NVIDIA´s Safety Initiative / The unique Volkswagen-Siemens joint venture

This podcast is presented by NXAI, your partner for time series. foundation models and physical AI. Hello everybody! And welcome to a new episode of our Industrial AI Podcast.

my name's Huard Viva... and it's a pleasure to talk too. Peter Sieberg. good morning Good afternoon Robert.

dear listeners wherever you are Peter We had some technical issues at the beginning of recording. Now we try again, you are in sunny South Europe I think so. would you share some AI news with us? It will be a pleasure.

I'm going to share some sunny AI news with you. I don't know what it was. Officially, it was called successfully processing more than a hundred thousand parcels without any human intervention some kind of parcel sorting. someone suggested they were repositioning the packages so they could be scanned.

Now, what was most striking to me that everybody were sharing? Everybody seemed to be so impressed and that's okay. but nobody was asking What Was Happening?, What Were They Doing?

, I Was Asking All The Time What Are You Doing?, Nobody Was Telling Me, you know there were just moving passes from left-to. right now i'll come back To That In A Moment. Two Weeks Before I Was Gonna Share You Were Running Off.

You Were in a Hurry A Figure AI at that time, so a couple of weeks back had just retired their humanoids from DMW right and Camilla Mazzolini. she looked into the numbers. She says they'd been up for fifty percent at the time. At the moment use case was helping build thirty thousand BMWs.

means They loaded sheet metal onto fixers what is very important. and she said the three words one task, one station. One line. that's important.

I'll come back to it in a moment. so then she compares the figure humanoid To the industrial robots under human workers because you know That's that's what they have been doing this type of work today right? And it turns out at the humanoid is currently about three times surprise of a robot that actually runs, so the actual robot or two times the cost of human. Now moving back to moving parcels from left-to right there was a competition between a figure humanoid and a human I believe.

now while in the BMW case baseline, you know because that's what today humans do or if it is a big site like BMW I'm sure. That's what industrial robots are doing right? Now the parcel moving use case...I am convinced!

That for our long fifteen to twenty years has been done fully automated companies You know, scanning past those automated data capture companies that you know by heart. we turn doing the same for food warehouse automation and many other companies. So I'm just not completely clear. Just to be sure i'm Not against per se humanoids in the factory right?

It's just that I do not see them working in brown field environments And they were built for humans and industrial robots. so I am very much looking forward To you, I thought it has something to do with the barcode or something like this. To find a bar code on the parcel wasn't that? the topic was humanoid?

Yes but that exists for ten-twenty years. three hundred sixty degrees scanning from all sides and bottom as well. You go over these lines. there are small blips in here.

That's my point. Who was saying that? I think Jakub also, was he sharing a piece on...I think it's about models.

And as I said and had to do with benchmarking. It is always like we are back in the nineteen nineties when there were processes. When you're going be proud of something your solution I believe you need to be honest and compare your LLM model with a baseline way of doing, let's say machine learning or statistical even. You can't say it is wonderful.

great technology is what automated solutions have been doing since a long time and that's why I said, i don't understand. What the humanoid is moving packages from left to right? What else do you have? pd because I had something related to safety and robots?

Maybe i go for, i need To explain a little bit. cause every three months. I get to host A meet up in linds And this time our guest was dr tim fischer from sick so From the company safety center company. And uh Our community knows him quite well Because he's always joins our events etc.

So He spoke about safety and AA, how to bridge the gap between safety and AI. And then he made a comment that I found very interesting in. now please sit down. NVIDIA is heavily pushing into this safety market.

Yeah! Tim made couple of comments. so i looked at it and NVIDia calls outside in safety And they seem to be very actively involved in all the relevant committees. You know that when you have safety regulation, there are standards and committees right?

Yeah exactly! They're hiring a lot of people into the safety market... This was interesting because Tim told me different approach than go via centers, going by hardware infrastructure. So Nvidia hardware infrastructure and I'm really looking forward to get somebody out of the NVIDIA universe to talk about this topic because This was very interesting for me that in video is now also actively Going into their whole safety a functional safety market in the industry sector.

Very surprising very interesting. Please do a podcast and it's as well. if we did our, when were there? Our autumn activity with SICK and I think at that time still the big topic was has been for quite some time that safety detect human you know is completely deterministic And all of discussion about how can probability based technology play role in security safety or concentrating on safety.

In this case, right? Exactly and I think we had a lot of discussion about that We did find out. the guys from SIGS said there is potential And they talked a lot. exactly what you say.

The ISOs where are moving If it's not everybody and their mother are announcing They work with NVIDIA Every day. there's ten big companies, right? And the NVIDIA stock keeps on moving and that is one side. If it isn't some company announcing their cooperation with NVIDia its NVIDA as you say or not announcing but somebody else I'm saying they are doing...

They're moving into fields which maybe we had not expected Exactly. I was very, very surprised so i really hope to get somebody from the nvidia team to talk about this topic because that's...I thought this is still a topic for all our manufacturing safety specialists since you know fifteen years working on safety and now nvidia also entering this market. but isn't there big development we see as well maybe where it would discuss again in that the companies, as we suggested exactly two weeks ago it was specifically OpenAI.

But not only OpenAI also and Tropica believe they had come up with their solution companies right? So integrators yeah Right. so they say We now have this wonderful LLM solution for you. We keep on updating and it's so awfully expensive, but is not enough.

