:00:43 – 00:00:46]** And with that being said, let's get started.

**[00:00:46 – 00:00:51]** For this next session, we're gonna explore how we go from intelligent

**[00:00:51 – 00:00:55]** robotics and into industrial operations with a great panel.

**[00:00:55 – 00:00:59]** We have Madison Huang, NVIDIA's Physical AI Senior Director

**[00:00:59 – 00:01:03]** of Product Marketing and Technical Marketing, and she will be

**[00:01:03 – 00:01:07]** introducing our panelists from Dexory, Agile Robots,

**[00:01:07 – 00:01:10]** Humanoid, Serect, and Wandelbots.

**[00:01:10 – 00:01:11]** Please give them a round of applause.

**[00:01:11 – 00:01:13]** Thank you.

**[00:01:18 – 00:01:23]** Thank you. All right, let's test our panelists, see if

**[00:01:23 – 00:01:26]** they remember where to sit.

**[00:01:29 – 00:01:31]** You've done it. Okay, good.

**[00:01:31 – 00:01:32]** All right, welcome everyone.

**[00:01:32 – 00:01:35]** We're super excited to have you here at GTC Berlin.

**[00:01:35 – 00:01:38]** Every GTC, especially in Europe, we always host our kind of

**[00:01:38 – 00:01:41]** hero robotics panel, where you can hear from some of the

**[00:01:41 – 00:01:45]** great founders and technologists of some of the leading robotics

**[00:01:45 – 00:01:50]** companies to tell you what's real, what's happening now, what's

**[00:01:50 – 00:01:52]** the frontier, what's coming next.

**[00:01:52 – 00:01:56]** And so I'm really excited to have all of these incredible panelists.

**[00:01:56 – 00:01:59]** I'll let you guys have 10 seconds max To introduce yourselves,

**[00:01:59 – 00:02:02]** the company, and then your hero robotics platform, and

**[00:02:02 – 00:02:06]** we have some images behind you that kind of introduces them.

**[00:02:06 – 00:02:10]** So a couple of seconds each, and then we'll get into it.

**[00:02:10 – 00:02:13]** All right, my name is Jiapeng Chen, founder and CEO of Agile Robots.

**[00:02:13 – 00:02:17]** As you can tell, we're making general-purpose robots.

**[00:02:17 – 00:02:21]** We're trying to make body, brain, and application all

**[00:02:21 – 00:02:22]** holistically together.

**[00:02:22 – 00:02:24]** That's what we do.

**[00:02:24 – 00:02:27]** So I'm Andrei; I'm one of the founders of Dexory.

**[00:02:27 – 00:02:29]** We are focusing on specialized robotics applications, so

**[00:02:29 – 00:02:33]** autonomous robots to digitize warehouses, and the intelligence

**[00:02:33 – 00:02:36]** to enable warehouses to operate autonomously.

**[00:02:36 – 00:02:38]** Hey everybody, my name is Sotirio Stasinopoulos.

**[00:02:38 – 00:02:41]** I'm the chief product officer at Humanoid, a UK-based company,

**[00:02:41 – 00:02:44]** but spanning across Europe and US.

**[00:02:44 – 00:02:47]** Our focus is in industrial applications, so although

**[00:02:47 – 00:02:49]** we have, let's say, our wheeled and bipedal platform, we're mainly

**[00:02:49 – 00:02:53]** focusing on mass commercializing the wheeled platform.

**[00:02:53 – 00:02:56]** Hi Marc, I'm CTO and co-founder of Sereact, and we're building

**[00:02:56 – 00:02:58]** Cortex, our robot brain.

**[00:02:58 – 00:03:01]** It's a world action model that we deploy to single arm

**[00:03:01 – 00:03:05]** industrial robots to do, like, bimanual stations and also

**[00:03:05 – 00:03:09]** humanoids, and we're doing this in warehouses and stores.

**[00:03:10 – 00:03:11]** So hi, I'm Christian Piechnick.

**[00:03:11 – 00:03:14]** I'm CEO and co-founder of Wandelbots, and we're building

**[00:03:14 – 00:03:17]** a hardware agnostic software layer for all sorts of robots across

**[00:03:17 – 00:03:22]** different embodiments, kinematics, and basically LED AI access robots.

**[00:03:22 – 00:03:26]** Amazing, thank you. We were just chatting before, and Christian

**[00:03:26 – 00:03:29]** actually called themselves the robot mafia because in the robot

**[00:03:29 – 00:03:33]** industry they all go way back to the era of when they were students.

**[00:03:33 – 00:03:37]** They all worked on PhDs together, and so hopefully they'll be

**[00:03:37 – 00:03:41]** a little bit of a spicy topic, you know, spiciness today.

**[00:03:41 – 00:03:43]** Well, anyway, I guess the number one thing that we always hear

**[00:03:43 – 00:03:48]** at NVIDIA from all of our ecosystem partners, from investors, from VCs,

**[00:03:48 – 00:03:52]** from customers is, have we reached the Chad GPT moment of robotics?

**[00:03:52 – 00:03:55]** Are these just demos, or is it real?

**[00:03:55 – 00:03:57]** So we'd like to know from your point of view.

**[00:03:57 – 00:03:59]** Maybe starting with Andre, since you've been in

**[00:03:59 – 00:04:00]** deployment the longest.

**[00:04:00 – 00:04:03]** Is it demos, or is it real?

**[00:04:03 – 00:04:04]** It's definitely real.

**[00:04:04 – 00:04:07]** I think there's still a lot of hype that you get in the market, but

**[00:04:07 – 00:04:12]** we have hundreds of these deployed across the world, which could be a

**[00:04:12 – 00:04:15]** point to brag about, but it's also a massive responsibility to make

**[00:04:15 – 00:04:17]** sure they keep delivering value.

**[00:04:17 – 00:04:21]** Customers' expectations are, I would say, past demo, past

**[00:04:21 – 00:04:23]** what they're seeing in movies.

**[00:04:23 – 00:04:25]** So when you get a robot, everyone wants it to run 24-7, everyone

**[00:04:25 – 00:04:28]** wants the reliability, robustness.

**[00:04:28 – 00:04:31]** So I think we're not quite at that inflection point,

**[00:04:31 – 00:04:33]** but it's getting closer.

**[00:04:33 – 00:04:37]** It's kind of like gathering that physical data layer;

**[00:04:37 – 00:04:41]** it's gathering the momentum to actually tip the balance.

**[00:04:41 – 00:04:43]** And we have, you know, two humanoid builders here.

**[00:04:43 – 00:04:45]** You know, are humanoids demos?

**[00:04:45 – 00:04:47]** Are they real?

**[00:04:47 – 00:04:50]** It's definitely real, I 100% agree.

**[00:04:50 – 00:04:55]** What we see right now is that we already passed the threshold.

**[00:04:55 – 00:04:58]** Our Like you just mentioned, that GPT moment.

**[00:04:58 – 00:05:00]** But of course, on the other hand, even for GPT, there

**[00:05:00 – 00:05:02]** are one, two, GPT three.

**[00:05:02 – 00:05:07]** So this flywheel of AI, I think we see, is running.

**[00:05:07 – 00:05:11]** And when that is running, I think it's unstoppable.

**[00:05:11 – 00:05:13]** Yeah, so I guess the question goes to Mia.

**[00:05:13 – 00:05:17]** So the idea is that in Humanoid, we separate different technology

**[00:05:17 – 00:05:19]** readiness levels into what we do.

**[00:05:19 – 00:05:22]** So we do have, let's say, internal MVPs; we have demos

**[00:05:22 – 00:05:26]** externally; we have POC, let's say proof of concept level;

**[00:05:26 – 00:05:28]** and we have a pilot, and then, let's say, mass deployment.

**[00:05:29 – 00:05:31]** For us, for a lot of the core use cases, we've already passed

**[00:05:31 – 00:05:33]** the stage of POCs.

**[00:05:33 – 00:05:35]** We did that last year, and now we're entering the pilot

**[00:05:35 – 00:05:38]** stage, which is at least 16 hours a day of continuous operations.

**[00:05:38 – 00:05:41]** So the idea is that we're at that stage now.

**[00:05:41 – 00:05:45]** Definitely not a demo stage anymore, almost, let's say,

**[00:05:45 – 00:05:47]** the time for continuous operations in customer facilities.

**[00:05:47 – 00:05:51]** And for that reason, of course, hardware has to be there.

**[00:05:51 – 00:05:53]** But AI is the one that has unlocked this potential.

**[00:05:53 – 00:05:57]** So for now, we're definitely very excited to be in the stage.

**[00:05:57 – 00:05:59]** Yeah, and both of you are kind of vertically integrated.

**[00:05:59 – 00:06:00]** You're building the brain and the bodies.

**[00:06:00 – 00:06:04]** And so what about from the brain's perspective, from Ciriact, Mark?

**[00:06:04 – 00:06:07]** So, first of all, I'm also happy that I can hand over the building

**[00:06:08 – 00:06:10]** the hardware to my colleagues here.

**[00:06:10 – 00:06:14]** I think building the brain is quite hard already.

**[00:06:14 – 00:06:19]** We're deploying single arm stations and bimanual stations to different

**[00:06:19 – 00:06:21]** use cases of different complexity.

**[00:06:21 – 00:06:25]** And how we tackle this is, we always go out there to the

**[00:06:25 – 00:06:29]** customers, we find the challenges, and then we solve challenges, and

**[00:06:29 – 00:06:34]** then we go from simpler tasks or even simpler items to more complex

**[00:06:34 – 00:06:38]** items, which is what we're doing in return handling, for example.

