Interview with Frank Park

Interviewer: I’d like to begin by asking you if you would share your full name, your date and place of birth and anything that you’re interested in sharing about your parents, family, background growing up, where you lived.

Frank Park: Yeah, okay. My name is Frank Park. I’m a professor of mechanical engineering at Seoul National University in Seoul, Korea. My education was, my college education, my grad school education was in the US. I did my undergraduate degree at MIT and my PhD at Harvard in applied math. I also did a stint at UC Irvine as an assistant professor. And after that, I’ve been at Seoul National University ever since.

Interviewer: Can you share a little bit about your earlier education, so your time in college, graduate school, any mentors that were significant to you?

Frank Park: Yeah, so let’s see, I was in college in the early ’80s, actually studying electrical engineering. I’m a faculty member in mechanical engineering, but I never really studied mechanical engineering until I became a faculty member. And I remember taking courses in signals and systems and various EECS courses. I was pretty sure I wanted to do graduate study, but I wasn’t quite sure what field to study. And one of my mentors at the time gave me some very good advice. And he said that your career as a researcher is going to span the next 30, 40, 50 years. So make sure you choose a field where it’s going to be relevant in 30, 40, 50 years, or remain relevant in 30, 40, 50 years. And that was very good advice. Because at the time, I’d taken an undergraduate robotics class and looked around at some of the robots at the time. They were fairly primitive. And so I said, OK, this is a field I think that’s going to be relevant for a very long time. So that’s how I kind of found myself heading toward robotics.

Interviewer: Had you had any experiences when you were younger that got you interested in robotics in the first place? Like any people, books, movies, anything like that?

Frank Park: Yeah. So a lot of my colleagues, they relate experiences where they were tinkerers. They tear apart the family television or radio set and build it back together. I was never much of a tinker. Well, I mean, I didn’t have things that I could break apart and put back together again. I remember taking a class in college taught by actually John Hollerbach in Chris Atkinson, who are actually now colleagues in the field, very senior, high respected colleagues, on sort of motor control, human motor control. And so this is basically about how human beings control movements. And so that was a very fascinating subject for me. And so as I mentioned, the robots at the time were very primitive. And so the first place that we turned to for ideas was humans. How do humans actually look the way they look, first of all? And then how do they move the way they move? And what are some of the ideas that we can take from the study of human motor control and apply to robots? So that was very inspirational for me.

Interviewer: Do you happen to remember the first robot you ever saw or heard about?

Frank Park: First robot, yeah, I think. First robots, I mean, back then the robots were all pretty much industrial robots. The robots in the lab, I think the most impressive one at the time was the one-legged and the two-legged hopping robots developed by Mark Raver, who was at CMU at the time. But yeah, there was not much in the way of hardware, especially compared to today, that really knocked your socks off.

Interviewer: What period of time are we talking about?

Frank Park: So this is the, I would say, the late ’80s, mid to late ’80s, to the early ’90s.

Interviewer: So when was it that you first started to work on robotics projects proper?

Frank Park: Yeah, so this is a little bit embarrassing for me to admit, but I never… So it was during graduate school, this is in the latter half of the ’80s, I never really tinkered much with real robots at the time. The robotics problems that I was thinking about at the time were, as I mentioned before, why do humans look the way they look? How do they move? And what are some of the ideas that we can take from humans to apply to robots? So this was all kind of maybe not surely theoretical, but it was in that domain where it didn’t require you to go into the lab and actually build and run experiments with robots. These days, of course, times have changed. And if you don’t do experiments with real robots, you’re not considered a true roboticist. But I think at the time, people are still trying to figure out what the foundations of the field were.

Interviewer: And as you moved on, what were… Can you paint a picture of the sort of core robotics-related work that you’ve done over your career? Like what different projects has been consistent through line or have there been different kinds of projects here and there?