So we're now gonna need the highly trained humans that we already have or are going to get them out of market? So we train him all we buy them That's the other thing right. And then go our customers show what they can do with it. Maybe did this as one those solutions I'm talking a little bit more about humans and work again.

That's always the kind of a red line that goes through all these things we do, And I'm not sure it is to grab for people who have a base until you know good knowledge doing something with the LLMs in our case Indian industrial environment I'm not sure. Maybe it's natural where its coming from, but of course NVIDIA has been moving with their simulation capabilities based on top of CUDA. and what is the other thing called Omniverse? Is that right?

So maybe it's an extension of their omniverse capabilities. because if you are going to simulate your factory and want have run... You know i am just thinking about my architecture training, right? It's the same.

You know... The architect has been doing that for two thousand years and you put a couple of lines on one piece of paper or what I did.. What we did in this screen And then your going to build it! That is of course with the NVIDIA with the Omniverse.

Now Dan, next step is make sure you can simulate it. such way your factory for all is going to be safe even if it's you know, I will do that. Okay, what else? Do you have?

second one is Andrew my friend andrew wang. i've actually not been following in That much lately because when i do most of the time i disagree with him. so And that's okay. That's actually important people can.

So love comes to an end is so much more important than I am in his market nevertheless. So being Andrew, my problem was a little bit the lack of discussion. It's maybe not even Andrew because Andrew or some those people do not always have the time, but sometimes I feel that Andrew is really playing the game by making statements such. you're not supposed to have a different opinion.

That's my feeling! So i'm not on a personal crusade against Andrew... I did shortly consider introducing a hashtag in this direction.. but I won't do that....

But im gonna share two topics from him. so the first one thats all. actually I didn't come to share that before. He has a new course out, it's called prompting for everyone and by the way now to forget i was one of in the meantime.

believe millions people did his machine learning class right ten years ago, I mean he's the number one on bringing technology to the masses of people. It wasn't he the Coursera guy that introduced it. so i was going too...I did actually start in this course and very soon thought again ...

and thats my specific point here. is AI for humans Or are we humans still there for AI? And it's perfectly okay if we're in an evolutionary mode here. But I'm convinced that eventually, you know...

AI will ask humans if and what needs to be answered with a request! Maybe until then people can learn something from this course.. I am not going to have them. go NOT TO THIS COURSE.

You should go to discourse if you believe you could learn something. but I really expect YOU a listener, if you are an AI product developer to have your A.I tell me what it is that needs from me rather than Me treating Your AI as If It Were Dump. So my point Is instead of me telling the AI Please do XYZ Oh and by The Way I Attach The Following Documents I don't Have To Be That Smart i think should tell me, oh I'm very happy to answer your request.

i can do a really good job if you give me access to abnc or in the meantime...I was really impressed here. this is not about google against andrew whatever..i don't care but it just happens to be that i use gemini and Google came to me ...

and i had asked Gemini couple of weeks back why? i always have to go back you know, so that it would know if could interact on the topic. And I think had to do with data privacy. and now its saying If You Want To...

you don't need go back any more than a theme! If you want to Gemini will know all of the discussions you've have..and you can just pick up from wherever and i liked it alot. So in general I'd say because I That's what I WANT And if you don't want, that's OK as well.

If you don' t wanna have technology AI involved in your life but if you've given AI access to a job or home for whatever then I expect it know where and find relevant information rather than the other way around. so that was one topic. talk about these things a little bit more high level, not always specifically industrial AI. You know I expect you listener to apply it your own environment right and then come back later as well.

maybe there's the second one that i would like to share is about. we talked about work so far. We Talk About Robots. Do We Want To See?

Factory that is for humanoids. Well, yes I'm interested in that but at the same time The next topic the next topic is that Andrew says there will be no AI job. Cope couple lips terrible word? I think he means a job apocalypse right?

quote-unquote, the story that AI will lead to massive unemployment is stoking unnecessary fear. AI like any other technology effect jobs were telling overblown stories of large scale unemployment as irresponsible and damaging. let's put a stop to it. now there's this other guy Patrick O'Callaghan at least he saying well someone who has taken your awesome course.

I did millions. what are people have? i loved it You know, because you know Andrew. He is the smart guy in AI but he's not an economics guy.

and I must say I see so many examples And i'm not sure if im going to go into details of any of those. But last time it was around May one right? So Im actually in Italy. uh put up these.

um what is it? may one polls right i would do a big party. are you still there robert? so peter oh peter peter.

okay and that's where our connection with peter ends. he's in thousand Italy enjoying the seaview and recording podcasts. So now we're moving on to the main part. I would say was in pro joint venture between Volkswagen.

Peter, greetings to Thousand Italy. Enjoy the sun enjoy this sea. and now let's move to the main part. bye-bye!

Today I have two guests Daniel Wolf and Florian Bohne. Daniel welcome to the podcast. Hello and good morning. Good Morning, good afternoon, good evening wherever you are in the world dear listeners.

And the second guest is Florian. Florian Welcome To The Podcast. Yeah Hi Robert thank You very much. it's a pleasure to be with you.

You both work at in pro. and the interesting thing about intro is that it's a joint venture between Volkswagen And Siemens. you will have to explain that to us later, but first please introduce yourself briefly To the listeners Daniel. Please start.

yes My name is Daniel Wolff. I worked with info. i started my professional career as a mechanical engineer. I focus specifically on factory design and operations, that's a discipline where you look at the design engineering of the factory building or equipment.

