Interview Transcript
Jean Jacques Slotine: My name is Jean Jacques Slotine. I’m French. I was trained in France, and later at MIT. I ended up at MIT,partly because my father always had some very international view of everything. And so I heard about MIT since I was a kid.
Q1: Oh, you’re training. You’re professional…
Jean Jacques Slotine I went to a school in France called SUPAERO, which is the aerospace school. And then I did my PhD at MIT.
Q1: And you can interpret this question as broadly as you would like, but how did you first become interested in robotics? Was it through a particular mentor or person or project?
Jean Jacques Slotine: When I was a kid, I was given, not by my parents, actually, a cartoon. And there was the word “cybernetics” in it. And I just liked the word, actually. So I wondered, what was this about? And so my father, at the time started showing me books about cybernetics, and Grey Walter, and Norbert Wiener and so on. And so that’s how it started when I was like seven or eight or whatever. I just wanted to know what this word was.
Q1: Yes. Do you happen to remember the first robot you’ve ever heard of?
Jean Jacques Slotine: It was Grey Walter, actually. Today his main papers and videos are easily available on the internet. You may be aware that just after the war, it was this big cybernetic movement, which was the ancestor of control theory, but also of computer science, of AI, of computational neuroscience and information theory, actually. So it was all cybernetics, you know. And Norbert Wiener wrote this book called science of control and communication in the animal and the machine. And so everything was there, actually. And so one of the first people who built robots was Grey Walter, who was actually a neurologist. There was a strong interaction between people studying the brain and people trying to build robots at the time. That was certainly the interest of people like Wiener and so on.
And so Grey Walter built these little robots in 1949 or so, and they became very famous because they were very, they were simple, autonomous robots, which could avoid obstacles. There were several of them. They could avoid each other. They could go and plug themselves
3 into the wall at some point if they were running out of power. And so those are the first ones I heard about. Years later, I found a video,on the internet, but, you know, there were books on them, there were pictures and so on. And it was very impressive because it was made from just, as Grey Walter was saying, a few neurons, you know, a few electronic neurons, a few transistors, if you look at the circuit diagram. And I thought that was really impressive. And also that also kind of piqued my interest. This came with articles in Scientific American that he wrote and that were reproduced in his books. And one of them was saying that as you put more and more neurons together,one problem you’re going to bump into, which he had bumped into, was stability. And to me, it was really interesting because it’s, oh, okay, you can start putting, I didn’t call it mathematics at the time, but you could try to put some formalism around this and so on.
Q1: So it sounds to me like this connection between kind of the human brain via cybernetics has been influential.
Jean Jacques Slotine: Oh, it has been very influential. And it has always been for me, you know, I’m in both at MIT, I’m in both engineering department and brain and cognitive science department. So, and I always kind of believed in that.
Q1: How has that connection affected the kinds of research questions that have driven your career?
Jean Jacques Slotine: Well, lots of things. So what I work mostly on is dynamic systems, differential equations and things like that. And, and of course, I mean, the brain is also made of differential equations, made of biochemical processes and stuff like that, right? And so you have, you have at least an existence proof that you can do this through a dynamical system. I’m not a neuroscientist by any means, but I’m in the neuroscience department too. And I talk to a lot of people who work in neuroscience and it’s very interesting point of view. And of course, I mean, the brain is very different from robots, in the sense that it’s dealing with elements which are really slow, right? And basic neurons that reacts about a million times slower than the transistor, right?
On the other hand, there’s about 10 to the 11 neurons in, in human brain and about 10 to the 4 connections per neuron synapses. So it’s about 10 to the 15 synapses. So first of all, as a system, you know, it’s a, it’s a network, it’s a very sparse network because you only have 10,000 connections per synapse in a 10 to the 11 set of element. But also you can see there’s feedback loops everywhere. So it must be these questions of stability that Grey Walter was talking about. It was really central. So these questions of dynamic systems have to be very central in that. And also, if you think about it, you know, so these 10 to the 15 synapses mean that, from the time a baby is born to the time she goes to college in a, in a simplest possible, very simplified model, she creates about a million new synapses per second. You have to create every new connections per second. And so from a stability point of view, and so on, from a dynamic system point of view, it’s very interesting. Also, from an information point of view, how do you meaningfully
4 create a million new numbers per second based on experience in the last second, you know? And so that suggests prioritization and things like that. So lots of very interesting things.
