[00:07] Hello, everybody. Welcome to the Azumuta podcast. This is the first episode in a series of three where we focus on one of the topics that has been widely discussed in modern manufacturing and also often misunderstood about dark factories or non-human factories. We will have discussions with business leaders, with experts, with the labor unions to look at this concept from all different angles. Today, we want to start with the science of it, so looking at the technological future, and that's why we've invited Professor Francis Werfels and Andreas Verlaenen, who I personally know as the winners of the cloth folding contest for robots some time ago.
[00:49] But I guess you're doing more than just that. Yeah. Today in our lab, we focus on all the technologies necessary to create robot helpers that can do everyday tasks in your home or in your company. Uh, so that goes from mechanics, electrical engineers, up to a- artificial intelligence. Research in AI and robotics, what does that actually mean? Or what are the things that keep you busy during the day and at night? Well, one, one sentence, uh, what, what keeps us busy is, um, research… every research technology towards a robot helper in your home.
[01:25] So it means a focus on electrical engineering, mechanical engineering, and of course, uh, AI. And we want to bring those aspects together so that robots can do more tasks. And with this more tasks, I mean, general, it's like giving tasks that humans are currently doing to robots so that humans can focus on things we like to do, like creative things and so on. To translate that to a, a factory environment, uh, there's been a lot of talk for decades around… about lights out or dark factories and things like that.
[02:00] Do you see a difference in how, uh, we used to talk or how we used to define dark factories, let's say, 10 years ago versus now? But from our perspective, uh, our focus is on, um, let's say lower volumes but high mix of goods, um, because then again, this generalization aspect is really important. And so the, the machine or the robot has to become smarter. Um, and while in the past, in, in '80s, '90s, the focus was more on high volumes and, um, let's say lesser goods variability.
[02:40] So this becomes more important and so hence our research becomes more important. Also, we used to look at how we can change the environment and the task so that it can suit robot technology in order to solve the task, while now we are looking more towards, yeah, how we can use AI to keep the environment and the task the same, but have the robot adapt to, uh, this task. Does it mean that the, the definition of a dark factory is different today than it was, uh, let's say five years ago?
[03:13] Good question. Yeah. Actually, I honestly, I, I, I don't know. As, um, uh, from our perspective, dark means okay, you can, um, be energy, uh, saving, saving, uh, by, um, by cutting the lights and, and, and heat. Um, but in our, from our research perspective, it doesn't have to be that case. Um, so dark doesn't need to be dark. It can also be just, um, adopting existing lines, um, while people can be still around- Mm-hmm … working together with these robots.
[03:49] Mm-hmm. And so that's an important part. Like, some people would, like, include a definition like no humans at all close to the robots, hence, again, changing this environment so that no humans will be around. Well, for us, we generally look at, like, yeah, humans will be around, hence also, like, currently the big, um, popularity of collaborative robots which share their workspace with, uh, humans. Oh, yeah. Yep. Yeah, and we of course never look at automation per se.
[04:18] It is like a broader… So, so it's going from automation to general purpose robotics- Yeah in human-driven context. Where do you think that people are at this moment overestimating or underestimating the, the capabilities of robots can do or can't do? You always hear the story about folding a towel, uh, that takes ages for a robot to do, or that's what's been said a year ago. I don't know. Are there other more general things that you think of where people expect too much or too little nowadays?
[04:56] Well, let's maybe first take the towel example, eh? Like, um, folding, let's say 10 years ago, and Pieter Abbeel, um, was one of the first to do for cloth folding, and it took a- approximately 20 minutes. Um, then three years ago, we won a cloth folding competition with our lab, uh, where we did, uh, folding of, of, uh, towels and T-shirts in under two minutes. Um, today I think we can even push it towards 30, 40 seconds. Um, but there is an important, um, condition there.
[05:31] And while, um, for example, our case three years ago was focusing on variable, uh, T-shirts and, and, and towels, you can push it, uh, faster, but then the diversity of towels, of clothing types becomes less. Um, but with imitation learning, for example, you can very well train a robot to fold this one T-shirt really, really good. Mm-hmm. But the question, of course, is can it then fold any T-shirt that's out there in the world? Mm-hmm. And then that's a good example on how people now think, "Okay, the robots can do it faster," but an important ex- aspect still lacks, and that's generalization capabilities.
