KW 41: Robotik-Häppchen

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

00:00:09: Guten Morgen, liebe Sorgen.

00:00:14: Nein, ihr habt natürlich keine Sorgen, denn ihr kriegt jetzt die wichtigsten Robotik-News

00:00:18: immer montags um 6 Uhr von Helmut und mir.

00:00:22: Hier ist der Podcast Robotik in der Industrie mit unserem neuen Format und Partner unseres

00:00:27: neuen Robotik-News-Format "Is Unchained Robotics".

00:00:31: Wir sagen vielen herzlichen Dank und Grüße nach Paderborn.

00:00:34: Willkommen zum Robotik-Häppchen in der Kalenderwoche 41.

00:00:40: Es ist gerade etwas mau, wenn es um News aus der Robotik geht.

00:00:44: Deshalb heute nur ein paar News und dann eine längere Geschichte im Häppchen.

00:00:48: Dazu später mehr.

00:00:50: Starten wir mit den News.

00:00:52: In den USA wird über die neue Robotik Roadmap diskutiert und die Themen ähneln doch sehr

00:00:58: stark der deutschen Diskussion.

00:01:00: Es braucht mehr Geld, ein zentrales Institut und KI soll eine wichtigere Rolle in der Robotik

00:01:05: spielen.

00:01:06: Aber es geht auch um die Ausbildung der Menschen und die Rolle neuer Wettbewerber aus Asien.

00:01:12: Interessant ist ein Zitat, heute ist uns der Rest der Welt voraus.

00:01:16: Wir müssen die Ressourcen, die wir haben, für Wettbewerbsfähigkeit und Technologieführerschaft

00:01:20: organisieren.

00:01:21: Sagt Hendrik Christensen von Contextual Robotics.

00:01:24: Er ist einer der Co-Autoren des Reports.

00:01:27: Und wir bleiben in den USA.

00:01:30: Kaum einer wird sich jetzt wundern.

00:01:32: Die US-Armee hat mindestens einen Roboterhund mit einem Geschützturm und mit künstlicher

00:01:37: Intelligenz in den Nahen Osten geschickt, um ihn als neue Fähigkeit zur Bekämpfung

00:01:42: von Drohnen für US-Soldaten zu testen, wie Militärangehörige bestätigten.

00:01:46: Zurück in die Zivile Welt.

00:01:49: Dexory hat eine Series B mit 80 Millionen Dollar für autonome Roboter abgeschlossen.

00:01:54: Das Start-up aus dem Vereinigten Königreich hat seine autonomen Lagerroboter im März

00:01:59: 2023 auf den Markt gebracht.

00:02:03: Seitdem soll ein großes Wachstum stattgefunden haben.

00:02:06: Dexory spricht von einer 5 bis 10-fachen Steigerung in den letzten 12 Monaten.

00:02:11: 60% davon stammen derzeit aus Großbritannien und Europa, die anderen 40% aus den USA.

00:02:17: Mittlerweile werden die autonomen Roboter inklusive dem Betriebssystem laut Dexory

00:02:22: in rund 50 Lagern auf der ganzen Welt eingesetzt.

00:02:25: Die Branchen reichen dabei von Logistik über Pharmazie, Fertigung und Automobilbau bis

00:02:30: hin zur Luft- und Raumfahrt.

00:02:32: 50 Roboter in einem Jahr, Schapo.

00:02:35: Und dann haben wir noch ein paar Event-Ankündigungen.

00:02:39: Am 15.

00:02:40: Oktober gibt es ein Webinar von Fruitcore zum Thema Kleben und Dosieren.

00:02:44: Link ist in den Show-Notes.

00:02:46: Wandelbots lädt zum Launch-Event nach Dresden ein.

00:02:49: Termin ist der 4.

00:02:50: November.

00:02:51: Link in den Show-Notes.

00:02:52: Ihr merkt selber, die News sind in dieser Woche ragesäht.

00:02:56: Deshalb haben wir uns noch was Besonderes ausgedacht.

00:02:59: Wir haben uns einfach ein Interview für euch geklaut.

00:03:02: Beim Industrial AI Podcast.

00:03:04: Robert sprach mit Baskonen über die KI und Robotikstrategie von Wanderlande.

00:03:09: Das ist der zweite Episode des "Ain't Hofen Special" und mein Gast ist Baskonen von Wanderlande.

00:03:18: Hallo Baskonen, willkommen zum Industrial AI Podcast.

00:03:21: Danke Robert, danke für mich.

00:03:24: Baskonen, ist das richtig?

00:03:26: Baskonen.

00:03:27: Baskonen, ich bin sehr sorry.

00:03:30: Nein, es ist alles gut.

00:03:33: Ok, bevor wir beginnen, bitte introduce yourselves briefly zu den Lesern.

00:03:38: Ich arbeite in Wanderlande seit 10 Jahren.

00:03:43: Ich habe einen Berat im Mechanischen- und Engineering-Gang, mit den Mastern von der Universität Einthoven.

00:03:49: Im Moment spezialisieren wir in Robotik und Autonomischen Vehikel.

00:03:54: Ich habe meine ersten fünf Jahre in Wanderlande mit Integrations- und Engineering gearbeitet.

00:04:00: In den letzten fünf Jahren ist das Innovation-Management.

00:04:04: Die meisten sind die Innovation-Managerin für unsere Innovation-Strategie, die auf einem Tag-a-day-Basis engagiert.

00:04:11: Wir haben auch viele Themen mit der AI-Räderung.

00:04:15: Wir sprechen über das.

00:04:17: Für alle Lesern von Wanderlande ist ein Generalkontrakter von Warehäusern, Konvierbauten und Airports.

