After the cucumbers AI - a robot for industrial applications

0:00 This podcast is supported by Siemens, your partner for industrial grade AI. Hello everybody and welcome to a new episode of our industrial AI podcast. My name is Robert Wieber and it's a pleasure to talk to Peter Sieber. Good morning, Robert. Good morning, afternoon, evening, dear listener, wherever you are.

0:29 Good morning, Peter. Welcome to a new episode of our industrial AI podcast. What do you have in the news part? What do I have? I still getting messages from CES. you as well. Oh, Las Vegas. Yeah. Right. Well, the next one is coming up. Let's talk about that later. But as far as the I have seen Roland Bush, the CEO of Siemens.

0:53 Yes, he's our partner. But nevertheless, or even because I think he was the star, right? introducing the industrial AI operating system together with Janssen many others. Exactly. I think he had the keynote, right? Yeah, that's what he did. I didn't look at it completely. I looked at pieces here and there and actually I think it wouldn't be a bad idea if we could ask Roland, you know, if he could.

1:19 Join us. Ten minutes. Why don't we ask him here directly? Roland, if you hear this, would you find ten minutes in your busy schedule to explain the industrial AI operating system to our listeners? That would be great. Thank you very much. But that's a new term, right? Operating system industrial AI operating system.

1:37 What is that? Exactly. And to be honest, I didn't get into the details. And maybe we are gonna be lucky. And in just in case Roland does not listen to our podcast, maybe someone close to him does and can bring our request forward to him. It's interesting, because I think one year ago, the big topic was industrial AI foundation model by Siemens, right?

2:05 And this year is now operating system. So Is it a new topic? Does it work together? Is it a new strategy? Yeah, I think it's developing. Yeah. Yeah, exactly. Or still searching for something. Well, yeah, but my feeling is they have a very good feeling what it is that they want to do. But then again, I haven't heard the details.

2:29 So we'll be great if, you know, if Rodent could explain it to us himself. Have you been there at all to see? Yes, you've been traveling a lot last year everywhere. No. No, no, no. I thought CES is more focused on consumer electric, right? It was. It was, yeah. And now with Roland Bush, with the keynote, it's now an industrial... Community I'm not sure.

2:52 Yeah, just everybody everybody wants to be there. I think I I was there long time ago at the time that I was kind of in the let's let's call it the consumer space. I mean the. the Intel inside mean appealing to consumers. and at that time I went there and I do believe that Intel hadn't been there yet and I was looking at.

3:15 you know, shall we be there? I was there a couple of times at that time. I was there later Later, I was there with Stefan Hoppe, who was Steinkraus, Tom Burke at that time. With OPC Foundation, you were? OPC Foundation, exactly. But that was not for CES. I think there was in Las Vegas. It's just a different show.

3:35 So it was, it was Las Vegas, but it was, it was a very special one, actually. Not going to get into the details. What happens in Vegas stays in Vegas, right? Exactly. I'm going to come, come, come back to that maybe in a minute. Yeah. But yeah, let's see, I thought of that as well. I'm preparing a podcast with Norbert Jung from Bosch.

3:57 They were there. And that's also from from manufacturing, industrial manufacturing capabilities. So let's let's see. Yeah, you're right. I mean, at that time, I was moving and it was a huge consumer show. And now maybe the industry is appealing like at that time. you know, Intel was appealing to end users, right?

4:21 Is that happening to the overall industry? Question mark? I don't know. Because you mentioned Siemens, right? And there was an article about Jan Le Koon's new startup in Paris and I comment on the article because he wants to build large industrial AI foundation models, right? all about images, time series, etc.

4:47 And I commented that won't be easy because small and medium-sized business are aware of their data. And then there was a guy and he wrote an answer and said, Federation Learning. And I said, Federation Learning, we haven't heard anything about that. in a long time right. so maybe now it's a time. it's a revival of federated learning.