[06:12] Um, and I think in that sense, uh, I don't think we have to expect, let's say, robot butlers, uh, robot helpers everywhere in the world in the next coming, uh, year, two years, like is promised by, yeah, many big companies. Mm-hmm. Okay. Yeah, so the general idea that lives right now by humanoid companies is that we can just plug and play a robot where a human is currently doing something. Mm-hmm. Hence the big dream of making a robot look like a human because we are trying to structure the world around us, uh, for us, for humans, for our form factor, and they assume, yeah, if we build this humanoid robot, then yeah, it can just take over these tasks.
[06:58] But I do think, like, in general that we are, uh, highly overestimating what these robots are currently capable, uh, of doing, uh, because of the, like, the generalization, uh, properties that Francis was mentioning. And for us as a research- as researchers, this puts a lot of pressure on us because, so we are known as like, uh, a textile cloth folding lab because textile is a really difficult task. It's like a holy grail task for robots to solve because if you can solve, like, cloth folding, we can do a lot of other tasks also.
[07:29] So that's the general idea. But we get a lot of phone calls when they see, like, um, the Tesla Optimus robot on a commercial folding clothes, and they're like, "Hey, but they can already do this quite well." Mm-hmm. Uh, "Well, what's the status? How, how is your progress going?" But- Mm-hmm. Then we are like: Yeah, but maybe if you look closer, you can see that this robot is being operated from a distance by a human in these, uh, commercials. So So you, you basically say like robots will probably do-- won't be able to do every task, uh, especially if they are, are more complex, and so they will need work instructions to know what to do basically.
[08:05] Yeah. That's at least our premise at, at, uh, Azumuta. They, they, they need working instructions. They also need a way to understand these working instructions. Yeah. And then also they need to have some kind of, uh, data, training data to learn how to do the task. Yeah. Or other ways, uh, for example, it doesn't have to be always like imitation learning with these, uh, large behavioral models. Mm-hmm. Can also be, uh, just, um, old school but good computer vision together with, uh, task planning and motion planning, for example.
[08:38] And are there, there, uh, s- specific things that's changed in the last, say, three years? Uh, as so many things changed with LLMs, are there specific things, technology changes, hardware changes that shifted in the last few years? Yeah. Well, many exciting, uh, uh, things of course, like, uh, to name one is on the mechanical side. Uh, long time, uh, mechanics had not that much progress, but like in the last two years, what we see coming up is a lot of anthropomorphic hands at actually, um, affordable prices.
[09:14] Yeah. Um, which of course enables us as a researcher in longer term, um, in automation industry to, to handle more variable goods. Uh, then of course, with advent of, um, vision language models and large behavioral models, uh, we can give, let's say, robots, so to say, common sense, like, uh, because that's still lacking today. A robot can do very good, uh, computer vision, it can recognize objects, but having common sense and this glass is full and that one is empty, uh, is way harder, uh, to deal with.
[09:52] And, and, and that's also something we, we try to research because- Mm-hmm … that's what humans definitely have is a sen-- a common sense. If something comes up, the human has to do something. Okay, I cannot do it because I'm lacking this or this components. Robot will not… If it doesn't see it in the past, will not act upon it and, uh, yeah, will just try to do the task. Mm-hmm. You have a lot of, of, uh- Public models like, uh, public models, I mean, uh, VLMs and VLAs from Groot and Pi and, and what have you.
[10:26] Um, do you use those in your research group or do you develop, uh, your own things? Like how do you make sure that, like there are big budgets, uh, uh, NVIDIA NVIDIA's budget is, uh, is, uh, I think almost unlimited. And so how do you, how do you say, "Okay, that's what we are doing ourselves and, and that's what we are using off the shelf"? Yeah. Fir- first of all, our lab tries to id- identify niches where we can be strong, and, and one point where we are strong is interdisciplinary, like combining, uh, sensing, electronics, uh, making customized, uh, fingertips, hands, skins, um, and leverage that with the state-of-the-art artificial intelligence.