00:04:24: Nicht nur das, aber auch unsere Systeme sind Integrator.

00:04:27: Wir arbeiten in drei Marktsegmenten, Airports, Parcel und Warehäusern.

00:04:33: Bei Airports, die wissen wir nicht, dass wir nicht sehen,

00:04:39: weil die meisten unserer Systeme unter der Basis oder in der Basis sind,

00:04:44: die Handlung der Bälle, die Sie in den Moment checken.

00:04:48: Ich kann aber den Logo auf der Backeclaim sehen.

00:04:51: Es ist bis zum Reklame-Prozess.

00:04:54: Es ist mehr und mehr über die Airplane.

00:04:57: Für Parcel machen wir die Automation von Distribution-Zentren.

00:05:01: Wir haben mit den Kunden wie UPS, DHL, FedEx.

00:05:04: Für Warehäusern ist es die Automation von Warehäusern-Prozessen.

00:05:08: Für andere generelle Merchandise, Fische, Faschen.

00:05:12: Das ist unsere Business.

00:05:14: Es ist die Konvei- und Römbau-Sortung.

00:05:18: Die AI ist in der Handlung.

00:05:20: In der Ende ist es alles über die Logistik-Prozess-Automation.

00:05:23: Das Material kann in verschiedenen Formen und Größen sein.

00:05:27: Aber in der Ende ist es über die Logistik-Prozess-Automation.

00:05:31: Wir haben ein bisschen über den Podcast geschaut.

00:05:34: Ich wollte über vier Stufen, vier Haupttopiken.

00:05:38: Der erste Thema ist die Forderung der Integration der AI-Solution.

00:05:42: Du hast gesagt, du bist ein Systeme-Integrator.

00:05:45: Und wie ist es mit der AI-Solution?

00:05:48: Wie groß ist die Komplexität für die Forderung der AI-Solution?

00:05:54: In deine Warehäusern, in deinen Automation-Prozessen,

00:05:58: in deinen Materialhandlungen, in den Materialflöten?

00:06:01: Wie integriertest du das?

00:06:03: Was ist für dich und was ist vielleicht schwierig?

00:06:06: Was ist nicht solved?

00:06:08: Das ist eine Frage.

00:06:10: Ich denke, für uns, in der Ende,

00:06:13: es startet immer immer, immer immer, immer immer.

00:06:17: Es startet immer mit unseren Kunden.

00:06:20: Wir schauen, wie wir die Kundenprobleme haben,

00:06:23: und wie wir sie solved werden.

00:06:25: Wie können wir die Businesses unserer Kunden besser erneut?

00:06:28: Das ist immer ein starting point.

00:06:30: Wir versuchen, wie kann die Technologie da hin?

00:06:34: Mit der AI-Solution, aus einer Systeme-Integrator-Perspektive,

00:06:38: haben wir schon viele Karten, die jährlich schon wieder gejagt sind.

00:06:42: Aber als Systeme-Integrator,

00:06:44: ist die Technologie schon genug bereit, um zu erlangen.

00:06:47: Und dann versuchen wir, die Integrations-Integrator-Perspektive zu integrieren.

00:06:50: In diesen Karten arbeiten wir auch oft mit dritten Partien.

00:06:53: Das ist besonders wichtig, auch mit Technologien,

00:06:56: die wirklich schnell mit der AI gehen können.

00:06:59: Ich denke, die meisten recenten oder die meisten dominanten

00:07:03: Anzeigen, die wir haben, sind die Applikation der AI

00:07:06: in der Maschinenvision.

00:07:08: Ich denke, dass viele der Systeme auf visuale Inspektionen

00:07:11: und dort haben wir schon ein paar Jahre lang starten,

00:07:14: um eine große Schicht in den traditionalen Maschinenvisionen,

00:07:17: die algorithms, um AI-traineden Systeme zu verabschieden,

00:07:21: für Detektion, Verkennung, Klassification zu haben.

00:07:25: Und diese Beispiele werden mehr und mehr entwickelt,

00:07:29: als die momentan in den Kunden auf die Kundenseite integriert.

00:07:33: Und ich denke, da sehen wir natürlich auch die Komplexität,

00:07:38: in dem Sinne, dass du die Moment, in die du die AI introduzst,

00:07:42: du natürlich auch etwas in einem Führung,

00:07:45: das in einem anderen Weg zu sein, also auch ein paar Jahre lang.

00:07:50: Also die Retraining des Models und so weiter?

00:07:53: Ja, also MLOps, also Maschinenverschleunigung in Operationen.

00:07:58: Ich denke, in diesem Sinne wird es auch geändert.

00:08:00: Ich denke, dass du sofort, zumindest wenn du das mal bekommst,

00:08:03: sofort, dass du große Potenzialen hast,

00:08:05: aber es braucht auch eine große Mindeststelle,

00:08:08: weil fast alle Geschäfts- und Kapazitäten,

00:08:12: die in der Ende des Deliveres von einer solchen Technologie zu den Kunden,

00:08:16: müssen in einer anderen Weise schauen, als sie vorher hatten.

00:08:20: Ich möchte zurück zu deinem Kustel,

00:08:22: weil aus meinem Punkt von view,

00:08:24: ob die CARE, ob sie AI oder nicht,

00:08:27: so lange die Passen sind im Motor oder im Plenar,

00:08:30: oder ob sie dich fragen, ob es AI-Bass ist

00:08:33: oder ob es eine traditionelle Maschinenverschleunigung ist

00:08:36: oder ob es ohne Maschinenverschleunigung ist.

00:08:38: Also, sind Sie wirklich interessiert über dieses Thema?