5:12 let's see i've wrote in my message. i'm quite skeptical. let's wait and see what. what do you think again in the preparation which i did yesterday and just for you listeners if we talk but one, two, three other things. From the moment you're going to be able to hear this Thursday, we're recording now Monday, I was preparing yesterday, and I can only share again in my preparation with the podcast for Bosch with Norbert, it's very clear that they have a very deep, founded understanding of the value of their data.

5:48 And that's, that's only, but that's just one example. But if, but they, they know about federated learning, Peter, federated learning. I'm not sure. But I strongly agree. I haven't seen the term for a long time. To be honest, I thought at the moment it was a great, great idea. It was about like, you know, instead of what was it?

6:10 Collecting data, sharing models, right? Yeah, but it was this wonderful quote, not the data doesn't come to you, the algorithm comes to you, right? So you, you keep your data where it is in the algorithm. So I'm not sure if that's still working here or there. I mean, to be honest, since ten years, typically, if we don't, if I don't ask it specifically, The issue, the opportunity always comes up, the question of, you know, cloud or local, right?

6:39 We've been doing it all the time, but we haven't been discussing this in-between kind of solution of federated. So I don't even understand how the person came up with that thought, not sure. But maybe we should keep an eye on that topic. Maybe really it's coming back. It's a revival of federated learning when it comes to build great big foundation models.

7:00 Let's see. Yeah, let's see. But then again, I think the very clear development in our world, in our industrial world is towards inference close to where things are happening. You know, that's what we talked about five years ago. And actually, you know, I keep on asking wherever I, and whenever I talk to people on the podcast, outside of the podcast.

7:30 I'm still really not completely understanding. Oh, I don't know. I was thinking who we did. I talked to last time. Actually, it was Gemini. I was asking what she thought of it and now I call her she the algorithm. Why is it a she? Yeah, that's that's as I say, she because it's a female voice. So I can.

7:48 Okay. But you can change that. Do you know that? The female voice is perfectly okay. In this case to me and completely still not understanding why we need trillions of trillions of new data center investments. Yes, I do understand that there is, you know, whatever five, six, seven big tech companies and each of them want to be the first and the best in training models and coming one, two, three percent better at.

8:23 what is there and I must say with what Gemini delivers is like really really very very amazing for what I do. but that's that's on the side but I still don't get it because at least in our world I mean everything is going to be on the edge. you know very close where things are happening. there's no need to do any inference in the cloud right.

8:45 so I'm not saying that I do not understand that we do want to continue. But with the numbers that we see that are supposed to be developed in cloud infrastructure, for me, I can't follow. Yeah. I think I mentioned it's four or five episodes. I talked to a manufacturer and he said, we cannot sell cloud products to our customers because there's no error eye on that.

9:12 And yeah, it's a topic for them. Yeah. Yeah. That's the other thing. I'm not sure. understood or always the difficulty of selling software basically right selling selling the AI service I guess so they're selling it as a solution. so here is what whatever I mean in the past it was here's my piece of hardware and there is certain software you know.

9:33 operating system for example. the operating system was never or was it sold? maybe I don't know. Yeah, I guess so. I mean, it has windows been sold. windows one, two, three, four, five, six, eleven, wherever we are. I guess so, right? It has. Sure. Do you remember the license keys? Everybody was sharing this license.

9:57 I just don't recall. Yeah. Yeah, let's see. Exactly. What do you have? Do you remember the guys from past robotics? We did an interview last year with them about the Again foundation model when it comes to welding the welding model. They are from Ohio so in the heartland of the US fly overstate and They have published figures today or now not today last week and they have over one hundred million annual contract value achieved.

10:31 I was very impressed. over one hundred million annual contract value. and you see the topic physical is AI is slowly taking off I would say. So it's a huge topic in the industry, in the US, when it comes to welding, especially a welding is a crucial domain. And I think we will see the same here in Europe when it comes to welding or when it comes to other domains with robots.

