MV Pro 006 – Visual Inspection Planning w/ Medabsy

1. Introduction to Medabsy and the Interview
Host Josh Eastburn introduces the episode and highlights Medabsy’s innovative approach to visual inspection systems.

2. Medabsy’s Founders and Their Background
Max Grau and Petra Gospodnetić share their academic and professional backgrounds, explaining how their expertise in mechatronics, computer vision, and research led them to found Medabsy.

3. The Impact on Industry and Ideal Users
Max and Petra identify the primary users of Medabsy’s platform—system integrators, R&D departments, and optical component manufacturers—and discuss how their solution benefits these groups.

4. The Challenge in Visual Inspection System Design
Max and Petra highlight the time-consuming, trial-and-error nature of traditional visual inspection system design, and how this process is inefficient for many companies in the industry.

5. Medabsy’s Solution: Virtual Design and Synthetic Data Generation
The core technology of Medabsy’s platform is discussed—how it enables companies to simulate visual inspection systems and generate synthetic data before hardware is even available, speeding up the design process.

6. Technologies behind Medabsy’s platform
Petra delves into the advanced technologies Medabsy uses, including Monte Carlo simulations for light propagation and stochastic geometry models for defect simulation, which ensure high realism and control over physical parameters

7. Medabsy’s New Product: Medabsy Light
Medabsy introduces Medabsy Light, a cost-effective version of their platform designed for smaller users or those looking for a simpler, more affordable inspection system simulation tool.

8. Medabsy’s Growing Presence in the Industry
Max and Petra talk about their participation in industry events, including the EMVA business conference, and the positive feedback they received, particularly for their Wide-Level Solution that accelerates hardware sales.

9. Future Plans and Upcoming Developments
Max and Petra share their excitement for upcoming developments, including their involvement at the Automatica Trade Fair in Munich, where they will continue showcasing Medabsy’s capabilities and advancing the machine vision industry.

Episode transcript:

1. Introduction

Josh Eastburn – Host

Hello and welcome to the MVPro Podcast. Back in episode three, Mark Williamson, our European Editor at large, gave a special mention to machine vision startup, Medabsy, for their new approach to front-end vision system design. I took that cue and reached out to Medabsy for an interview. In today’s episode, I share my conversation with Medabsy’s co-founders, Maximilian Grau and Petra Gospodnetić. Max graduated with a Master’s degree in Mechatronics from the Munich University of Applied Sciences and worked for 10 years as an application engineer in machine vision. His hands-on experience optimizing inspection system hardware, programming image processing software, and commission functioning real-world systems for a number of different customers, familiarized him with the recurring problem of missing necessary components late in the project life cycle, which led him to form Medabsy. He currently serves as CEO. Petra is Head of Visual Inspection Planning and Synthetic Data Research Group at Fraunhofer Institute for Industrial Mathematics, ITWM in Germany, and Co-founder of Medabsy, where she serves as CTO. Her research group is focused on developing virtual tools for visual surface inspection and rule-based synthetic data generation for AI in inspection. She has received her master’s degree from the Faculty of Electrical Engineering and Computing at the University of Zagreb, Croatia, in 2016 for application of computer graphics methods for surface inspection in cooperation with Fraunhofer.

This work led to the introduction of virtual inspection planning, for which she received her PhD, summa cum laude, in 2021. Her research on inspection planning joins several major research fields such as computer vision, computer graphics, and robotics, with focus placed on industrial quality inspection. Please enjoy my interview with Max and Petra.

2. Medabsy’s Founders and Their Background

Josh (01:54)
Okay, so now that I know how to pronounce the name of your company correctly, tell me in one sentence, what does Medabsy do and how do you do it?

Maximilian Grau
Yes, okay, I will try it. Perfect. We are the startup Medabsy from Germany, and we help companies planning or implementing visual inspection system to do up to 90% faster by simulating the system and generating synthetic data before you have the hardware or data.

Josh
Perfect. Well planned. Petra, do you have a take on that?

Petra Gospodnetić
I fully agree with the sentence. What we do now is we virtualize all the steps that people had to do manually until now and collect the data and test stuff. We virtualize it and give people a possibility to sit by their laptop and just design the whole system and simulate it with a few clicks.