Frank Park: Yeah, so later on in my career, after I joined the faculty and started getting funding and doing research projects, there were a series of sort of industrial or applications-oriented projects, some of them involving what are called parallel manipulators, and then a lot involving sort of just motion control of single arms, dual arms, typical robots at the time. But again, I think my main focus throughout at least the early… The first half, maybe even more of my career, was actually kind of trying to sort of develop sort of foundational tools for robots. You know, while pondering the question of why are human arms… The proportion of human arms… Why are… For example, why are fingers… Why do they look like this? If you look at the lengths of the fingers, they’re roughly in the proportion of 3 to 2 to 1, right? Why is that? I came to the realization with the help of my mentors, of course, that this is actually a very fundamental problem that’s actually… There’s a core theme, not just robotics, but I think in just all of science and engineering, that is that many problems can actually be formulated as trying to fit like a square peg into a round hole. You know, trying to make things that don’t fit fit. And so mathematically, these problems end up being problems in geometry, differential geometry to be more specific. And so I spent a great deal of my time in graduate school studying differential geometry. And I think that was… Those were the tools that really I relied on a lot. And they were very helpful in sort of developing tools for robot kinematics, robot dynamics, motion generation, motion planning. And I think they’ve had somewhat of an impact today.

Interviewer: Are there any specific robots or projects that stand out as particularly interesting or significant over the course of your career as you reflect?

Frank Park: Okay. Well, I can tell you about some sort of things that I was really excited at the time. And then it kind of… It was a lot harder than we thought. And one of them was, I mentioned parallel robots. There was a lot of activity in parallel robots, I think, in the early 2000s. And then I kind of was hopped on the bandwagon. And there was, I think, a hope within the community that parallel robots were actually much stiffer and as maybe paradoxical as it may sound, actually easier to build than serial arms. So people have been trying to develop parallel robots for machining. For machining, stiffness is everything. And they thought that, well, a parallel structure of the robot would make the robot much stiffer. But it turns out it’s not necessarily the case. It’s actually quite hard to make these parallel robots very stiff. And there was a lot of sort of belief at the time, even companies were developing parallel robots. Now you don’t see anything in that domain. I think they’ve all gone back to sort of the traditional architecture. And of course, the early walking robots, the hopping robots have now evolved into these amazing humanoids that you see walking around here on YouTube. So those are also very inspirational.

Interviewer: As far as your own research goes, what would you say are some of the main research questions that have guided your work? Have there been consistent research questions? Has the agenda evolved over time?

Frank Park: Well I think the elephant in the room, and I’m sure you talked to any roboticist, nine out of 10 will probably say the same thing, is the impact of machine learning and AI. Deep learning has been around for over 10 years now. When it first started making a real impact, and even in robotics, I think it was like early 2012, ’13 or so when people started using deep learning for grasping. So the old school method for grasping was use a camera, figure out what the object looks like and then calculate what the best locations are to place the fingers and grasp the object in a stable manner. Of course deep learning just completely changed the paradigm. And then so I think there was pretty successful, much better than sort of the classical methods. Still not good enough to really work in a factory setting. You have to get 99% accuracy in grasping. I mean just even 90% is probably not good enough for real factory settings. But there was a lot of excitement and at the same time a realization that these deep learning and machine learning methods for robotics, although they work well, they are hard to make work. There’s a lot of data that needs to be collected. Retraining is involved. You have to monitor performance. And so that was I think a big kind of a realization that it works, but it’s actually a lot harder to really make it really in practical settings. And then of course in 2022, foundation models, the advent of CHAT GPT came along and then that got everybody thinking, well, why not build a foundation model for a robot? So robot foundation models are just about all that anybody can talk about these days. There’s still a lot of ongoing debate whether this is feasible. But the community, at least for now, seems to be sort of maybe not converging, but there seems to be sort of a maybe consensus slowly building that robot foundation models are a very long term project. And these challenges make developing a practical robot foundation model very, very challenging. No, I think people are saying, well, if you’re going to build foundation models, you have to do it in a different way, different from what’s been done in vision and language. You have to solve the data problem. But you have to also think about the learning model itself, the architecture. And so there’s a word being tossed around these days called inductive bias. So why does deep learning work so well for image and vision problems? And a lot of people will point to the convolutional neural network, the CNN. Why does that work well? Well, it works well because it’s a shift invariant. So if you show an image of a cat to a computer, you can move the image left to right, up or down, and it still recognizes that it’s a cat. So there’s this notion of shift invariance. So that’s an inductive bias that the CNN captures. And there are other inductive biases. So CNNs were sort of the hottest thing. They thought that CNNs could solve all the world’s problems. And of course, now all people can talk about are these transformers, transform architectures. So what’s the built-in inductive bias for transformers? Well, it’s actually something called equivalence with respect to permutations. So basically, transformers are very powerful for understanding language because it’s less the sequence of the words is less important than whether the words are in a sentence or not, as long as they’re in a sentence. The order doesn’t matter that much. So it’s this idea of equivalence with respect to permutations. So what are the proper inductive biases for robotics? And I think that’s where the focus should be on. And if you can find the right inductive biases for robotics, then you can actually construct learning models, architectures, that will work well for robotics. And that will make the data requirements much smaller. It will make training easier. But in order to do that, you need to go back to the old textbooks. You have to start looking at pulling up all the old models of robotics kinematics, dynamics, contact, impedance. So I think people are just finally kind of dawning on this realization that, oh, wait a minute, we better not throw out the old textbooks yet. What if I need this to actually build foundation models the right way? That was a long answer. Sorry.