00:08:41: Ich denke, es lebt zwischen den Kunden,

00:08:44: aber es ist nicht, dass sie uns beobachten,

00:08:47: um etwas mit der AI zu Deliveren.

00:08:49: Es ist es, wie du sagst, dass sie eine Lösung für das Problem wollen.

00:08:52: Und das ist, ich denke, was wir immer auf den Weg sind.

00:08:55: Und ich habe auch gesagt,

00:08:56: dass das ein starting point für uns immer ist.

00:08:58: Und nächstes Mal schauen wir,

00:09:00: ob wir das Problem auf einer AI solveren,

00:09:03: weil es einen großen Potenzial hat,

00:09:05: aber nur, wenn es auch in der Ende erfüllt kann.

00:09:07: Also, ja, wir starten mit dem Kunden.

00:09:09: Du hast die MLOps Infrastruktur, Pipeline,

00:09:13: und so weiter gesagt.

00:09:14: Wie handelst du das?

00:09:16: Im Moment, wo wir realisieren,

00:09:19: dass wir ein Innovationprojekt haben,

00:09:21: und wir das Potenzial,

00:09:23: zum Beispiel, mit AI,

00:09:25: mit Maschinenverschleunigung,

00:09:28: das Moment, wo wir realisieren,

00:09:30: dass das feasible ist,

00:09:32: und wir das über die Entwicklung-Zentren zu handeln,

00:09:34: um diese Lösung zu produzieren.

00:09:36: Wir haben auch realisieren,

00:09:38: dass wir die same Kapazität der MLOps

00:09:40: in der Entwicklung-Zentren haben.

00:09:42: Wir haben also auch starten,

00:09:44: um zu überprüfen, wie wir das tun.

00:09:46: Und wenn wir es lernen, wie wir das tun,

00:09:48: dann haben wir das auch

00:09:49: das Training in unseren Entwicklungszentren zu verhindern.

00:09:52: Aber dann haben wir auch, ich glaube,

00:09:54: schnell auch in alle Fragen,

00:09:56: alle kinds of challenges,

00:09:58: as we realized that,

00:10:00: so how do we keep training this algorithm,

00:10:03: with customer data coming in,

00:10:05: customer feedback coming in,

00:10:07: who has the responsibility on site.

00:10:09: So we've had quite a lot of discussions

00:10:11: with our development teams on how to approach it,

00:10:13: but in the end indeed,

00:10:15: it settled in the process,

00:10:17: in the tooling that we use.

00:10:19: So it's not an easy road,

00:10:22: I would say, to take.

00:10:24: But I think with this always exploring

00:10:27: first the alternatives that you have.

00:10:30: And next together with your stakeholders

00:10:32: deciding on what best fits

00:10:34: was our solution moving this forward.

00:10:36: Can you describe how do you approach?

00:10:39: How we approach the MLOps topic?

00:10:42: Well, I would not say out of first hand,

00:10:44: because that was really,

00:10:45: or that is still really in our development centers.

00:10:48: But I think we use quite a lot of tooling around this,

00:10:51: also supplied by, as part of the Microsoft Azure pipeline.

00:10:54: Microsoft is also a strong development partner for us.

00:10:58: And I think also in,

00:11:00: basically training the people around it,

00:11:03: also in using it,

00:11:04: and also at the customer site,

00:11:06: that they know how to use it,

00:11:08: when to use it,

00:11:09: I think that is in the end what we did there.

00:11:11: And how do you integrate the MLOps solution

00:11:14: in your, what is it, warehouse management system,

00:11:17: or how do you integrate the solutions

00:11:19: like MLOps in this system?

00:11:21: Well, I think in the end we,

00:11:23: you mean towards our customers?

00:11:25: Yes, yes.

00:11:26: Yeah, so what we in the end also foresee,

00:11:29: or for so already quite some time ago,

00:11:31: is that our customers really need a digital portal

00:11:34: to access this kind of technology in the end.

00:11:36: So we have created a MyHonderLanda portal,

00:11:39: which is the portal by which a customer has access

00:11:42: to the operations on site.

00:11:45: And that is in the end where we need to integrate

00:11:47: all these capabilities to make sure that also

00:11:50: the customer can, where needed,

00:11:52: get access to it.

00:11:54: I want to come back,

00:11:56: because from my point of view,

00:11:58: when we talk about the future of automation companies,

00:12:00: and automation companies like Funderlanger,

00:12:03: I think this MLOps business,

00:12:05: the combination of domain knowledge

00:12:08: combined with AI and infrastructure,

00:12:11: knowing the customer,

00:12:13: that you are a reliable partner,

00:12:15: that you ensure that your solution will run,

00:12:19: and that you deliver industrial grade AI applications,

00:12:24: including updates MLOps.

00:12:27: Is that a new business?

00:12:29: Is that maybe in case instead of selling more hardware?

00:12:34: Or what is your opinion on that?

00:12:36: Yeah, I think it's, we do see our business

00:12:39: shifting towards more software intensive solutions,

00:12:44: more digitalisation,

00:12:46: and indeed, our customers still, of course,

00:12:48: expect the same reliability as always.

00:12:51: So that shift is still ongoing in our company,

00:12:54: and it will keep going, I think,

00:12:56: for the coming years.

00:12:58: And that also should be, I think,

00:13:00: supported by these kind of technologies.

00:13:03: So we do see that, for example,

00:13:06: in our technological developments,

00:13:09: we have streams dedicated to building

00:13:12: these kind of solutions and services

00:13:15: towards our customers.

00:13:17: And that is also, I think,

00:13:19: what our customers expect from us in the end.