Josh
Up to 90% faster is quite a claim, and we’ll dive into that. What I’m wondering, first of all, is who really needs that right now in your mind? Who’s your ideal user?

3. The Impact on Industry and Ideal Users

Maximilian Grau (03:00)
I think there are two groups, also what we identify. The first group are the people who are implementing the inspection system or planning the inspection systems. These are, for example, R&D departments of bigger manufacturers. They have also applications engineers and implementing this inspection systems. Also like smaller or middle-sized companies like you call it system integrators who are just specified to integrate this inspection system. The second group is our manufacturers of components, of optical components, as it means their hardware provider for cameras, lenses, illumination, because they see also, Hey, this simulation tool is a great sales tool. It helps to accelerate the sales.

Josh
Okay, because it lets them easily put together a demo system. Is that the right way to say that?

Petra Gospodnetić
Yeah, so, if you think about their sales process right now, it’s basically face-to-face. It’s trade shows, face-to-face, conferences, networking, talking to people, and sending their products for testing or taking in smaller feasibility studies to show that it works. That all requires manual work and takes a lot of time. Having a tool that can virtualize your components gives you a possibility to free up a lot of that time and not only free up, but reach a much wider customer base.

Josh
I can respect actually the amount of time that goes into making a sale, all of that front-end work. That just leads to a proposal which may not turn into a sale in the end. All of that effort. I can see why you’d want to reduce the time spent on that. Tell us a little bit more, Max, about the current state of the industry. What effort is involved in the visual inspection system design and planning today?

4. The Challenge in Visual Inspection System Design

Maximilian Grau (04:46)
Yeah, today there’s a lot of work to implement this inspection system. Normally, the process for a machine vision project is normally that the manufacturing or a company say, Okay, I have a new product or something, and I have to inspect my product. I have to look if the quality is every time good. Then the design, they think, Okay, should I do it myself with our R&D department or should we contract a system integrator, for example, to implement this machine vision system? Okay, let’s say they go to a system integrator. Then normally the next step is that the system integrator says, Okay, I need some products, some okay parts or some with defects because I have to first identify which hardware I need to get a really good image. Because if you have a good image, then you can also analyze it with a software like image processing software or AI software. This part, it takes a lot of time. It’s in the most cases, it’s unpaid. Why? Because they have to look with try and error, Okay, which camera, which lens, which illumination do I need? Normally, in the most cases, the tricky part is the illumination because you need with the same hardware, you have to analyze many things on the product.

5. Medabsy’s Solution: Virtual Design and Synthetic Data Generation

Maximilian Grau (06:02)
For example, if they have defects or maybe you have to measure the size and everything or the color, there are many tasks in machine vision. There are some projects that only take some hours to get the inspection system, but there are also projects that take months to get the right hardware. They have to train a lot of stuff. Then finally to say, Oh, with this setting, I can solve this machine vision project. Then they make an offer and then it’s continuous. But this take could take a lot of months.

Josh
Like you said, that’s potentially all unpaid work.

Maximilian Grau
Could be, yeah. If it takes more time, then they know, Okay, I have to do. It’s called a feasibility study. The customer pays. But in the most cases, what we’ve heard from our customers, there’re also system integrators or departments to say, No, it’s unpaid.

Josh
For system integrators, that makes sense. I could see if you’re an internal engineering team, probably that’s not the case, right. But there’s certainly a financial incentive to shorten that time, particularly for a certain class of users. Are there other specific problems that you’re targeting in that process you’re trying to solve for people?

Maximilian Grau
Yeah, I want to add one problem. Maybe, Petra, you have also some thoughts about it. One thing is also if you know which hardware you need, then it could be that after implementing the inspection system into the factory, into the assembly line, everything, that afterwards you are checking your inspection system with real products, with real parts from the production which are fresh, produced. It could also be that it doesn’t work. Why? Because the first parts are no representative, like the parts which are produced from the factory. It could also be that you have to rethink your inspection system, or maybe if it’s just an easy error, then you can just fix it with the software, if you change the software. But it could also be that you have to change the hardware, or it could also be that you can’t inspect this product. This is a huge risk.

