Podcasts›Automated›Amanda Prorok on robot teams, physical AI, and trustLike💬 Be the first to comment →AudioSeptember 14, 2026Amanda Prorok on robot teams, physical AI, and trustFrom Automated, hosted by Brian HeaterRecorded at the Davos Tech Summit. Why the next step may be teams of specialized robots, not one machine that does everything.Play episodeSaveOpen the audio file TranscriptTranscript from the show's feedSearch the transcript0:00There's a lot of stuff that's being deployed that has empirical guarantees at best. We know through data that it works pretty well, but we know that, you know, also through data that sometimes it doesn't work well, and we don't know why. I really connected to the idea of the whole world is about collective intelligence. It's about the fabric, it's about how everything has to interact with one another. That's the only way you can solve the really big problems because everything is interconnected. One of the things that really drew you to the field was very ones and zeros. And now we're talking about physical AI, and I think things are getting maybe a lot blurrier.0:33Are you finding that in your work? You know, when I got into robotics, there was this kind of other side of robotics where people are like, well, no, no, you need to have algorithms with proofs because you just don't know what's gonna do. And now with all this AI, I feel like, well, you guys were the ones who were saying you cared about the MPs and but we're not living up to those standards right now. We know from human soccer that we've gone down the route of heterogeneity and role specialization, but we didn't really know, okay, if you give an AI the same problem, is it gonna figure out the same thing? So we coded up a heuristic football team.1:04After running our system long enough, we had converged to a stable solution, and it turned out that they formed themselves into positions, they held those positions, and they also acted with that intent. We never told the system that's how humans play football, and yet it had learned exactly that behavior. You feel like at this point, like you've got a pretty good idea of how that specific evolution happened within the system.1:43My name is Brian Heater. I am the managing editor at the Association for Advancing Automation. We've got a special episode for you today. I think I say that effectively every episode, and I honestly mean it most of the time, including this one. This is the first of two conversations that we recorded at the Davos Tech Summit in Switzerland over the summer. Professor Amanda Perak gave a fascinating speech on the tech stage that I received at the event. So this was a great opportunity for us to dig into a lot of that a bit further, a bit into her history and the work that she's doing over there.2:17Thanks to her and the event organizers for helping set this one up. Thanks to you for tuning in. If you're enjoying the show, you can help us out by liking and subscribing, checking out the newsletter over at automated.fm. In the meantime, please enjoy this conversation with Professor Amanda Kurok. From quality inspection and predictive moments to robotics, logistics, medical devices, aerospace, and manufacturing.2:48And that's important and why conference will work from industry workers to see real-world AI applications connection, exploring some computers, computer version, sensors, and computing and all the technology. You'll also be able to connect with the people who are different what comes next. Whether you're building, buying, or implementing these systems, this is a chance to stay ahead of the curve and see how vision and AI are moving from promise to practical impact. Learn more and make sure to buy your ticket early at the Advanced Vision and AI conference September 23rd and 24th in Santa Clara.3:25So I was initially going to ask where you were from. I was trying to figure out it says British British and Swiss nationality. And then I realized I like I couldn't I couldn't quite figure it out after hearing you talk, especially after talking to multiple people in different languages. And then I realized I think that might be part of being Swiss. Like you're kind of like a bit of a shapeshifter when it comes to speaking different languages. I just grew up in many different countries and I had I was exposed to many different English-speaking environments, each time a different accent.3:59So I kind of soaked up a bit of each one along the way. I actually was born in Brazil to a Brazilian mother, a British father. And when I was very young, we moved to Switzerland. Okay. And I picked up, I first was, you know, embedded in the Swiss schooling system and then moved on to I did a short short stint at an English-speaking private school, and then got back into the Swiss educational system, did university in Switzerland. but actually the language we talked at home was always English. I'm probably a typical American.4:29I I like I speak Spanish very poorly. Okay. And otherwise I you know don't know any other languages. how do you feel that I I mean, did it being multilingual, has that kind of like reprogrammed your brain? Has that taught you to think and to learn differently? Yeah, I I totally think so. I think that speaking multiple languages is not only about the actual functionality of speaking multiple languages, but it also teaches you to think in different ways because every language comes with a culture and you have to understand the culture to actually communicate with the people.4:59So when I'm talking in English, I'm a different person than