**[00:06:38 – 00:06:41]** We can do shoes, we can do shirts, we can do jeans.

**[00:06:41 – 00:06:44]** We will later do big winter jackets, and that's how

**[00:06:44 – 00:06:47]** we can scale and generalize our model while still

**[00:06:47 – 00:06:50]** bringing customer value.

**[00:06:50 – 00:06:52]** And Christian, you've always had this vision, I like what

**[00:06:52 – 00:06:56]** you said a couple of years ago to me, that every company

**[00:06:56 – 00:06:58]** will have a zoo of robots.

**[00:06:58 – 00:07:01]** There's all these heterogeneous robot deployments, and you

**[00:07:02 – 00:07:04]** sit at a really interesting layer that's connecting kind

**[00:07:04 – 00:07:09]** of the traditional industrial programmed robots to this kind of

**[00:07:09 – 00:07:12]** new general purpose era of robots.

**[00:07:12 – 00:07:13]** And so what about you?

**[00:07:13 – 00:07:14]** What do you think?

**[00:07:14 – 00:07:17]** Yeah, so then it's my job to bring the spiciness in.

**[00:07:17 – 00:07:18]** So I partially disagree.

**[00:07:18 – 00:07:22]** So when we look at, and it depends on what we are speaking

**[00:07:22 – 00:07:26]** about, I fully agree that for, like, the low-end simple

**[00:07:26 – 00:07:29]** use cases, definitely we have passed the point, and we can

**[00:07:30 – 00:07:32]** apply the technology today.

**[00:07:32 – 00:07:35]** However, the vast majority of things we see on LinkedIn

**[00:07:35 – 00:07:38]** and YouTube and stuff, it appears that I now can buy

**[00:07:38 – 00:07:42]** a humanoid robot, get it out of the box, turn it on, and then

**[00:07:42 – 00:07:45]** it magically starts doing stuff.

**[00:07:45 – 00:07:47]** It will take a couple of years still, from my assumption,

**[00:07:47 – 00:07:49]** until we are fully there.

**[00:07:49 – 00:07:52]** But we are getting there, and that's the important point.

**[00:07:52 – 00:07:55]** If we are looking at more conventional robot arms, so

**[00:07:55 – 00:07:57]** applying AI technologies, especially from industrial

**[00:07:57 – 00:08:02]** metaverse at body shops of large car manufacturers, to optimize

**[00:08:02 – 00:08:05]** cycle time, to basically... Increase the versatility of

**[00:08:05 – 00:08:06]** robots, we are there today.

**[00:08:06 – 00:08:09]** So that's, I think, it works.

**[00:08:09 – 00:08:12]** And just now it's a question of industrialization and

**[00:08:12 – 00:08:14]** scaling the technology.

**[00:08:14 – 00:08:17]** Yeah, and so I guess that kind of brings us to our next question.

**[00:08:17 – 00:08:19]** So what really is working today?

**[00:08:19 – 00:08:22]** What do you guys have running in customer operations and

**[00:08:22 – 00:08:24]** real production environments today to kind of give people

**[00:08:24 – 00:08:26]** an idea of what is possible?

**[00:08:26 – 00:08:28]** And maybe you start again since you said no.

**[00:08:28 – 00:08:32]** Yeah, so we differentiate today basically in three main

**[00:08:32 – 00:08:34]** pillars where we're driving.

**[00:08:34 – 00:08:39]** In brownfield, it's all about optimization of what you already

**[00:08:39 – 00:08:42]** have, and that's mainly a software update most of the time.

**[00:08:42 – 00:08:46]** So improving cycle time of existing production machinery, reducing

**[00:08:46 – 00:08:49]** downtime, that's basically...

**[00:08:49 – 00:08:52]** The process is creating a digital twin, then shoot with

**[00:08:52 – 00:08:58]** AI and engineering on digital twin in terms of optimization, and then,

**[00:08:58 – 00:09:03]** once you have a digital twin, which is precisely acting as the real

**[00:09:03 – 00:09:08]** environment, you can get to results where, with a week of work, in

**[00:09:08 – 00:09:12]** extreme cases, we doubled basically the production output of a

**[00:09:12 – 00:09:16]** machine without changing a screw of customers and automotive suppliers.

**[00:09:16 – 00:09:18]** And in Greenfield, it's very much about reducing the total

**[00:09:18 – 00:09:20]** cost of ownership.

**[00:09:20 – 00:09:23]** Having now all the capabilities on precise digital twins to

**[00:09:23 – 00:09:27]** applying AI models to real manufacturing robots, I think the

**[00:09:27 – 00:09:32]** ROI on new production equipment now goes down below 24 months, close

**[00:09:32 – 00:09:36]** to 16, sometimes even 12 months of ROI of new production equipment.

**[00:09:36 – 00:09:39]** That's a no-brainer for almost all of the industries today.

**[00:09:39 – 00:09:42]** And that hasn't been possible in the past, so we are working

**[00:09:42 – 00:09:45]** with the large automotive OEMs, automotive suppliers,

**[00:09:45 – 00:09:49]** but recently also tapping new types of industries.

**[00:09:49 – 00:09:53]** And this is, for instance, chemical and pharma, and we've

**[00:09:53 – 00:09:56]** been very successfully now start to deploy robots in,

**[00:09:56 – 00:09:57]** for instance, lab automation.

**[00:09:57 – 00:10:01]** And there, it's more about versatility of robots, so

**[00:10:01 – 00:10:04]** doing many different tasks and being extremely fast and

**[00:10:04 – 00:10:07]** efficient in kind of teaching new capabilities, which, at the end,

**[00:10:07 – 00:10:12]** then, in industry, we also have this 99.9% of up-time availability.

**[00:10:12 – 00:10:15]** So that's the challenge, and I think that's what we are

**[00:10:15 – 00:10:17]** currently seeing.

**[00:10:17 – 00:10:20]** And maybe Xiaopeng next because Agile builds, well, you have

**[00:10:20 – 00:10:24]** many, many, many companies in your company, but from Agile robots

**[00:10:24 – 00:10:26]** only, you have many different types of embodiments, and you're

**[00:10:26 – 00:10:29]** already deployed in industry, and so what are you seeing?

**[00:10:29 – 00:10:32]** What works now? What's a great customer story?

**[00:10:32 – 00:10:35]** Well, we have been deployed more than 30, 40 thousand in

**[00:10:35 – 00:10:41]** different factories, so what I can tell is that This actually depends

**[00:10:41 – 00:10:43]** on what customer wants to have.

**[00:10:43 – 00:10:46]** Most of our customers, because some of them wants to get

**[00:10:46 – 00:10:49]** their phone produced, some of them want to get their servers

**[00:10:49 – 00:10:52]** to be produced, some of them want to get their car produced.

**[00:10:52 – 00:10:57]** So eventually, what the customer cares about is a couple of

**[00:10:57 – 00:11:01]** things, productivity, throughput, year rates, and whether you can

**[00:11:01 – 00:11:03]** reduce the cost of the production.

**[00:11:03 – 00:11:07]** So as long as you can fulfill those requirements, actually

**[00:11:07 – 00:11:11]** what they care more is about The ultimate outcome instead

**[00:11:11 – 00:11:16]** of individual sort of individual layers of accomplishment.

**[00:11:16 – 00:11:20]** So what I was trying to say is that basically, if you,

**[00:11:20 – 00:11:24]** if you view the whole production or logistics warehouses or

**[00:11:24 – 00:11:30]** at home or offices as one gigantic sort of picking scenarios, then

**[00:11:30 – 00:11:32]** whole factory is basically a robot.

**[00:11:32 – 00:11:35]** Then, which part is AI driven, which part is probably

**[00:11:35 – 00:11:38]** deterministic programming driven?

**[00:11:38 – 00:11:44]** That's the part which we are seeing gradually driven from automation,

**[00:11:44 – 00:11:45]** actually, to AI-driven part.

**[00:11:45 – 00:11:48]** That's why I mentioned that actually we see that GP moment

**[00:11:48 – 00:11:51]** is already happening; it's just not 100% replaced.

**[00:11:51 – 00:11:54]** I would not say, you know, it magically just snap a finger,

**[00:11:54 – 00:11:57]** then you got a magically, you know, all the factories

**[00:11:57 – 00:11:59]** are automatically AI driven.

**[00:11:59 – 00:12:02]** That's not going to happen. But we see where many things

**[00:12:02 – 00:12:06]** are available right now already with VLA, with World Action

**[00:12:06 – 00:12:10]** Model, together with certain reinforcement learning, we can,

**[00:12:10 – 00:12:13]** we see already, like, close to 100% of a successful rate with a word

**[00:12:13 – 00:12:17]** action model, working reinforcement learning with specific tasks,

**[00:12:17 – 00:12:21]** but whether this word action model ways in reinforcement learning are

**[00:12:21 – 00:12:25]** working for the whole process, I would say no, because at a certain

**[00:12:25 – 00:12:29]** point, people will just say, well, the action model probably is too

**[00:12:29 – 00:12:32]** slow, probably the tick time is not enough, probably a six-fifth rate

**[00:12:32 – 00:12:33]** for some plastics is not enough.