Petra Gospodnetić (08:05)
If I can jump in. The thing is, we are human, right, and how do we, at this point, design systems? We go into the lab. There is usually an expert who says, Okay, I’ve been doing this for the past 15 years. I know what kind of light and camera setup I need for this specific type of product and defect. They will run through their already known setups that will mostly work, do the job, and they will test them. Now, the thing is every product is different. Even though there is a set of rules which one could apply and take from experience, there is always this whole untested area which has not been tested because the person doesn’t have time. Because every time you need to remount the camera, remount the light, reposition the product, you don’t have the eternity to find the perfect solution. You find a feasible solution. A feasible solution is not the perfect solution. Now, imagine, and then you ended up with the system that you say, Okay, this will cover, let’s say, 80% of the required specification. You go to the customer and say, If 80% of the defects, and I mean, and 80% of the types, for example, are covered, is this okay for you?

Petra Gospodnetić
They say, No. Okay, then you find another one that covers 95%, maybe. Is this okay for you? They have to go back and forth and agree on customers, what defects is possible to find and whatnot. This takes time. Now, imagine if you had the possibility to save that time where you are physically remounting the light and the camera and retesting and retaking the images. Rather, you can just say, Hey, generate, I don’t know, 15 different setups. Let’s see, I think these setups will work, but can you propose some other ones? Oh, wow. Maybe it works. It gives you an opportunity to not only speed up your process, but actually test things that you haven’t tested before. This is where we are going with the part that we call ‘inspection planning’. We leverage this capacity to simulate things, to actually explore the solution space. Say, is there a system that can do better with the same hardware? Or if you would add a different hardware with how much impact would it have on your system?

Josh (10:31)
Very interesting. Okay, so you’re saying the typical process is iterative. We’re trying different setups and trying to maximize as much as time and money will allow, and we’re trying to approximate this ideal by an ideal design. But the reality is that maybe we’re getting to 80%, and if we’re at a stretch, we’re getting to 95% with the system, with the tools that you’re putting together. The goal then is to have a design that better fits the problem space. It’s just a better solution overall.

Petra Gospodnetić
Exactly. And think about this, you have a test benchmark, you have a test bench in the lab. And while you’re testing one product, you cannot use this bench for another product. So that means the next product waits, A, for the bench to be released or B, for your expert to be released. And this is another benefit for a customer. So you have virtual benches. You test as many products entirely as you want.

Josh
There are multiple types of efficiency that we’re talking about. The release of expertise makes a lot of sense also because we’re always constrained for within automation in general, finding engineering expertise. I imagine when you’re talking about machine vision, that’s even more the case. You have a smaller pool of people that you’re drawing from. Maybe you have one expert that you’re relying on in your company to do that work. So freeing them up is a win for the user. I’d like to talk a little bit about Fraunhofer and what took you there, and then how did you get connected with Max and Medabsy?

Petra Gospodnetić (12:01)
What took me to Fraunhofer? I originally come from Croatia, from Zagreb, and I studied computer vision at the Zagreb University at FER. At the point when I was finishing computer vision, at that point, there was not a big computer vision industry in Croatia. I wanted to do computer vision. I definitely saw the potential in it. That was about, let’s say, 10 years ago. I saw the potential in it, and I saw how it can be used for different kinds of things, and I wanted to continue working in it. As there was not a lot of companies that could give me this experience, I I learned outside of Croatia to look for potential internships because I was just looking for some practice for experience. This is how I accidentally found Fraunhofer. Funny thing, I had no idea what it is, and it is one of the biggest German research institutes. What’s that? The institute over 75,000 researchers. I had no idea! I got accepted. I probably came to my father and said, Dad, I got accepted at that fraun, fraun, fraun. He was like, Fraunhofer. Oh, you know what it is? And he said, Yes, Petra. I know what it is. And he explained it to me nicely. He was also an engineer.

Petra Gospodnetić (13:15)
I went there for an internship. They liked very much what I did and how I worked. I I had the background as a photographer, so cameras and lights were nothing strange for me. I was helping out on the inspection system for another company, so doing exactly this, testing which setup, which camera goes where? And I, Come on, as a computer graphics and combination of computer vision and light, expert, I said, If there is a better way. I mean, we know the rules of light. We know what angles work for which situations. There must be a better way. So this is how they invited me to continue my work with them in inspection planning for my master’s and then further as a PhD. So you can say that I have a Fraunhofer stamp on my forehead. And already in my PhD, I started leading other PhD students, and I already started in my research group on inspection planning. There we developed tools for designing systems, where should which camera optimizing it and so on, on different sets of parameters.