00:13:21: So yes, there is a big shift,

00:13:23: and I think it's also industry-wide happening.

00:13:25: So more system integrators

00:13:27: are making the shift towards

00:13:29: much more software intensive solutions.

00:13:31: What about data sharing or sharing data?

00:13:34: Are your customers familiar with how to share data?

00:13:38: Are they able to share data?

00:13:40: What do you see there?

00:13:42: It's tricky.

00:13:44: And I think we can recognise

00:13:46: different types of customers in this.

00:13:48: So I think we have customers

00:13:51: that really are protective with their data.

00:13:54: So they have typically also their own capabilities

00:13:58: to make use of that data.

00:14:00: And I think therefore also

00:14:02: that they are really geared towards

00:14:04: developing their own solutions

00:14:06: to get the most out of that data

00:14:09: to optimise their own processes.

00:14:11: On the other end of the spectrum

00:14:13: you also have customers

00:14:15: that really also rely,

00:14:17: for example, on Van de Lander

00:14:19: to provide them with those insights.

00:14:21: And I think with those customers

00:14:23: it is easier to have the conversation

00:14:25: about sharing data

00:14:28: and to, for example, generate insights

00:14:30: out of that data.

00:14:32: And that is a real spectrum.

00:14:34: And I think that also varies,

00:14:36: in den sehr per-market-Segment.

00:14:38: So, for example,

00:14:40: we have very extensive service contracts

00:14:44: with many of our customers.

00:14:47: Well, that's good for you, right?

00:14:49: Yes, and it's also allowing you,

00:14:51: I think, to optimise the operation of a customer

00:14:54: within that service contract.

00:14:57: But there's also customers

00:14:59: that do that service really themselves.

00:15:01: So I think it's really about

00:15:03: what type of customer you have in front of you,

00:15:05: and how you can talk about data sharing in that sense.

00:15:08: And also the level on which you share data, right?

00:15:11: So is it at the level of a product

00:15:14: so really close to the equipment?

00:15:16: Or is it more operational data,

00:15:18: really system-wide?

00:15:20: Is I think another access, if you will,

00:15:23: that determines how much

00:15:25: and how easily data can be shared.

00:15:28: So, you mentioned when you are in the lead,

00:15:31: when it comes to service,

00:15:33: when it comes to technical service in the warehouse,

00:15:36: you are able to collect data.

00:15:38: So you will have a huge data set of time series, right?

00:15:42: You can have months, years of time series.

00:15:46: What do you do with this amount of data

00:15:49: of time series for your drives and for your components?

00:15:52: Do you have any ideas what to do with that?

00:15:55: Yes, well, a lot of this is actually

00:15:58: also at the moment under exploration

00:16:00: because the potential of that is indeed huge.

00:16:03: I think we see, already have seen,

00:16:06: quite a lot of good use cases

00:16:08: on predictive maintenance of our equipment.

00:16:11: Because there is indeed a vast amount of data

00:16:14: that is being produced that we can use,

00:16:17: I think, in a good way to predict whether certain equipment

00:16:20: will fail and to prevent the maintenance in time.

00:16:24: For various pieces of equipment all across the segment.

00:16:27: But it is also easier, I think,

00:16:29: because on that level we also own the data.

00:16:32: Because it is really close to the equipment.

00:16:36: But next there is also, I think,

00:16:38: on a more process level use cases,

00:16:41: where we are also looking into at the moment

00:16:43: to, for example, find root causes of certain events.

00:16:48: So, especially, for example, warehousing business,

00:16:52: complexity of warehousing systems is huge indeed.

00:16:56: So, far more complex if we would compare it,

00:16:58: for example, to our airports or parcel business.

00:17:02: And there also these AI techniques

00:17:05: and the availability of data really allows us to,

00:17:08: and that is what we will be exploring soon,

00:17:11: is can we find a root cause of a certain event

00:17:14: by using AI techniques and the data that is available.

00:17:18: Earlier on in the process already identify

00:17:21: and tackle the root cause to avoid that it causes

00:17:24: some sort of side effects later on in the process.

00:17:27: And this is, I think, one of those use cases,

00:17:30: which now suddenly becomes solvable.

00:17:34: But still, it starts with,

00:17:36: and I think it always starts with,

00:17:38: good data representations to be able to unlock it.

00:17:41: Can you share a little bit more details

00:17:43: how you proceed with this concrete approach?

00:17:47: Which approach exactly do you mean?

00:17:49: The root cause approach?

00:17:51: Well, we're still intending to,

00:17:53: we haven't really started the exploration yet on this,

00:17:56: but what this is, so just give an example,

00:17:59: what can happen is, so in a warehouse

00:18:02: where order picking is being done,

00:18:04: orders are being picked and it can happen

00:18:07: that certain orders are being delayed.

00:18:10: If we analyze in real time the lay of those orders

00:18:15: and find which, which is for what might be,

00:18:19: for example, an item in every order that is being delayed,

00:18:22: we can find that item might be the root cause of those delays.

00:18:26: For example, that item is too low on stock,

00:18:28: or it's too late restocked,

00:18:30: or maybe it's in an aisle of our storage

00:18:34: that is facing issues to retrieve in the time,

00:18:38: but bringing that data together in a way

00:18:41: that we analyze it by means of data analysis techniques

00:18:45: with AI to find the root cause earlier on

00:18:48: and have it run in real time in the control room

00:18:51: allows the operators in the control room

00:18:54: to spot those potential root causes earlier on

00:18:58: and next focus basically their effort to resolve it

00:19:03: as soon as possible.