**[00:12:33 – 00:12:36]** So that's something, I think, in the industry that matters

**[00:12:36 – 00:12:39]** in manufacturing or logistics, that's still improving.

**[00:12:39 – 00:12:42]** But we see this is actually the trend; as long as the

**[00:12:42 – 00:12:47]** wheel is running, we see the data is coming back, we see the

**[00:12:47 – 00:12:51]** model training, and the growth of the intelligence level of the model

**[00:12:51 – 00:12:53]** is actually dramatically increased.

**[00:12:53 – 00:12:58]** I see Andre nodding vigorously, or sometimes, so what are

**[00:12:58 – 00:13:02]** some of Dextry's, what works today for Dextry in a real-world

**[00:13:02 – 00:13:05]** production environment?

**[00:13:05 – 00:13:09]** One of the big expectations that customers have when we started

**[00:13:09 – 00:13:13]** deploying these robots is the technology, is that it continuously

**[00:13:13 – 00:13:17]** scans and gives them this real-time visibility over the warehouses.

**[00:13:17 – 00:13:20]** So we operate in logistics; we focus on warehousing.

**[00:13:20 – 00:13:22]** These are big spaces.

**[00:13:22 – 00:13:25]** We're talking a million, 1.5, 2 million square feet.

**[00:13:25 – 00:13:28]** To be able to collect that information, to be able to make the

**[00:13:28 – 00:13:32]** data available, to then be able to provide value to customers, that's

**[00:13:32 – 00:13:34]** something that we do very well.

**[00:13:34 – 00:13:38]** That's kind of why customers decide to work with us.

**[00:13:38 – 00:13:41]** But I think when you look at this at scale, when you

**[00:13:41 – 00:13:45]** look at deploying this technology across the world, that's something

**[00:13:45 – 00:13:48]** that works well, but also that's something that people expect.

**[00:13:48 – 00:13:52]** So people, and what we're seeing in the industry, they'll

**[00:13:52 – 00:13:56]** look at specific facilities, but then they want to have a very

**[00:13:56 – 00:14:00]** seamless deployment and upscaling of the capabilities that you

**[00:14:00 – 00:14:02]** have from one facility to another.

**[00:14:02 – 00:14:06]** So it's really difficult to make an argument for, oh, we

**[00:14:07 – 00:14:09]** are spending a lot of engineering time in facility one, and

**[00:14:09 – 00:14:11]** then we will spend the same amount of engineering time for

**[00:14:11 – 00:14:16]** the next 10, 15, 20, 100 factories or warehouses or whatever you have.

**[00:14:17 – 00:14:21]** So I think that's working well because it's scaling rapidly.

**[00:14:21 – 00:14:24]** And what we're really excited about is starting to build this

**[00:14:24 – 00:14:26]** context around the decision-making.

**[00:14:27 – 00:14:29]** So you can scan an environment, you can scan a factory, you

**[00:14:29 – 00:14:32]** can scan a warehouse, you can take robot-specific actions

**[00:14:32 – 00:14:36]** that are really impactful, but bringing that context

**[00:14:36 – 00:14:39]** and bringing the ability to close the loop and say, okay, I've done

**[00:14:39 – 00:14:44]** a certain set of actions, and then I've driven a certain outcome into

**[00:14:44 – 00:14:47]** the customer's operation, and then I can verify this, and I can learn

**[00:14:47 – 00:14:49]** how to do it better next time.

**[00:14:49 – 00:14:52]** That's really the bit that's going to completely catalyze the industry

**[00:14:52 – 00:14:54]** because that's the context that...

**[00:14:54 – 00:14:58]** It's reduced scale at the moment, but it's also where

**[00:14:58 – 00:15:03]** AI comes into play and where it's accelerating that knowledge

**[00:15:03 – 00:15:06]** that you bring in the organization and that the technology learns

**[00:15:06 – 00:15:08]** from the organization.

**[00:15:08 – 00:15:09]** That's a really good point.

**[00:15:09 – 00:15:12]** I think Jensen yesterday in his keynote, one of his slides

**[00:15:12 – 00:15:16]** says that useful AI has arrived, but that he's primarily talking

**[00:15:16 – 00:15:18]** about digital AI.

**[00:15:18 – 00:15:22]** And so we're wondering if useful AI robotics has arrived, and that's

**[00:15:22 – 00:15:24]** what I'm hearing from you is a yes.

**[00:15:24 – 00:15:28]** It's not necessarily a full-scale production, but useful AI

**[00:15:28 – 00:15:29]** robotics has arrived.

**[00:15:29 – 00:15:32]** And Humanoid is in Scheffler.

**[00:15:32 – 00:15:34]** You guys, I think there was an article, you guys operate

**[00:15:34 – 00:15:37]** your bimanual manipulator Humanoid mobile bases for

**[00:15:37 – 00:15:41]** 16 hours a day or longer.

**[00:15:41 – 00:15:44]** What tasks are working, what jobs are working today?

**[00:15:44 – 00:15:47]** Yeah, so for us, of course, we have service and household

**[00:15:47 – 00:15:50]** applications in our roadmap, but that's more mid to long term.

**[00:15:50 – 00:15:52]** For now, of course, industrial applications are the

**[00:15:52 – 00:15:53]** most mature ones.

**[00:15:53 – 00:15:57]** So we're talking retail, logistics, manufacturing, energy sector.

**[00:15:57 – 00:16:00]** And for this reason, we actually went very extensively and

**[00:16:00 – 00:16:04]** identified use cases from more than 150, mostly Fortune 100 companies.

**[00:16:04 – 00:16:06]** So a lot of, like, hundreds of thousands of employees.

**[00:16:06 – 00:16:09]** And we have identified four different phases for us that, let's

**[00:16:09 – 00:16:13]** say, Separate use cases in terms of the technology readiness level that

**[00:16:13 – 00:16:14]** they require and their complexity.

**[00:16:14 – 00:16:17]** So we have maybe, in the first phase, applications

**[00:16:17 – 00:16:19]** that have to do with picking, kitting, sortation, let's

**[00:16:19 – 00:16:23]** say, palletizing, depalletizing, moving parts, boxes, totes.

**[00:16:23 – 00:16:26]** So this is the first big category that we started working on

**[00:16:26 – 00:16:28]** last year, and we expect to have very close to human level

**[00:16:28 – 00:16:30]** performance by the end of this year, early next year.

**[00:16:30 – 00:16:33]** This is one of the use cases that we're actually deploying

**[00:16:33 – 00:16:35]** at the Schaeffler facility that you mentioned.

**[00:16:35 – 00:16:38]** So this is already kind of working at this stage, and we're

**[00:16:38 – 00:16:40]** taking that into the pilot stage.

**[00:16:40 – 00:16:42]** Next phase for us is more about dynamic objects that

**[00:16:42 – 00:16:46]** move on conveyors, AMRs, let's say, machine feeding operations,

**[00:16:46 – 00:16:48]** greater accuracy of placement.

**[00:16:48 – 00:16:51]** We have as well some of the packaging operations, machine,

**[00:16:51 – 00:16:52]** open machine machinery operations.

**[00:16:52 – 00:16:54]** So all of these are a little bit higher complexity.

**[00:16:54 – 00:16:56]** We're starting to tackle them now, and we're seeing already

**[00:16:56 – 00:16:57]** promising results.

**[00:16:57 – 00:16:59]** And then from next year, we're probably going to start some

**[00:16:59 – 00:17:01]** applications that are a little bit more complex that have

**[00:17:01 – 00:17:06]** to do with, let's say, soft objects, wires, let's say,

**[00:17:06 – 00:17:07]** tool operation, assembly tasks.

**[00:17:07 – 00:17:10]** So these are a little bit further down the line, but

**[00:17:10 – 00:17:13]** we're already seeing some proof-of-concept projects

**[00:17:13 – 00:17:15]** working quite well for that.

**[00:17:15 – 00:17:19]** Excellent. And Mark, you're in a unique position because not only

**[00:17:19 – 00:17:22]** do you not suffer from the same thing that these guys do, which

**[00:17:22 – 00:17:26]** is it could be any environment, any task, you also have the problem

**[00:17:26 – 00:17:28]** of it's any body, any embodiment.

**[00:17:28 – 00:17:31]** And so what are you seeing that works today for your

**[00:17:31 – 00:17:33]** cross-embodiment brain?

**[00:17:33 – 00:17:36]** So what works well is, of course, industrial arms.

**[00:17:36 – 00:17:38]** They've been deployed for 30 years, 40 years.

**[00:17:38 – 00:17:43]** It works well, and we added the new AI technologies to them.

**[00:17:43 – 00:17:47]** What also works well is biomanual stations with cobots.

**[00:17:47 – 00:17:50]** That's pretty good and pretty robust.

