KW 49: Robotik-Häppchen

00:00:00: Robotik in der Industrie, der Podcast mit Helmut Schmidt und Robert Weber.

00:00:09: Guten Morgen liebe Sorgen, ihr habt natürlich keine Sorgen, denn hier ist das Robotik-Häppchen

00:00:17: um 6 Uhr am Montagmorgen.

00:00:20: Mein Name ist Robert Weber und wir suchen einen neuen Partner für das Robotik-Häppchen.

00:00:26: Wenn ihr also Lust habt, Partner zu werden, dann meldet ihr euch bitte bei Helmut oder

00:00:30: mir und jetzt geht's los.

00:00:33: Hallo liebe Zuhörerinnen und Zuhörer, das ist das Robotik-Häppchen und wir gehen in

00:00:40: den Urlaub.

00:00:41: Mitte Januar sind wir wieder mit News am Start.

00:00:43: Aber es kommen noch reguläre Robotik-Folgen, wir haben noch ein Jahresrückblick, auf den

00:00:49: ihr euch freuen könnt.

00:00:50: Und für nächste Jahr suchen wir auch noch einen neuen Partner für unser Häppchen.

00:00:54: Heute haben wir noch einen Happen für euch aus dem Industrial AI Podcast zum Thema Safety

00:01:00: und AI.

00:01:01: Ein sehr cooler Ansatz.

00:01:02: Das war es von der KI, euch eine gute Zeit.

00:01:05: Und jetzt geht es los.

00:01:08: Hi there, welcome to a new episode of the Industrial AI Podcast.

00:01:13: My name is Peter Sieberg and I'm your host.

00:01:17: Today we'll be talking to Stefan Milz.

00:01:20: He's the founder, managing director and also the head of R&D at Spleenlab.

00:01:26: And Stefan and I are going to be talking about SAVE AI and also towards the end their first

00:01:34: endeavor with XLSTM.

00:01:36: Hi Stefan.

00:01:37: Hi Peter, thanks for having me.

00:01:39: You're welcome.

00:01:40: To start, please introduce yourself to our listeners.

00:01:43: I keep it short, so I'm Stefan Milz.

00:01:46: I have a technical background.

00:01:48: I'm the CEO of Spleenlab, also the founder, a technical background in physics.

00:01:53: So I have a PhD from the Technical University of Munich.

00:01:57: And then I worked several years in the automotive industry in highly automated driving for the

00:02:03: continental introduced.

00:02:05: They are working on the first NVIDIA series production project ZFAS and then for Vallejo

00:02:11: on highly automated driving.

00:02:13: During that time, I had the idea to introduce something new on chip AI with a focus on functional

00:02:21: safety and I guess founded Spleenlab.

00:02:24: And this is the reason why we are sitting now here.

00:02:26: That's right.

00:02:27: On chip AI, I guess we're going to be talking about that.

00:02:30: So, to start, what does Spleenlab stand for?

00:02:35: You mean the name itself?

00:02:37: Yes, exactly.

00:02:38: Okay, that's a funny story.

00:02:39: I believe so.

00:02:40: Yeah, I thought it this morning when I looked.

00:02:43: It's a very funny story.

00:02:45: And originally I started the company in 2015 as a pure site project, right?

00:02:52: And Spleen is a translation of my surname, Milz.

00:02:56: So it only makes sense if you understand the German language.

00:03:00: No German, right?

00:03:03: And then this was needed in a fast way a name, right?

00:03:08: And then when we really started with Spleenlab, we wanted to have a different name and then

00:03:13: we thought, okay, this is so unique.

00:03:15: You find it everywhere.

00:03:17: Let's keep it because there is no mismatch with other companies.

00:03:22: And then we left it.

00:03:23: So literally my surname.

00:03:25: Right, your surname.

00:03:30: Also we not excluded from future, right?

00:03:41: As safe AI is a really important topic also in medical, but it's no focus at the moment.

00:03:46: Oh, you got.

00:03:47: Maybe we come to that.

00:03:48: I mean, I've been saying time and again that everything is growing together.

00:03:52: And that's an example of it.

00:03:54: So when we talk about, and that's what we are, industrial AI podcast.

00:03:59: We do talk about industrial.

00:04:01: Today we talk maybe a little bit outside even.