10:59 Let's see, but over a hundred million for a startup. Congratulations to the guys from past robotics. That's what we are going to. Or the time that you hear this, dear listener, we will have talked about a lot during our one, one and a half day Amsterdam AI conference, right? Exactly. We talked with Stratus about physical AI, scaling simulation versus reality, but we will do an special episode on AI and Amsterdam, I think, in the coming weeks.

11:29 So we can summarize there the whole topics, right? Yeah, right. Exactly. Okay. Ohio. Oh, interesting. We have two, two very, very, well, very good players in our marching band. That is the actually, actually term for what our Bavarian band is. I looked at it and they are from Ohio, actually. Talking of Stefan before, as I said, CES.

11:56 Stefan Hoppe. Exactly, Opus UA. There was a post by Ken Wenzel. He works at Fraunhofer Machine Tools. And he says, data without context is just noise. In a modern factory, you don't need sensor underscore seven, four, two, whatever kind of readings. you need to know which machine it is attached to, which production line is that machine on?

12:18 Who's a technician on shift right now? And he says, this is where linked in factory, interesting term on a linked, sorry, linked factory changes the game. And then I, I answer and I say, uh, isn't that what OPC away was supposed to take care of? Yeah, exactly. And by the way, I'm not confusing now the OPC away.

12:41 foundation podcast with our industry podcast because at some point in time and exactly here of course they do come together right to which then can replies. great question it's a common point of confusion but we actually see opposite way and what he calls linked in factory. as a perfect pair rather than no, not linked in.

13:01 It's a linked factory. Did I say again, linked in factory? Yeah, it's not a linked in factory, linked factory for your postings. You can have a linked in factory, but it looks exactly. Yeah, but okay, it has to do some linked factory. And then he says OPC UA is mainly used as a tree. Exactly. It's actually an object oriented data model, because that's the information model, right?

13:23 That I talked about many times. Excellent for a hierarchy. access to specific machines variables in real time. But as soon as you need to query across silos, e.g. find all the bearings from vendor X across five different factories, the tree structure becomes a bottleneck. And he says then the resource description framework is a web, the linked factory takes those OPC UA notes and turns them into a global knowledge graph.

13:53 And that's what I've talked. So you are the OPC. foundation expert, right? Peter, you're running the podcast on OPC. Do they have a strategy now on agent? I don't know. I'm not that close with it. I would say I would say exactly and I have linked exactly, you know, to what he does here, you know, he says like, and I've said in the past, the information model of OPC is exactly like this global knowledge graph, right?

14:23 And he says, this allows for horizontal queries that OPC UA wasn't designed for, et cetera, et cetera. And my point is over now to Stefan, Stefan Hopper, OPC Foundation President. And yes, I think I have suggested, and maybe I've actually asked exactly that question. I'm very interested in. asking it again.

14:43 And the interest, of course, is, you know, what is the role that OPC way will be playing? It's the same thing with MCP, MCP that may be sitting on top. And now there's another one, the resource description framework, or who call it, what is it, linked, was it a link factory? So very much looking forward to a second direct request, Stefan, very much looking forward.

15:10 As a topic for both I think the OPC foundation podcast as well as for us here because for us here at the industry iPodcast It's very interesting anywhere. We are all the time We are being suggested that is important to get your data organized in one or the other way and one or the other way may be and very importantly through an OPC way information model.

15:36 and at the same time we see development that where we then have questions and we want to understand what the continued role of OPCUA in an agentic world is going to be. Okay, perfect. What else do you have, Peter? I have one, two more maybe. Oh yeah, very quick one. It's still from CEF when Roland was the human star at CEF.

16:01 I think there was a second that was the industrial humanoid star. You know which one that was? Atlas. Yeah. Very good. Atlas buy by Boston Dynamics and Hyundai, I think Hyundai is a part of Boston. They bought him, right? Yeah. They've been moved from one to the other to many different companies in the last ten years.

16:22 But but we all recall Boston Dynamics, they were the ones where already three or four years ago, they were jumping in the fact in the factory holes, right? And figure with BMW, I think. All right. But at that time, that's what they were. Now this time and I did look at it maybe just like two three minutes this one looks like almost like a machine?