when I'm talking in Swiss German or French with my husband, or you know, I lived in Japan for two years, or Japanese with the people I was working with when I was in Japan. It's a completely different different way of thinking. And you actually only really speak a language when you shift into the mentality of the people that you're with, and it can be a very different one. Oh, that's interesting. So you actually have to be to a certain extent immersed in the culture. I think it's also about how you think about things because the way you organize your sentences is is different.5:30And I actually Japan was extreme for me because I was taking a language course and I was kind of like, I was learning the words, I was learning the grammar, but I wasn't kind of really I hadn't made the I hadn't, you know, felt that make-break point where I'm like like, okay, it's flowing now. Japan, Japan is is hard for everybody, right? Well, yes, but then I suddenly, after I had been immersed, I think it was at the three-month point, I suddenly had this point. I I think I was at home preparing some rice with my homestay family, and I'm like, oh, you have to start at that end of the sentence with that rationale, and then it'll come kind of flow better.6:05And when I had that aha moment, I started thinking in a different way, kind of the reverse way of the way I'd normally think. And it just came. At that point, I was just speaking Japanese suddenly. And that was a really important aha moment for me because I was like, yeah, you just have to understand that different, you know, societies, communities have different ways of organizing rational concepts and you know, actions and consequences. They see them differently. So once you embed that in your system, that's when language flows. So so that that that's the analytical mind right there.6:35Like that's that's that's part of being, you know, that's a scientist or or technologist is that moment when things really like click into place. Yeah. It's suddenly now I understand the concept of something. And now I because I understand the concept, I can apply it in a generalist way. Did you start coding too? when I was quite young in here. I started young-ish. My father is a computer science scientist by training. so he they never kind of forced anything upon me, but I was you know in an environment with a very rational dad and you know, had that kind of influence.7:13But I was gonna ask, what's that like having a rational dad? No, kind of like, you know, very systematic in in in you know, making decisions and what the facts are that are flowing into decision making. I think that I got from him. My mom is actually a psychologist, so I got kind of got that interesting side as well. Actually, it's funny because that combination would actually mean I should go into HRI or something like that, you know, blending, you know, you know, robotics, coding with more the soft humanity side of things, but for some reason I didn't. Actually, what I do is really, you know, multi-agent stuff and multi-agent systems.7:46And that actually connects really nice, nicely to the language thing. I I really love this idea of I understand something about how humans communicate and how, you know, us, how we function and how we interact. And some of that does, at least in the beginning, did flow into some of the ideas we were having about how to train agents to communicate with one another. And understanding, you know, the principles of language is actually really helpful because it kind of runs the world, language. Yeah, I I I was also gonna ask, you know, how you ended up in so many places, but it sounds like you kind of like hit the ground running.8:18You know, m my love for robotics was very indirect. It happened very suddenly. I was studying computer science, and to be entirely honest with you, I was like on my three-year mark or almost into the fourth year, I was like, oh my gosh, this is so terribly boring. I don't have no clue what I'm gonna do with myself. Because at the time when I was studying computer science, it was about, okay, you graduate, then you go on, you know, you design databases, you go designing websites, and you work for some company doing advertising or whatever. I don't know. Google had barely existed at the time.8:49Lots of exciting stuff wasn't happening yet. And I was like, I don't know why I'm doing this. I love the concept of, you know, computer science is rational, it's zeros and ones, and there's no kind of fuzzy middle ground. And that's why I studied computer science. I was really kind of in between computer science and math, and I really liked that. But then when it came to I'm now finished with all of this, what I'm gonna do next, I was completely at a loss. And what happened was this really awesome professor came in. He had done his training partially in Switzerland, but had done the stint in in in the US at Caltech. his thing was swarm intelligence.9:20And we also had another professor at my university at the time who was teaching bioinspired robotics. And I took both courses in my fourth year, and then suddenly, like, my world opened and the lights went off. And I was like, this is fireworks, this is amazing. I really connected to the idea of, you know, the whole world is about collective intelligence. It's not about one individual being and how I am going to be autonomous and how I'm gonna be, you know, if I'm an agent, how I'm coded or how I function by myself. It's about how what's the fabric of the world? I really believe that fundamentally it's about a fabric.9:52It's