[11:10] And, and on that side, uh, um, we start from, for example, uh, LBMs or these large VLAs, um, but then we fine-tune them with local data, um, often in collaboration with, uh, industry or other stakeholders. But is this true, like with research lab we- Might, might be difficult for us to find the resources to train these very big models. Mm-hmm. But also i- in truth, we knew that for like, um, for large language models, we needed language as an input to get language as an output.
[11:46] Yeah. For robots, to be honest, it's still an open question. Which inputs do we need? And work instructions will be one of them, but what other inputs do we need to have as output robots doing actions? Yeah. So it would also be, like, very costly for us to just try around, um, train very big models for months, and also collect the data for it. Yeah. So we're more like, "Okay, but we need to research what type of data we need to collect." Yeah. Hence the… Hence why we were focusing on, okay, let's make these fingertips, uh, feel multiple types of things just as a human do, so.
[12:20] Yeah. Yeah. Yeah, and then we can adapt these LBMs so that they can also understand tactility or, or maybe temperature or other, uh, signals from the hands, the skin- Yeah … and to- What are LBMs? Uh… Uh, large behavior models. Yeah. So, so the, let's say, the ChatGPT, uh, models for robots. Yep. Okay. Uh, they have their own LLMs, the robots. Yeah, because you have vision, the actions, uh- Yeah. Okay … and hopefully, uh, uh, uh, we are also convinced that you also need tactility and, and other information.
[12:53] Yeah. What is your view on, on using, like you need to use training data to make those models? Uh, I hear, like, the assumption, like, a, a humanoid's, um, modality is useful because we have so much training data on YouTube, and, like, we, we have so many examples, and we can also, uh, put on a mock-up suit or w- whatever, uh, to generate this, this, um, example data. Uh, so a, a humanoid robot modality makes most sense. Do you agree, or do you think, oh, that's not necessarily the case, it can also be something else?
[13:32] Well, I, I would say I tend to disagree because, uh, collecting all that data is still very time-consuming, huh? Like, um, uh, large language models, uh, to train these models, there is a huge, uh, amount of data available on the web and, and, and everywhere. You have digital libraries. Uh, you can get data internal from internal companies. Um, while for the robot, what we still see is someone has to teach the robot by executing the task remotely to- together with the robot.
[14:06] Um, so that's the dominating trend, I would say, today. And that's, yeah, just as time-- so time-consuming because it has to happen at real time. And the question for us as a lab is how can we, let's say, reduce the training time? How can we do that more efficiently by adding other sources of inputs, like Andreas said, uh, apart from the video data? Can we add other tactile information, for example? Yeah. Um, and can we also use other representations than pixels to accelerate the process, for example?
[14:39] Yeah. And so concerning this data collection We know that like large language models, they have very good performance right now, but they still make mistakes and we can ask a question, yeah, can we afford these mistakes on robots? But putting that aside, if we like converse the data set size of something like ChatGPT and compare it to like the largest robots data set size we have right now, and if we take into account like the data collection farms that are currently running around the world, and these are like, uh, all cubicles of people next to each other with a VR headset and these remotes operating a robot.
[15:20] If we take that data collection rate into account and we want to get to a data set size that's like ChatGPT is using, then it's more, it's more than, more than 100 years of collecting more data like this. So if we take the current tech and just keep on doing what we're doing now, then I would say, yeah, we don't have general purpose robots in the first 100 years, so. One, 100 human years, so you can- Yep … parallelize it or- Yeah, while we're paralyzing.
[15:43] Yeah. So the, the current- Yeah. The current data collection rate, if you ex- if you keep on having the current baseline of like what we're collecting at, the rate we're collecting at- Yeah … it's more than 100 years of, uh, collection. So. Yeah, yeah, yeah. Okay. And on the other hand, it's very tempting, huh, because you have-- this becomes also very accessible. So you see many companies, let's say, fine-tuning or, or training their own, uh, small, uh, VLA, uh, um, to do a task.
[16:12] But then again, the question is, how does it generalize? Uh, to, to show the robot it has to pour water in a cup. Mm-hmm. That's easy, and you can train a robot within 10, 20 hours to do that. Uh, and even an 18-year-old can do that. But the question is, can it pour in any glass in the world? That's the question. So while we do have like experiments from like medical sciences where they, um, paralyze like, uh, people their fingertips and then see how well they can take a match out of a box and light it, and then they see like success rate of people drops to like 25% because you cannot feel anything- Yeah more with your fingertips.