16:45 Well, it looks like almost natural, because, you know, as, as you know, as listeners probably know, I've never been a real big fan of humanoids in the factories. And I'm still, I haven't been sure until I now saw, well, there is a, there is really like a machine, there is a, is a machine and the machine, yes, has maybe looks like it's a machine.

17:09 maybe like with arms and legs and and it's not so clunky anymore it almost it can you know it? can you can really imagine that this thing this machine is capable of taking in its hands whatever up to thirty forty kilos I believe or something like that. so so can really be doing useful things. I think that it was.

17:32 that's what it was showing and I thought it was somehow impressive. and I must say this time I didn't care that it was like whatever you call it for me. it was a robot almost and I didn't care that it looked like as a as a humanoid and that was the first time. maybe the humanoid as the humanoid as in those for consumers with their you know eyes looking at you.

17:55 and can I help you? dear dear poor human that's this one was not doing that right. i mean this one is supposed to be now for the shop floor i believe right for doing hard tattiest work. that maybe is better for a machine than for humans. absolutely let's see what's going on. I will visit the European Robotics Forum in Starvanger in, I think, in, what is it?

18:20 in March or February? I'm not sure. Let's see, because there's a Norwegian humanoid robotics startup. There are two German AI robots and Neura robotics. So let's see what is what their answer is when it comes to Atlas. You're traveling up north again. Exactly up north to Starvanger. Many people traveling north.

18:42 Europe these days, but it's another topic. Although I do want to finish if I may. I do want to finish with that. They're all coming back. The topic is actually very close. You wouldn't understand. But if you, as the world is traveling to Davos, and Davos is not in the North, Davos is in Switzerland. I think, starting today with the World Economic Forum.

19:05 Of course, lots of AI in industry, interest activities, but I will not travel there. But I just started rereading. I should say listening this time to the famous book by the late German author, Thomas Mann. Oh, the so-called book. Of which, you know, in English, it's called The Magic Mountain. I'm listening it in... in German, right?

19:30 Now that is set in a Swiss sanatorium, right? Just above Davos. Exactly. Main character visits his cousin. He plans to stay three weeks, but stays for seven years. And if that's the, that's the story, very good, very interesting story. But at the, in the background, this is all about. it plays like more than a hundred years ago, right?

19:55 And it was just in the years before a breakout of World War One at that time. Exactly. We are the industrial AI podcast. We are also the education podcast. I just happened to read because... In the saw you sometimes hear about that they are going down, coming down the mountain and you go to Davos village.

20:18 and as I hear that and then at the same time coming out of the, out of listening to the book, I was here and this morning I actually heard that everybody's traveling to Davos. Let's see, maybe in two weeks we will tell you. Exactly, and I'm certainly, I'm currently reading a book about Thomas Mann escaped to France.

20:36 I will share that also in the show notes, very interesting. about his escape. First he was in in Lugano, then he went to France. So it's very interesting. Yeah. Oh, you got as you said, here you go. And that's the iPodcast. You can learn all about literature. What's what's happening in the past? Education podcast, literature, education, whatever you want.

21:01 Exactly. Robert, we'll we'll see each other tomorrow. I have one more. We will see each other tomorrow in Amsterdam, but I have one more reading tip from the Financial Times and there's a connection to your humanoid. And the message is or the question is why we are not seeing the chat GPT moment in robotics now.

21:20 So this is very interesting. I will share this link also in the show notes. And in the main section now, we are in Amsterdam and I have a episode from the Netherlands. I met Wouter from VVTI and he's waiting with his cucumber robot. you remember right? oh really yeah. and now the cucumber robot serves industrial applications and that was exactly what you said.

21:46 and back then we will see how how to transfer development from one domain to another domain. And this is an interesting approach because the cucumber come robot is doing now different stuff when it comes to industrial application. I'm very much looking forward to, you know, because the thought was, yeah, me being from the Netherlands, very much looking forward to travel there tomorrow.