Petra Gospodnetić
But out of that came another point of saying, Okay, you can optimize all you want, but you need simulation. You need the photorealistic simulation. You need to figure out what you’re looking at. Out of that part came the synthetic data generation. This thing aligned perfectly with the vision of Max and Medabsy that was kind of a great coincidence. I had been working in this field for 10 years, and then about two years for the first time, Max reached out to me, Max and Richie, and let’s talk about inspection planning and what it is. From that point on, I became aware of them, and I was just looking at them like, Okay, they’re having some good ideas. But I mean, I’m a researcher, so in my heart, I am focused on solving problems, and that was always my thing. When I came for a PhD interview, my supervisor asked me, Why do you want to do a PhD? I looked him dead in the eyes and I said, I don’t. He’s sitting there and he lets this slight giggle out. Looks at me, looks at my boss who was also sitting in the room and like, Oh, I’m sorry, maybe I should make myself clear. I really want to solve important industrial projects and do good work. If this gets me a PhD along the way, that’s great. He liked it. We definitely had a great relationship throughout my PhD. I earned it with best grades

Petra Gospodnetić (16:00)
But that was always my point. I’m a techie. I like solving stuff. In order to bring technology from a point where it is in research and it is solving problems, you need the whole team, like Max, like I actually like Thanh. We can only work together to bring this technology properly to the market. I really like the team and we talked some more, and now we’re here.

Maximilian Grau
We are about two years ago, and I think one I think what was, it’s called the Heidelberger Bildverarbeitungsforum. This forum is a thing about every three or four times a year, you can do a presentation about the news in machine vision technology. Then we get to know, oh, they are researching in this area because for us, it was really new. Before, we did a research, of course: are there any companies who are doing this? We didn’t find it. But one reason is because they are still researching like Petra. This is what we thought, Hey, it’s really interesting. Then we got in touch and now we are working together.

Josh
Okay, so you’re very early to this industry.

Petra Gospodnetić
Yeah, you could say so.

6. Technologies Behind Medabsy’s Platform

Josh (17:08)
Tell us a little bit more then about the technology behind inspection planning and defect simulation, Petra. What are some technologies engineering might be familiar with?

Petra Gospodnetić
It all starts with the good old computer graphics. What you would see in the movies and the video games, this is where we started. Because the simulations there, of course, you need to develop them in a specific directions to match what you really need. But the approaches are very much feasible, and this is what we built up on. We are using Monte Carlo simulation to achieve proper light propagation simulation. When it comes to defect simulation, here we are using stochastic geometry models where we really model different classes of defects, and we have our mathematical models for it. This gives us a to have this high degree of realism as well as a full controllability with physical parameters.

Josh
You take as inputs to your tools CAD drawings?

Petra Gospodnetić
So, the assumption was Every company that produces something will have a CAD drawing. That was the initial assumption. The reality is most of the companies that produce something have a CAD drawing. But the good thing is all of them said, It’s like, okay, maybe we don’t have it directly, but we can actually provide it in a way. For those that don’t provide it, we always provide an option. Okay, look, we give you potential primitives or surfaces that resemble what you have. It will not be one-to-one, but you can test it. For example, if you have product with high curvature or surface, you can always create a resembling model to be used as an place.

Josh (19:01)
I feel like there’s a question there. So you talked about the precision of the geometries that you’re able to create. Certainly going to be better if you’re working from a model, a CAD model. How is it that you’re able to assure users that the synthetic images you generate because each material is hopefully going to be unique in some way, whether that’s the shape of the object itself, its geometry, or the materials that it’s made from, its reflectivity, interactivity and so on, how can you assure them that those images that you generate are useful or are representative of what they’ll see in the real world?