Interviewer: Not at all. It’s actually– it leads really nicely into the next question, which you’ve outlined a number of different challenges in terms of application of LLMs or different kinds of AI to robotics, et cetera. Have you, in your own research, faced specific kinds of challenges? And if so, were they technical, social, institutional? If and how any specific challenges affected your research, it would be interesting to hear about how that evolved.

Frank Park: Yeah. So some of the tools that I work with to develop robot techniques for robot modeling and control are ideas from what’s called a theory of Lie groups and just reminding geometry. So it’s hard, actually. There’s kind of a steep learning curve. And it’s hard to convince people that, yeah, it’s worth investing the time and effort to go up this steep learning curve to learn these rather abstract tools. So I think we developed model techniques based on Lie groups for developing models for robots. But it’s one of the challenges trying to get the community to buy in. They’ll say, well, it’s nice. It’s beautiful, this connection with this elegant theory. At the end of the day, what does it buy me? Why should I learn this? This hammer seems to work pretty well for all my needs. Why should I learn how to use this new tool? So that was a challenge that I faced for, I think, persistently, even today, to convince in this ethics that, yes, there is something practical and useful in this. But I think it’s not just me, of course. It’s the small but growing community of people who use these tools. But I’d say that we’ve had some measure of success. For example, words like, well, terminus like SC3. It’s fairly standard now. You go to a robotics conference like ICRA. Most people will know what SC3 is and use it to describe whatever part of their work is relevant to that. Screw theory, modern screw theory, things like that, Ramanian metrics. It’s now become part of the vocabulary of standard mainstream roboticists. So still a challenge, but I can see slow, well, maybe more better than slow, acceptance of these new geometric tools.

Interviewer: So by looking like a kind of classic picture of technology adoption in a way where you’ve got to persuade this group, whether it’s farmers or business people or whatever, that the cost of transitioning is worth it from where they’re coming to whatever it is you’re offering. So on the topic of tools, have any popped up for you along the way that have affected the trajectory of your research that you found particularly significant, helpful, made an impact in whatever way for your research?

Frank Park: Again, it’s the elephant in the room. It’s AI and machine learning. And of course, for me, it was– so I think maybe this might be an oversimplification, but if we were to divide the community into two camps these days, they’re the people who believe in end-to-end learning, scaling, that eventually if you have enough data and enough computing, then all these elaborate and difficult to learn modeling techniques are not needed. This learning model will take care of everything for you. And then there’s out of camp. No, not so fast. So I think the way to go is sort of the– I would say the bottom-up approach. To solve real problems– so I mean, in theory, I would not argue against scaling. If you had infinite data, infinite computing, yeah, of course, why not? Things would work. It’s kind of like the– I guess people call it the infinite monkey theorem. So if you had a monkey typing randomly for a billion years, then eventually it would produce Hamlet. I’m kind of skeptical of that. So it’s that argument. But it’s not practical. I think somebody mathematically disproved that at some point. I think so, yeah. I may have been. So I mean, it’s one analogy, I think, of this debate about scaling. But I’m of the school that you should use physics and mechanics-based methods and ideas from planning and control, see how far you can go. And then when things get difficult or you get stuck and you have some data, useful data lying around, then that’s the time to start using sort of data-driven machine learning methods, augment the existing models, and in that way, kind of go from a bottom-up way. And I think that’s much more effective, especially for real-world practical applications. That’s a way to make progress.

Interviewer: Have you had any hardware tools, so to speak, pop up along the way that have been significant at all?