**[00:17:51 – 00:17:55]** Last year, I would have said humanoids will be tough to

**[00:17:55 – 00:17:58]** scale because our humanoids we had last year, we bought

**[00:17:58 – 00:18:01]** them last year, they failed after 30 minutes, overheating,

**[00:18:01 – 00:18:03]** falling over, even on a demo.

**[00:18:03 – 00:18:07]** But this year, I would say this has changed dramatically.

**[00:18:07 – 00:18:11]** We can run, we're testing, of course, many different

**[00:18:11 – 00:18:14]** embodiments, many different humanoids we can buy, and we

**[00:18:14 – 00:18:18]** can run easy 16 hours of operations on a humanoid these days.

**[00:18:18 – 00:18:22]** Incredible. I saw Christian nodding vigorously with you because of the

**[00:18:22 – 00:18:23]** industrial arms, so that's good.

**[00:18:23 – 00:18:25]** You're in agreement this time.

**[00:18:25 – 00:18:30]** Yeah, so it's on the one side, so technology is advancing

**[00:18:30 – 00:18:31]** fast, I fully agree.

**[00:18:31 – 00:18:35]** That's why I said we're not there today, so the picture

**[00:18:35 – 00:18:38]** people have in their minds are also the expectation that was

**[00:18:38 – 00:18:40]** maybe set by the community early.

**[00:18:41 – 00:18:45]** AI was higher than we all could deliver at the time, but

**[00:18:45 – 00:18:46]** now everything's coming together.

**[00:18:46 – 00:18:50]** So there's still like some hurdles to be taken, but I

**[00:18:50 – 00:18:52]** also expect this to be live.

**[00:18:52 – 00:18:55]** But I fully agree, like with industrial robot arms,

**[00:18:55 – 00:19:00]** it's basically, it's proven technology, it's rock solid.

**[00:19:00 – 00:19:02]** The way how those things were programmed for decades is

**[00:19:02 – 00:19:08]** ridiculous, so, like useful AI.

**[00:19:08 – 00:19:10]** AI is a tool being applied to robots.

**[00:19:10 – 00:19:14]** You can equip them with better perception.

**[00:19:14 – 00:19:19]** AI can solve those harder algorithmical problems much more

**[00:19:19 – 00:19:21]** efficient than any engineer can do.

**[00:19:21 – 00:19:24]** And this is where those models can really be applied with

**[00:19:24 – 00:19:25]** a high value for the customer.

**[00:19:25 – 00:19:29]** But at the end it all boils down to a positive business case.

**[00:19:30 – 00:19:33]** And what I really liked is a couple of you mentioned it across

**[00:19:33 – 00:19:36]** the way, is kind of how we got here or the different approaches.

**[00:19:36 – 00:19:42]** You know, Zhao Peng, you mentioned VLAs, Mark, you're working on

**[00:19:42 – 00:19:46]** a brain, a world action model, and so, you know, this, I remember I

**[00:19:46 – 00:19:50]** attended CORAL, the robot learning conference, and the moderator asked

**[00:19:50 – 00:19:53]** the question, the first question was, what's better, classical

**[00:19:53 – 00:19:55]** methods or learning-based robotics?

**[00:19:55 – 00:19:58]** And the whole audience was like, oh, no.

**[00:19:58 – 00:19:59]** You know, we're going to have this big battle.

**[00:19:59 – 00:20:02]** And so, you know, what are you guys seeing?

**[00:20:02 – 00:20:04]** There's so much has changed, especially with the introduction

**[00:20:04 – 00:20:07]** of Astra, you know, and Opus 5.5, you know, the whole world

**[00:20:07 – 00:20:09]** is changing in AI all the time.

**[00:20:09 – 00:20:10]** All the methods are changing.

**[00:20:10 – 00:20:14]** All the discoveries are, there's something new every day.

**[00:20:14 – 00:20:16]** And so, you know, what are the different approaches that

**[00:20:17 – 00:20:21]** you guys started with, and what are you guys now trying to explore?

**[00:20:21 – 00:20:23]** You know, maybe, Xiaoping, you mentioned a couple.

**[00:20:23 – 00:20:27]** Yeah, we basically experienced the whole, I think Christian

**[00:20:28 – 00:20:31]** and the other guys probably very similar trajectory in the past

**[00:20:31 – 00:20:36]** 10 years, more than 10 years, about the whole development of robotics.

**[00:20:36 – 00:20:41]** So from our experiences, we started with all those in

**[00:20:42 – 00:20:46]** the past, let's say, technologies like reinforcement learning,

**[00:20:46 – 00:20:48]** deep reinforcement learning, we tried many things.

**[00:20:48 – 00:20:51]** My PhD was making robotic hands, and then we're trying

**[00:20:51 – 00:20:54]** to make the work crazy.

**[00:20:55 – 00:21:01]** And when we change the industry, the starting point obviously is not

**[00:21:01 – 00:21:06]** only deterministic programming, it is based on multi-sensory systems,

**[00:21:06 – 00:21:12]** what we would call talk and force tactile vision, multi-sensory

**[00:21:12 – 00:21:15]** feedback, multi-sensory fusion-based approaches.

**[00:21:15 – 00:21:20]** And then gradually, thanks to the breakthrough of the transformer

**[00:21:20 – 00:21:24]** and the foundation model, then we started a VIA, we tried that.

**[00:21:24 – 00:21:28]** And then we turned to, because the generalization of our

**[00:21:28 – 00:21:31]** VIA is not good enough, then we turned to word-action model,

**[00:21:32 – 00:21:33]** which we are working on right now.

**[00:21:33 – 00:21:36]** And when we found out actually word-action model probably

**[00:21:36 – 00:21:41]** sometimes is not robust enough, then we start trying to Compensate

**[00:21:41 – 00:21:44]** all those flaws or, let's say, drawbacks of the world action model

**[00:21:44 – 00:21:46]** and with reinforcement learning.

**[00:21:46 – 00:21:49]** But then reinforcement learning, it's a circle, right?

**[00:21:49 – 00:21:51]** And then you started with reinforcement learning, and now you

**[00:21:51 – 00:21:55]** want to fulfill the last mile also again with reinforcement learning.

**[00:21:55 – 00:21:58]** But reinforcement learning is also the drawback of, you

**[00:21:58 – 00:22:00]** know, generalization part.

**[00:22:00 – 00:22:03]** Actually, I really like what you mentioned because of the

**[00:22:03 – 00:22:07]** recent progress, like in, for example, like GBD-Astra.

**[00:22:07 – 00:22:11]** So we also see a very exciting breakthrough in the domain,

**[00:22:11 – 00:22:13]** again, in the past bounces.

**[00:22:13 – 00:22:17]** Basically, as you can imagine, you can use agent to drive,

**[00:22:17 – 00:22:21]** to generate, to drive physical systems or simulations.

**[00:22:21 – 00:22:25]** You can basically generate an anonymous, unlimited environment

**[00:22:26 – 00:22:27]** to train the model.

**[00:22:27 – 00:22:31]** That means the traditional way of reinforcement, for example,

**[00:22:31 – 00:22:34]** is going to be alive again.

**[00:22:34 – 00:22:38]** We used to have to create all the simulation environment

**[00:22:38 – 00:22:41]** and try this reinforcement algorithm one after another.

**[00:22:41 – 00:22:45]** Now you can basically generate it and then use agent to basically

**[00:22:45 – 00:22:49]** drive that in the environment, and then go back and fit in all those

**[00:22:49 – 00:22:51]** results into the gigantic model.

**[00:22:51 – 00:22:56]** So we see this whole flywheel running pretty well, and then

**[00:22:56 – 00:22:59]** the next step will be basically scale deployment, and then

**[00:22:59 – 00:23:01]** we'll see the outcome.

**[00:23:01 – 00:23:04]** That's what I experienced, and that's why I also sincerely

**[00:23:04 – 00:23:06]** believe that GPT moment is coming.

**[00:23:06 – 00:23:10]** Amazing. And maybe Mark, what's Ciriac's view on

**[00:23:10 – 00:23:12]** this and your approach?

**[00:23:12 – 00:23:16]** So I think there are two things that have changed.

**[00:23:16 – 00:23:20]** One thing is that across use cases, across embodiments

**[00:23:20 – 00:23:23]** even, the code stays the same.

**[00:23:23 – 00:23:27]** Like, basically the whole code controlling the robot, feeding

**[00:23:27 – 00:23:30]** the outputs of the model into the robot, it all stays the same.

**[00:23:30 – 00:23:31]** And I think that's quite interesting.

**[00:23:31 – 00:23:36]** It changes how you define your organization, even.

**[00:23:37 – 00:23:39]** That's why we can tackle many use cases at once.

**[00:23:39 – 00:23:42]** Before, it was a robotic company has to focus on one single

**[00:23:42 – 00:23:44]** thing, nothing else.

**[00:23:44 – 00:23:50]** And now, this enables us to tackle many use cases with the same code.

**[00:23:50 – 00:23:53]** But I think the most fundamental change that came was, like,

**[00:23:53 – 00:23:55]** with these new imitation learning methods, doesn't matter if

**[00:23:55 – 00:23:57]** it's VLAs, world models.