And all the infrastructure in logistics are these topics. so what? You quickly understand that factories is. it's very complex system to complex ecosystem.

Yeah It's a very complex ecosystem or jungle some call. Yes, it's an organism. Anyway to understand this better I already identified that simulation digital tools are really a great help in these areas. so i quickly moved into simulations specifically material flow simulation and was doing for a couple projects in the Volkswagen Wolfsburg planet looking at line integrations of golf back then.

And I was fascinated by the idea of Impro very quickly when i saw that Impro joins use cases, problems applications from Volkswagen with digital solutions. From different vendors and also Siemens... and then yeah! I contacted Impro.

it wasn't a match and.. I've been here ever since. The last years I co-led our business unit for joint projects and have also been involved in industrial AI on the course of this topic. Okay, and Florian please introduce yourself?

Oh yes! My background is in numerical methods. I find element analysis and computer fluid dynamics. after the masters i did my PhD at the Institute of Formic Technology & Machines in Hannover with Professor Berndens.

And well, during my PhD I already came into contact with the VW environment and automotive industry of course because it's quite significant for forming. processes. and yeah, but during my PhD I mainly looked into the field of numerical simulation in material characterization. And what we also did was back then already to look into this field off AI.

if you can build up surrogate models an order two speed ups in numerical simulations. there's a lot going on when it comes to numerical stimulation at that moment right? When he comes through AI Yeah Yes You definitely do. That's definitely right In having all these world models and ongoing because in order to enable robots It is really fascinating field.

And I would like to talk with you a lot about that, but i think it is not enough time because its quite fascinating. Exactly! It's new podcast we need. Yeah, you invite and we will be there.

That's not a problem perfect. And when we looked into that I think it was then back in two thousand eighteen. Back then the all these methods encoding message and so forth who haven't been set strong. So um It's much stronger right now actually?

This is why You can also make lot of program Much more progress. Then i did my PhD. I moved to impromptu. I worked for Two years as project lead was tasked to build up a group around the topic of next generation of computer-edited engineering methods.

And then I looked much more closer into AI, because it's really rich for message. so when you have numerical messages and work on that topic... When we approach this topic in AI gives us much more opportunities to approach problems efficiently with different angles than before Because normally you do rule based programs. But what to that later, and then I went into the topic.

And my thing two years ago we together with in our project We also addressed started to address these topics and To make it really than the heart of the topic off The projects and as a software? That's What i'm going to talk about Later. It's A lot About Geometric Deep Learning and Using it for the Design Of Production Processes. Perfect Normally We do not Speak a Lot About the Company.

but Let's Talk Briefly About the company because It's a joint venture by Siemens and Volkswagen. How did it come about? What is the focus of the company, Daniel? Yes that's great question!

And its really interesting story I was mentioning...I was fascinated with idea maybe as fun fact or little bit. looking into history. you remember Berlin was divided city in World War era.

A lot industry had been in Berlin before left the city due to situation. There was a time in the early eighties and nineteen eighty two where politics And big companies together thought about how can we change that situation? In nineteen eighty-two in December. so they called Deutschland agi right, So everybody was involved to gather.

Yeah That's it wasn't spirit Right. It Was A Different Time Right so different a little bit different than today. so anyway This Politics and Economic The Companies They Enjoyed for Innovation Conferences, one of them being in December and Berlin. In December of two years ago there was an idea born between the board members of Siemens of Tagla BMW to found a company that would address the need for innovation and digitalization in factories and manufacturing and build it specifically in Berlin.

so this approach Today, it sounds natural in some ways to some of us. In many areas we have a lot of industry collaborations today back then and was kind of revolutionary right? So to create the joint venture that would bring together big manufacturing and technical providers to address innovation in manufacturing. That's the birth story very short of IMPRO.

And a dress topic that when you look into this, it's really interesting. They started to address topics with us today. so it was about sensorics and introduction of robots in the factory, robot programming simulation still with us now and interestingly development of expert systems. This is certain branch of artificial intelligence.

Essentially, today we are still in many areas. We have the same topics moving us in manufacturing right? So that's really interesting. But to be clear Daniel you do not work exclusively for Volkswagen or am I wrong?

No! We are organized as a joint venture. That's an organization of the company. so we're independent company but...

we are very closely aligned with both these coffers. Today is a joint adventure as mentioned by Volkswagen and Siemens. Volkswagen for us is the partner that mostly challenges us with manufacturing problems and needs to make production and manufacturing more efficient, more digital. Bring it up to speed at safe costs in time.

both engineering operations and Siemens in many projects activities brings a lot of technical background expertise hardware and specifically also software, you know the entire automation portfolio seven and all this tier but also simulation tools. So this is a combination that we work in And We are In between both with them in an orchestration layer where we form the basis for cooperative projects pre-competitive project Where we look at potential applications of digital approaches, technologies also AI specifically AI in the last years of course to bring it and how to bring into production.

We do not only work as a think tank and scouting company an exploration company. we also work as development companies. so to the end, to the shuffle operations with Volkswagen and beyond. So you scale solutions for your customers right?