about how everything has to interact with one another. And that's the only way you can solve the really big problems because everything is interconnected. So that was something that really, really clicked with me. And robotics came in because the the medium with which the professor taught swarm intelligence was through robots. So we got our hands onto actual robot platforms, we would program what we thought. So our hypotheses of okay, this is how the singular agent is behaving. And then let's observe what the emergent global behavior is going to be.10:24What does that local behavior, what is its impact on the fabric essentially? And that was such an exciting moment for me because that's when I started understanding, okay, you can tell an agent to do very simple things in very specific scenarios, but when they're doing it in a group of, you know, many others or just you know, multi- you know, five, ten, six, whatever robots, then unexpected things can happen. Yeah. It's interesting that you mentioned that one of them was bio-inspired robots, because it sounds like insects are probably, I assume, a really big part of the inspiration there.10:59Yes, yes. So swarm intelligence, that whole field is strongly connected to bioinspired robotics and bio-inspired computing in a much broader way. And I think it's so that actual thread of research emerged, I think, in the early 1980s, maybe early 1970s. There were these scientists, a couple of really interesting scientists, one of whom was Brightonberg, you might have heard of him. they would observe natural organisms and try to understand, okay, why are they behaving this way?11:31What are the rules by which they operate? And how is it possible that the commute then does such fantastic, phenomenal things? One of the most famous examples, I love citing it, is it's called stigmergy. It's how like ants use stigmae a lot in nest finding. So the ants actually communicate through the environment with each other. They lay for a pheromone trail. That pheromone trail evaporates as a function of time. Because shorter paths have stronger pheromone trails, the ants automatically find the quickest path to a source of food.12:03And that algorithm is actually one of the most successful heuristic search algorithms that ever was invented. It's a phenomenal algorithm, extremely efficient, and people still use derivatives of that algorithm to solve search problems that are very hard that typically can't be solved with first principles uh-based methods. So so how would a swarm differ from what we would refer to as collective intelligence? Is is there a distinction between the two? It's a whole kind of I think these def there's no I've never read precise definitions of all these different fields.12:36You know, also in the robotics community, some people call it multi-robot systems, other people call it connected robot systems, other people directly say, oh, it's a robot swarm. some people say connected yeah, connected robots and so they're essentially the same things. Yeah, I think, you know, as I said before, there's no real precise definition. I think that as soon as you have more than one robot in a group, they're interacting by some medium. You already have a multi-robot system. I find personally I like the terminology multi-robot system because it's technical, it's pure, and it's down to the point.13:06Multiple robots, right? And it's a system because they're communicating. It's more than just, you know, the one single entity. Yeah, I I I I'm probably really off on this. But you know, I remember going and seeing you know, maybe like 10 years ago in a lot of research facilities. I don't think it was Michigan, but probably like I think CMU, they were doing a lot of joint research at the time. a lot of the swarm-based stuff was really like synced up robotics. They were synchronized in that often they were doing the same action. Oh, I see homogeneous.13:37Yeah, exactly. And and the work that you're doing, I I suppose it'd be heterogeneous in that it it's a lot of different robots, often in different embodiments, doing different tasks. So heterogeneity is something we started exploring quite recently. I'm assuming that the probable reason why you were encountering a lot of like homogeneous systems originally is because from an algorithmic standpoint, it's very hard to develop algorithms for heterogeneous systems because you suddenly have a combinatorial explosion of options that each robot can take at any point in time as a function of what role that robot is supposed to take and as a function of what everybody else in the system is doing.14:18And given that everybody is different, there are there's a bigger set of possibilities at every single time step. So heterogeneity is a complexity that you have to deal with, and not many scientists dealt with that because already there was enough work to do in studying homogeneous systems. Often, you know, when you look at the work in bio-inspired robotics or in general bio-inspired computing, the assumption is that all actors are the same, because one of the underlying premises is that homogeneity leads to or confers robustness because you can swap them out, one can break, it can have an impact on the system.14:53And that's true to a certain extent. But actually, in our work, we've shown that heterogeneity confers robustness in a special way as well, because you have complementary roles, and when something unexpected happens, somebody else in the system can step in because that other actor has that capability or has been paying attention to