Frank Park: So like I said, I’m not much of a hardware tinker, but I’m in awe of hardware designers. People are just– and one of my, I guess, my big regrets is that I never found a colleague who might have worked very closely with, over the course of my career, a hardware expert who I could work closely with and kind of work together with my sort of algorithmic software background. But these days, hardware is getting better and better. And again, I’m not an expert, but some of the remarkable progress that we’re seeing in today’s hardware is as a result of sort of these custom actuators, electric motors that are being designed, some sensors. And actually developing the models for these actuators and sensors, I think, is important. And so I’m kind of hoping to sort of look into these problems.

Interviewer: So for the most part, that’s been happening kind of off in a different domain than you’ve been observing it, but it hasn’t directly influenced anything that you’ve been working on personally?

Frank Park: Yeah, but I’m hoping it will. And I think it will.

Interviewer: So can I ask a little bit about funding? Who’s funded your research over the years and how you develop the relationships with those funders? And we’re also interested in the different ways that funding can influence research. So anything that you can share about the impact that that’s had?

Frank Park: Yeah, so the funding situation is different. It’s so different across various countries, different funding programs, priorities. So it’s hard for me to kind of, I guess, translate my experience compared to other researchers from other countries. But I think I’m very fortunate to be in a situation where government funding has allowed me to just do whatever I want without too much pressure on delivering deliverables, deadlines, things like that. I’ve done a few industrial projects of that nature. So in terms of funding, I’ve not had that much pressure. I’ve had a lot of freedom. But again, I think I was kind of in a lucky position. And again, it’s so different across the world.

Interviewer: Largely public funding then for the most part?

Frank Park: Yeah, I’d say about 70 to 30 public versus private.

Interviewer: Now, when you said industrial, would that imply like private sector or companies that you worked with?

Frank Park: Yeah, private sector companies.

Interviewer: Do you want to share anything about that or how those opportunities developed?

Frank Park: Interestingly, I’m getting a lot of even more. I’m being approached even more by companies, industrial robot companies, trying to solve assembly problems, manipulation problems in factory settings. And so they have very large robot hardware, software teams. And yet I’m kind of puzzled. Why don’t they do this in-house? Why are they coming to me to solve these problems? I mean, they’re not easy problems, but certainly they have the talent and the resources to tackle these problems. It turns out that I think machine learning is changing the game a little bit. So I mean, one application that I can tell you about is there’s an appliance manufacturer, also refrigerators, air conditioners. There’s a lot of sort of cable insertion going on, harness cable insertion. And people still do most of those tasks. There’s like a two-shaped cable. You have to, the person has to grab each and plug them into the right places. Sometimes the cables are tangled, so you’ll have to untangle them. It has to be done in 10 seconds. Failure is not an option. You have to get it right every time, 10 seconds. So to get a robot to do that every single time correctly in 10 seconds, pretty challenging. So it turns out you need machine learning. You need some AI. Well, you need perception. You need some data-driven algorithms to help you with the manipulation, to figure out whether it’s tangled and to untangle it. So I think we’re now starting to see sort of these, the machine learning, the algorithms and tools actually seeping into the industrial domain. And so there’s some new interesting challenges there.

Interviewer: We’ve been talking about what has unfolded and things that have been happening. But have there been any paths that you didn’t have an opportunity to take that you wish you had? So questions or problems that you would have liked to pursue or haven’t had a chance yet? Anything along those lines?

Frank Park: I see. Well, I think I told you that I wish I had more hardware experience or work with somebody with a more hardware experience. But as far as things that I wish I had been able to explore more, I think I’ve been kind of lucky in that respect, in that I haven’t had a lot of funding constraints. I’ve been able to kind of sort of look at what the challenges are, the time, and kind of explore problems. So, well, going back to that course in human motor control that I took early in my college career, I guess I kind of wish I had a chance to maybe go down that road a little bit more, study problems in the biological motor control more.

Interviewer: What would have propelled that interest or what do you think it might have yielded if you’d been able to look into that?