That's a target! The target is to provide value form-infection from real production generate real business values that are always in focus. so we have expirated. freedom of course to find the approach...

the right approach how to generate value. but the target clearly is value cost savings time savings and increased quality in manufacturing, that's where we are statistically looking at. And yeah when it happens and what makes sense? We integrate other customers of the companies as well so we can work with other customers who have projects or solutions running to other customers because there is a problem that Volkswagen has in manufacturing.

Do they really have problems? Not Problems but Challenges. It's always a problem that arises in quality or something. Always challenge for the engineer to improve it, obviously exactly.

We don't want to go into details when it comes to challenges, but. When you talk to your customers Florian what are their current concerns or What other current topics do they all just want agent on the shop floor? Or They also ready to use time series simulation and AI And stuff like that. What Are The Current Topics of Concerns?

When I'm truthfully answering That they're not interested in AI so they are Interested In Having Their Problem Solved And AI, I think it's a really big thing to do and gives you a lot of capabilities. So no engineer is going to tell your use AI in order to do something. they have the clear problems which they want to have solved? Now having all these methods like having large language models or geometric deep learning what i'm gonna talk about later Or having agents These are all give you capability to solve their problem.

This is whats there heading for and what they're striving to do. And there's a lot going on, right? How do you separate the good from the bad? What really makes sense and doesn't make sense for customers?

This is something we go into discussion with our customer. How can we solve the problem? And then every year, I don't know when you're planning for projects. Survival of the fittest process so to say and where there's a good return on invest that all colleagues are really involved in this.

also management is involved with solving problems. these are real indicators. Just starting something and hoping that AI is going to solve it, It's not really a good approach. You should make sure the path is clear with all of these methods which there are like having normal rule-based programming or numerical simulation but also AI...

And when all this is cleared maybe you add something. how we do the service fitters process. When this is clear and all the stakeholders from all participants then agree, than we start a project. Daniel I think it's good when you give some more details if you can think of something.

Yeah as mentioned obviously We are working on a specific area of cooperation between these large corporates right? So there has been common interest Has to be a common interest in addressing an infection challenge with a specific technology. so This is scoping that we do very early-on and a very important fact of the most important one is how do we see real potential to build an innovative approach, to generate real value in manufacturing or engineering. So you write business cases at the end?

Exactly! We calculate them by looking at technologies like AI where many things are happening News every week coming out new technologies to explore. we sometimes do not know exactly how We will address the challenge in the future. So, we work in a fixed budget flexible scope way of mode working mostly and we Define the target.

we define the value potential that we see but you continue to iterate this In the course of our walk obviously to be sure that we look at all the dimensions of the projects, off-the-later implementation and also for cooperation potential between companies. We do this continuously along with project. if we identify that we are drifting off course then you can correct quickly in the carousel stock approach. That sounds like a crystal clear solution or crystal clear process.

but why is there let's talk little bit more common? Why de-illusionation, why is there so much? We thought AI could solve more. So what's the problem for the German European industry Florian from your point of view?

Oh that's a really great question! Do we overestimate this topic or... No It's a process where you are doing it in the wrong way I think. when.

So where does it start? Where is, Does It Start. It starts at like software companies and here in Germany we do not have Software Companies really Like In The United States. We Have Also A Different Market.

Its Not Consumer Based Its More Having Industry. so when you have a consumer based market You Need Other Techniques Technologies In Order To Evaluate What Kind Of Product Is Really Like Searched For. If You Have A Lot Of Customers You Have to Address That Differently When Your Here And When You Are In Germany. You're looking into the typical industries, you have engineering problems.

problems which have to do something like with rigid materials. Solving these problem is different, so of course you've got a lot data as well. but this data comes from machines and what we want to control the machine in order for all those approaches existing already there are lots... In Germany now we know how our stuff works when it's coming into the industry.

I would say and how to create that, make sure they are safe. Now when AI comes with it must bring a value for them. not all of these companies do programming The one which large-language models help right now a lot. They also speed up our working processes at Inpro, but in the industries you do not program so much.

You are doing more stuff with physical things like looking for a CAD or simulation and go into shop floor to solve problems. These things must be... technologies for that and I think this is we are moving to there. so it's a building up foundation models which was for these purposes, but what for us really important then.

These large language models? of course they support us tremendously when it comes to finding an answer in the stuff. big differences and Daniel and I, this is what we are working on. We're working on introducing these new methods into the industry And i think that will come up.

but in all innovations it takes time to fulfill. There will be a failing, there will be problems. This is some challenges. this is normal.

it's totally normal. It's frustrating but that's normal and we have to deal with them. I think We are on good path here in Germany And this is something i want to say Is we also then have to fix all the other things? clear that this is going to be in place, then we can when the technology of large language models and agentic working meets really engineering work.

And makes it really fast than you can also speeded up and scale it here in Germany even more When we have all these energy and data center in place. Yeah This story. for a long answer. Can you please share two free projects?

You already mentioned geometric deep learning. What are doing your daily business Florian? Yeah, I can do that. Or you could share two.

where Florian won by Daniel please flow and maybe you can start. yeah perfect thank you very much. so when we are talking about your deep learning We're using this for automatizing the design of production processes. So what is it?

Design of a production process? And what do I understand under this? It's like defining the sequence operations. What kind of equipment is used to do a layout or tooling, so ever make everything then happen that when they really ramp up their production and every thing in place it works but you don't need time for fixing problems.

So all these things are happening beforehand of the industry, so to create these processes. And in automotive industries this is where you have a special situation. I would say it's not so special but The geometry of the parts shape of the part was established. they do not change So much actually when you look at the body and white?