00:04:04: You already mentioned here autonomous.

00:04:07: We're going to be talking about your visionary approach.

00:04:11: And at the same time, I would do the same thing.

00:04:13: I would talk about how a physician, a physical doctor may be using very similar technology.

00:04:20: So we will be talking about safe AI.

00:04:24: Let's start with a quick definition of what is safety maybe as opposed to security.

00:04:31: Exactly.

00:04:32: So this is also a fun, funny side note in Germany or in German we have safety as Sicherheit

00:04:42: and security also as Sicherheit.

00:04:45: So this question is quite good.

00:04:47: Right.

00:04:48: And when we talk about safe in the context of what we are doing at SpleenLab, then it's

00:04:52: meant to be functional safety.

00:04:54: So functional safety is something, if we automate processes with heavy machines or with sensitive

00:05:01: processes, then all the time a human is in the loop or there is a probability human could

00:05:08: get hurt, then we talk about functional safety.

00:05:12: And this is where we are working on.

00:05:15: There are real regulations in that field.

00:05:18: They are there for many, many years.

00:05:21: So in ADAS we have ISO 26262 and now for the software we have SOTIF.

00:05:28: But there are similar things in aerospace like the DO178.

00:05:33: These are real process definitions that define how to develop a software or a system that

00:05:40: needs to be safe in the context that it interacts with a human and no one gets hurt.

00:05:46: What I recall when I was more strongly involved is was saying that safety is making sure that

00:05:57: a machine does not hurt a human and security is more about a human maybe doing something

00:06:04: bad to a machine or to other humans I guess.

00:06:08: Yeah, and exactly.

00:06:09: And security is in the context of software really more meant what we classical know as

00:06:14: cyber security.

00:06:16: So this is the typical differentiation at the moment.

00:06:20: So what does SPLEENLAB do then?

00:06:24: Is SAFE.AI your central thing or is it one of many things?

00:06:28: What is your top approach?

00:06:31: So SAFE.AI is a very important thing where we are working on in the context of AI.

00:06:37: We all know that it's a really unsolved topic and there are many buzzwords out there like

00:06:42: trustworthy AI, debugging AI, whatever.

00:06:47: But this is a really important pillar because we develop software for autonomous machines

00:06:54: like autonomous drones.

00:06:56: For example we work together with DroneUp, a big company in the US that do delivery,

00:07:04: last my delivery of packages for Walmart.

00:07:07: We do precision landing and their goal is to remove an operator in that automated process.

00:07:13: So they need to fulfill specific requirements regarding safety when it automatically lands

00:07:18: or automatically drops a package over some area.

00:07:23: And here our product comes into the game.

00:07:27: So we have a redundant approach to deploy robotics software always with a part of AI.

00:07:35: Sometimes 20% AI, sometimes 40% AI, but at the end it's a robotics application that fulfills

00:07:42: a partially critical task and needs to fulfill some requirements regarding safety.

00:07:49: So that's why I would say it's a very important pillar, what we have here, because our business

00:07:54: at SPLEENLAB could only scale if we solve it.

00:07:58: Our approach is to do it iteratively.

00:08:01: So we don't want to solve the full autonomy yet.

00:08:05: So we say okay let's break it down in some partially automated tasks and take the big challenge

00:08:12: iteratively.

00:08:13: And this is our big vision and we're working on more, we're working on full safe AI algorithms.

00:08:19: We work on the full AI deployment which needs to be safe.

00:08:22: So there comes everything into the game.

00:08:24: So the most important thing is we're working on an embedded solution, but with a very, very

00:08:29: strong focus on the current safety regulations.

00:08:32: Right, we'll get into a little bit more detail just in a moment.

00:08:37: Maybe first tell us where, as in what markets are your customers.

00:08:44: You just mentioned Walmart, you could almost say delivery logistics.

00:08:49: Maybe also in industry, but not necessarily the majority.

00:08:54: So where are your current solutions, your current customers?

00:08:57: Yeah, this is a very good question.

00:08:59: So this also aligns with our vision.

00:09:01: So we really deploy cross industries.

00:09:05: So could be drones as a work vertical and industrial drones.

00:09:09: It's also logistics, last mile logistics was long term logistics with tracking.