**[00:23:58 – 00:24:03]** I think, like, for my life, I spent so much time designing

**[00:24:03 – 00:24:07]** control systems, like state machines, huge state machines,

**[00:24:07 – 00:24:09]** combining control with perception.

**[00:24:09 – 00:24:12]** And that's, like, most of the time I spend in robotics.

**[00:24:12 – 00:24:16]** And now this has been replaced by a single model call.

**[00:24:16 – 00:24:18]** And I think that's quite exciting.

**[00:24:18 – 00:24:22]** And you can Teach the robot recovery from failure modes

**[00:24:22 – 00:24:24]** just by showing it.

**[00:24:24 – 00:24:28]** And before, this would have been weeks of work to recover

**[00:24:28 – 00:24:29]** from a single failure mode.

**[00:24:29 – 00:24:32]** And I think that's quite exciting.

**[00:24:32 – 00:24:36]** And André, I can see you semi-nodding.

**[00:24:36 – 00:24:39]** So I think it's a little bit different based on the type

**[00:24:39 – 00:24:40]** of robot you have.

**[00:24:40 – 00:24:44]** So we also run a lot of arms in our production facility

**[00:24:44 – 00:24:44]** as a full-stack business.

**[00:24:45 – 00:24:48]** We manufacture these robots in-house so we can benefit from all

**[00:24:48 – 00:24:53]** these advances with our static arms with what production looks like.

**[00:24:53 – 00:24:55]** We obviously have an autonomous robot that goes around this

**[00:24:55 – 00:24:58]** environment, so I think, at the end of the day, it's really important

**[00:24:58 – 00:25:03]** to use what brings most value to the customer at the lowest cost.

**[00:25:03 – 00:25:06]** Because cost is a very important factor, you know, people deploy

**[00:25:06 – 00:25:09]** automation, and we were talking about ROI, we were talking

**[00:25:09 – 00:25:12]** about cost of ownership.

**[00:25:12 – 00:25:14]** It's fine, you can buy the hardware, but if the cost

**[00:25:15 – 00:25:18]** of ownership of running the models, or you have to install

**[00:25:18 – 00:25:22]** or run a large infrastructure just to be able to get benefit from

**[00:25:22 – 00:25:26]** the automation or the autonomous systems, it can become prohibitive,

**[00:25:26 – 00:25:29]** and you have to be also very mindful of how this scales.

**[00:25:29 – 00:25:32]** So in our case, for example, the robot will go around the

**[00:25:32 – 00:25:36]** warehouse, it will collect 40-50 terabytes of data an hour.

**[00:25:36 – 00:25:38]** There's no infrastructure that will enable you to push that

**[00:25:39 – 00:25:40]** data and process it somewhere else.

**[00:25:40 – 00:25:43]** So we do a lot of local edge processing, and we

**[00:25:43 – 00:25:45]** want to make sure that we keep only what's relevant

**[00:25:45 – 00:25:48]** to the customer, and that's the data that gets presented.

**[00:25:48 – 00:25:51]** On the flip side, when you have a huge autonomous robot,

**[00:25:51 – 00:25:53]** it's 18 meters tall.

**[00:25:53 – 00:25:56]** The way it navigates in this environment and the safety

**[00:25:56 – 00:25:58]** components and everything else, you have to have something

**[00:25:58 – 00:26:00]** that's deterministic, you have to have something that's

**[00:26:00 – 00:26:03]** very, very fast, not something that decides to suddenly try

**[00:26:03 – 00:26:07]** something new because it would be cool to take a different path.

**[00:26:07 – 00:26:09]** And also we're scanning, so we want to make sure the coverage

**[00:26:09 – 00:26:12]** is assured for the customer, so they get their complete

**[00:26:12 – 00:26:16]** visibility over their facility.

**[00:26:16 – 00:26:18]** But when you bring this data into the cloud, you

**[00:26:18 – 00:26:21]** then have access to a completely different suite of models.

**[00:26:21 – 00:26:24]** So you can use the latest Frontier model, you can use

**[00:26:24 – 00:26:28]** specialized smaller models to answer the questions.

**[00:26:28 – 00:26:31]** So it really, I know this is a fairly long and vague

**[00:26:31 – 00:26:34]** answer, but I think the important bit is that we have different

**[00:26:34 – 00:26:37]** tools for different value that we bring to our customers.

**[00:26:38 – 00:26:41]** And if it's going to be a small vision model that immediately

**[00:26:41 – 00:26:45]** recognizes the parts and enables the grasping, Use that, and

**[00:26:45 – 00:26:49]** then a much more complicated model that enables recovery from a

**[00:26:49 – 00:26:54]** failed state or a clash or anything like that, or unknown objects in

**[00:26:54 – 00:26:57]** your operating area for robot arm.

**[00:26:57 – 00:27:00]** And so it sounds like it's not a one model architecture

**[00:27:00 – 00:27:02]** fits all problem.

**[00:27:02 – 00:27:07]** And from Humanoid's perspective, this industry trend of, are

**[00:27:07 – 00:27:10]** you sticking with VLAs, are you moving to world action

**[00:27:10 – 00:27:12]** models, what's your view?

**[00:27:12 – 00:27:15]** Yeah, I mean, me personally, I've also seen like the whole,

**[00:27:15 – 00:27:16]** let's say, transition.

**[00:27:16 – 00:27:19]** I remember back in my PhD days as well, where CNNs and

**[00:27:19 – 00:27:21]** RNNs were the bomb and everybody was talking about them, and

**[00:27:21 – 00:27:23]** I was excited to use them.

**[00:27:23 – 00:27:26]** Now, of course, the world has changed.

**[00:27:26 – 00:27:28]** Humanoid was lucky enough to be established because

**[00:27:28 – 00:27:31]** we're a young company, about two years old, right after VLAs were

**[00:27:31 – 00:27:33]** actually, let's say, introduced to the market, so we went

**[00:27:33 – 00:27:35]** directly and used that framework.

**[00:27:35 – 00:27:38]** So we didn't have to pay the price of switching from a more

**[00:27:38 – 00:27:41]** traditional to this newer, let's say, imitation learning approach.

**[00:27:41 – 00:27:44]** So we're very big believers of the imitation learning in general.

**[00:27:44 – 00:27:47]** Now we have tried VLAs, we have tried world action models

**[00:27:47 – 00:27:51]** for us because, as André said, you need to be very much independent

**[00:27:51 – 00:27:54]** up to some degree when you deploy an industrial application.

**[00:27:54 – 00:27:56]** So we need to be able to run things on the edge.

**[00:27:56 – 00:28:00]** We don't have So good connectivity at all times, the bandwidth

**[00:28:00 – 00:28:01]** is not, let's say, guaranteed.

**[00:28:01 – 00:28:04]** So for these reasons, we're sticking more to VLAs in this

**[00:28:04 – 00:28:07]** perspective because with world action models, you do need, let's

**[00:28:07 – 00:28:11]** say, higher computational power, which we may not have on the edge.

**[00:28:11 – 00:28:13]** So although our next, let's say, platform is, let's say,

**[00:28:13 – 00:28:17]** Thor-powered, we definitely have to manage our resources correctly.

**[00:28:17 – 00:28:20]** So for this reason, we do believe that for bootstrapping a policy,

**[00:28:20 – 00:28:24]** the VLAs and doing an inference on this is very good already.

**[00:28:24 – 00:28:26]** And then we're building on top of that with reinforcement

**[00:28:26 – 00:28:30]** learning; the idea for us is that whatever way you use to bootstrap a

**[00:28:30 – 00:28:33]** policy just needs to be good enough to get you, like, to start this

**[00:28:33 – 00:28:36]** data flywheel and the reinforcement learning, let's say, path.

**[00:28:36 – 00:28:39]** So for that reason, we're sticking to what makes sense for us to work

**[00:28:39 – 00:28:41]** in these, the same environments without having to, let's say,

**[00:28:41 – 00:28:44]** depend on the network too much, so that's the main approach.

**[00:28:44 – 00:28:47]** And then, Christian, I can imagine that one day you're

**[00:28:47 – 00:28:51]** going to have a customer that has one robot or multiple robots of

**[00:28:51 – 00:28:54]** each one of these guys, and you're going to have to go integrate

**[00:28:54 – 00:28:56]** that into a factory brain.

**[00:28:56 – 00:29:01]** And so is it okay that each one has a different approach,

**[00:29:01 – 00:29:03]** a different model, different model architecture, different approach?

**[00:29:03 – 00:29:08]** Yeah, so I think the question, what's the best, is impossible

**[00:29:08 – 00:29:12]** to answer, especially in these times where things change so fast.

**[00:29:12 – 00:29:16]** So the answer anyone can give only is true for maybe

**[00:29:16 – 00:29:19]** the next 10 minutes.