So this is actually. you have all that already really similar, but every time you want to design a process. You start more or less from the beginning not totally of course and there's a lot of manual work processes going on. And what we are looking at is these working process where you defining geometries or defining points, lines or sections on geometries.

to be little bit more concrete here for example clamping point when you have part which want measure or part that you want weld You must be really sure that it's properly clamped. And this is because we want to be, you want to repeatedly measure it reduced warpage when it comes to the welding process. and well already talking about welding You have to define where are the weld points? Where are the tech wells?

Tech wells other wells which you place before The real world process in order To keep it together as a whole assembly. Well, where do I do that? The way i put these were tech wells and you have to do that cautiously because if we don't Do that. You will have a lot of warpage in the geometry.

So this is then affecting the whole the warpage or the shape deviation Of the whole part. We have inspection points. Where do you want to measure the part? how do you Want to make sure that it's really in shape And they cook.

good go on. Let's just shortly jump to injection molding processes for example defining the gates. All These are things where you have defined in the Geometry manual processes and these haven't been automatized. So this is really strange, you have a lot of data because we've done that many times but not automated.

however when looking to the field of AI robotics there many methods which work with three D objects, also an autonomous driving where you are processing a lot of geometric data. And this is why me and my team we said hey let's have a look how maybe these AI methods can really make a difference here? So is able to, by the training process you are able to identify geometric features on the geometry and we save over based on these features there's a classification or segmentation done.

So but when you have that it's little bit difficult in the field of computer-aided engineering because thirty parts at your thirty different variances of a part. And because, well you do have the specific amount of cars that are produced and this is maybe thirty or fifty parts but it's not enough actually for non-standard AI methods which normally start working when they're like hundred thousand something. This challenge we discovered there When you're using this typical AI method, and have a big model.

You push all the information into weights or parameters of models... you cannot explain results but it's really important to know. why. is there a clamping point at specific location?

Just because it has been somewhere in the embedding space? This will not convince anyone who wants to sell that These are challenges which are quite difficult and spicy. And also when you have a new, well work on something new and want to put it into the AI model You need training and having a training process in software is something which is difficult challenge because data must be properly prepared. you as a data engineer might not be able to look at it before it comes into the model.

And in my, somehow distort the models result and so on. So this is something which has really difficult and challenges but I think we are under good parts of solve these problems. What did us? We looked at that.

then we took with little bit older approach case-based reasoning Which i think have been there already like ten years or twenty years maybe even more. And what you do there is when you have a problem, you search and retrieve similar case which we have in your database. You're reusing the solution - you revise this solution, you adapt it to your purposes - and then change the solutions for improving the algorithm." We said okay I think that can be used.

so... What are doing? We've got a legacy data database And we identify the most similar part. So having this in more similar fender or having some more similar side panel, we identified it.

This is what you do with a set of geometric deep learning? We then use that data transfer this data to the new path which depart and query inquiry. This was also something we did with dramatic deep learning. Then when we do them as we used local rule based checking so go into every point, we say okay.

We have some company routes which apply here to take this and that's in the account. if you check it And when they've done then another layer. It becomes really interesting checking If its necessary with respect of numerical simulation. So for example having a tolerance simulation and so on, combine these AI approaches with numerical simulations.

Really exciting. what your dream? Is the numerical simulation as a backbone or is it second layer? So this is a second layer of checking.

First, having good guess which comes from the database and transferring all knowledge to the most similar part we have here then checking it and seeing does that really hold up or the requirement? And if in case you do so go on saying okay these are perfect clamp points or weld points whatsoever The engineer can also look at them say I'm convinced working with it. And in case of doesn't, you can still have like doing all the I would say normal stuff than do an optimization and so on.

but we already had a really good starting point. that's always. this is going to start from delivered by their eye. however over when you know change perspective and when you say okay um i think this message using his really well nice also to build up digital digital twins of to build up a numerical method or numerical simulation models and using these for digital twins.

And when you want to build a surrogate model, we need a lot off data. so We have two very lot the boundary condition also that geometry. in there. You need automatic processes which set up the model because you cannot forever.

When do one to build upper surrogates model? You can not spend your time like when it costs two weeks to set up a simulation model, which is not something too highly overestimating. It can even take much more longer. so you need a lot of automatized methods in order to get the data then to train and surrogate model and make this data available.

And also I think that method we are applying here's really interesting because they help build-up these simulation models with help transferring points based on all data. But the next step is to use an AI simulation instead of a numerical simulation, this Is The Next Step or Am I Wrong? Yes it's what i would like To Do This When I Can Convince My Customers. This Is A Way Where We Would Like To Go Because This Is The Whole Topics Really Exciting.

And Did We Work Really Well Together? but There Must Be The Need. If We Have Good Methods There And if a normal simulation fulfills it, we do not have to change. Because the normal simulations they've been there for a long time and established in a lengthy process that many people trust in these.

They have built up trust. all of their AI must be done the same because AI has always like this myth of not being really precise or making things up. This is something which it's really toxic when you want to check some thing and be sure that there are a lot of skepticism, the process of convincing people what we have to do in projects? I have one more question because you mentioned welding and injection molding.

right. so how do you generalize now your solution through different domains? Well, what you have to do is look into the formats of the numerical simulation model. software companies and established software.

when it comes to the single process. So injection molding, you have a mode flow When it comes forming processes your outer form And they are not alone. but these other like them big players These I liked that ones which are mainly used on. They have an when do want now too?