00:09:16: We are also in the field of automated agriculture as we talk about automated tractors.

00:09:22: Also we are working in ADAS on the streets.

00:09:25: So we see us here complimentary, but there's also applications in the defense pillar.

00:09:32: So this is our vision.

00:09:33: So we can say that we are something like Mobile Eye in the drone industry.

00:09:38: We want to become it off-road.

00:09:40: And we are also complimentary in the ADAS street domain.

00:09:45: So these are our markets.

00:09:48: And here we have everywhere very famous paying customers.

00:09:53: Which maybe we come to in a moment.

00:09:55: So let's get into maybe a little bit more detail of your offer, which I understand.

00:10:01: Or one of your offers is called Visionary, Visionary Safe AI.

00:10:07: What is it and how does it work?

00:10:10: Exactly.

00:10:11: So the big challenge, if it comes to autonomous robots, is that every robot has a different

00:10:17: design.

00:10:18: But every robot has a different sensor.

00:10:21: And our approach is to keep it as generic as possible to support as many sensors that

00:10:27: are possible.

00:10:28: So we have solutions on industrial RGB cameras, but also on IR radar.

00:10:35: And then we have different solutions in the different verticals like agriculture, drones,

00:10:43: ADAS.

00:10:44: And deliver to many, let's say niche markets.

00:10:47: And our offering is to deploy all this software stack to different chips.

00:10:53: So there's NVIDIA, Qualcomm, TI.

00:10:56: So there's a very heterogeneous target market of robotics, but we serve them all with our

00:11:01: solutions.

00:11:02: And when it comes to the offering, then we can tell, okay, we can provide you an RGB-based

00:11:08: SLAM.

00:11:09: We can provide you an RGB-based visual navigation or a lighter-based sensor fusion.

00:11:15: This is exactly the offering.

00:11:17: We offer a spatial AI software stack for autonomy, in visionary sensor fusion, visionary object

00:11:24: detection, visionary free space and visionary SLAM, which then ends in real functionalities.

00:11:30: Let's imagine a drone does collision avoidance, a drone does GPS denied operation.

00:11:37: The parking car has the capability to predict free space on a very low-cost sensor, like a

00:11:44: low-cost camera.

00:11:45: Or we do sensor fusion on L3 autonomous tracking.

00:11:51: So this is our offer.

00:11:52: And where we're really strong is that we are focused heavily on the embedded part.

00:11:59: So we have a high-performance team.

00:12:02: And that's why we can serve many different chips, like what I told before.

00:12:07: And I guess this is a little bit the strong USB that we offer very fast in the integration

00:12:12: part.

00:12:13: Okay, yeah, we come to that a bit later.

00:12:16: Now, I understand that working with artificial intelligence, specifically neural networks,

00:12:23: because of their statistical property, and I read about it again this morning, a huge

00:12:30: discussion on LinkedIn.

00:12:32: Is your machine learning?

00:12:35: Is your large language model whatever element of the AI world we choose?

00:12:41: Is it really statistical?

00:12:43: Or is it maybe deterministic?

00:12:44: I understand that because at least it typically has the statistical property.

00:12:51: The neural network-based approach to safety has not yet been approved for safety requirements.

00:13:02: How do you deal with that?

00:13:04: That's a very good question.

00:13:06: This is exactly why we started Splenelab.

00:13:09: To explain this a little more, you fully introduced it right.

00:13:16: So classic algorithms that are in current automated, for example, driving systems, or for example,

00:13:26: an approved other industrial robots are highly deterministic.

00:13:31: And the deterministic nature of their system design helps to approve towards a, we call,

00:13:38: safety integrity level.

00:13:41: And this is also the pure idea behind functional safety, that you literally design a system,

00:13:48: that you can really say any possible situation that could happen.

00:13:55: You can deterministically describe what you do.

00:13:58: Very quickly, deterministic, same set of inputs always gives the same output.

00:14:04: Is that the right definition here?

00:14:06: Exactly.

00:14:07: Describe what happens, right?

00:14:09: And with a specific probability, but you know in which end, let's say, state you will land

00:14:17: and then something needs to be triggered.

00:14:19: And a safety function needs to be triggered, for example.

00:14:23: And this is the nature of how the current regulations are designed, like what I told you before,

00:14:28: right?