Frank Park: Well, going back to this idea of inductive bias again, so I said that finding the right inductive biases for robots is, I think, crucial to building robot foundation models. And one of the really important sources of inductive bias is actually human motor control. So, you know, take learning a skill, for example, you know, learning how to play the piano, or maybe playing tennis or golf. Let’s say golf is an example. So, you know, when you’re first learning how to swing, you know, a golf club, right? If somebody’s teaching you, they’ll tell you, “All right, just stand this way, you know, make sure you look at the ball this way, swing a certain way.” And as your brain is trying to process all these instructions, there’s a lot going… There’s so many muscles, degrees of freedom that you have to control that your brain is overwhelmed. So what does the brain do? Initially, you lock some degrees of freedom. I’m just going to worry about my shoulder. And, you know, while you’re doing this, you’re also relying very much on, “Am I gripping this… gripping, you know, properly?” You’re kind of noting, you know, the angles, everything. So there’s a lot of reliance on vision, on other sensing modalities to make sure that you’re following the instructions correctly. What happens as you improve, right? You know, we talk about your fingers learning how to play the music, your muscle memory, right? Developing muscle memory. So what happens is, like, at first you lock up some degrees of freedom, but as you get better and better, these degrees of freedom start to free up. And they kind of learn how to move on their own, right? And at the same time, you also rely less and less on, you know, if you’re learning how to play the piano. After some point, you don’t need to look at your fingers, you know, when things are going, right? So there’s actually, you know, a lot of work in human motor control on, you know, how people and animals learn motor skills. And so this is an important inductive bias, you know. So, for example, you know, if you look at baseball players when they’re catching a, you know, a fly ball, right? If you look at, you know, how they track the ball, usually if the ball is hit, they’ll take a quick look at the ball, and in their head they’ll calculate roughly their trajectory. And then they start running. They don’t need to look at the ball. And then as the ball approaches, right, then they start looking and they rely more and more on their vision and use their vision to guide the glove to catch the ball, right.? So there’s this transition from, you know, very low reliance on vision and feedback sensing to greater and greater reliance. So initially there’s like open loop, or feed forward, you know, move over there, right? And then at the end, now you’re trying to guide and find, you know, feedback and sensing. So this is, you know, motor skill learning, you know. And these days people are trying to teach robots how to do this and that, right? And so they’re collecting all this human data and throwing it at this large neural network model and hoping that it’ll process and magically spit out, you know, some recipe. We have all this beautiful theory from human motor control. So we have to figure out how to embed that into the learning.

Interviewer: So because an outfielder, I probably would have benefited from looking at the ball a little more sometimes. So looking back over your career so far, is there anything that you would identify as your most important contribution to robotics?

Frank Park: Oh, yeah. Wow. No. I don’t want to sound, you know.

Interviewer: You should. This is a good opportunity.

Frank Park: Well, I would say probably maybe sort of the geometric tools that, you know, not just myself, but, you know, our community of people who have used different geometry in lay groups. I think those tools for robot modeling and control and planning, I think, are now becoming more and more closer to the mainstream. I wouldn’t say they’re quite mainstream yet, but I would not be surprised if, you know, somewhere down the road, you know, they become part of the mainstream. I think that would be maybe a modest contribution.

Interviewer: Any in-modest contributions you want to share?

Frank Park: Oh, I wouldn’t go that far. But, you know, OK, well, maybe going back to this whole idea of, you know, problems, many problems in science and engineering boiling down to, you know, trying to fit a square peg into a round hole or maybe a round peg into a square hole. You know, there’s this very beautiful theory in geometry, you know, of how to fit these things, things that fit, don’t fit, make them fit in an optimal way. You know, I think I’ve been kind of trying to think of what a good example would be. So I came upon this recently, you know, this while shopping for jeans, you know, these are like stretchable, right.? They’re all arranged these days. They’re extremely comfortable. So even if the size isn’t the perfect fit, if you have these stretchable jeans, you know, they somehow conform nicely to your body. Yeah. And once you wear them, you can’t wear anything else. That’s kind of the idea, right.? The jeans are… So if you’re trying to wrap something that doesn’t fit into something here, you know, so these stretchable jeans kind of like settle on your body in a very relaxed way. So it’s in a minimum energy way, in an optimal way. So a lot of problems, I think, in robotics and not just robotics, you know, can be framed as trying to map some space, some thing, object into another thing. And the objects are fundamentally have fundamentally different geometric properties and shapes, curvature, like that. So I think I’ve been able to use that tool, that theory, to… Explain why our fingers are shaped the way they are, you know, why the ratio is three to two to one, roughly, for example. Why we have equal length, you know, upper length. And also problems in… So this is, I think, maybe less robotic than machine learning, but, y’know, representation learning. So there’s this idea of, y’know, this is a machine learning problems are usually very high dimensional problems. But, y’know, so the manifold hypothesis posits that, y’know, although these data points are in a very high dimensional space, they actually lie on a very low dimensional manifold, right.? Which is why you can actually do something because they’re just randomly distributed in a high dimensional space. There’s not much you can do, but they lie on a low dimensional manifold. So what is the structure of this low dimensional manifold? And then can you find a good representation for this low dimensional space? Maybe a good example would be the maps of the Earth, for example, right.? So the Earth is curved, it’s a sphere, right.? But if you want to make a flat map of the Earth, there are many ways to do it, right.? So, y’know, the classical map is where the North Pole is, right.? Y’know, the standard map, the grids, right.? So that’s why the old Soviet Union looks three times larger than the United States, for example, because the North Pole gets stretched. So there’s a lot of distortion there. So can you make better maps with less distortion? So that’s what this theory tries to do, explain, construct, right.? And so machine learning problems fundamentally, it boils down to that, y’know, this idea of trying to come up with a latent space representation. Y’know, it’s basically the idea of trying to make low dimensional maps that have less distortion or minimum distortion. So I think, y’know, at the risk of sounding modest, I think some of these tools are actually quite useful.