To use your methods with that You need to know the data format. so you have to access this David data format. This is not easy because he's often closed formats which you cannot access. and when, uh, where it can not access that.

You cannot prepare the data for these simulation models. but in case you can do that And you have like other simulation methods than all of fields of simulation For example Where use our course? or we'll use Alice, you know Which are really general purpose simulation model say? We've opened standards of open formats really work with this data format.

And when you can work with his data formats, You can prepare the date. I was help of other software tools and these software to then can be enriches as I am Then you are able to access these and them too Really? So this is no way to go. Daniel we haven't lost you but Florian so so enthusiastic about a solution.

Maybe you could do on with a second approach. Well it's nice here that your employers also enthusiastic where they come through their work. I hope all this was understood that we are an enthusiastic team because technology and the challenges that you find to bring back together is really, it's an exciting work. It is a privilege to be able to do that anyway.

so yes Florian uh Is at a specific point in his work where basically The product world throughout the geometries And the equipment world of the factory join it. The manufacturing concept that he mentioned, is in a critical stage where we define how an assembly is built from single parts. and I would like to jump into the other two directions. first maybe beginning of the entire engineering process.

when you look at a carburetor assembly It's typically an automotive monocoque shell manufacturing made up out of a lot of metal right. so mostly we have metal sheets That has to be joined together with different welding and joining techniques. And the first step always, this is where we have successfully applied AI in the past... is to decide which types of metal can actually be joined Of course In industry and company like Volkswagen has a lot knowledge about it.

This is what I've done right now Newly! This long-standing experience. So We thought this particular niche or starting problem some years back and when we started working with AI, actually have applied machine learning to the topic. So you must imagine that when an engineer is designing his assembly he has to choose different metal grates And industry is always providing new materials, new material grades and different thicknesses of sheets.

So you have to decide about this. so You have to know will I be able To join them? Will i Be able to perform a welding process on it? This Is where we built an automated checking tool basically A feasibility checker you might call It that evaluated the feasibility Of joining two types of metal sheet together.

There was a successful application of AI And specifically machine learning rebuilt maybe to explain a little bit how it works. So from the experience of material testing, or engineering our customer had a lot of data off course trial and test data specifying the type materials thickness physical and chemical properties technical properties. on the other hand you have information if certain material combination can be welded together these two information sets we joined. We built a machine learning model out of this and tested it, brought to the internal production cloud.

This is where we bring in our software expertise that can be deployed into an internal platform. so everyone has access material combination, data and then get a result. if this is the probable combination that can be welded well or there's maybe some technical challenges. That would require some additional testing And If you do this You will suddenly start automating A very basic process.

It has taken up a lot of effort and time before. Now it's running automatically. Get an automatic recommendation Of course! The Stuart engineer has to validate and check this result if it sounds reasonable.

But because the most part of paces, It works very well And thus economical benefit is basically reducing that effort for physical testing That has to be done in the lab right? So formally you had to take the material sheets really go through the lap or Contract someone that operates a lot make the test wait for the test to be finished, get out of that data and so on. So it also delays the overall engineering process. It's not only about saving costs but getting faster.

I have one question because from my experience you talked about data right? And normally in the automotive industry they have good data They don't have bad data do not have enough bad data when it comes to material, When It Comes To Welding, When it Comes To welding spots etc. So how Do You Handle That? In that specific case we actually had data about non-feasible combinations.

Okay great you're lucky. Yeah We were Lucky. I mean of course where you start such a project you Have to check Is the Data Available? Can i use pre-processed?

Thats A Major Part Again and again also Of Our Work. thinking about AI would be nice if you could develop the most part of a project. But in fact, we all know that there's a lot of pre-processing work to done organizational work around so pure development workers only share. but yeah I mean as i mentioned We tried to identify problems challenges and use cases.

where can actually working generator success In that case were lucky had data identified it. And we have also encountered cases of course where it was not. But you know the phenomenon, right? So they're missing off bad data or bad results?

Yes yes yeah It is an issue Of course Yeah but do I look into synthetic data for that robot? That's my question. But its tricky with synthetic data. Am i wrong?

It is tricky. We've made some activities and experiences specifically in the area quality assurance and testing So generating visual defect data for different applications like PressShop or Functional Services. And we found that given a certain amount of experimental test data, of real-data available today mostly you are able to generate very good synthetic data out of this. and if you combine the synthetic data with just a small amount then suddenly your data set and your models become much, much better.

It's not in relation to each other but just adding a small amount of real defect data through a synthetic data batch. we have found that our cases has really improved the quality of prediction at the end if apart or feature this defect or not At least for visual-quality assurance topics. We've done a little bit work on it And I think now is technology gaining traction and being distributed across industry more, so of course that's a very interesting topic. And you can use GenAI for this too right?

Yes! That is the future direction. GenAI will operate these environments systems on also to detect with data. obviously it's a really interesting approach.

I mean we are experimenting in developing us around co-pilots in several areas around simulation to integrate co-pilots and assistance into complex engineering and simulation systems? That's the first step. Who is paying for that, who is willing to pay for co- pilots? what should be the benefit to...

to pay at the end? because when you talk to everybody in their whole automation space In the industrial automation space Everybody is working on automation On co-Pilots For engineering etc. but everyone Tells me how it's so difficult to sell the customer because they say, oh this is common technology we know. How do you handle that?