Q1: Have you experienced a lot of change in your research agenda over the past decades? Or has there been continuity?
Jean Jacques Slotine: I have a tendency to try to build on things, but I have very broad interest in…I also worked in quantum control. And I also worked in many other things. I have worked some biology, and some very broad interest. What I’ve always believed and so one of the first things I tell the students who are interested in working with me, nature doesn’t care whether it’s math, physics, robotics. It’s just there, right? And so you have to be prepared to play with all these fields and put whatever is needed as part of that, because nature couldn’t care less whether it’s robotics or anything else.
Q1: Can you talk a little bit about who has funded your research and how you created connections with your funders.
Jean Jacques Slotine: Okay, so the short story is nobody, basically. So, you know, I was given some kind of nice award a couple of years ago and, you know, I told the person to introduce me as the researcher who has the most citations per dollar, because basically the per dollar is very, very small. I got tired of writing proposals very quickly because, probably because I’m very lazy. But I thought it was not a good use of my time. Okay, so I’d much rather spend, have much less money and work on what I’m interested in and work with my few students…I’m not interested in managing 20 or 40 students because I might as well be in industry then, right? From my point of view, of course. Everyone has their own style. But my point of view is, I work with one or two students, hopefully super bright, and I was very lucky from this point of view. Whatever I say about my work is always the students and mine, because I have had some really brilliant students.
Q1: So it sounds like that’s a very deliberate decision that you—
Jean Jacques Slotine: It was a deliberate, I mean, the first two years I did the usual thing, you know, I wrote proposals and so on. And it was not was not encouraging. I’m not going to give details or names, but basically with one of the things which I became reasonably famous for, adaptive robot control. So it was rejected several times. So, you know, it’s like, yeah.
Q1: Have there been avenues of research that you haven’t been able to pursue because of this deliberate decision?
5
Jean Jacques Slotine: Oh, absolutely. So the few things which did work is I interacted a lot of companies a lot, comparatively a lot. I interacted with Fujitsu, with Sony, with…and so I got money, small amounts of money, but very easily from companies, you know, and, and, you know, the last, you know, I worked also with Google for five years as a sort of consulting and so. But of course, of course. So very often it happened that you had some very bright students and he or she wants to work with you and you would really like to work with her, but you can’t do it. So of course, there were lots of missed opportunities. There’s no question about that. But overall, there was this deliberate choice. And I’m not sure I could give this advice to young students today, although MIT and Stanford, and so on are very good at that actually. When you come to promotion or tenure,you’re not allowed to put any reference to funding. Okay. But a lot of universities are not like that.
Q1: Right. Right.
Jean Jacques Slotine: So I am very much more interested in, you know, what impact you had.
Q1: Yes. So you were clearly able to function without funding and do your research. Have there been kind of major challenges in your own career and are those institutional, technological?
Jean Jacques Slotine: Yeah, so it’s not without funding. It’s without significant funding. And then of course, there were students who came with their own grants and so on. One of my first students, Weiping Li, had a Chinese grant.
Q1: And how about other kinds of structural or other kinds of challenges to the kinds of research that you would like to do?
Jean Jacques Slotine: What do you mean by that?
Q1: So I’m just thinking about challenges that have faced you in your career and what were the nature of those challenges? Was it the state of the technology at the time? Were they conceptual challenges?
Jean Jacques Slotine: No, it is one of the difficulties in robotics, of course. But hardware is really hard. Hardware is really complicated and it’s fragile and so on. So I was lucky to interact with some absolutely top hardware people like Ken Salisbury and so on. So, you know, we did robotic catching and stuff like that because, Ken and Bill Townsend had built this super robot. And so it was capable of doing that. Nobody really knew how to control it. Well, which was our
6 job, with Gunter Niemeyer and others, but it should be capable of doing that. So hardware has always been a little of a…not of a problem, but, you know, it’s not my specialty at all. So I rely on what everybody else is doing. And I know it’s super hard. I know it’s super hard. People have always been very impressed by Ken, by Hirose, by a few people who do fantastic hardware. Marc Raibert.