22:09 I don't come there that often. and having been kind of raised in an environment of agriculture and all these plans that grow. And that's what he was explaining the last time, one or two years ago, right? And these plans and then the robot needs to decide if, you know, what, what am I going to take? Let's stick with the cucumber and needs to be careful.

22:31 And then these, these leaves need to be taking off, but there is a very important leaf. And so there's understanding as well. And that was the thought at that time. Not sure it was my thought. It was, I thought it was. his thought maybe it was your fault. you mentioned that in the podcast. oh really okay yeah because it's so complex and and if and if they can solve these complex things then now they can move it to you.

22:52 okay i'm very much looking forward to him. i don't know the english name but it's all about biegel schlaffer teile. oh i don't know the english name either. what is biegel schlaffer teile? i don't know. i have no idea. if you don't know it in german robert's. okay perfect peter thanks bye. see you tomorrow bye bye.

23:10 My name is robert weber and my guest today is water from vbti. Hello water welcome. Hello robert Thanks for having me. Glad to have you with us. Yeah, it's nice to be here as well. I've never been on a podcast before, so it is quite exciting for me. Oh, great. But some of our listeners are already familiar with the company.

23:28 I would like to remind you guys of the episode with the cucumbers and the robots. But today we want to talk about Adam one. I'm really looking forward to what is Adam one now. But before we do that, Walter, please introduce yourself briefly to the listeners. Yeah, yeah, my name is Wouter von der Bämpt.

23:47 And I've been working at VPTI now for four years. My background is in engineering. I did mechanical and also software engineering. But now I'm working also on product development, specifically our inspection robot for high tech called Adam One. I'm really looking forward. So what is Adam One? I read your LinkedIn post and it says Adam One doesn't just automate inspection, it understands inspection.

24:13 What does that mean? Yeah, it's something really new. So what we see a lot is that basically anyone can make a inspection robot, right? But our robot is very different because it's very smart. So compared to other robots, you don't have to program item one. You can just place any part, press start, and it will just inspect whatever you put on the table.

24:39 And that's something very new. Mm-hmm. What did you learn from the cucumbers? or is there no knowledge transfer from the other projects? Yeah, there's actually a lot of knowledge transfer here. So in the cucumber project we had to deal with a very complex environment where we had to track which part of the cucumber plants are leaf stems and which part is the main stem so that we can cut them.

25:02 And this was very challenging because normal treaty estimation technologies like stereo vision wouldn't work in this complex environment. So in the greenhouse project we designed our own treaty algorithm which is based on stereo vision. But on the background, it replaces the algorithm with AI. So we do the three D depth estimation.

25:25 And this is something we are using now in our inspection robot to really scan what is on the table. And that's why Adam is so smart. OK. Why Adam? Is this an acronym for something? Or is it only we need a name for a robot? Yeah, actually, we have two reasons for it. So first reason, Adam is one of the optimizers used in deep learning.

25:48 It's one of the first optimizers used. So this is very cool to reference to that optimizer. And the next thing was, yeah, we wanted to make a robot that's so smart that you cannot give it a number or a name, but we had to give it a real name, a person's name. OK. Great. Can you please share a little bit what you did when it comes to AI about the model and how robust it is?

26:12 How do you deploy it? What is the model capable of? So what kind of cameras do you use? So can you please share a bit about that? Yeah, maybe a question back then. Which model? Because we have quite a lot of models in Adam. How many models do you have? I think right now we're using five. Isn't that too complicated to run five models on a robot?

26:34 How do you handle that? It's okay. So we have one quite beefy computer on the robot and there's a powerful graphics card in there. Okay, GPU and the robot? Yeah, and this graphics card can run all those models with ease, I would say. And that's because vision models are not as big as those language models everyone is using nowadays.

26:56 So you can put quite some models on there. Okay, let's talk about the five models. Start with the first model. What is the first model doing? Yeah. So first of all, if we press start, the first model that will do his job is the depth estimation model. So we start by taking some pictures and our depth estimation model will use those pictures to make a three D point cloud of the object that is on the table.