00:19:05: So this is one of those use cases

00:19:07: that at the moment we're looking into,

00:19:09: like is it feasible, that's one.

00:19:11: Is it desirable, so also,

00:19:13: does our customer also see a working in this way?

00:19:17: Is it really solved their problem?

00:19:19: Is it also one of the bigger problems

00:19:21: that they really have in the control room today?

00:19:23: And next also, is it viable?

00:19:25: So how much faster will an operator

00:19:28: really be able to detect and solve such a problem?

00:19:31: So we try to look at all those three lenses

00:19:33: when we start to explore these kind of use cases.

00:19:37: And the business model is also to relive

00:19:40: you as a service provider, right?

00:19:43: You save not directly money, but you save time.

00:19:46: Ja, exactly.

00:19:48: It's sometimes quite hard, of course, to quantify this,

00:19:50: because it's very, I think, typical to then start looking at,

00:19:53: okay, it makes an operator that much more efficient.

00:19:56: So we will save x amount of operators on a yearly basis.

00:19:59: But in the end, this is also again,

00:20:01: of course, the overall performance of a system,

00:20:03: which is harder to measure,

00:20:05: especially when you're in the early stages.

00:20:08: But we do try together also with,

00:20:10: for example, process engineers in this case,

00:20:13: we will have at certain customer sites,

00:20:16: and also if possible directly with customers,

00:20:19: to quantify this by building a minimum viable product

00:20:24: that we can't trial at a customer site to measure this also.

00:20:28: Because in the end, that is always a bit the challenge that we have.

00:20:32: Our systems are at very large scales.

00:20:36: And if we really want to get a certainty

00:20:39: that it will work as a product,

00:20:41: we need to also deal with that scale.

00:20:43: So we can only test it internally up to very small scales.

00:20:47: But we also need to rely on customer trials early stage

00:20:50: to identify that viability,

00:20:52: to be also sure that we build down all the risks

00:20:55: that we need to build down before we actually start real product development.

00:20:58: But your currency is a package at the right time

00:21:01: when the truck is there at the warehouse, right?

00:21:04: The customer pays for that in the end.

00:21:07: Ja, well, not yet as a service-based model,

00:21:12: but in the end, that is what our systems do.

00:21:16: So, and then you can calculate a return on invest

00:21:20: when you do route courses like that.

00:21:22: So, typically we make these kind of calculations

00:21:26: to see, is there in the end a good business case,

00:21:30: and that we do of course through also a lot of internal collaboration

00:21:34: with our product managers.

00:21:36: So, with my team, we're working really on the innovations,

00:21:40: so really on the front end, early stages.

00:21:44: But we know that we also need to collaborate heavily with

00:21:47: with product management and also with the development teams in the end,

00:21:51: if this turns into something that becomes a product.

00:21:54: So, when we talk about warehouses or airports or stuff like that,

00:21:58: these kind of buildings or these kind of warehouses

00:22:02: does not operate alone in the world.

00:22:05: So, you have supplier, you have the shops,

00:22:08: you want to sell something in the shops.

00:22:10: So, the warehouse maybe is in the middle of the whole supply chain.

00:22:14: And what about thinking about demand forecasting for the warehouse,

00:22:20: because that would be very interesting from my point of view,

00:22:23: that you can see what will come into the future and react on that.

00:22:29: Or is that not a discussion at Fundalunder?

00:22:31: Oh, it definitely is.

00:22:33: I think you're spot on with this.

00:22:35: It's also one of the main AI capabilities that we foresee,

00:22:40: that can generate a lot of value.

00:22:42: And again, this can happen on various levels.

00:22:44: So, I think you're fully right.

00:22:46: These warehouses, also parcel distribution centers and also airports,

00:22:49: they're all part of a network.

00:22:51: So, there's incoming good, outcome good,

00:22:54: but they also go to a next spot into this chain.

00:22:57: And I think there's a lot of potential in doing these

00:23:01: demand-driven optimization Techniques in the chain.

00:23:04: But then again, the question is, do you also have the access to the data?

00:23:08: Are all those parties willing to share their data in the end?

00:23:12: And that is, I think, the tricky part, which is much more geared towards,

00:23:16: do you have the ecosystem to make this happen?

00:23:19: And do you have it?

00:23:21: I think often it's not enabled yet to do it.

00:23:23: So often I think it's about getting the trust in that ecosystem to share that data.

00:23:30: But that is, I think, very, very hard to get done.

00:23:33: And I think also, if you look at maybe more specifically to the parcel business,

00:23:38: those parcel companies already have a network of their own.

00:23:43: So, it's also in their own best interest,

00:23:46: if they will figure out themselves how to do it.

00:23:49: And I think companies like DHL, UPS,

00:23:52: they are doing this themselves already,

00:23:55: because there is a huge potential if they can connect these dots basically.

00:24:01: But for us also, even within the scope of Vandalanda today,

00:24:06: we can do this.

00:24:08: Also, we can optimize the operations within the warehouse,

00:24:14: within the airport, within that parcel distribution.

00:24:18: center by just connecting to the warehouse management or enterprise resource planning systems of the customer.

00:24:25: If we know also within our system what is coming,

00:24:29: we can also optimize our own operations to better fill in the need and also better predict towards the outgoing process of the customer

00:24:37: what will be the potential impact and that is something that those use cases

00:24:42: we will be definitely also focusing on more and more in the future whereas at the same time and that is

00:24:47: especially I think to the interest also of innovation strategy so my scope is to see

00:24:52: is there also in the future solutions that Van de Landa might provide a future for Van de Landa in this

00:24:59: optimization over the value chain is there a future for Van de Landa in it.