Q1: How do you think about the field of robotics as having changed over the past decade and how are you thinking about the changes that are occurring now in the next 10 years?
Jean Jacques Slotine: Well, we’re kind of at an interesting time, right? So, because people like me, you know, coming from cybernetics and so on, we’re raised with the idea that intelligence comes from the refinement of sensory-motor properties. So that basically, you know, predators learn to be better at catching preys, and preys learn to be better at running away from predators. And you know, in the process, they get more “intelligent.” And the poster animal of this is the sea squirt, a little thing that swims in the sea. At some point, it fixates itself on the rock. And once it fixates itself on the rock, it swallows its brain, because after that it doesn’t need a brain.
So, this notion that, you know, the brain is really the result of the refinement of sensory motor propertiesis something whichI’ve always assumed was right. Now we are at a very interesting point now, because the question is, what happened to sea squirts? Now we have LLMs, large language models, and ChatPT and all these things. And suddenly, you know, we can pass the Turing Test. I mean,computers can pass the Turing Test on language very easily. And they can speak 100 language and translate from each other, you know, super impressive feats.But we’re nowhere near a robot plumber.
Okay, so in other words, somehow we solve some of the problems from the top. Because you think intelligence, you think, perhaps human intelligence and then elaborate language. Forget about that. Elaborate language we have now with machines. And it didn’t require solving the motion control problem. And the motion control problem in interaction with environment is still to be solved. And that’s what robotics is about. So it’s an interesting time. And I still think, you know, and I think that’s why these machines are very…these programs, how very impressive they are, are also very bad at common sense and qualitative physics and so on. But because precisely, they were not built from, you know, the contact with physical reality, but they were just built for statistics. Right?
Q1: Let me just follow up on what you said about kind of motion control being at the heart of robotics. Can you maybe say a little bit more about what you think that essence of robotics is, what it what it is and what it isn’t.
Jean Jacques Slotine: Well, yes. So I think, you know, robotics is basically, now people call it embodied intelligence. Grey Walter was already doing that. But you know, it’s basically intelligent machines that interact in the physical world, right, and having to do complicated
7 things in the physical world, which are not necessarily easy to describe in words. Like, I can insert, parts and things and then, maybe even if the things are flexible or whatever, I can do that. I would be incapable of describing exactly what I measure when and so on to do that. And nobody would be.
And so so yes, robotics is basically intelligent machines, but in their interaction with the physical world, and learning from the physical world and being constrained by the physical world. Now there’s a flip side to this, right? So we, evolution trained us to work on Earth, right? So we really good at 3D, at imagining things in 3D and so on. If imagining things in 4D is really hard and imagining things in 10D is impossible, right? Now, if a machine has been trained completely differently, you know, maybe it had a harder time getting to deal with the 3D world as we do, but from abstracting to a 20 dimensional thing, it has zero problem. And so in finding patterns there is nothing special about 3 for this machine. And so that’s interesting. You know, that’s an interesting difference. The other one, of course, is, as I said, you know, the machines deal with electronic hardware, which are, which is, you know, millions of times faster than neurons.
The other thing is, you know, when you board an airplane. So these days, and not all not all airplane makers are as good, but the airplane is supposed to be reliable at 10 to the minus 9. So this means that all the systems of the airplane and everything, including the control system has been designed so that there’s only a chance in a billion that something goes wrong in the next hour. It’s 10 to the minus 9 power. Now how would you like to come and to have the pilot say, “Welcome aboard. You’ll be glad to know that the control system of this aircraft has been designed by the latest deep network from Deep Mind, Open AI, whoever. And as a result, we have a 97% chance to land in San Francisco,” because that’s where we are now. And so you have to find kind of the right mixture between formal techniques, which allow systematic, safe, guaranteed safety and so on, on the one hand, and the power of learning and so on.