27:19 That is not rocket science, right? So this is a common way. Yes, exactly. But afterwards, we have all our inspection tasks running. So there's various functions that anyone can complete. And one of them is scratch detection, for example. And for scratch detection, we use those pictures and we process them with our models to find all the scratches in the image.

27:40 Okay, that's also not a rocket science. also common approach, right? Yes. Let's dig a little bit deeper when it comes to very, let's say unique approach. Yeah, I think one of the more interesting technologies that we are using is for the development of our missing parts algorithm. So missing parts is quite a complex problem.

28:01 It's basically you have some assembly and maybe there's a bolt missing somewhere. And we want to be able to detect that with just one sample. And we've seen that you can train AI models if you have a thousand annotated images. There's no issue there. But it's a bit more complicated if you go to one shot learning where you have like one golden sample.

28:22 And then afterwards, you have to already judge is this OK or not OK. How do you do that? Yeah, so right now we are using approach, we're using Siamese twin networks. What is that? I've never heard that. Yeah, it's actually a little bit of an older method and in this network you train a model to basically tell if something is the same or different.

28:50 Yeah, and we train this model across all kinds of images in the inspection and manufacturing domain. And afterwards, we can give it two images and ask the model, is this the same or is it different? Okay, that's interesting. And what else is unique in your approach? And how do you combine all these models?

29:08 And how efficient is that running a GPU next to a robot? Can you share a bit about that? Yeah, which question first? How efficient is it to run a robot with a GPU? Yeah, that actually works quite well. So in general, we see that our robot is a lot slower than the GPU. We have no issues processing the images in time.

29:30 But is it a standard robot? Or is it a specialized robot arm? Or what is it? Right now, we're just using a cobalt. Okay. But from an arrow eye perspective, you have a GPU and you have to sell the GPU. the robot? isn't that a problem at the end? No actually no because most of the cost for us to develop one system goes into the software development.

29:53 So the GPU and the robots together, they have a price, but most of it goes into the R&D to develop all the models and all the smart systems that we make. Yeah, but when you want to sell the robot, right? So I think it's a challenge to sell a robot and the GPU, right? It's actually okay. So usually we sell the whole thing as a solution to a client.

30:14 And that works actually quite well. Okay. Can you share a bit more about the USP? So you already mentioned the first USP with the model. Do you have more AI USPs on the robot? Yeah, so our main unique selling point compared to the other robots is that you don't have to program it and that with one press of the button you can get your part inspected.

30:38 So we've not seen any robot on the market and we've also not seen any partner who has such a robot. So we think this is a really unique selling point of our application. But how do you do that? Because that's a very challenging problem, right? And you mentioned nobody solved it yet, but how do you solve it?

30:59 Yeah, so we solve it just by the combination of techniques we have thinkered on in the past few years, right? So one is the depth algorithm, but other are some algorithms to determine where to take a picture most efficiently and combining all this knowledge into one product. Yeah, made it work. And maybe it also has to do with that.

31:21 We are based in Eindhoven and Eindhoven, we have a lot of the semicolon companies, and they are all dealing with these high mix, low volume parts that are they produce. And this is maybe something also a little bit unique to the region, because you also see a lot of manufacturing sites where parts are produced in large quantities.

31:40 And for that, our solution is totally not interesting. So your focus on high tech industry or am I wrong? Yeah, right now we're focusing on a high tech industry, but We do see that the robot can be applied in many industries. But high makes low volume is the main case where it's very interesting to use this technology.

32:01 Because that's the moment. If you only need to inspect maybe ten parts, you're not going to automate it and make a complicated program for it that takes maybe weeks to develop. You're just going to inspect it by hand. And we have a solution that you don't need a program, which makes it... now for the first time I think feasible to inspect high mixed low volume parts.

32:26 You're also connected to the ASML guys there. Is there a connection or partnership with ASML on this topic? Not yet. So I want to come back to your five models. I think we stopped with the third model. Maybe you can share also what is the fourth and the fifth model. Yeah, so we're also running a vision language model on our robot.