00:25:05: So we looked a bit into the future. What are you and your team currently working on Bas?

00:25:11: Give me two or three examples.

00:25:15: One I think interesting one for us at the moment is around applying AI assistant technology.

00:25:22: Oh everybody is doing that.

00:25:24: Yeah of course.

00:25:25: In the control room.

00:25:28: Okay.

00:25:28: We have made a vision on how the control room of the future

00:25:33: would look like and in that vision we see well basically the control room could operate anything whether it would be autonomous.

00:25:43: Well first of all whether independent of the business that we would be in it it could look the same.

00:25:49: It's the same challenge but next indeed ultimately it is also autonomous because we see step by step we see automation coming in

00:25:57: and we see that next also the decisions that are being taken right now data comes to the operator that will become much

00:26:04: and more descriptive next it can become prescriptive up to the point that it turns into decisions that are taken

00:26:12: and processed automatically where we already see quite some value on the short term as if we for example take incident reports

00:26:19: that are happening for example in an airport control room and that those incident reports are also fed to an LLM to make sure

00:26:26: that next operators can make use of it basically use the collective knowledge of all the other operators in that control room

00:26:33: to find the right maybe also root cause of your all the best way forward to solve an issue and that that is already

00:26:41: where you see quite a nice potential for this technology also interesting I think with it is that you realize how easy is it is

00:26:51: suddenly to sort of cross the language barrier because just a flick of a button you could switch it to another language

00:26:58: which might be more native to the person that is working at the moment so this collective intelligence in such a control room

00:27:05: is I think very interesting because it's a lot of information is being centralized there but operators traditionally have quite a long learning curve

00:27:16: so at an airport control room it takes about one to two years before a control room operator is really experienced and seasoned enough

00:27:26: to handle all kinds of incidents and we believe that if we can introduce this technology and we're actually trialing it at the moment

00:27:34: that that learning curve that that that can be a lot shorter and maybe the LLM can translate alarms or warnings because sometimes it's very

00:27:45: crypto for beginners what does it mean now and maybe LLM can then explain so this means exactly one two three four five

00:27:55: exactly so it could really guide you in the steps to take and I think also the fact that it can but that is also of course the tricky part next

00:28:04: that it can translate given the context what to do based on prior experience in a new situation is of course also powerful but also kind of tricky

00:28:14: because it could also come up it could also start hallucinating up to the point that it comes up with a certain scenario how do you make it robust

00:28:22: yeah exactly and that is something I think yeah we yeah we're sort of at the moment also experiencing together with customers which is I think really nice

00:28:31: because you you also need to get I think at some point some hands-on experience and also the understanding of the user in the end

00:28:39: so how will it be used how will the user react to this and I think especially with generative AI it has a yeah almost a direct interface with the user

00:28:47: that needs to on the spot and also decide what to do with that information

00:28:51: so the user wants explainability yeah so the user needs to be informed how the model decides and why the the the solution provides this event or provides this idea

00:29:05: or is this the topic exactly yeah I think also with the new GPT releases it's also easier of course with with directly referring to the original source

00:29:17: I think also that is important the explainability to enhance the the learning

00:29:21: but nobody will read this PDF right no but I think the moment so if an incident happens and of course there needs to be a quick immediate response

00:29:30: but learning can also of course happen afterwards and I think the explainability by also using the references to okay where was this advice coming from

00:29:38: what can we learn from it is also something that can happen at a slower time constant than the on the spot decision to intervene with the system in some way to solve

00:29:48: the incident but I think that this is in that sense really interesting technology where generative AI can already already provide a good solution a valuable solution to our customers

00:30:00: second example yeah not genai please yeah not genai is I think mainly for that we use for example to classify what I mentioned already in the beginning

00:30:10: as so classification of the material that we need to handle is I think crucial for us in in next also the integration of other components in our systems

00:30:20: and this happens we noticed especially when we started to introduce robotics so item picking but also so in warehousing but also picking off parcels loading bags into containers that go into the airplane

00:30:37: it's all over the place what we noticed the moment that we switched there from a human to a robot is that yeah robots cannot handle everything yet so yeah you cannot just simply swap them out

00:30:50: but as a systems integrator we still can use robots but then we really need to know what do we send to the robot because we only want to send stuff to the robot that it can actually handle

00:31:00: because if we don't yeah you will you know at some point you will get an intervention and a human needs to go to the workstation again to fix it so that led also to the fact that we really needed to have classification in our systems to identify

00:31:15: is this piece of material that we're handling here is it is it can the robot handle this yes or no and we notice that very often it depends on the packaging material the shape the size and with AI it open up I think up the possibility to really start identifying this in a reliable way on the fly such that we can really take that information and bring it forward to to basically the robotic modules that we have

00:31:43: and also use our system design is such a way that we next also only basically send to the robot what it can pick so with the technology we now can for example take an airport we can identify that it's a hard shell suitcase it's a soft shell it's a backpack it's an old shape and based on that we can start to take decisions how the robot should handle it if it if it should handle it at all

00:32:07: but if you if you talk to the robotics manufacturers or if you listen to them I thought oh I hear everything is solved in this field and now you tell me oh no this is not so it's not it's not no no it's not I think the point in the end is also what you didn't mention in the beginning it has to happen with a with a with a very high reliability so that robot always needs to be able to solve it and and we know it is not possible yet to do it so and at first I think it will

00:32:36: be for a very long time also because I think the challenge with with robotics is it's not only the reliability but it's also the the speed with which it needs to do it so there are many facets where in the end I think the benchmark will always be the human picker for example yeah before I think those robles will really get that level that capability it will take quite some time and I think up to that time having those classification