Petra Gospodnetić
That’s a great question. That is a very important question because in reality, there is no good measure for image quality or image realism. In the end, only measure is the observer, whether that observer is a person looking at the images or that observer is the network training on those images. There are metrics for measuring, quote, unquote, the realism in the images, but they always reside on certain features, which are not necessarily the features that you want to measure when measuring realism of, for example, inspection images that go in much further details. Those features in natural scenes will be different because you’re looking at cars, houses, people, and so on. Those features will be different from features of a closeup of a milled surface. Before I go really deep into this, I’m going to try to avoid this. But the point is, The metrics that exist for realism are not adequate for close-up images of inspection. Let’s go with that. What we do, we say, look, you have parameters of the light intensity, you have parameters of the sensor, you have the physical You have the parameters of the aperture, you have the parameters of the material reference, you have the parameters of the texture, which texture it is, what pattern it is, how it is produced.

Petra Gospodnetić
These are all physical parameters that influence the appearance. This is what can guarantee you that your image has been created based on physical appearance. You can say how close it is once you actually Like put side to side real image and the simulated image, but already based on the parameters, you know ‘I am there’. You can be assured of that.

Josh
It sounds like there is a quantitative aspect that you’re able to verify. You could say, were those properties retained, the properties of texture and so forth, that we can also measure in the real world. We can take a physical object that has been manufactured and we can, a human, visually can compare that to the generated image and say, yes, this has parity. This looks like an accurate depiction of the real thing. Is that right? Okay. As well as some quantitative aspects that you’re able to control, for example, like you said, camera properties, the size of the aperture and so on. Those are things that can be modeled accurately. Okay, I think I’m starting to understand. I think a technology that people will feel like is familiar to them is AI or machine learning. That concept has certainly become popularized quite a bit in the last couple of years. Help us understand what aspect that plays within your technology.

Petra Gospodnetić
One thing that we very, very, very clearly state is we are not using any AI or generative models for our system.

Josh (23:02)
Okay, great. Now I’m super interested because I had lots of assumptions in my head about how that might play a role. If it doesn’t, I definitely want to learn more.

Petra Gospodnetić
This is where we have models of different materials and different textures and different defects. When I say models, I mean mathematical models, so close solutions. We have specified physical parameters which are used to control the outcome of this model, so-called realization. What you do is you generate appearance with variation of these parameters. You give the parameters in ranges and you say, generate, for example, and you generate in ranges and you add stochastic part into it. Saying, Give me some randomness. We have some randomness in those models, and you can control that randomness. Saying, Okay, how big should it be? How strong should it be? But exactly this randomness in parameter selection, this slide jitter, you can say it, gives you this reality because every time you generate, it will have a different appearance, slightly different appearance, just like reality. Of course, you can nail it down for that every single time you click a button, it will be exactly the same. Of course, you can do that. But in general, if you’re generating a data set or you’re testing solutions, you want this jitter, and you can control how big are your ranges. Do you want to keep your data set only in the range of super realistic, what is likely to happen?

Petra Gospodnetić
Or you want to go super mad and say, Generate stuff which is physically plausible, highly unlikely, but yet physically plausible. What has been shown is that when generating data sets, for example, you want to go with both. You want to have your realistic data set, but then you want to have several data sets which have different sets of parameters It’s way out of range, saying like, Okay, this is crazy. Chances of this happening are basically zero. But it turned out it helped the network. It’s called ‘domain randomization’. It’s like you have a child saying, Look, now you’re in a completely different situation, and this situation should not bother you at all. So just focus on this one. It helps networks have to stay focused on the things that are actually important.

Josh
I see. You’re able to define the common case. If we’re imagining that there’s a normal distribution, you’re also training on what would be considered the outliers, the extremes of the potential data.

Petra Gospodnetić
Important thing, we are not training. We are providing data sets. We can advise for training, but we are not providing models. We will in the future, probably, but not at this point. At this point, we are just the tool, the tool for simulating and designing the systems and the tool for generating data system. Everything that you need for training and developing your system is in our tool. But we are not doing this because as a startup, we want to get the core well done, and this is what we’re working on now.

7. Medabsy’s New Product: Medabsy Light

Josh (26:06)
Okay, now I get it. Let’s see. A question for you, Max. Now that we understand the underpinnings of the tools that you’re developing, what is your tool called?