**[00:29:20 – 00:29:24]** And we're seeing, and it all boils down to, first, is the

**[00:29:24 – 00:29:28]** application stably running with whatever approach, and how much?

**[00:29:28 – 00:29:30]** So then there are two core KPIs for us.

**[00:29:30 – 00:29:34]** So one is basically cost per task, so how much money do

**[00:29:34 – 00:29:37]** I need to spend in order to train an algorithm, program something

**[00:29:37 – 00:29:40]** statically, whatever approach I'm taking, and time to skill,

**[00:29:40 – 00:29:45]** so how long is it actually taking, and how good can I abstract what

**[00:29:45 – 00:29:48]** I've trained to other environments, to other eventually embodiments,

**[00:29:48 – 00:29:51]** to other, kind of also use cases?

**[00:29:51 – 00:29:55]** And what we are seeing is that there is no silver bullet

**[00:29:55 – 00:29:57]** solution, not even in approaches.

**[00:29:57 – 00:30:01]** So there are different VLAs, there are different approaches, also like

**[00:30:01 – 00:30:05]** we've seen reinforcement learning approaches like DeepMimic to

**[00:30:05 – 00:30:08]** perform very well in certain very specific types of applications.

**[00:30:08 – 00:30:12]** So at the end, in order to manage your robots and get

**[00:30:12 – 00:30:15]** robots being done, for us it's super for mission critical

**[00:30:15 – 00:30:20]** that customers get agnostic not only for the embodiment they are

**[00:30:20 – 00:30:24]** using, but also to the, so to say, robot brain which is underneath.

**[00:30:24 – 00:30:28]** And it's not one brain, it's a shared brain, and you basically

**[00:30:28 – 00:30:33]** want to choose the best of breed solution for a given specific task.

**[00:30:33 – 00:30:37]** And what we are still today seeing in many applications

**[00:30:37 – 00:30:43]** is that those two core KPIs, so cost per skill and cost

**[00:30:43 – 00:30:47]** per task and time to skill, are not in the right balance yet.

**[00:30:47 – 00:30:51]** So we are emerging there very quickly right now, so we are

**[00:30:51 – 00:30:55]** seeing now real applications in industry where you have

**[00:30:55 – 00:30:57]** this incredible ROI right away.

**[00:30:57 – 00:31:02]** And I think it's all of our job right now to kind of Move

**[00:31:02 – 00:31:06]** from research to industrialization, so to say, to really be able

**[00:31:06 – 00:31:12]** to scale that not only in single production lines, but

**[00:31:12 – 00:31:14]** across factories globally.

**[00:31:14 – 00:31:16]** And then on that point, you know, what, and a couple of you

**[00:31:16 – 00:31:20]** touched on it, and, you know, about we're not in full scale production

**[00:31:20 – 00:31:22]** or production scale at the moment.

**[00:31:22 – 00:31:25]** And so what do you, what do each of you think about what's

**[00:31:25 – 00:31:28]** the main block or the main challenge from us getting there?

**[00:31:28 – 00:31:31]** What will get you to the next click?

**[00:31:32 – 00:31:35]** Any of you to start?

**[00:31:35 – 00:31:37]** I'll take it for now, and then we'll see who else

**[00:31:37 – 00:31:37]** wants to add things.

**[00:31:37 – 00:31:41]** So for us, of course, we identified different, we'll say, challenges.

**[00:31:41 – 00:31:43]** I won't say that everything is blocked because, as we

**[00:31:43 – 00:31:45]** said, we think we're ready to start this, let's say,

**[00:31:46 – 00:31:47]** at least for the industrial sphere.

**[00:31:47 – 00:31:51]** So definitely reliable hardware that is safety certifiable

**[00:31:51 – 00:31:54]** is a challenge, especially when you need to get some certification for

**[00:31:54 – 00:31:56]** that to deploy it in facilities.

**[00:31:56 – 00:31:58]** That's why we're saying with, let's say, mobile robots,

**[00:31:58 – 00:31:59]** it's still a challenge.

**[00:31:59 – 00:32:02]** That's why also as a company we decided to go to the wheel-based

**[00:32:02 – 00:32:04]** approach first because there are some challenges with doing

**[00:32:04 – 00:32:05]** that with bipeds.

**[00:32:05 – 00:32:09]** But for us, we're seeing that as a key, let's say, first

**[00:32:09 – 00:32:11]** baseline that you need to have to start these deployments at scale.

**[00:32:12 – 00:32:15]** If you don't have hardware that can go up to these 16 hours or more a

**[00:32:15 – 00:32:18]** day, then you don't have a winning case, like the ROI is not there.

**[00:32:18 – 00:32:20]** So that's one thing for us.

**[00:32:20 – 00:32:22]** Of course, there are other types of automation, but myself,

**[00:32:22 – 00:32:24]** I'm talking more about the humanoids that are mobile

**[00:32:24 – 00:32:26]** platforms, so that's one thing.

**[00:32:26 – 00:32:29]** On top of this, you need to have an AI that can actually,

**[00:32:29 – 00:32:31]** let's say, improve itself.

**[00:32:31 – 00:32:33]** We did mention that any kind of, let's say, AI that you deploy

**[00:32:33 – 00:32:37]** now will be old news in a matter of, let's say, a few days or weeks.

**[00:32:37 – 00:32:39]** Ten minutes, apparently. I don't know about ten minutes,

**[00:32:39 – 00:32:40]** but yeah, definitely very soon.

**[00:32:40 – 00:32:43]** We're going to have to start the panel over.

**[00:32:43 – 00:32:46]** But the idea is that you do need to have AI that can upgrade itself,

**[00:32:46 – 00:32:49]** and that's why we're saying for us, after we deploy using reinforcement

**[00:32:49 – 00:32:53]** learning, like after you deploy, like real world reinforcement

**[00:32:53 – 00:32:56]** learning, as we call it, to, let's say, keep the robust learning on

**[00:32:56 – 00:33:00]** the task, to improve that, that's like a key thing to, let's say, get

**[00:33:00 – 00:33:03]** your robots to deal with edge cases that will only appear, like, during

**[00:33:03 – 00:33:05]** the actual application scenario.

**[00:33:05 – 00:33:08]** Because no matter how much Same you do in the beginning to bootstrap

**[00:33:08 – 00:33:10]** something, or no matter how much, let's say, you try to train

**[00:33:10 – 00:33:13]** you with your data, there's always going to be a new edge case that

**[00:33:13 – 00:33:16]** will appear only in the actual, say, scenario where you apply it.

**[00:33:16 – 00:33:19]** So you do need to have a way of integrating that

**[00:33:19 – 00:33:20]** data back into your model.

**[00:33:20 – 00:33:23]** So that's another thing, and on top of this, of course, but a lot

**[00:33:23 – 00:33:27]** of the companies, I mean us as this panel, things been covering it very

**[00:33:27 – 00:33:30]** well, but the actual application layer and how you integrate

**[00:33:30 – 00:33:33]** is very important because we're not talking anymore about, we're not

**[00:33:33 – 00:33:36]** talking about single robots doing single use cases, we're talking

**[00:33:36 – 00:33:39]** about fleets of robots that need to go in a facility and cover multiple

**[00:33:39 – 00:33:41]** workstations or multiple use cases.

**[00:33:41 – 00:33:44]** So in order to do that, you need to have a very good coordination

**[00:33:44 – 00:33:47]** system, not only among the robots themselves, but you

**[00:33:47 – 00:33:49]** also need to coordinate with other automation systems,

**[00:33:49 – 00:33:52]** so with AMRs, with the existing ASRS systems, with anything

**[00:33:52 – 00:33:53]** else that goes into the facility.

**[00:33:53 – 00:33:57]** So you do need to have a very good integration level for this part.

**[00:33:57 – 00:33:59]** And again, a lot of the companies here are talking about it.

**[00:33:59 – 00:34:03]** I know, like, they have experience in this, but most of the industry

**[00:34:03 – 00:34:06]** is not thinking about that yet because they're just thinking

**[00:34:06 – 00:34:09]** about, OK, let's just do one task, OK, and then one robot

**[00:34:09 – 00:34:11]** do it at the time, and then I'll see how to scale it.

**[00:34:11 – 00:34:14]** But we already are thinking about these things already.

**[00:34:14 – 00:34:17]** And, of course, supply chain needs to mature, but I think

**[00:34:17 – 00:34:17]** we're getting there.

**[00:34:17 – 00:34:20]** I think that's why we need to have, let's say, strategic partners

**[00:34:20 – 00:34:23]** to help us with the scaling, and NVIDIA definitely is one

**[00:34:23 – 00:34:27]** of those, but we have also in our, let's say, group, other strategic

**[00:34:27 – 00:34:30]** partners that help us, let's say, clear that, let's say, challenge.

**[00:34:30 – 00:34:32]** So, from my part, of course, there are a lot more, but

**[00:34:32 – 00:34:34]** I think I would focus on these.

**[00:34:34 – 00:34:37]** You don't need to traumatize us.