We are not willing to pay Because I get it at entropic or Claude for fifteen dollars a month. So isn't really so difficult To sell it in the end? That business and distribution question probably not fully answered yet Okay, so it's difficult. Yeah I guess its an exploration process at the moment.

how you market these functionalities if they are part of the tools is their general purpose or specific purpose. i believe from the experience we have made with selected co-pilots especially food seedlings in their tools that a copilot can bring value, very specific value to users of the tools. Not only explaining the tools and helping me on board learn and study using that tool which can be a huge hassle I mean look at the effort and volume of documentation or training you have go through if your want a standard simulation tool, like for example plan simulation from Siemens.

For automotive I mean you don't have just to learn Clark's simulation. if they understand that modeling philosophy You have to observe certain coding rules inside the tool... you have to observed specific automotive libraries. so all these aspects need be considered and they slow the onboarding time when it costs money of course.

So If your build..if are able to build co-pilots at address specific domains It helped me to solve my engineering task in a specific domain faster. I mean, there is tremendous value on co-pilots but these would not be general purpose copilots like we see from Microsoft or as you mentioned from Anthropoc and so on. that can help me code anything.

But they are... They need to be embedded specifically intertwined with the simulation tool. if they understand My intention of modeling And my intention of analyzing complex technical system Then they can help me much more than a general purpose agent or co-pilot. But then, I need to have some certain understanding of what i'm doing both for this view and the system viewer.

This is where we are standing now in industry Or as I think In this technology Where there's The challenge Of building such things Right? Because this Is not the first use case that you address when You build General Co-Pilot That try To solve how to model let's say, clapping system for the cowboy assembly. Right? So this is not the first use case you haven't released if your and tropic.

so This Is a road ahead. And yeah maybe we can conclude to one or other topic. Daniel have One more question To management Topic. right when I talk to industry companies they do A lot of PUCs.

They are saying Oh We Do a PUC in this But at The end When it comes to scaling A Lot Of People are failing or a lot of companies fail to scale our solutions through the whole company. How do you scale your solution? Through production plan by Volkswagen, how did you scale it? and what is magic behind scaling?

That's exactly core issue that we address in the most interesting part because as you say... With a good idea, you'll see the potential. You start developing and find good results in your first pilot that can even apply it to the real plant. It runs for some time And of course...

it is a lengthy process. Often takes more than what we initially expect to validate the solution technology. Let me give an example from our preventive maintenance solutions We have integrated into The first book, Volkswagen Planet Car Body. This is AI models that basically monitor the performance of certain assets in the car body line like weld guns as mentioned or clamps.

you know we look at them using those Siemens Edge technology for example. and then AI model that will monitor all these systems and analyze smart faults with small delays and predict a health score. tell us okay sir it's not a problem. And that is a good example for the scanning issue, because this is certainly a topic where you have great potential to fail.

You know? We've got hundreds of stations in our poverty line. Israel's station has several many clams so we can have thousands of clams and thousands of workguns at factories. In the world there are other factories like operates I think hundred or so factories and there's many other.

So this is basically from the business side solution or a problem that sounds like you can scale it, but how do you really validate? That's our first challenge. so we typically undergo validation as early as possible in our projects. yeah We try to bring into the shop floor and through a pilot which was test send maybe first.

then they get selected at my island when we can operate with low risk And then we validated for certain time obviously classical engineering approach. We get results, we improve and then we enter the phase where you think okay it can work? We tweak and improve the solution at this time to work on both of them. a business in the technical model.

how is the Siemens Edge ecosystem something that we can use to scale. When will it be integrated? Where, how do we connect the edge devices? through data in the controls and in line then see okay maybe some areas are covered by an automation standard on which you can dock onto.

Some areas may not yet exist. so these factors have been considered. Is there a magic recipe for this? I believe It's hard work to do that, to achieve it.

It takes more time than expected. yes our approach is that we try to consider all these factors from the beginning. We tried to involve the stakeholders The IT operator environment Elevation specialists, the AI specialists and data specialists involved All of them as early as possible And we tried keep them on boarded and involved during their process. I think this was a road to be successful because things change over the course of such a project and such developments.

And as Florian mentioned earlier, you will always find some unexpected rocks that you think about like accessing certain control devices or non-availability of certain components, you don't have access to a certain line due to different reasons. So things like that can always stay there and then we need to stay on the topic as keep working not to solve it. but You need a commitment right? We needed decision.

let's do at end. so you need decisions from sea level. Then you need a decision from operator. The commitment for AI needs be in all stages.

Am I wrong? Yes yes That is true. So, and we are very lucky. We have it in most cases or many cases.

so talking of Volkswagen maybe it's known. there is a strategy or commitment to go into the ILS for IT but you also need that. the commitment may be two to change processes But this is difficult when it comes to let say traditional machine building automotive industry processes. And how do you handle topics like this?

Yeah, with our made customer we have process for that a specific one. they operate and we operated together. it's called the technology development process where We talk about integration of new technologies or change of technologies in running manufacturing lines production. This is um process where you will have to run several checks it, ask yourself a lot of questions about how the future solution will actually work and run.