Interviewer: I want to say something about the maps you’ve been looking at if the Soviet Union is still on them or not. We’ve talked about this a little bit. This has been sort of embedded in many of the points you’ve raised. How do you think that robotics have changed over the years? So any kind of… It doesn’t have to be a meta narrative, any kind of picture of like how things have unfolded. And then the corollary of that being, y’know, do you have any instincts about where things are going to be going in the next 10 or 20 years?

Frank Park: Yeah, I think that’s… Again, it goes back to the… Y’know, we’re a good friendly elephant in the room now. And it’s this data, data, data, right.? If you go to a typical robot conference now, half the talks, maybe more than half the talks about data. I wish I had more data. What’s the data we need to solve this problem in robotics? Some of my senior colleagues say they’re worried. Y’know, I’m not so worried. I think it’s okay. It’s okay. Let people follow this obsession with data and see how far they go. Roboticists, y’know, just applied scientists are practical people by nature, y’know, and if they find out that something’s a dead end or it won’t work, they’ll quickly abandon it and go back in. And I think once you realize that, y’know, data won’t solve all your problems and you need something better, inductive biases, then they start again going back to the old books, studying mechanics, physics, geometry, and then realizing, oh, these are actually quite useful for building more efficient, y’know, models that scale better. So that would be the big change. But again, I’m not too worried. I think, y’know, people come around pretty quickly.

Interviewer: You’ve talked a little bit about ways in which your work has played a role in other research, other work. Moving even beyond that, are there ways in which your research has been applied in sort of, I hesitate to call it like real world situations, but in industry, in everyday life, anywhere along those lines?

Frank Park: My work or the work of our group. Well, y’know, we have done a number of sort of, I guess, research projects with companies, y’know, where they’ve been trying to solve, for example, problems in manipulation. And so we’ve had some success with those things. I wouldn’t say that, y’know, are shattering, but, y’know, there are reasonably challenging problems.

Interviewer: Any other collaborations with industry or the commercial world that you’d like to highlight that have been significant?

Frank Park: Maybe I’ll relate my sort of experience with starting a company.

Interviewer: Yes, please.