00:14:29: The ISO 262 or industry, it's the IEC 61508.

00:14:37: They really want some deterministic three, where all probability states are defined and

00:14:43: everything that could happen with the system will be described in the safety mitigation.

00:14:48: And if it comes now to a neural network, you typically have so many parameters.

00:14:53: So you all know this large language models, right?

00:14:56: Even 70 billion parameters of a medium-sized large language model is almost impossible to

00:15:04: tell in terms of safety mitigation what could happen if you try out all permutations.

00:15:10: So at the end, it's a statistical model.

00:15:12: You're fully right.

00:15:14: And then you need to design something that we call safety by design.

00:15:18: You need to design watchdogs.

00:15:20: What's typically done, you have a different hardware where a deterministic algorithm runs

00:15:26: parallel to the AI and that assesses the AI at the end of the day.

00:15:33: So these are typical safety patterns.

00:15:36: And only if it's really in the same area of their output, then the output is validated.

00:15:45: And what we do is exactly similar.

00:15:47: So we do safety by design, but we design software functions that assess the output of a neural

00:15:54: network.

00:15:55: And you may have seen this depth estimation post I have done also with the use of XLSDM.

00:16:02: It's a good example to explain how we do it.

00:16:07: So we include temporal consistency.

00:16:10: So we designed it also patented.

00:16:13: We designed some assessment function that deterministically describes the output of

00:16:20: this depth map, which comes out of a neural network, the confidence.

00:16:24: But this confidence comes not out of the neural network itself.

00:16:28: It's a parallel path and it's done with temporal consistency.

00:16:33: So let me explain it a little more.

00:16:35: So right.

00:16:36: It sounds like you made a step going from probabilistic towards deterministic.

00:16:43: Yeah, it's kind of a mixed thing.

00:16:46: Because we take over the approach of self supervised learning, where you have a loss function during

00:16:50: training, which needs no human input.

00:16:53: You transfer this idea to the runtime we call inference.

00:16:58: And then you use this deterministic loss function as deterministic safety watchdog.

00:17:04: And this works out quite well.

00:17:06: We have shown it for depth estimation in free space, for example, for partially automated

00:17:10: driving function.

00:17:11: And we are very sure to achieve an SLB, we call it safety integrity level B.

00:17:17: And this is quite a very important milestone.

00:17:20: Right, because that was then, of course, going to be my question.

00:17:25: So you're very close to receiving, and you are good hopes that you will get an approval.

00:17:31: What is that from one of those specifications that you will get?

00:17:37: What is basically a what neural networks based system will get a safety approval?

00:17:43: Exactly.

00:17:44: But really in a specific context, right?

00:17:47: Ich verstehe das.

00:17:49: with your pay-tented technology, you can use that technology in different other environments as well,

00:17:55: I assume. Exactly. That's a big thing then, actually, because the way we started discussing this,

00:18:02: and I think you're using the use case of autonomous in this case, but it sounds like you have kind

00:18:11: of solved this, what I guess has been an issue, a problem for a long, long time, that as you said,

00:18:20: the regulations have always been written on the basis of whatever, 50 years of deterministic

00:18:26: programming, and it sounds like you have found a solution towards getting approvals even with,

00:18:33: on the basis, probabilistic neural networks. Yeah, but I wouldn't call it so, it's not

00:18:40: for my point of view, it's not rocket science. It's really understanding what we are doing there.

00:18:47: I know it's a big thing, but this is maybe something you asked in the beginning, so where do we stand

00:18:52: for? And I believe we're very strong and embedded in the engineering, very strong at the algorithm

00:18:57: design, but we also have knowledge about the current safety regulations. And this triangle is

00:19:03: very important if you design a system. And there's another important thing we faced, right,

00:19:08: when it comes to AI on chip, you need to somehow execute the AI. And this execution of the AI,

00:19:17: for example, could be an ONX model or something, this execution, this software piece is also not

00:19:25: certified yet. And this is also something we are working on, because if we solve the overall

00:19:30: thing, then we also need to have a safe execution on the chip. I will come back to this, I will come

00:19:35: back to SLAM you mentioned, I will come back to XLSTM in 10, 15 minutes, if that's okay with you.