[16:51] But then you still rely on the senses you have in your arms and your muscles to do this thing. These are things like currently we are starting in the robot world to take these things into consideration, but- Yeah … hence, again, uh, interdisciplinary is very important for robots. Yeah. Yeah. And yeah, the question is how much are these fundamental research topics taken into account- Yeah … in the current humanoid trend, which Frans is mentioning, just use the same trick as, yeah.
[17:17] Yeah. Large language models, yeah. Yeah. Yeah. Yeah. Do you see that as a part of your teleoperation so that you can feed back this, this tactile information back to the, the teleoperator? Ideally, we can, yes. So that really a human can feel what the robot is feeling and vice versa. So, uh, what's for, for example, something that's really difficult for robots to do is opening doors. Opening doors, it's such a tactile thing we're doing. Like, we can feel where the hinges are just by the f- the, the friction we feel, uh, in our whole body when opening a door.
[17:47] Um, it would be nice if we have a way when teleoperating this robot that we can communicate what the robot is, like, on a joint level, uh, feeling the resistance on the door. Yeah. Um, but it's hard to develop these- Yeah … uh, devices. Yeah. Okay. 'Cause there was… Uh, I had the question on, like, the human-robot collaboration, and for training purposes it's obvious, but once in production, for instance, is it… I think you already answered the question, is it like an, a necessity because the world is like that and you still need that collaboration?
[18:21] Or is it more from design that we still want, uh, humans somewhere in control or in support of robots? Well, I think, I think it depends a little bit on the use case, eh? If you have, uh, high volumes and, and simple basic objects or repeating objects, I think in the end the robot can do it autonomously. Mm. That can be the goal. Of course, if you have a high task variability, a high goods variability, then you might want to be able to show the robot this new task without having to go back to the company who developed the robot- Yeah.
[19:01] Yeah … or the whole setup, eh? One, uh, one big question I think is still a robot can do 100 task, how can you teach the, the, the 101st task? And for me, it's a really technical thing also. Mm, because, you know, I'm not a, an ethical expert or anything, so I don't know about whether it's a design choice or anything, or at least there are more people suited to answer that part. But, like, on a technical side, the general premise is that we won't need humans anymore if we have a robot that can do the same things as humans can, which is, like, our hands are magical, like, dexterous manipulation.
[19:35] That's, like, the thing we're trying to achieve, having the same, uh, dexterity and force that we have, um, and the, the knowledge on how to operate it, um, as humans can. This is, like, we're still f- very far away from that, so… And we're not going to see that the next couple of years, so we will need this human in the loop, uh, to collaborate with, to, like, uh, the robot is that, for example, does the heavy lifting of an object, but giving it over to a human on the table or whatever, or directly to the human to do more, like, a dexterous task with their hands, for example.
[20:09] Uh, so this human-robot collaboration will remain an, a important part. Yeah. Is, is there any, uh, approach of the, the, uh, companies that are, uh, developing humanoids, like Unitary, Figure 01, uh, Neo, uh, Optimus? Is there- A difference in those and like do they approach it in a different way and, and do you think like, okay, they are approaching it in a way that might be better than another? Well- You have a favorite basically? Oh, we, we have favorites- … of course, but, uh, let's say, and what they all do in common or all have in common is, is a focus usually on humanoids- Mm-hmm … which I think is a little bit strange because in most places you don't need legs, for example.
[20:56] Right. It's way cheaper to have a, a, a wheeled base. Um, and another focus they always have is imitation learning. And so they are leveraging, um, VLA's, LBMs and, and, and, and trying to push towards that. Um, what differs sometimes is the way how they are collecting the data. So, so some use their repos. I think, for example, Google has YouTube, so it has access to a lot of human, uh, videos. So they are, uh, we don't know actually, but we might think they do it like that.
[21:32] Uh, others, they, they just use, uh, these, these, uh, VR controllers and a VR headset to remotely control the robot and collect the data. Mm-hmm. Um, some companies, uh, for example, I was in China in, in, in August to a conference on robotics for industry. They even sell this, eh, like this, eh, like they-- you can buy a humanoid robot together with the VR headset and the remote controllers- Mm … so that you can train this robot in your own context.