something different, or et cetera. In terms of like a single embodiment can do different things and step in for different roles. potentially. Potentially. for us, heterogeneity has also been about role specialization in some cases.15:29In other cases, it's been about being more parsimonious with the capabilities that you have, because if you have a heterogeneous system, you can get the individual actors to specialize more with what's given to them. So you basically memorize a narrow part of the task, and hence all your resources are dedicated to that narrow bit, but then you're still working with all the others. And everybody together then does something much more sophisticated. I I really suck on this this cross-embodiment idea. you know, I do think it's really interesting, and and we're having this conversation a lot around physical AI.16:05That the the two holy grails right now are just this sort of like generalization in terms of task and cross-embodiment. in terms of when you're doing these these swarms or or these collective robots, how difficult is it to introduce, you know, a mobile robot and a drone working together effectively? Yeah, that's a really good, a really good point. So cross-embodiment has many facets to it. So you have to think about cross-embodiment. I think the the hardest part about cross-embodiment is actuation.16:37There's also the perception side of things, but usually you you can say, you know, okay, a drone is carrying a camera, a legged robot is carrying a camera, a wheeled robot's carrying a camera, and the only different thing is the height at which they're carrying that camera or maybe the point of view. But ultimately you can still use the same world model. But the output of that model is going to be drastically different if you're in the cross-embodied system, right? You're you the output of that model for a legged robot is going to be, you know, take steps. The output of that model for a drone is gonna be, you know, the thrust of your rotors.17:12So so so the so the way they sort of like navigate around the world is the same, but the way they actually move through the world is different? Is different. I mean, there are different ways you can abstract from that cross-embodiment problem, but the decision-making might still be quite different because different embodiments also have different holonomicity constraints. So basically in robot lingo, what that means is a drone can move through an environment with slightly more agility than potentially, and also obviously it's in this in a different dimensional space than potentially Lego robot or a wheeled robot. if a wheeled robot has acroman steering, it's very constrained because it has a turning radius.17:46But if you're using Swedish wheels, you can turn on the spot and do, you know, left, right, forwards, backwards, like a human can, essentially you can move in any direction. So I think knowing what your degrees of freedom are in space, that has an impact on what your observation that what how your observations influence your action. I I suppose one way to make it really uniform is to decentralize the cameras themselves, right? And we saw this in some of the work that you were doing in terms of actually, I think you called it like a blind robot, taking the cameras off the robot, sticking them around, and using that for the robot to navigate.18:21Yeah, that's that's that's a really good connection. So what we did in that piece of work was to say, well, we are, yeah, in effect, we're separating the actuation from the perception. the robot itself is only the actuator, and the perception is just simply distributed in your environment. the nice thing about that piece of work was the perception was super flexible. You could kind of move it around, nothing had to be calibrated, given the fact that it was completely feature-based. So using again a type of world model that, you know, your cameras would were seeing stuff happen, they were able to interpret that and communicate that piece of information to the robot in a way that the robot could understand it, and the robot could then action it with its own private, like local action model.19:04Right. and then you could think about cross-embodied systems where each individual actuator, like robot embodiment, has its own way of interpreting what's coming in through the environment and then can act upon it. Now, when you you used the term decentralized, are are you are you effectively referring to the fact that there isn't like one central brain controlling everything? That's right. that's the definition of what a decentralized system is. In a centralized system, you would have at least one component that is centralized. I'm not saying that decentralized systems are better than centralized ones.19:36I just say it gives you a little bit more flexibility in terms of what you want to do if you allow at least a combination of both, so hybrid systems. and obviously the nice thing about decentralized systems is that if you can do compute locally as a function of local observations, that is computationally much more efficient than if at every time step you have to make a decision that's based on all of the information in the whole system. Now, the the caveat is that often in many algorithmic problems, the guarantee of optimality, you only get it if you compute as a function of global information.20:14So it's centralized. There's there has to be some centralized component. And that is often the challenge that we have when we're trying to solve very difficult coordination problems. We can only