00:19:42: Before that, I do want to look into one or two other use cases, maybe you talked about, you know,

00:19:51: specifically autonomous here, I understand there is this inflight use case, yeah, for well,

00:19:57: and maybe then let's move to SLAM, because SLAM is a technology or an approach that I know,

00:20:04: typically from robotics, but you've been using it, or one of your customers has been using it for

00:20:11: inflight use case, can you talk about that a bit? Exactly. So yeah, when it comes to drones,

00:20:16: operations, automated mission, for example, also for delivery, but also in surveillance drones,

00:20:23: right, we have a big customer quantum systems, or it's a big partner. For the inflight use case,

00:20:30: so this SLAM solution is something we call visual automotry, or if you map everything,

00:20:37: it's a SLAM. And at the moment, the system only relies on on GNS signals and IMU. And this makes

00:20:46: the system dependent on on outer, outer conditions on an auto signal, or let's say on an active signal,

00:20:53: and to do the operation more safe, and then we are back into the safety case, right, we need one

00:20:58: more input. And that's where we come into the game. We provide SLAM as a functionality that can

00:21:05: provide a geo reference position. And all together makes the overall inflight of an automated drone

00:21:11: much more stable, much more precise in its mission planning, because it's a very, very stable

00:21:17: geolocalization. And SLAM is helping the vehicle, maybe the drone in this case, to know where it is,

00:21:27: or to first map exactly an environment of what is around it and of where it itself is in it,

00:21:35: is that right? It's an exact definition. So, so typically we differentiate between perceptions

00:21:41: of what is around me and localization where I am in the robotics field. And SLAM directly points

00:21:49: to the second one. But as you told fully correct, so you use signals, like a pure camera signal,

00:21:57: and you try to create a 3D understanding of the surrounding due to video stream. So, let's say,

00:22:05: temple approach. And based on finding good features, you design a 3D map. This map could be sparse,

00:22:15: or could be very dense, so that different AI approaches out there. And based on this, you can

00:22:20: estimate where you are. And then you have a kind of understanding in your map, where you are. And if

00:22:27: you match the measured map, you get out of your SLAM algorithm with a global, global map, then you

00:22:35: can get a position of your robot and global coordinates. So this is at the moment a very

00:22:42: important approach, because it enhances safety and it also improves the accuracy.

00:22:48: One customer maybe you can share with us is understand Deutsche Bahn, the German railway.

00:22:54: What are they using this approach for? At the moment, they are going into surveillance, right?

00:23:01: They want to see something on their rails, right, to see if there could be a tree or something. And

00:23:09: therefore, they need an operation system that is kind of very precise in its fully automated

00:23:16: perception approach. Therefore, they use for short drones following the rails,

00:23:21: and they need to be precise. And there are many, for example, many, how could we call it,

00:23:27: signals that are toxic for a pure genius approach. And this makes it more safe and more precise.

00:23:35: Okay, yeah, I live actually close to a railway track. And every day I do actually pass it a

00:23:41: couple of times a day. And typically, I've always been on the outlook of one of the drones of Deutsche

00:23:47: Bahn to see. So I'm looking forward to seeing them one of these days. You mentioned already,

00:23:54: XLSTM. So you had a first endeavor, I think your colleague maybe yourself, XLM, and how to apply it

00:24:02: in industrial setting. I was going to ask you also for the SLAM because you have been using the word

00:24:08: robot a couple of times, maybe in combination. Let's first, what have you been, why did you get

00:24:13: into XLSTM and what did you get out of your deep dive? Yeah, I guess it's, again, a great question.

00:24:22: So also, if you take into account what we discussed before, as I told you, it's all about

00:24:28: temporal consistency, how we treat the safety, right? And we have seen transformers and for us,

00:24:36: especially vision transformers outperform CNNs and the results are really great in the recent years.

00:24:44: But a little drawback is how the transformers are designed. So they split all the input into

00:24:50: embeddings and you either focus spatially on some image where you have some image patches,

00:24:57: you order and put into the transformer, or you have some temple signal like video stream.