The other thing is, as I’m working on dynamical systems, the, I think there’s a lot to do for just people interested in dynamical systems to understand the dynamics of learning. If you take a deep network, you get input and you get some output response, like it’s a dog or it’s a dog running in the snow, whatever. And so that’s a dynamics because you’re going through layers. And of course, gradient descent, which is what’s used in most of these learning algorithms is also a dynamics. Okay. And so, understanding this from a dynamics point of view, I think it’s very interesting because, maybe exaggerating a little, a lot of these algorithms that you have for learning now are very, very simple, very hacked, almost comical. The only thing that makes them not comical is that they do work. But they work because you can throw at them a very efficient over parameterized gradient descent or multi parameterized gradient descent and it will learn your function. It means that understanding that kind of learning, specifically that algorithm, which is a sub part of all the algorithms we’re learning, understanding it much better in terms of data efficiency and speed of convergence and just the general nonlinear facts. That’s, for me, it’s a very interesting question, which is not particularly a robotics question, it’s a dynamical system question as applied to learning.
8 Q1: Thank you. You answered in that several of my next questions about the future of the field. Can I ask you about your own contributions to the field and which contributions do you see as most significant?
Jean Jacques Slotine: Well, sure. But so we did a lot of work in the mid 80s, actually, on adaptive control. So this became the standard adaptive robot control algorithms. And, you know, there was, so, we as my students, Weiping Li and I, and that particular paper, I think, had a lot of impact because, first of all, it introduced three new ideas at the same time, right? And choose the idea of using sliding variables and then the high dimensional thing, the right representation, if you want, for error signals to try to replace a second-order problem by an equivalent first order problem. Something we had done a little earlier in the single input case with another student, Joseph Koitsi. But here we applied it to robots. But it was mostly also this whole notion of using energy to build the ebb and the functions and using proving, things about energy like, you know, there’s a standard result, since we’re talking to roboticists, you know, standard results in all textbooks about H dot minus 2C being a skew-symmetric matrix, H being the inertia matrix and C being the Coriolis matrix. So we prove that actually. Now they say it in robotics as if it was, you know, everybody’s always known. But actually, this was the first paper that proved that. Okay. And so this paper had a lot of impact and, you know, everybody is kind of using it as kind of a default, if you want.
We did a paper on task priority with Bruno Siciliano. And that actually was actually just a conference paper but has also had a lot of impact. That’s basically what allows you to specify. Do you want to say, “well, you know, I want to do something, and if I can, I would like also something else to be optimized. And if within that framework, I can do something else, then I could…” So this notion of multiple task priority, which extended an idea of Yoshi Nakamura in a very general sense. And we worked a lot. So we’re reasonably famous also because we worked a lot on catching using Ken Salisbury’s and Bill Townsend robot with some really good students, Barbara Hove, Gunter Niemeyer, Jesse Hong. So we had robot catching objects. You know, that was sometimes, I may even show it at this conference, but sometimes I show it because, you know, it was reliably catching like 50balls or whatever in a row, right? 50 balls in a row. So it’s not that something happened to work at three o’clock in the morning and we happened to have a video, right? It really worked. And we played with throwing. We also played with catching paper airplanes.
Actually, the, some of the work we did later on contraction analysis, which is this kind of very general way of looking at dynamical systems, which we did with an exceptionally brilliant student named Winni Lohmiller,. was inspired actually by trying to predict the motion of a paper airplane, which we were trying to catch. And we realized it was nothing in dynamical system theory which could do this thing well, right? We knew what to call, it was a nonlinear observer, right? A nonlinear predictor, but there was no technique to do that. And so that’s actually what motivated all the work, originally, originally motivated all this work on contraction and so on.
And so there’s been, so we did, you know, we did some work on SLAM. You know, we wrote this paper, maybe 10 years ago on analytical SLAM, you know, so in other words, a simultaneous location and mapping, but with equation without linearization, without
9 approximation. You know, and one of the reviewers said, “you know, when I saw the title, I was sure it was bullshit. But actually, when I read the paper, it wasn’t.” So it’s, you know, so it is basically that people didn’t realize it was a simple trick, which said that you didn’t need to approximate anything to do SLAM, you know. You didn’t need to linearize. You could use directly the nonlinear equations as they were, okay, the nonlinear geometric equations as they were. So so we did that.