[21:59] Uh, for example, you have a supermarket, small one, okay, you can train the robot to do that task. Really? Yeah. Yeah. Yeah. And that's, that, that's their business. Mm-hmm. It's actually very funny. And then for example, another example is, is Sunday Robotics. They, they recently launched, uh, very impressive videos of, uh, uh, a wheeled robot he-helper, uh, or humanoid robot. Uh, you can discuss about that. Uh, what they use is, is, um, the humans, they collect the data with the same gripper as the robot does.
[22:30] Uh, so they have like this, yeah, they, they li- they, they don't use the un- the hands of the human, but they have this, uh, gripper, the glove, you can say it, with some sensors and, uh, the vision data and, and that's collected, uh, together with the robot. That kind of makes sense because the closer you are to the morphology of the robot- Mm-hmm the easier the robot can exe-execute the task. And it seems that they, they also have some, uh, yeah, nice performance in, in some household contexts.
[23:01] Yeah. A company that used to be a little bit different, but it's actually the company that's been around for a very long time, and most of us know, like is Boston Dynamics. They built, uh, the Spot robot, the Atlas robot, but many more. Um, it's really, they make great technology Now they have an AI institute behind-- around Boston Dynamics, so they are incorporating, like, uh, imitation learning and also, uh, things like reinforcement learning and such.
[23:26] Mm-hmm. But like couple of years before this trend, they were more like, "Oh, no," like, "we don't believe in cheap robots, so robots need to be very expensive, and we'll put crazy sensors in there, crazy hardware." Um, and they were more like traditional control theory, um, to control their robots and let it do tasks. Mm-hmm. Do mind, now they are incorporating, like, modern AI techniques. Um, but just to compare this, like some humanoids-- humanoid companies try to let their, uh, humanoid walk, for example, just by doing reinforcement learning, uh, which is different to like how Atlas, uh, is, is running around, so.
[24:07] Yeah. But they are also like converting like now, we think, but once again, it's proprietary, we don't know. Um, but now they use like a combination probably. Yeah. Yeah, and the same is true for Toyota Robotics Research. And there's a lab, um, we actually really like because they have an, an interdisciplinary approach, um, rooting in, in control theory and mechanical engineering, uh, and expanded that to, um, human and social sciences, um, to, to electrical engineering, to computer science, to combine that in, um, a robot helper and really helper that can do tasks together with the humans.
[24:49] Uh, and, and that's something I, I find that more impressive than yet another, uh, company just, uh, copy-pasting, uh, tasks. Yeah. When do you expect us to have a robot helper in our house to do random tasks or the, the things we don't like to do so that we have time to do the things we do like to do? The- Well, actually we already have robot helpers. But we just have many, uh, tiny robot helpers. Okay. But like one generic that does all-- that can combine all these tasks and- Yeah … put things in the dishwasher.
[25:27] It's always hard to put the date of course for that, but I would expect- But do you see that in, I don't know, three years or 10 years or 20 years or- Yeah, earliest 10 years. At least a decade. Yeah, at least a decade. Because at home we will need probably legs to walk around, but- Yeah at this point in time, look at the videos, like you don't want to get close to these humanoid walking robots that weigh like 100 kilograms. And you see it in the videos, they keep their distance except for when there's like a big table in between, then they come close to the table where the humanoid's like one meters from them.
[25:53] But in general, they try to stay like two to three meters away from them because these things, if they fall, they first try to recover by pumping a lot of energy in their legs. But if you get like a robot, 100 kilogram, that's for a couple of milliseconds putting a lot of energy into these legs and you get it on your feet, yeah, I think you would need a, a medic at home for so. Yeah. And apart from that, I would say again, task generalization, huh?
[26:17] Um, let's take the example of pouring drinks in a glass. Uh, in, in our lab we are confident we have a very good system, uh, um, that has 99.3%, uh, uh, success rate. Yeah. But there's still this 0.7% success rate where the glass breaks. Oh, yeah. And then what? Yeah. Yeah. What does it mean? This- What does it mean? Yeah. And of course, we are confident in that number because… and, and we can deal with any kinds of glasses that are used in houses and, and elsewhere.