00:25:04: But here it's only a fixed focus, right? And it's very small. And the XLSTM has the capability to

00:25:12: have a much wider look into the past, which makes it very, very, very designed, very, very,

00:25:18: relevant for the approaches I told you before. So processing of video signals over time with

00:25:24: the scope of functional safety. And this was our main idea to use it because we think this

00:25:29: temporal consistency approaches, they match very well with the nature of XLSTM. And that's why we

00:25:35: moved into it. And we see lots of challenges are still out there. But what we found that, yeah,

00:25:41: on our first glimpse, it's on par, or at least in the same range with vision transformers for

00:25:48: our first trainings, for example, for the depth estimation. But we also want to check it out

00:25:52: for free space and others. But then there's still work to do. But the main reason is really the

00:25:57: temporal consistency, because as I explained, this is the nature of our safety approach.

00:26:03: And this matches very well with the design of XLSTM.

00:26:07: Okay, sounds good. And in addition, I understand XLSTM may be more suited for your embedded

00:26:14: deployment as well and faster actually in a time of inference or also training time.

00:26:21: Yeah, it's more suited. It's more suited. There's a very, very strong advantage. So it's the kind of

00:26:27: processing dynamic size of input parameters. So if we come back to the SLAM, for example, if you

00:26:33: have imagined you're in an environment, less number of features, right? For features, you can

00:26:38: match with the next frame, or you're in an environment with many, many numbers of features.

00:26:43: You have some different size of input into a specific neural network based on some backbone.

00:26:50: And here it could suit very well because it's per design capable of processing dynamic input sizes.

00:26:57: Again, when it comes to different resolutions or all these things, so there it really makes

00:27:02: sense because on a transformer, you only have a fixed size of inputs, you go workarounds and

00:27:07: this works quite well. But here for it's very, very well suited. And on top is that's what the

00:27:12: paper proposed. It could be more efficient on the onboard execution and because transformers are

00:27:19: expensive, expensive in the sense of costs we need to execute. And this is at the moment shown in

00:27:27: the papers. At the moment, it's not true on the device because it still needs works to do that

00:27:32: make the XLSDM more efficient on execution. I mean the implementation, right? Because there's

00:27:39: lots of work done on transformer side, but per nature, it's more efficient. And then it could

00:27:45: run faster, then we have more safety that makes it very well suited. Okay, one final question

00:27:52: towards us being the industrial AI podcast. As we talked about SLAM as you've been using it,

00:28:00: and you've been using the term robotics a couple of times, maybe in a different sense, but sure,

00:28:05: I mean, talking about robotics in an industrial environment, could your approach be used in such

00:28:14: industrial robotics environment as well? Yeah, it could be fully abused. At the moment,

00:28:22: we don't have so many customers in that field, but it's for sure an idea we want to go also in

00:28:28: warehouses. We already had projects also together with drones to be honest, but really on fixed

00:28:34: robots, it could be also used. So our approach is agnostic. So we are agnostic towards the sensor.

00:28:40: We are agnostic towards the deployment on the chip. So yes, it could be used. And the good thing,

00:28:46: the safety is also towards industrial safety standards, which I told before is very similar

00:28:53: to automotive safety or medical safety, or even aerospace safety. And that's the fun fact,

00:28:58: these mitigation strategies are very similar. Yeah, you explained this, that your pay-tent

00:29:05: approach allows for moving from a probabilistic to a deterministic approach. Now, as you mentioned,

00:29:14: I had in my, I think I've seen maybe some, some, some videos somewhere of drones in warehouses.

00:29:22: What would be if at all you say you're agnostic? So which means in the end, you don't really care

00:29:29: as a big word, but assume you're open for customers also for listeners, customers in a typical

00:29:35: environment, industrial setting. But what would you say? Because we have heard lately so many

00:29:42: times about the humanoids and even humanoids in an industrial environment. And we've had this

00:29:47: discussion a couple of times. So I have a fixed robot or I have a robot, which is like a human

00:29:53: or have a robot, which maybe I put on wheels, doesn't look really human anymore. But then

00:29:58: there is the, there is the option of drone. As you say, you have any view on where you see

00:30:05: the market moving kind of thing is drones going to be a the thing. And because we have drones,

00:30:11: we don't need kind of humanoid or fixed robots. It's a really tough question. So what I have seen

00:30:18: is that drones are always underestimated and simply by the reason they could fall down.