We did a lot of work on quorum sensing. So that’s something that’s bacteria use when they reproduce. Suppose it’s a bad bacterium, right? So it reproduces. At some point, there’s enough of them to overpower the immune system by itself. It couldn’t do anything. So how do they know there’s enough of them? Because, if they make a mistake, there’s not, then they will be swallowed by the immune system, which is good for the host, us, but not for them. Right? And so how do they know? Well, they send chemicals in the environment and they measure also the same chemical. And this way, they know how many they are. Okay, so that’s called quorum sensing because it’s a way, in biology, right? It’s a way to basically synchronize behavior through the environment. And so we’ve done a lot of work on that, on synchronization of oscillators and so on, which is directly related to things like locomotion. Right. But through the environment. And we also did work with some colleagues at Stanford on showing that if you manipulate an object with multiple robots, it’s also a quorum sensing problem. And you can you can actually do similar things.
Q1: As you think about how your research has moved out into the world and to apply technologies, what do you find most interesting? Have there been surprises to you in that process?
Jean Jacques Slotine: Right. So, I think it has moved around in many ways. One of the main ones, of course, was adaptive control, because that’s everywhere. That was in the original adept robotics, direct drive robots. Adaptive manipulator control is also the basis. A lot of things happen, for instance, Amazon warehouses, right, where you have, you need to manipulate objects, which, of course, changes mass, inertias and so on. You know, these are boxes with lots of things in there. Right. And so you need to be adaptive. And also, you know, as the day goes on, the motors heat up. So their properties changes. So, they’re using some form of adaptive manipulator control everywhere in these systems. But of course, it has been, this general trend of ideas has been used in many other things. So the contraction has been used a lot. The contraction actually is one of the things we used for for the SLAM as well.
And then, sometimes you know the outcome, sometimes you don’t. The following is a joke, but it’s also a bit of a true story. I was having lunch one day with a then Soviet historian who was at MIT, who was spending a sabbatical. And and we talked about the impact of science on politics. And I didn’t know anything about that and politics in the Soviet Union ? Because he said, “well, you know, in particular, there’s been a strong impact of control theory (which is what I do) on Soviet politics. Because in Russian, the word for control is “upravlenya”, which also means government. By the way the same is true in Greek. So they played with this kind of, and they tried to use a lot of control theory there. And he explained to me that every year for the
10 anniversary of the revolution, the head of the Communist Party had to give a six hour speech, you know, where they were applauding all the time and so on. And of course, he didn’t write it all himself, but they had to fill it. Right. And so to fill these six hours, they started putting more and more ideas from control theory, in particular, Kalman filtering, Kalman filtering was everywhere in these speeches.And then I said, jokingly, “You know, when I did my thesis in the early 80’s, one of the points was that the correct way to control large, complex nonlinear systems is to have decentralized, but robust controllers.” And he says, “Yes, decentralized control was the end of the Soviet Union.”
Q1: Yes, yes. Well, this has been wonderful. As we conclude, are there, is there anything that you would like to share that we have not yet had a chance to talk about?
Jean Jacques Slotine: No, not really. I think, I hope it’s still possible for students to do things the way it was possible for me in, other words, to pursue what they wanted, to care less about funding than they used to, but to be really excited about their, to be really excited about their work, to, I mean, to be really excited about showing up at the office and so on. Right. To also be, you know, if I had to give any advice would be to learn a lot about physics, about math and so on. And then we, as I said, realize that nature doesn’t care what’s what. And so you should play with whatever is there. After all, you know, the adaptive controller was really putting things like Hamiltonian physics directly into adaptive control, in very simple way, but still. And so I would highly recommend this, and then there’s a there’s a bridge to do. There’s definitely a bridge with machine learning.
For instance, recently we worked on the variation of the adaptive algorithm where, in the over-parameterized case, so basically when you have many more parameters than you need. And we showed that, with a very simple modification, you can get your data controller to work, but at the same time optimize sparsity. So, in other words, get the smallest number of parameters for this to work. And then actually, if you do that, it has the potential to say, okay, so now you may have multiple models and you don’t know which one it is. And maybe it’s 100 multiple models, but now you put them all. They add. There’s a mechanical model. And now you let the adaptive controller deal with the sparsity. And the adaptive control is a real time thing, right? So it’s not like machine learning in the sense you need, you know, you need a bank of computers for months to learn something. It’s something which happens in real time. And so now, you can pick the most relevant model in real time. And so, I suspect, when you start dealing with compliant motion and moving objects around and so on, that’s going to be a central part.