Frank Park: Yeah. So some former students and I started a company in 2016. We wanted to solve problems in robotics. So, y’know, by the time we developed quite a sort of library of tools for motion generation and optimization for industrial robots. And we thought it would be pretty practical and useful. But most companies, electricity, we’re not too worried, y’know, we don’t need to really minimize energy consumption of this robot. It’s doing fine now. We don’t need to get there in minimum time. So it was kind of, wasn’t a high priority for them. Nice to have, but not a must have. So we said, OK, maybe now is not the time. So let’s see what will work. And of course, at the time, vision using deep learning was, y’know, the hottest range. And so I visit a lot of factories. And one of the common features of the factories that I visited was that I had never seen a factory that was completely in the man. There are always people, always people, y’know, even a factory that looks almost completely automated. If you look very carefully, there are people in blue suits way in the back, y’know, walk around doing something. And if you look at what they’re doing, most of the time was inspection, inspection of parts for defects and things like that. And so we thought, OK, so maybe this is what industry needs. So we said, well, this is perfect for deep learning because this is what deep learning is good at. Y’know, so the type of defects were actually very qualitative, y’know, stains and kind of cracks and things. And some are actually acceptable. Y’know, so there’s like a very sort of a qualitative threshold. Y’know, what when something is considered defect. So we thought, OK, great. We’ll just take our favorite, y’know, the CNN architecture and just ask the factory for, y’know, the images of the fact data. And we’ll just train the model. And here we go. Didn’t quite work that way because it turns out that the factories didn’t have a lot of data. So you think that, oh, they must have just like just piles of data lying around and defects. Well, I mean, if a factor had a lot of defects, then, y’know, clearly that factor has some problems. So, y’know, there and, y’know, these deep learning models only work if there’s enough sufficient good quality data in balance. So you need data for not only just regular non-effective images, but also, y’know, defective image samples. And there was a huge imbalance. So that was one challenge. The other challenge was, y’know, you got to get it right. Nine, nine percent of the time. You cannot miss, for example, in batteries, lithium ion batteries, EV batteries. Y’know, if you miss a defect, y’know, it can be catastrophic. So what some people did at the time was that if they even just even slightly suspected that this was defective, they treated it. And this drove the factory managers crazy because, y’know, they looked at the pile of parts that this inspection system had, y’know, classified as defective and they find that more than half are fine. So not only do you need to be 99 percent accuracy on the defective products, but you also need to get misclassification down to one less than one percent. I mean, y’know, that is, y’know, we don’t want to classify perfectly fine products as defective. So getting that down. And then they don’t like GP. Well, y’know, the factory managers will ask, so how much are these GP? How many do we need? And also you need PC with at least this specification to run the GPU. How much does that all cost? And then, y’know, they do the they add up and say, I don’t know. So you got to get all this. You’ve got to get it down. So it’s like it’s life. Nine, nine percent accuracy using minimum number of GPUs. If a new defect comes along, you can’t just stop the line and retain retrain for another hour. It just has to be on the fly. So it turned out to be quite challenging. So we kind of turned into a business. So, y’know, figured out, OK, so this is not something that anybody can do. So that led to, I guess, going down this hole.

Interviewer: Would that be kind of like a it wouldn’t be a consultancy, that kind of a business. It would be more like technical, technical advisement as your company’s goal, like her as your industry as a company.

Frank Park: Yeah. No. So it’s actually not a real company with a real product or a portfolio of products that actually, yeah, software products that actually do this image inspection using our code.

Interviewer: Addressing like QC concerns and things like that.

Frank Park: So we’ve expanded that from from images of industrial parts. Now we can do the same for videos from low resolution videos, safety monitoring, y’know, kind of expanding the application.

Interviewer: It’s fantastic. This is the penultimate question. Is there anything that you would recommend to students who are interested in getting into robotics? Any advice you may give?

Frank Park: Yeah, I guess maybe the same advice that I got, y’know, when I was young, that is that, y’know, your career as a richer researcher, y’know, is going to span the next 20, 30, 40, 50 years. And so make sure that you have the tools to remain relevant for 20, 30, 40, 50 years. And I think that’s especially, y’know, in this time where people are talking about, y’know, artificial gender intelligence being just right around the corner. Y’know, these foundation models actually, y’know, proving theorems and solving, y’know, math Olympiad problems. I think that’s even more of a challenge and a concern. But I think, yep, to make sure you’re ready to find problems that, y’know, will not be solved into three years, but find problems that will, I think, remain open for a long time. And try to develop your own set of tools, y’know, your sort of secret weapons. And in my case, it was, y’know, geometry, different geometry, the lay groups. It was a steep learning curve, y’know, at the time, y’know, there are many occasions where I said, wow, is it, am I really spending my time wisely learning this? My advice would be it’s much easier to learn when you’re in school, rather than when you’re out in industry or in the real world trying to learn this. So even though, y’know, it’s difficult and, y’know, you may not see exactly where this is going to lead, I think it’s important to invest the time and effort in learning those maybe difficult theoretical tools because it will serve you well for a lifetime.

Interviewer: That’s fantastic. The last question I have for you is whether or not there’s anything you’d like to share we haven’t covered in our conversation.

Frank Park: Well, I don’t know if we’ve covered a lot, actually. No, nothing comes to mind, anything comes to mind for you. It’s hard to answer a negative. Yeah, I think we’ve covered.

Interviewer: Terrific. Thank you so much. It’s been wonderful. Thanks for sharing your time and expertise with us.

Frank Park: Thank you.