[26:49] Most companies, what they show is just one type of glass or maybe two or three. Mm-hmm. Um, but, but okay. And what's the, the success rate and what's- Mm-hmm … uh, the generalization capabilities? Mm-hmm. And they, they, they are very quiet about that. And then, yeah, when something breaks, then what? Uh, the robot cannot recover on that. Mm. Uh, so I think if you would buy a humanoid robot now for your house, you have a lot of cleaning to do. Yeah.
[27:16] But so not to catch us on the number, like- Never with it … we think a decade, but we also think like As a society, I think we will slowly change the definition of a humanoid. Like, we'll probably, given the marketing budgets right now, we'll still call it a humanoid, but they will probably give it a set of legs, wheels, and they will still call it a humanoid. Yeah. And yeah, so it's also hard to, like… The current form, like legs and the humanoid hands, which, which, of which they claim they can do the same thing as human hands or at least try to achieve, not the first decade.
[27:45] Yeah. Yeah, and we, we don't want to be pessimistic, eh? Again, there are a lot of opportunities- Mm-hmm in, in what we like to call semi-structured environments, eh, because the house is a total chaos, eh? But there are many environments where there is some structure, eh? Like the SMEs, eh, small companies, they have some structure to work there. The same is true in, in healthcare, eh, um, hospitals, um, elderly care centers. They have very similar rooms, sim- similar layouts.
[28:17] So I truly believe that you can mean something in those locations- Mm … first, uh, and that makes sense. But the household robot, yeah, it's, uh- Yeah. Still, yeah … it's still very hard. So you, you expect, like, a, a more, uh, dedicated laundry folding robot sooner than a general, uh, cleaning robot in your house? Yeah. Yeah. Yeah, I, I expect, or as a lab, we expect that, uh, today, yeah, you have, like, these wheeled cars which drive goods around in, in even hotels if you go, for example, to China.
[28:51] And there are also some, um, in Belgium in restaurants already. Mm-hmm. Mm-hmm. Um, I truly expect that they will get some arms or some manipulators and, and do some basic tasks, like maybe cleaning up the table, but not in your house, but, but in a- In a- in a restaurant or in another- Yeah, standardized environments where- Yes. Yeah, yeah, yeah. Mm-hmm. Okay. Yeah. And that's already challenging enough, eh? Yeah. Okay. How do you see, uh, open source versus closed source approaches?
[29:19] I know that there are, like, uh, projects like, uh, Le Robot on the Hugging Face and things like that that collect, uh, huge data sets because they are open source, cheap. And, uh, do you think that's also an approach that can win, or do you think, you know, it will be from the privates that have larger budgets, uh- Well, what I do like about these open source approaches is the same with, um, some YOLO implementations for, uh, vision, uh, that they make it very accessible.
[29:49] And, uh, some, let's say, very creative bottom-up approaches might pop up, huh, because now you see a lot of, uh, even school kids working with that, eh? I, I, I know kids of 16 years old who, who bought Le Robot, eh, one, an AZERO 101, eh, and, and they use the imitation pipeline, and they do creative- Uh, tasks with that. Yeah. Um, so that's really interesting because, uh, it, it gives, um, creativity to the people and, and, and, and we might get ideas from that.
[30:24] Yeah. Um, but on the other hand, again, the accessibility comes with also sometimes over expectations. "Look, I-- my kid, my 16-year-old kid can train a robot to do this task." Yeah. "Oh, yeah, we-- next year I-- we will have much more in, in companies." Yeah. Uh, so that's a counter side. And the same I think we saw in, in computer vision, uh, with these deep learning models, um, who became very accessible for basic tasks, but still, uh, not all vision problems are solved, I would say.
[30:56] So, um, even there, uh, there are some difficulties. Yeah. Yeah. The difficult thing with closed source, it's It's hard to find a moat, like, around your business model, like, to protect your business from just people copying the thing you're doing. Same with humanoids now. We have so many humanoids, um, companies right now because, you know, you can just copy-paste the thing. You can also collect the data. You can use open source models. So even we, we've seen this thing with, um, uh, with language models.
[31:30] Um, open source was just, you know, always a couple of months behind the closed source models. They were not as good, but, you know, there's also a business case to make, that if you can run these models free locally, privacy-wise more interesting, then, you know, it has, it has, it has a, it has a place in the landscape. So, um, so that's, like, the really difficult thing, and we're seeing it now with the whole, yeah, large language models, uh, out there.