**[00:34:37 – 00:34:39]** Just give us the top one.

**[00:34:39 – 00:34:42]** I think if we go a little bit back into the supply chain,

**[00:34:42 – 00:34:47]** I think that's a very interesting point because it's not about

**[00:34:47 – 00:34:50]** the core component or generic component of the supply chain.

**[00:34:50 – 00:34:53]** It's the advanced sensors, the advanced compute, the

**[00:34:53 – 00:34:56]** latest technology in stereo cameras and LiDARs.

**[00:34:56 – 00:34:59]** People are just now starting to scale the production

**[00:34:59 – 00:35:01]** of these things up.

**[00:35:01 – 00:35:03]** You know, LiDARs, we've had them for 20 years, but they've

**[00:35:03 – 00:35:05]** been incredibly expensive.

**[00:35:05 – 00:35:08]** They've been manufactured in very, very small batches

**[00:35:08 – 00:35:09]** and small volumes.

**[00:35:09 – 00:35:12]** And now, to be able to push the autonomous systems to another

**[00:35:12 – 00:35:16]** level, suddenly we need to go into the millions of these units.

**[00:35:16 – 00:35:19]** And LiDAR as a sensor is a very complicated

**[00:35:19 – 00:35:20]** electromechanical system.

**[00:35:20 – 00:35:23]** It's got spinning components and so on and so forth.

**[00:35:23 – 00:35:26]** You need them to be able to scale, so you could have the

**[00:35:26 – 00:35:28]** best production facility in the world; if some of the

**[00:35:28 – 00:35:30]** core suppliers are not there.

**[00:35:30 – 00:35:34]** Battery technology is another good one, another good example.

**[00:35:34 – 00:35:39]** Again, you can't have infinite power consumption on the systems

**[00:35:39 – 00:35:42]** and then rely on battery technology that is 10 years old.

**[00:35:42 – 00:35:46]** So I think those are two of the bigger blockers.

**[00:35:46 – 00:35:47]** And we are in production.

**[00:35:47 – 00:35:50]** We are scaling this technology, deploying it across the world,

**[00:35:50 – 00:35:53]** and we're continuously working very closely with our manufacturers

**[00:35:53 – 00:35:58]** and our key providers to help give them feedback and help

**[00:35:58 – 00:36:00]** make their systems and their products better, because that ends

**[00:36:00 – 00:36:02]** up benefiting the entire industry.

**[00:36:02 – 00:36:04]** I think that's a very important point.

**[00:36:04 – 00:36:08]** So our experience is probably a little bit different, I would say.

**[00:36:08 – 00:36:12]** So I think the blocker is mainly in the environment.

**[00:36:12 – 00:36:15]** So, because if you really have like tens of thousands of robots

**[00:36:15 – 00:36:20]** working in a factory like us, then you realize the robots with the AI

**[00:36:20 – 00:36:24]** system in a factory or warehouses, they're not automatically

**[00:36:24 – 00:36:25]** walk in and then start working.

**[00:36:25 – 00:36:27]** You have to think about how to connect your system to

**[00:36:28 – 00:36:29]** their ERP systems.

**[00:36:29 – 00:36:30]** What about Mesh systems?

**[00:36:30 – 00:36:33]** What about everything underneath that?

**[00:36:33 – 00:36:36]** There are a bunch of Siemens PLCs, and how are you

**[00:36:36 – 00:36:37]** going to control that?

**[00:36:37 – 00:36:39]** There's nothing...

**[00:36:39 – 00:36:42]** The whole industry was created in the last 200 years, and

**[00:36:42 – 00:36:47]** we're basically trying to get a very bumpy new kid into

**[00:36:47 – 00:36:50]** a world with stable systems, and then people tell you what

**[00:36:50 – 00:36:52]** to do, what not to do.

**[00:36:52 – 00:36:54]** So this is something which we experience, and that's probably

**[00:36:54 – 00:36:58]** one of the biggest challenges, I would say, on the AI-driven robots.

**[00:36:59 – 00:37:04]** And so, in contrast, aside from what the colleagues were saying

**[00:37:04 – 00:37:07]** here, I'm not too worried about the supply chain and the other

**[00:37:07 – 00:37:12]** things, because if you look at robots from actuators, batteries,

**[00:37:12 – 00:37:15]** to compute, it's exactly the same.

**[00:37:15 – 00:37:18]** DRIVEOS, NVIDIA, Mesh Shading.

**[00:37:33 – 00:37:38]** So, I would say, I think the constraint is like how the

**[00:37:38 – 00:37:41]** partners, how the customers in the industry, they should open

**[00:37:41 – 00:37:46]** their mind, give a little more room for the new technology, embrace it;

**[00:37:46 – 00:37:49]** that's something which I would say.

**[00:37:49 – 00:37:52]** And Mark, what are your thoughts? We've had hardware; well, I saw

**[00:37:52 – 00:37:56]** it had four challenges, but we have hardware, there's supply chain,

**[00:37:56 – 00:37:58]** there's an integration level; you know, what do you think is going

**[00:37:58 – 00:38:00]** to be the one biggest challenge?

**[00:38:00 – 00:38:05]** So I think for us right now, we have a couple of products

**[00:38:05 – 00:38:08]** that work full-scale production, and of course procurement,

**[00:38:08 – 00:38:11]** supply chain, and all of these things are challenging.

**[00:38:11 – 00:38:18]** But I think we have models that can potentially do everything.

**[00:38:18 – 00:38:21]** And how you get there is by getting that data and getting

**[00:38:21 – 00:38:22]** the failure modes.

**[00:38:22 – 00:38:26]** And, like, historically, in robotics companies, you build

**[00:38:26 – 00:38:29]** a product once, that thing runs, then you hand it over to

**[00:38:29 – 00:38:33]** the deployment team, go to market team; they've been scaling it.

**[00:38:33 – 00:38:36]** You build something and technically it's done; you

**[00:38:36 – 00:38:38]** just need to add more data.

**[00:38:38 – 00:38:42]** So meaning, your product team will be a customer of the

**[00:38:42 – 00:38:46]** deployment operations, getting the data back, then you can iterate

**[00:38:46 – 00:38:50]** on the model again, and then you roll out a new version of it.

**[00:38:50 – 00:38:55]** I think that is quite a challenge to organize that well, because

**[00:38:55 – 00:38:57]** if you add the right data, you will end up with a robot

**[00:38:57 – 00:39:01]** that can do everything, but it's just about how to get

**[00:39:01 – 00:39:02]** that data and how to iterate fast.

**[00:39:02 – 00:39:04]** Mm-hmm, yeah, it makes sense.

**[00:39:04 – 00:39:09]** And I guess, so it says only one minute, but I'm gonna be greedy.

**[00:39:09 – 00:39:12]** I'm gonna take seven more minutes because I have two

**[00:39:12 – 00:39:14]** more questions for you guys.

**[00:39:14 – 00:39:17]** Number one, because we're here in Europe, and overall, we came

**[00:39:17 – 00:39:22]** to Europe, so we can show Europe what incredible assets and ideas

**[00:39:22 – 00:39:25]** and entrepreneurs you have and what incredible energy, industrial

**[00:39:25 – 00:39:27]** energy and AI energy that you have.

**[00:39:27 – 00:39:30]** And so, do you think, across the whole board, all of you started

**[00:39:30 – 00:39:35]** your companies in Europe, all of you are frontier labs, frontier

**[00:39:35 – 00:39:40]** startups in Europe, do you believe that Europe is uniquely positioned

**[00:39:40 – 00:39:41]** to lead industrial physical AI?

**[00:39:41 – 00:39:44]** Do you think it has a chance?

**[00:39:44 – 00:39:45]** Yeah, I do.

**[00:39:45 – 00:39:46]** I think we have it.

**[00:39:47 – 00:39:48]** I think there's a chance in Europe.

**[00:39:48 – 00:39:51]** Well, I have to say, in the past wave of AI, I think

**[00:39:51 – 00:39:53]** Europe is far behind.

**[00:39:53 – 00:39:56]** But I think physically, I think we have a chance.

**[00:39:56 – 00:40:01]** If you look at different territory on the planet, US, basically, they

**[00:40:01 – 00:40:04]** are based center of innovation, and if you can see, if you

**[00:40:04 – 00:40:06]** look at all the innovations in the past, they're all coming from

**[00:40:06 – 00:40:08]** Silicon Valley, from that land.

**[00:40:08 – 00:40:13]** China has very systematic industrial systems, full supply

**[00:40:13 – 00:40:16]** chain, really incredible speed.

**[00:40:16 – 00:40:21]** But Europe has a very special, unique stance here, I would say,

**[00:40:21 – 00:40:27]** because of 100 years of industry, profound soil of all those

**[00:40:27 – 00:40:29]** different manufacturing logistics.

**[00:40:29 – 00:40:33]** And I think that's very unique here, and especially for physical

**[00:40:33 – 00:40:39]** AI, because we're not only talking about data, compute, and software.

**[00:40:39 – 00:40:42]** We're talking about something which is coming from AI, but has to

**[00:40:42 – 00:40:44]** be realized in the physical world.