00:33:06: abilities in the system allows you to use robots in a much more efficient way what's a third topic maybe let's let's pick one which is more internal yeah so for our own processes now a gen A I use case yeah in the end yes yeah I think there you well of course we touch already on predictive maintenance so it would not be fair to pick that one but yeah I there I think more in terms of there there's

00:33:35: quite a lot of effort for example going into writing test cases requirements etc which is most often pretty straightforward and I think their gen AI can can do a lot so basically on requirement management and translation test cases or test case automation if you will but also in the domain for example of of of manuals that need to be written manuals that need to be interpreted I think everywhere where you basically produce a bunch of text need to

00:34:05: interpret it in different ways I think there it is really powerful but then again also I think crucial to really be in touch with your suppliers right because we of course rely a lot also on content management systems requirement management systems and tooling and I think many of those providers are also scrambling to integrate gen AI into their offerings into their products into their services and and I think it's

00:34:35: especially their crucial to to really be in close contact with your suppliers and their roadmaps on when they plan to integrate such services and how they want to do it because I think the possibilities are endless but it's of course it would be a pity if you start to invest a lot of effort to build all these kind of services if they next also are provided by the company that you're relying on for requirement management for example so do you have an example of a requirement tool you already have a lot of

00:35:02: an example of a requirement tool you already started to work on or you're using one by by a supplier or what is what is behind no not not not yet in that sense but if we if we just make it more closer to home right so if if you so I think Microsoft it's it's it's integrating it all over the place right so yeah of course it would be handier if you just rely on I don't know Excel with gen AI capability word with gen AI capability rather than

00:35:31: developing your own tools and implementations but I think that the question is so is it usable for us if it's there so you really need to engage on a strategic level I think with those suppliers to see okay how are you looking at this how is it usable for us and then decide okay should we should we build this ourselves or do we rely on a third party knowing that it might not be perfect but okay you might get there in the end a lot faster if you do how difficult is it

00:36:00: to to build robust AI solutions from your point of view that is quite the question I think it is it sounds easier than it is yeah and I also wonder so how robust should it be right so with those gen AI cases how yeah how robust should it be right in the end I think the target is to for example increase the efficiency with which an operator can work but but how

00:36:29: how much wrong answers are you willing to accept yeah 80% or 10% yeah is it 80 is it 90 is it 95 should it be should it be all the way what I like is too early on also already ideally with the end users to have that conversation because I think it can vary a lot from application to application yeah what what that robustness level or accuracy level should be and I think that's a good question

00:36:58: and I think that is really important because before you know what you're developing something you're really goal-plating maybe something which which in the end makes it become completely unusable and I don't believe that all use cases require it to be 100% robust and I think it's also something that can grow over time right so also on that journey from let's say having an assistant up to complete autonomous decision making this will change and I think if you approach it

00:37:27: that way and you can do more stepwise automation so you introduce the technology that that that generates basically experience it also generates I think more trust on both sides so also add our customer in in in making use of it I think from that point on it's easier to gain more robustness and go towards full autonomy sooner than if you say no it has to be 100% accurate always so getting I think the technology into the field and I think that's a good question.

00:37:56: I think the technology into the field at an acceptable level with I think also a customer that that that understands this I think gets automation happens sooner than if you if you don't so I think that I think defines also for me robustness yeah what level do you want to how robust do you want to be.

00:38:18: I want to come to a little more internal perspective because the Netherlands also lacks specialists in the field of AI.

00:38:27: Do you make domain experts fit or do you train them to develop smaller solutions.

00:38:36: It's called Auto ML Auto ML tools or anything like that is it on your agenda.

00:38:42: Well, at the moment, we are mainly focused on identifying also as part of our strategy.

00:38:50: What are the main AI capabilities that we need to get more in house.

00:38:57: What tell me what are the main capabilities.

00:39:01: A lot revolves around predictive maintenance a lot also revolves around enhancing this this this productivity and effectivity of operators we have in the field also a lot around machine vision as so so this this item inspection classification and then I think we have this this scope of optimization so optimization of flows of the the the the month the discussion we had earlier on.

00:39:29: So reinforcement approach for every warehouse and it will go automatically.

00:39:33: Yeah, exactly.

00:39:34: So yeah, yeah, the dream.

00:39:37: Yeah, right.

00:39:38: We talk about dreams and at the end.

00:39:40: Yeah, yeah.

00:39:41: And I think that that is the type of capabilities where I think we can make huge steps and then it's about indeed finding the right finding the right talent again and training our engineers to do it.

00:39:54: And also, I think you need to look at the competences you have because I potentially affects everything from a system integration perspective.

00:40:05: When we for example work with third parties that maybe supply that classification algorithm based on AI, then also I think other competences that we have so for example in architecture and integration need to be more aware of of AI.

00:40:21: It's opportunities, it's risks and how to integrate it in order to in the end make it an efficient to make it into a working product.

00:40:28: So I think on the one end, that is what we're doing at the moment you need to look at what capabilities do I need that best basically provide customer value.

00:40:37: And how do I actually realize that that with those capabilities I can also build a product that is integratable into a system.

00:40:47: So the competences that you need to realize it.

00:40:52: When we are talking about AI and new models, we talk about large language models, maybe the industry needs more language models or smaller or smarter.

00:41:02: We always talk about deploying then because you need to deploy it on, I don't know, on a PLC on a server next to the shop floor.

00:41:12: Or do you need GPUs in the warehouses in the airport? What is your opinion on that? Do we need more GPUs in the warehouses?

00:41:20: Do we need smaller, smarter language models? What is your opinion?