It involves technical feasibility criteria, it involves IT related criteria. that involves the question who would provide maintenance improvements later on. so business side financial aspects is all in there. It's a company framework.

Exactly, it's not as strict process but is the framework where you have to ask yourselves questions and critically ask yourself them. And we do together with our customer and all people that are involved in providing and generating solutions The shop floor operator, innovators who work with innovation experts technology drivers And also, of course not only the senior but also middle management. The people that have to decide are we ready? Are we really ready to go with their technology?

Those who said they had a responsibility even after you introduce new technology model your line will still work right. in an ironic factory You could introduce a new software or hardware module Only little money, only a few thousand euros and you could break the entire line. This is not what we want. We wanted to inject new technology And still everything needs be running.

this Is so very different approach from In The way of let's go on break things right? So we do NOT wants to break. things will improve six while having been running it. It's a little bit like operating on open heart, as we say at Germany.

So the patient is already living. actually he's walking and running. We have to make improvements during that process. The Steward to Deplaner I have already referenced.

We talked about generating synthetic data, it's just actually that we used Generative AI and its really working well to do that. You have use cases for this computer vision systems but It is always important that the introducing these computer vision system has been done And then you can settle on do generation of synthetic data if it's necessary. Sometimes when you have quality assurance issues, this is wrong data or this is defect data and this is hard to get because you have to wait for a long time because it takes time until there really something better cures in the good running process.

so um... For that we need the synthetic data. As I said, the narrative AI helps. And when it comes to then scaling that is maybe also connecting this with what Daniels had.

from my point of view you need platforms within a company where we can bring them on and make it easily accessible for everybody. When you have these efforts in order to get a solution... And when you have then a good solution combining with easy accessibility, I think these are good starting points for scaling. But you should also not be too early with the technique technology, because when all the other process haven't been realized yet.

You will have a problem or you might have scaling problems. so yeah that's this second use case I can think about. Yeah i want to come back to the synthetic data stuff. Can you share in three four sentences?

How do you use gen AI To produce synthetic data? For example, when you come to computer vision systems. You need images and you need images from surfaces for example And there you can use diffusion models in the normal techniques In order to generate these data. you have of course together some data which is already of bad data or defect data But then you can enable the model To generate you the defect.

It takes its work Of course going into that but it's possible as images are really good. They help set the help to train a computer vision system. So guys, what is on your agenda for the next coming months when it comes to customer projects? Well one of our focus points at that moment as I talked about quickly before this really working or continuing all work in bringing preventive maintenance AI to cover operations and folks.

but we are strongly involved here with activities And we are making success and progress there, but this is some up to do obviously. We see that it's a very interesting direction. in the first thought you might think that preventive measures... Oh!

It solved. everybody thought predictive maintenance was solved But still it does not. Yeah Exactly You think its solved? Its no problem.

Just talk with vendors and see many solutions on market. What were doing? as we bring into reality, right? We bring it to a real factory.

To be operated on real lines and this is as I mentioned before there's a lot of work after just generating the software or AI solution itself integrating into the real system that they're actually working very strongly in this way. we have good cooperation also with Sheroda and parent company Siemens. This is one strong direction we're going forward into. So this basically our work stream industrial AI and digital twin in the factory, We have other work streams that are covering too like AI, digital twin supply chain logistics And circular value chains.

That's a second direction we are exploring. How can AI help us in producing these processes, making predictions involved? For example with the Katina X consortia. In the past a little bit explored things there.

so that's using digital twins in combination with AI for this kind of purpose I think is very interesting direction as well. and what is continuing Is for us in the shuffle operations. we've made some successful projects with chatbots and co-pallets on the shop floor. And we will continue this as well, so helping operators maintenance personnel and factory engineers to improve running lines in order understand quickly now issues getting transparency about the machine data, reacting to issues, diagnosing them making root cause analysis.

We believe that or I believe that you're just chatbots and co-pilots for this entire approach of a GenDKI can help us here too make workflows faster integrate access more as the data we have available on the shop floor. in the running factory, but it's not always in the formats and accessibility that we need to employ AI on top of. So this is a challenge how make this data available accessible to AI? On the other hand making scalable across shop floor providing solutions easy-to use where user says hey!

This is cool... I want have actually using and then scaling it across the platforms that we find. So I'm really surprised, but i didn't hear the word agents in your agenda for next months Daniel? I did!

I said agentiqi will help us with this addressing of data and functionality. It's a technology. Well, we have to see how far will it carry us right now. We are on a certain evaluation type cycle as you know and yeah And where we have evaluated that there is some industrial setting an agent has observe certain restrictions.

maybe You don't want to make the nation but make decisions totally autonomously and in source of data. So this is something that you have to figure out, what kind of degree of autonomy do we give the agent? And which parts of the workflow does it support? where are its limits?

at which technologies do actually employ right? so will be like these claw systems or would be smaller agents? how does it actually work together and orchestrate? And another important question when you try to build this industry is, How do you scope these solutions with the general purpose solution or very specific ones.

It's their functionalities where individual departments groups plans and users can configure their solutions. Or Do You Have More General All-Service Solutions? That's all part of a question discussion. It was a pleasure, guys.

Thanks a lot! I keep my fingers crossed for your agenda... for your agentic approach but also for the predictive maintenance because i think it's underestimated in the whole industrial setting at this moment still underestimated and it is a pleasure to talk with you. thanks so much.

all the best from you. Thank You very Much.