provide optimality guarantees if there's at least something in the system that has collected all the relevant global information and can give us that guarantee, that proof essentially mathematically. but what we've shown, at least in our work, is that empirically we are near optimal because the models, the way we've trained them, they're so good. We just don't have the proof. We're not quite there yet in understanding how to prove that with a purely AI-based system.20:47The upside is actually that what we're now considering is actually hybrid systems, where we take classical solvers, we use that skeleton, we know the skeleton provides us all the guarantees we want. And then we just take the little bits and bobs inside of those classical models where we know, okay, that part is really computationally expensive. Let's take that bit out and replace it with some AI-based heuristic, speed it up as much as we can without breaking anything around it, and we still get the guarantees, but now much faster than we did before. So is it it's also hybridized in the sense that that it's hybrid there is centrality and decentrality both working together?21:20Yeah, exactly. Yeah. Yeah. It's funny you were talking before about one of the things that really drew you to the field was you know, that it was very, very black and white, that it was very ones and zeros. And now we're talking about physical AI, and I think things are getting maybe a lot blurrier. Are you finding that in your work? Yeah, absolutely. I think you know, when I got into robotics, and I I I was coming from the side of swarm intelligence, and I really thought I was fascinated by emergent behavior and all that. But there was this kind of other side of robotics where people are like, well, no, no, you need to have algorithms with proofs, and we don't want this fuzzy stuff, and fuzzy stuff is no good because you can't put a fuzzy robot or fuzzy Logic or something that does emerging behaviors into the real world because you just don't know what's gonna do, right?22:06You have to have the guarantees, you need to know what it's gonna happen. Safety safety and all that made sense to me. but at the time it was just really a bit of a big deal. And now with all this AI, I feel like, well, you guys were the ones who're saying you cared about the entries and everything, but we're not, we're not, you know, we're not living up to those standards right now. There's a lot of stuff that's being deployed that has empirical guarantees at best. So we're still kind of just like on the hype wave of we're trying out the stuff, we're running it, we know through data that it works pretty well, but we know that, you know, also through data that sometimes it doesn't work well, and we don't know why sometimes, right?22:47Is is there a sense in which a physical AI, generally speaking, the physical AI that you're working with is still very like black boxy? Well, we're trying to make them as gray as possible. Yeah. or white. You know, there's there's ways of of clear box, I think. Absolutely, absolutely. And there's a lot of work around you know, providing robustness. and it's not like people aren't working on this. And and you know, we're taking it seriously. And there are ways of of trying to you know, scaffold around these, even if it's a completely black box scaffold around it so that you do have certain guarantees.23:24And you can always, you know, clamp an output and say, okay, if your robot says you're gonna do this, you're just not gonna do it because this doesn't make sense. You can always have that kind of you know, if your robot wants to accelerate at X and we know that X is just high, you're not gonna do it. There's always ways of doing simple stuff like that. But sometimes it's just a little bit more subtle. Sometimes the robot wants to make a decision that s looks right, despite all your checks and balances, but ultimately then isn't. and those are the hard cases to catch. Yeah, I I I was I can't remember who I was speaking with, but somebody who was doing really like large-scale physical AI training, and it really is like negative reinforcement.24:03You know, they stop the robot every time it does something wrong. Every time it does something right, you know, it doesn't get a cookie. They just it just sort of moves on into the next task. I guess it depends on what part of the model you're focusing on right now. If you have a lot of data with positive examples, you probably want to focus on, okay, now we have to feed it more some of the negative examples. So that kind of so so that's sort of more like edge case kind of thing. Yeah. Yeah. The thing the the part that like really jumped out at me during your talk was the soccer, we would say, or the football example.24:33And that strikes me as something that really gets back to what we were talking about for one of those aha moments, but also something that seemed to be genuine surprise on your own. Yeah, absolutely. So we we had we really did not know what the what the agents were gonna learn because we essentially so we were we were using reinforcement learning in that study. It was part of a broader study led by one of my PhD students where we were analyzing the role of diversity in multi-agent systems. Is it good or bad?25:03And if it there is actually no answer to good or bad, but when is it useful to have diverse and heterogeneous multi-agent systems? And the football example was one of the case studies we did. And we thought, well, we know from human soccer that we've gone down the route of heterogeneity and role