[31:57] Um, open source is always catching up, so it's still to see whether this will be true for, uh, the robotics models. Mm-hmm. Um, and we don't know yet, but if it's also the case, then I would think, like, uh, uh, it forces, like, companies to make things open source, which I think as a society is a net benefit. But we're academics, so, um, makes sense that we… Yeah. Do you think there is a, a difference in, like, uh… Of course, we are on the, the SME side and working with factories and working with manufacturing environments.
[32:32] Uh, uh, I, I think, like, our thesis is that those large enterprise, uh, manufacturers, they have a lot of people, a lot of data. They will actually have a moat compared to up-and-coming, because, uh, up-and-coming companies. Because they have s- uh, large amounts of, uh, training data. They… Do you think this is true, uh, first of all, and do you think, like, will this be the case going forward, or will the foundational models become good enough that the smaller enterprises can also start using them?
[33:06] Like- Well, one of the problems is, of course, that, that data and, and automation is a bit similar to, like, in, in the field of robotics. It's… There is plenty of data, but nevertheless it's still Not that much data. It's a niche, so to say. If you look at all the foundation models, if you would look at, at the proportion of data from, let's say, the make industry, uh, it's a very small portion. And so the generalization capabilities to use it in a small company will be, uh, will be limited.
[33:41] Uh, and that's why I think it's really important for companies to value their data and their data sources and to leverage that. And I think that's also where we can play a role, uh, uh, as a regional, uh, entity or like as Belgium or even Europe. Uh, we can maybe put efforts in, uh, leveraging our automation industry by maybe combining sources, data sources together, and maybe we have their models trained on that. Yeah. Um, because yeah, we cannot win on the LLMs, but maybe we can, uh, by combining efforts and, and on niche data, we can maybe at least have a good, um, yeah, model, foundation model on, for the automation industry.
[34:31] The same is true for language, uh, like, um, okay, we, we, we cannot make a large language model for English and reasoning capabilities, but we can make maybe one w- that handles, uh, dialects very well for basic conversations. Mm-hmm. And that's really useful for our local, uh, uh, industry and, and hospitality, um, because for instance, people in West Flanders, uh, if you would put a robot in elderly care or in a hospital- Mm-hmm yeah, it doesn't need to speak English, yeah?
[35:01] Yeah. It needs to speak, uh, the, the local dialect. That's what the people want to hear, and it doesn't need to talk in very complex ways and, and needs a lot of stuff. No, it needs to do basic conversations, and that's feasible, I would say. And the same is true for the automation industry. Maybe don't want to catch it all, uh, but a smaller niche, and I think we have their data available, high quality data, because, uh, we have a very, uh, good industry here.
[35:30] So, so we have high quality data, and maybe we should le- be leveraging that. Yeah. If you had a, a, a manufacturing factory yourself, would you today already start collecting data, or do you think that's not necessary, the robots are not yet good enough, it can wait? Well, I personally would have started 10 years ago with collecting the data. Yeah. I f- of course, it's little bit a biased opinion. Yeah. But I would start, uh, I, I would feel very late if I would start collecting data now.
[36:04] Yeah. I think they, they should, and, uh, collecting, uh, be collecting data. And if they are not yet doing it, uh, they should start definitely now, because that's huge. Uh, there's a lot of value there. Yeah. Thank you very much. Very interesting. Thank you very much for all these, uh, these interesting discussions. For the people, uh, watching or listening, with what questions can they approach you? How would you be able to help them? Is there anything you would want to say to them, uh?
[36:31] Yeah. So as a lab, we are always interested in collaborations with industry. Any company can come to us, especially on, uh, problems in robotics and AI, with a focus on, um, let's say the manipulation, handling of, um, challenging objects like deformables, glasses, um, shiny objects. We're always willing to, to help. Then we actively also try to put some common sense into the current inflated expectations, uh, around robotics to give, like, a very clear view on what's currently possible, because there's a lot of new possibilities out there, but also what, what is currently impossible.
[37:11] Mm. Yeah. Great. Thank you very much. Thank you. See you. Thanks. All right.