32:49 This is quite an interesting model because this allows actually to just let normal people write what they want to inspect in just English or Dutch or German text. It's simple as that and that's very cool. because Yeah, you see what chat GPT can do. You can give it an image and you can get all kinds of questions answered.

33:10 But we want to make a vision language model that understands all domain knowledge of manufacturing. Okay. Yeah, exactly. But how do you train that? Yeah, so we bought a very powerful GPU here at VBTI and we are actually obtaining comments from inspectors and images also as well. And we use them to train our model to understand better the domain of quality inspection and manufacturing.

33:43 Do you use a standard VLM out of the box or what is behind this VLM? Yeah, we're using. So for this one, we're using some out of the box VLMs at this moment. So it's also a little bit of a research topic. I think at the moment we're using Moondream, not quite sure. And how good does this model perform to train when it comes to domain specific topics?

34:04 Yeah, that's an interesting question. So this is right now for us a little bit of research. So in the base version, we can already get some quite some things done. But right now we're, yeah, we have some small data sets that we made ourselves. But yeah, we hope that if the robots gets a wider adoption in the market, we can build some very cool tools around this by, yeah, basically collecting data and making the models better.

34:27 So is your plan to build a let's call it if inspection foundation model or what is the idea? Yeah, that would be amazing. Yes. So that you collect so much data that you can build at the end the inspection foundation model on high tech industry stuff, right? Yes. And it really understands all the specific names that people use here.

34:49 And yeah, you won't find that in chat GPT because that's trained on public websites and not not on this internal knowledge of all these companies. Sure, exactly. But how open are your customers to share this? comments and names and topics and with you to build or to collect data for for your inspection robot and at the end building a foundation model for inspection topics.

35:13 Yeah, so until now, the consensus has been a bit that most companies are okay with it as long as their data is handled very securely, right? So we don't want this data to end on any server, not on any cloud provider. We keep it very securely, and we only train with it. And no one else ever gets to see this data.

35:35 So you make some very clear contracts about how you use the data. And you mentioned the VLM, and now the fifth model. What is the fifth model? Let me think. We almost had everything. So it's only four. Yeah, we are working also on a model to detect stains and that is also just a simple supervised model in some ways.

35:58 But yeah, the optics are quite different here because you want to use a bright light or maybe some UV light. But when it comes to hardware, let's talk a little bit about the cameras. Do you need special settings when it comes to the camera? Yeah, so right now we're using actually a very simple... RGB camera, five megapixels.

36:16 Oh, wow. Yeah. And actually, most of the magic is in the software. And we can already find very small scratches with this, like hundred mu or something, like hundred microns. That is okay. But we are doing research on how we can improve our detection. So we are looking at maybe using monochrome cameras or other types of cameras, but especially lighting is important.

36:38 Exactly. Yeah. Because what we see with scratches, for example, yeah, a scratch. You can see it usually from a certain angle, but if you rotate the part, maybe you cannot see it. So you really have to have a couple of lighting directions to really find all the scratches. And that is something we really want to upgrade in the next iteration of this product.

37:02 Coming back to train the model or to train one of the five models, how difficult is it for you to get enough data to train the model? because In the industrial sector, we do not have so much scratches on the parts, right? So do you work with synthetic data or generated data? What about your approach there?

37:24 Yeah, we did encounter this problem. So first, it's not so easy as cucumbers. No, you cannot just go around and grab a couple of assemblies and put them full of scratches. That's going to be very, very costly. So yeah, we are using a combination of synthetic and real data. And synthetic has been proven pretty useful, but always you have to be a little bit careful because sometimes it seems realistic to you.

37:51 to a human but AI models they can really sometimes look at very weird patterns in the data that yeah you don't really see as a person. so sometimes it works very well and sometimes it doesn't. but this combination of real and synthetic has proven very useful. How do you generate this kind of data? Yeah so we have our own pipelines there.