specialization, but we didn't really know, okay, if if you give an AI the same problem, is it gonna figure out the same thing? And so we actually did tests where we said, okay, if you give this multi-agent system the possibility of becoming homogeneous, so everybody doing the same behavior, as well as the option of specializing in roles, what's it gonna do?25:37Because RL, if you let it run for long enough, will converge to something that's supposed to be near optimal, right? and all the only information we gave to our robot team was you got to be. Actually, we should back up a little bit. Can you describe what the experiment was? Well, basically we're not we didn't win at a game of football. We coded up a heuristic football team, hand-coded. So that we called our red our red team. We said this was all in simulation. Yeah. We had a red team that was coded up just to play football according to simple, like, you know, shoot up shoot the ball toward the goal.26:12And we had a blue team that was supposed to learn its football strategy as a team. And we gave it the possibility of either all of them learning the same behavior or specializing. So we have an algorithm that allows them to explore those options. And also the level of heterogeneity was something that the team could then decide. Like they could go into extreme role specialization or just a little bit different behavior. what heterogeneity actually formally means is if we two see the same thing, we either take the same action or a different action, right?26:43So, like if we stand in front of a door, both of us, one of us will push down the handle and one of us might turn and go the other way. Right, that's role specialization. So we have the same thing in our football example. So for the same observation, I agents would either take the same action or a different action. And after running our system long enough, we had converged to a stable solution. And it turned out that that stable solution was really one of the players was a goalkeeper, one of the players was an attacker, and one, I think, a set where a set of the players were attackers and a set of the players. They essentially formed themselves into the positions. Absolutely.27:14They formed themselves into positions, they held those positions, and they also acted with that intent. So the goalkeeper was in the goal and would only defend the ball and shoot it out if it was shot at itself. It would never move forwards, never attack, never dribble. And that was totally astounding to us because we never told the system that that was a human prior. We never told the system that that's how humans play football. And yet it had learned exactly that behavior. Just through we had very sparse rewards. And the rewards were simply if you shoot a goal, you get a you get a score of one, and that's it. No densification of rewards, no nothing.27:46We really wanted to leave it as pure as possible. And so that was really quite amazing to us. it kind of speaks to also maybe the way humans have evolved the game of football. They also, we also somehow figured out that that made sense for us on the players on the playing field. but the AI figured out the same thing, and that was really quite interesting. Yeah, I I was gonna ask you, so it was kind of like a philosophical question whether this is like speaks to anything like innate with sports in and of themselves, that that it's something that just kind of like occurs, almost like occurs in nature.28:21So yeah, so the idea of role specialization is something that's very common in nature. communities, there are some communities of organisms that do it in a very obvious way. So you have amp communities, you have bee communities. there it's sometimes a temporal heterogeneity as a function of their lives. Sometimes it's anatomically defined. So for bees, for example, you have big bees, small bees, and they take different roles as a function of their physical capabilities. and scientists had observed already in those communities it makes a lot of sense because that's how they best function as a team.28:51And then there are more subtle forms of heterogeneity when agents look the same and kind of do various roles throughout their lifespans and it's kind of mixed. But there always tends to be some form or another of heterogeneity and diversity in in natural communities. And I think it's actually an understudied field in nature as well, so science as well as in engineering, simply because it's quite complicated to study and analyze it. Yeah, for part of the reason why I ask the the part the thing that was really interesting and surprising and astounding to me about that specific experiment, you know, wasn't just the fact that you gave it one input, but the fact that that one input was to score a goal, there as far as I'm aware, there was no input to not block the goal.29:35And that's why I'm wondering if, again, there's just some inherent understanding that you're going to not want the opponent to score on you. Yeah. I mean, they figured that out over time. Yeah. Yeah. Is it like the is it coming from to a certain extent like the the the data that it's trained on, the models that it's trained on are using some of like the real world data? There's no there's no like football data in it's a really sort of simple model that you're using. Yeah. Yeah. Yeah. So I it's but it like genuinely surprise you, and you feel like at this point like you've got a pretty good idea of how that specific evolution happened within the system.30:18We don't know how it happens. so in the sense why you mean