00:30:23: And please trust me, this is really the thing. Everyone thinks drone is very easy, but this is

00:30:29: not to be honest. And in the industrial setting, you rely purely on, on slam, right, slam in the

00:30:36: sense of, of localization because there is no geo, geo reference, the GNS signal. So yeah, I, I

00:30:44: believe there's still some work to do with drones in the industrial environment. Some drones are

00:30:50: wired drones and they are combined with a ground robot. And this was also something we did in the

00:30:56: past. And I believe some teaming in the sense makes, makes sense. But you also told the human

00:31:04: need approach is, is a thing. And I don't have a clue, but I feel they turns it, it turns to the

00:31:12: human needs because nothing else needs to be changed. I have no clue how expensive they will be.

00:31:18: But this is my opinion on that. Yeah, I've heard that, that reasoning for why, but it's a different

00:31:25: discussion. I'm not sure I, I can understand that in a transformation approach, you know, for

00:31:31: whatever the next couple of years until so we can use existing factories, but I'm a stronger

00:31:37: believer in, you know, if we, if we have a new, a structural new way of doing things, we can make

00:31:44: new, you know, factories. And if in the new factories, we're going to have, it doesn't matter

00:31:48: what we're going to have. Maybe we're going to have drones could be, and we're going to design

00:31:52: that factory with drones. So then yes, in the meantime, we can have humanoids because, you

00:31:58: know, factories have still been built for machines plus humans. Different discussion. Tell us about

00:32:05: your team, where your base, maybe you're looking for new colleagues. If so, what should they bring?

00:32:09: Yeah, oh, this is something I need to emphasize, right? The team is the greatest part of Splenelab.

00:32:16: And we're really talking about our software engineers. We are currently about 45 people at

00:32:23: Splenelab located in Jena. Jena is a tech city in central Germany. You can come with a very,

00:32:30: very traditional university and a very strong, let's say university landscape around. And this

00:32:38: is where we founded Splenelab because the core team settled down there. And then it was obvious

00:32:46: to start there, but it was the right decision because there's a good location. There's a strong

00:32:52: university that other big companies like ZEIS, Yeen Optik, but also in video certified camera

00:32:59: producers like Light Vision. So there's also a good ecosystem. And the most important thing is

00:33:04: that we have a loyal, strong team. And yeah, it's an international team, but mixed of robotics

00:33:11: engineers, AI engineers, and pure software engineers. But this is mainly the feedback.

00:33:18: And we are always looking for new people. We have no specific roles open because we believe in a

00:33:25: way to insert people that they can really put their most important strength into the company that

00:33:35: it's most worth for us and adds the most value. So it's most efficient. And this worked out most

00:33:42: of the time very well. And if someone is really interested, then he can always apply. We have

00:33:47: some open roads, but the main meaning is to attract it to people that may feel that they're

00:33:53: good in the space. I understand. Thank you very much for that. So if one of you dear listeners

00:33:58: feel maybe you would like to work together with Stefan and his team in this great safe AI

00:34:06: environment, you can best contact him on LinkedIn, Stefan Milz, M-I-L-Z. And that's the German word

00:34:15: for what we say, spleen, right? Exactly. Yeah, it's one of your ways of people. I think you're

00:34:24: right. It's one of the ways that people that are going to know exactly who you are. So otherwise,

00:34:30: if you're dear listeners, if you have any other question, comment as always, please send a short

00:34:35: email to peter@aipodde. I'm very happy that you stayed with us so far. Looking forward to have

00:34:42: you with us again. Stefan, thank you very much for your time. One side note, maybe. Please. We are

00:34:49: really looking for some salespeople. This is where we are. Oh, okay. Not too strong. So if

00:34:54: someone is in the sales field, either at us or agriculture or whatever field or has really the

00:35:02: ambition to work with us. So this is something where we are not well occupied. And I think

00:35:09: it's a great opportunity. Thank you, Stefan. Noted. Take care and bye-bye. Thanks a lot for having me.

00:35:15: Was great discussion. Really good questions. Thanks a lot. Really good answers from your side.

00:35:20: Thanks. Thank you, Stefan.

00:35:23: Robotik in the industry. The podcast with Helmut Schmidt and Robert Weber.

00:35:31: (upbeat music)

00:35:34: [ GENTLE MUSIC FADES OUT ]