**[00:40:44 – 00:40:48]** That's something I think you'll stand a chance.

**[00:40:48 – 00:40:49]** Christian, what do you think?

**[00:40:49 – 00:40:51]** I feel like you have to say yes.

**[00:40:51 – 00:40:52]** Super simple to answer.

**[00:40:52 – 00:40:55]** So we are already like, do we have a chance?

**[00:40:56 – 00:40:57]** Of course, we are currently doing it.

**[00:40:57 – 00:41:02]** So you have now here a panel of five companies which are

**[00:41:02 – 00:41:04]** globally leading this topic.

**[00:41:04 – 00:41:09]** So, and I think, so Wonderbots is now, for instance, nine

**[00:41:09 – 00:41:12]** years old, so we're not so young, and when we founded the

**[00:41:12 – 00:41:15]** company, in every financing round, I was challenged by investors,

**[00:41:15 – 00:41:19]** like, what's about that Palo Alto-based company that was founded

**[00:41:19 – 00:41:21]** by ex-DeepMind people, and so on.

**[00:41:21 – 00:41:24]** They are all gone.

**[00:41:30 – 00:41:34]** The thing is, I think especially in robotics, we are today

**[00:41:34 – 00:41:38]** mainly speaking about industrial use cases and home cases later.

**[00:41:38 – 00:41:42]** There's more than just pure AI technology; it's a combination

**[00:41:42 – 00:41:45]** of bringing everything together so at the end it provides a value

**[00:41:45 – 00:41:46]** to the customer on the shop floor.

**[00:41:46 – 00:41:49]** And I think we can do this pretty well, so we have smart

**[00:41:49 – 00:41:52]** people here in Europe, we know the industry quite well, we

**[00:41:52 – 00:41:56]** are all young, driven people that want to do something different; I

**[00:41:56 – 00:42:00]** think we are already leading that.

**[00:42:00 – 00:42:02]** Incredible.

**[00:42:02 – 00:42:05]** Maybe I'll ask the last three of you the last question.

**[00:42:05 – 00:42:07]** You know, what are you most excited about?

**[00:42:07 – 00:42:09]** It could be technology, it could be an industry, it could

**[00:42:09 – 00:42:15]** be, you know, any sort of trend or application.

**[00:42:15 – 00:42:18]** What are you most excited about in your area of robotics in

**[00:42:18 – 00:42:20]** the next, in the next 10 minutes?

**[00:42:20 – 00:42:24]** No, I'm just kidding, in the next, let's say, 6 to 12 months?

**[00:42:26 – 00:42:30]** So from our side, there's probably two things.

**[00:42:30 – 00:42:34]** One is the customer excitement for the technology.

**[00:42:34 – 00:42:37]** I think, and again, sorry, we're not a very young company

**[00:42:37 – 00:42:41]** either, but doing this for a period of time, you now see real drive

**[00:42:41 – 00:42:45]** from the customers, so there's real momentum, and people are genuinely

**[00:42:45 – 00:42:49]** a lot more open-minded around deploying these new technologies

**[00:42:49 – 00:42:55]** and scaling, not just trying a POC or looking at a small test

**[00:42:55 – 00:42:58]** in a part of their facility, but really looking at how can we deploy

**[00:42:58 – 00:43:01]** this across all our facilities.

**[00:43:01 – 00:43:04]** And the other big thing is this combination of data and

**[00:43:04 – 00:43:09]** models and scaling physical deployments that enables you

**[00:43:09 – 00:43:13]** to to have an understanding of decision making and how you can

**[00:43:13 – 00:43:15]** create a much more robust system.

**[00:43:15 – 00:43:18]** I think AI really unlocks the ability to go very deep

**[00:43:18 – 00:43:21]** into your data sets, and and that obviously brings a lot

**[00:43:21 – 00:43:23]** of value to customers, but brings a lot of value to all of us as

**[00:43:23 – 00:43:27]** companies because we can make much much better products a lot faster.

**[00:43:27 – 00:43:30]** So those two things definitely in the next 10 minutes, absolutely.

**[00:43:30 – 00:43:33]** I like it. Top of excitement.

**[00:43:33 – 00:43:37]** For us, for sure, in the transition towards this, let's say, pilot

**[00:43:37 – 00:43:40]** era, because I said, like, past the POC era, we're now

**[00:43:41 – 00:43:44]** again, let's say, starting to deploy at, let's say, larger

**[00:43:44 – 00:43:46]** volumes, larger scales, and reaching human-level performance,

**[00:43:46 – 00:43:50]** so seeing that, let's say, fleets of robots in customer facilities,

**[00:43:50 – 00:43:51]** this is what we want to be doing.

**[00:43:51 – 00:43:53]** And of course traditional automation has done this,

**[00:43:53 – 00:43:56]** but I'm talking about like the new age of, let's say, humanoid robots

**[00:43:56 – 00:43:58]** doing that as well, so that's what I'm more excited about at Humanoid.

**[00:43:58 – 00:44:01]** But personally as well, since you already mentioned like

**[00:44:01 – 00:44:04]** Astra or the frontier models, let's say, helping us, let's

**[00:44:04 – 00:44:08]** say, speed up how we add more use cases to our, let's say, arsenal,

**[00:44:08 – 00:44:11]** for me that's a very big bet, because up to now I've had to spend

**[00:44:11 – 00:44:14]** a lot of times, let's say, with our engineering trying to get them,

**[00:44:14 – 00:44:15]** okay, I want this new use case.

**[00:44:15 – 00:44:17]** I.

**[00:44:17 – 00:44:19]** And they were like, okay, let's see when we'll have time for this.

**[00:44:19 – 00:44:20]** Now it's getting much, much faster.

**[00:44:20 – 00:44:22]** We can bootstrap policies much easier.

**[00:44:22 – 00:44:25]** So for me, adding new use cases on the fly is something

**[00:44:25 – 00:44:26]** I'm very excited about.

**[00:44:26 – 00:44:28]** Yeah. What about you, Marc?

**[00:44:28 – 00:44:30]** I think the same thing.

**[00:44:30 – 00:44:35]** One year ago, if you want to get any VLA to do something to 90% to

**[00:44:35 – 00:44:38]** 95%, you had to add a lot of data.

**[00:44:38 – 00:44:41]** And today, you might need half an hour of data collection,

**[00:44:41 – 00:44:44]** and then you're already at 90% of a new task.

**[00:44:44 – 00:44:47]** And I think that is going to continue.

**[00:44:47 – 00:44:50]** How the world, the whole world, produces the data we need

**[00:44:50 – 00:44:52]** is quite exciting.

**[00:44:52 – 00:44:56]** And I'm looking forward to just instruct our robots with

**[00:44:56 – 00:45:01]** language and have agents doing all the subtasks that are possible.

**[00:45:01 – 00:45:03]** You know, you guys can maybe take a vacation for once,

**[00:45:03 – 00:45:05]** you know, in robotics.

**[00:45:05 – 00:45:07]** You've been working on this for decades.

**[00:45:07 – 00:45:08]** What about you, Christian?

**[00:45:08 – 00:45:10]** So if I look back, what would really change in the last

**[00:45:10 – 00:45:16]** 12 months for us was this... Now we're at a stage of really

**[00:45:16 – 00:45:20]** scaling the technologies across whole production lines, factories,

**[00:45:20 – 00:45:21]** or even across factories.

**[00:45:21 – 00:45:25]** And currently the speed in which progress is happening...

**[00:45:25 – 00:45:30]** So we have 5 million robots around the globe today, roughly.

**[00:45:30 – 00:45:33]** So I think that number will easily tenfold, maybe not

**[00:45:33 – 00:45:36]** in the next six months, but...

**[00:45:36 – 00:45:39]** I would bet that that's happening in the next two to three years,

**[00:45:39 – 00:45:42]** a tenfold in terms of applications.

**[00:45:42 – 00:45:45]** And it's a pretty exciting place to be right now.

**[00:45:45 – 00:45:47]** And Zhaopeng?

**[00:45:47 – 00:45:50]** So what I would be most excited in the next six months is

**[00:45:50 – 00:45:54]** about scale deployment and RSI.

**[00:45:54 – 00:45:56]** So, because the two things in combination,

**[00:45:56 – 00:45:57]** the flywheel is running.

**[00:45:57 – 00:45:58]** Incredible.

**[00:45:58 – 00:46:01]** Well, I want to thank you guys all for being here today, and

**[00:46:01 – 00:46:05]** for all everyone in the audience, they each have at least, there's a

**[00:46:05 – 00:46:09]** talk, a session, another panel, and demo booths out there, so you guys

**[00:46:09 – 00:46:11]** got to go see their robots at work.

**[00:46:11 – 00:46:14]** So thank you so much for joining.

**[00:46:20 – 00:46:22]** Thank you.

**[00:46:27 – 00:46:29]** Right, hello, thank you.

**[00:46:29 – 00:46:32]** While we, when you guys exit, please exit through this door

**[00:46:33 – 00:46:36]** to the right, or you can stay for the next session, which

**[00:46:36 – 00:46:38]** we're doing on automotive.

**[00:46:38 – 00:46:38]** Thank you.