00:41:25: Well, I always go by learn what is hard to model and model what is hard to learn, which is I think in the end the most efficient solution.

00:41:35: And next with that I think your deployment architecture can also be organized in that way, right? So I think also what you can do for example on the edge, do it there.

00:41:45: So I think it will in the end be always a mix of them both. So doing it really close to the equipment on the edge or doing it centralized.

00:41:55: And also even up to the point where you bring it completely remote and put it in the cloud. And I think again there also the discussion with the customer for us is really important.

00:42:06: So what is his or her view on this? And I think also there we noticed that customers take various approaches.

00:42:14: So some are really focused on on-premise solutions, others are much more already working towards the cloud.

00:42:21: So I think that is also a very important aspect to take into account when you start doing product development.

00:42:27: So what I hear you haven't ordered GPUs in large quantities yet, right?

00:42:31: No, no, and I think especially that type of solutions also there we often also rely on third parties still.

00:42:40: So it can be that it even goes goes via them, but up to now not yet indeed.

00:42:45: But as said, I think the majority of the AI that we're at the moment using is a lot of that is still an exploration phase.

00:42:54: So I think the most mature applications and I think you will see it maybe even industry wide will be on I think predictive maintenance and these machine vision objects that you have in the field to classify or inspect things as in customer facing.

00:43:10: Right. So because I do believe that there's a lot of things going on, of course, also in data analysis, which is happening also more internally.

00:43:17: But I think those are the more dominant use cases also in the field, but it's not it's not industry wide accepted.

00:43:23: I don't think there's a lot of gen AI, for example, going on at customers at the moment.

00:43:28: I think a lot of this is still in an expert exploratory phase.

00:43:32: So you're planning your warehouses and we have already talked about reinforcement learning and all the nice stories we hear about.

00:43:38: Organizing by themselves and the process and stuff like that.

00:43:42: But what about AI based simulation to not to simulate once again using the old simulations building an AI based simulation for warehouses for process.

00:43:53: Is this an option? Is it a dream? Is it a perspective?

00:43:57: What is your opinion on that?

00:43:59: I think you're making the transition out to dreams.

00:44:01: But yeah, exactly.

00:44:03: Yeah, what we see is indeed that.

00:44:07: And that is, I would say, indeed the dream is that if we have digital twin representations of our systems, if we can then use AI to put forward various scenarios, basically, so to do some forward playing of various scenarios to solve certain operational disturbances and have it then pick the best one.

00:44:31: I think that would be that would be the dream because this is at the moment we do, of course, a lot of simulation, a lot of emulation in our systems.

00:44:41: But often these tools, models that we use, they're not yet used throughout the lifetime of a system.

00:44:50: And I think that in the end is the dream if you can build it up in such a way that those twins evolve over time and you can connect AI powered services to it, I think it would unlock tremendous value because essentially this is, I think, really core also to our capability of a system integrator that from simulation.

00:45:14: Basically, we already start in the very early phases when a customer approaches us with requirements.

00:45:20: We start with building simulation models to see, okay, how can we best fulfill these customer requirements and throughout the lifecycle, we keep building basically virtual representations of those systems in various ways, but they're not always connected throughout the flow from selling all the way up to delivering.

00:45:43: And in maintaining such a system.

00:45:46: So having, creating those living digital twins would be, I think, a huge step forward.

00:45:53: And if you then have those AI capabilities on top of that, that is how I also would foresee the control room of the future be looking like.

00:46:03: What else do you dream of?

00:46:04: Many things.

00:46:05: I will go on a holiday next week now.

00:46:07: Two cases of industrial AI.

00:46:11: Well, what I dream of in that sense is that I, what I dream of is to do these kinds of explorations really together with customers.

00:46:24: Well, I mentioned also, I think earlier on is to get into this mode with customers and all the other business capabilities at Vandalanda to do or to reach this sort of stepwise automation.

00:46:39: So I really am a, I would call myself a technology realist, but finding what is real about technology, I think only happens the moment that you put it into practice, that you put it in the customer environment.

00:46:52: But this is of course one of the hardest things to do because that customer environment is working with mounts on reliability, on accuracy that often it's hard to judge.

00:47:05: Will it be possible, yes or no?

00:47:08: So creating such an innovation ecosystem in which you can do this with customers that understand this and also all the connected processes on our end to do it.

00:47:19: I think that would be fantastic.

00:47:21: And then I think also maybe in line with that it would also be great.

00:47:25: Like if you would have now a clean sheet of paper, you take the latest technologies, it would be a dream.

00:47:32: If you could say, let's bring some people in our company together and start building basically this warehouse of the future or airport of the future or parcel distribution center of the future from the ground up.

00:47:45: Because I think one other thing is that we always need to deal with this legacy, right?

00:47:49: Exactly.

00:47:50: And being able to start with a clean sheet of paper, I think would also be a dream, right?

00:47:54: So we're doing how rapidly technologies develop today and what could be possible if we just forget about the legacy for a moment and do such an initiative, that would also be a dream.

00:48:06: Bas, let's start with your vacation.

00:48:08: Yeah.

00:48:09: I think that's a good plan.

00:48:10: Yeah, and then I can dream a bit more.

00:48:12: Yeah, you can dream a bit more.

00:48:14: Enjoy it.

00:48:15: Thank you very much.

00:48:16: It was a pleasure to talk to you all the best and greetings to the Netherlands.

00:48:20: Likewise, Robert.

00:48:21: Thanks for the talk.

00:48:24: [MUSIC PLAYING]

00:48:27: Robotik in der Industrie, der Podcast mit Helmut Schmidt und Robert Weber.

00:48:32: (gentle music)

00:48:34: you