38:13 we're using mostly open source tooling connected together. One of the core components is Blender. So in Blender, you can apply all kinds of scripts to basically automate how you render certain parts. Yeah, we're using this to make all the accurate real-world lighting. So you mentioned it's a cohort application, right?

38:36 So you do not need a fence or something, right? yeah so right now we're shipping it without a fence but in the next version we may include a box and the reason for the box is two-fold one. you cannot accidentally hit a person. i mean a cobalt is pretty safe but still, it's not nice to get tapped on your shoulders from time to time.

38:57 But also we want to block out some of the external light. So we see sometimes that if people have factories that are pretty dark, some factories are very light, sometimes next to the window. Okay, that's interesting. But you are robotic agnostic, you can work with every robot on the market or am I wrong?

39:13 Yeah, yeah, pretty much. We actually had a really interesting conversation today with someone about maybe we should use a humanoid robot instead of this robotic arm so we can inspect very big objects. Oh, that's a long way to go. Yeah, that's a long way to go. But in the end, if we can control the humanoid and just take a picture at a certain point, there is no nothing holding us back from from doing this with the same software stack.

39:38 Yeah. What are so special? From my point of view, it's a software stack. And how do you combine these five AI models? This is your, your usp at the end, right? Yeah. So is it an option for you, as you mentioned, humanoid robotics to only to sell the software stack in the end, because this scales like hell when it works, if it works, because if you sell a robot or a robot in a box, and you have to integrate the whole safety stuff, etc, etc, etc, and you have to integrate the solution into the shop on the shop floor, etc, next to the machine, blah, blah.

40:14 So is it an option to say, okay, let's scale with the shop only with the software stack? Yeah, that could be an interesting idea. So the reason why I Yeah, I think we don't really do. that is because we are solving customers problem. And for that, robots is essential parts, right? So we wouldn't be selling to the manufacturing companies we are selling the robot to right now, if we would only offer a software.

40:43 So the full solution is where the value is for those. But it could be an option to sell the software stack to the robotics companies or today to the camera guys or something like that. Yeah. could be interesting approach also maybe to spread further than Europe. Exactly, because because in the moment you are more let's call it an integrator, right?

41:05 So you you do projects with your customers and you have a great idea. But the next step is the idea how to scale this software stack at the end, right? Yeah, that's gonna be interesting. So one of the nice parts about is that most of the hardware we can just simply buy because as I told we're right now just using a simple camera off the shelf cobalt and a frame that is manufactured for us.

41:27 so Yeah, potentially we can already ship quite some systems right now, even though we don't have a very big manufacturing department. Absolutely. Coming back to the beginning of our discussion, talking about the cucumbers, are you still focusing on let's call robots for farming greenhouses? Or is it now really fully industrial AI applications?

41:49 Or are you still in the greenhouse field? Yeah, we're actually in both fields. And sometimes I say jokingly that here in the region and in Eindhoven region you have farmers and high tech and that's what we do as VBTI. But yeah we do apply a little bit more focus now. so we do always choose projects in the agri sector but also in the inspection sector and manufacturing.

42:14 preferred types of projects that we can reuse a lot of our existing technology in. Thanks. Thanks a lot. Well, it was a pleasure to talk to you and to get an insight how you transfer your knowledge from the Cucumber's case now to an industrial AI application. It was a pleasure. We keep fingers crossed for you and Adam one.

42:35 Do we will see Adam two or is it Adam three, four, five? Do you have some plans to do that and to go in different domains also? Yeah, we have plenty of plans, but you have to stay tuned and subscribe. You can share one idea. What is Adam Two doing? We are considering setups with more robots or more automated systems where you have conveyor belts, drones or humanoids, those kind of directions.

43:06 Mostly a different form factor because we see the table with the cobalt. can address quite some parts, but sometimes you need to go bigger or more mobile and then we can just reuse the same technology on a different form factor. Exactly. I'm really looking forward to hear more from you. It was a pleasure of ours.

43:23 Thanks a lot and greetings to the Netherlands. Thanks a lot Robert. Thank you.