Maximilian Grau
Medabsy is the core product. What we are offering is, as we call it, a platform. But what we are now doing is we have specific products or license models, and they have a specific name. For example, now we published a new product. It’s called ‘Medabsy Light’. This is a reduced form of our platform, which is really cheap. We publish this. It costs €150 per person and per month. With this license, you can simulate different inspection system with our database and get, it’s called a preview image. It means you don’t get the highest synthetic realistic image, but you get a middle realistic image, which is also good and gives you a rough number, where I have to put the hardware, how the image would look like. This also reduces a lot of time for the inspection planning. It’s called the ‘Delight’ version. Then we have next product, it’s called the ‘Pro’ version. With this license model, you have the access to all functionality of the platform and you get the high synthetic realistic images. The last thing is it’s called ‘Wide-Level Solution’. It means that our A customized version of our platform runs on the website or on the server from our customers with their products, for example.

8. Medabsy’s Growing Presence in the Industry

Josh
Then obviously, Light just came out pretty recently, right in the last month?

Maximilian Grau
Yeah, this was a really huge thing. The next big thing, because for example, two weeks before, Peter and me, we were in the EMVA business conference in Rome. It was really interesting. We’ve got a really good response and feedback about the wide-level solution because in the business For instance, there’s a lot of hardware providers. They say, Hey, this could solve our problem or this could accelerate our sales. This is really important. This is also what we are developing and thinking. This is going to be huge. Now we go to the different trade shows, for example, this month in Automatica.

9. Future Plans and Upcoming Developments
Josh (28:22)
Tell me more about what you’ll be showing off at Automatica this month. I think it’s the 24th to the 27th, like that? Yeah.

Maximilian Grau
We don’t need too much stuff. You only need our laptop. Then we can show our product and also show some showcases. We have to honestly have to say we have to look which things we want to show. I think it would be a metal project where we can see how the reflections are on everything. What I can say is that I don’t know how we did that, but in this trade show, we will have three presentations.

Josh
Oh, wow.

Maximilian Grau
In one tradeshow. Yeah. I think on Tuesday, Wednesday. Petra I think you are pitching on Friday.

Josh
Is that right? Oh, yeah. As a startup, there’s always a startup stage, right? You’ll have a pitch there. Then are you each speaking separately or are the people from your company as well?

Maximilian Grau
In two presentation, I will speak. One is the vision huddles from automatic, from VDMA, sorry. They have this stage for Machine Vision. One is called Funding Box. This is more for startups.

Petra Gospodnetić
I will be in Smart Manufacturing. It’s a little bit off the direct. I mean, the target audience is there, but it sounds a little bit side because it’s not directly machine vision. But the idea is to give them a little bit more insight into what is possible with machine vision and how can one actually include some more intelligence also in the design process.

Josh
Absolutely. I can completely see how that makes sense. Where can our listeners learn more about Medabsy, about your work? Where can they find you online?

Maximilian Grau (30:00)
Yeah, it’s the first, of course, of our website, www.medabsy.com There is a lot of information and also direct contact to us. The second thing, I think, where is really good from our side, very good Where we get the good feedback is our LinkedIn page because we are updating every week or we try every week to updating what’s going on with Medabsy, what are the new things. I think LinkedIn is one of the best channels to get the news from Medabsy. Of course, we want to try to make more videos also with a YouTube channel where there’s interview or also product video is there, for example, online or some presentations. Yes, these are the three channels.

Okay, fantastic. I hope we send a lot of interested listeners in your direction. This is fascinating. This has been such a great conversation, and I’ve learned so much, which I really enjoy. So thank you to both of you.

Josh Eastburn – Host
A big, big thanks to Max and Petra for coming on the show. Doubtless, there will be lots more to report on in this nascent field in the coming year. Email info@medabsy.com or visit medabsy.com to try Medabsy Light yourself. You can hang out with Medabsy this week at the Automatica Trade Fair in Munich in the Startup Arena, Booth 328, Hall B4. MVPro’s European team will also be on site at the show, so reach out to editor@mvpromedia.com if you’d like to meet up. For tickets and information on the show, visit automatica-munich.com. For more on the latest developments in machine vision software, listen to our previous episode on the Best of Automate 2025, as well as recent coverage on AI in the March issue of MV Pro magazine. For MVPro Media, I’m Josh Eastburn.

.