Yeah, if effectively, you know, they got really good at one thing and how they moved on to how or why they moved on to the other thing. I think the team just figured out that it's more efficient for them in roles because it takes time. If everybody Yeah, I mean, you have a certain speed you can run at. You're taking time out of the clock if it's going into the goal. So if if you have all like if the ball if the ball is doing a certain thing and you have all the agents kind of think the same thing, oh, I have to chase that ball and shoot into the goal, it doesn't seem like it's efficient because ultimately if the other side kicks the ball to the other side of the field, the ball is kind of flying faster than you can run, you gotta have somebody ready to catch that ball on the other side of the field, and that's gonna be your goalkeeper.31:04So I think intuitively it makes a lot of sense. And that's kind of why humans probably came up with this strategy. But that the I also kind of figured that out was was was interesting. Yeah, that's yeah, and that there then there's just time lost, just generally speaking, in them storing a goal and like having to return the money. Yeah, yeah, yeah. Absolutely. something I I've I've been trying to speak with people about, and I wonder if you have any opinion on this. when we're talking about decentralization, I'm really interested in this idea of decentralization within one system.31:40You know, in in terms of not having a central brain within like a given robot, you know, using it to control different joints. Is that something that you're exploring or looking at? So this is actually something that's been that was originally explored in the 1980s, where I don't know, you might be familiar with Rodney Brooks. Yeah, he's been on the show. So he he was probably, as far as I know, probably the first to propose that that concept because he essentially was controlling individual actuators on the robot in a decentralized manner. So they were all kind of interlinked, but each actuator on the robot, so each like motor was run by its own processor.32:18And the processors were event-driven. So as a function of something happening in the environment, any actuator was triggered to do something. yeah. Apologies to Rodney if I misrepresented it in the work, but that concept exists since the 80s. It's I think you know this idea of decentralization with one within one system today still exists. I would pin it to the subcommunity of robotics that's called modular robotics, where people are actually thinking more about how to construct robot parts and how they then come together to form more powerful, like bigger robots.32:53And I think that's really interesting in environments where you have to deploy robots in parts and then they have to, because you can't deploy them in bigger parts, and then when they get there, they assemble and do something, and then they can also disassemble and assemble into something new and do something different, right? So this idea of multifunctionality and modularity, I'm a great fan. I think modularity is a great, great concept, but it's just super hard, especially if you didn't combine it with actual like mechatronics and stuff. One of the things that we were talking about at the top of the conversation, or actually before the conversation, was this idea of you know, of like be being your own boss and the positives of that.33:28Obviously, like there are some negatives when you actually like put your name on something. something I really had to ask you. I guess I haven't seen a lot of how do you how do you get a lab with your name on it? How does that happen? Well, you join a department and you start bringing in grant money and you start building a team. Okay. I I've seen I've seen a lot of people build departments, but like you like that's impressive that you actually like a guy like this is this is your this is definitively your department and that you're it it is the pro rock lab.34:03So, you know, when I was thinking about you know young faculty members at some point, they're like, oh, and I have five people in my group, my group needs a name. And I was thinking a lot about what I was gonna call my group, right? so technically I'm I'm a tenured professor, I'm professor of like four million professor of collective intelligence and robotics. But at the time when I was creating my group, I was like, I need something that's really future-proof because I'm gonna be in this job for the next 30, 40, 50 years. and what I call my group today, I want it to be the same group.34:34I want to have the same name at the end of my career. But what am I gonna, what am I gonna be doing in 50 years? Is it multi-robot systems? Is it collective intelligence? I don't know. Let's just call it ProBock Lab, because I'm gonna be here. and so it was just kind of the simplicity and the purity of the solution, and that was it. Yeah, that that it's interesting because I you you had a post, I think it was a LinkedIn when you're talking about maybe like the eighth-year anniversary and something. Yeah, because it effectively it really just started with your studies and then ballooned into something larger.35:06Yeah, yeah. It starts with, you know, a gem of an idea, a seed of an idea. you then start working with a small group, and that then grows into, I mean, my lab's own is roughly 20 people now, and it's not just my ideas anymore. It's really, it's it's a lot of it is bottom up. I have a brilliant team, and they they also go their directions, and I'm there to steer the ship. We do have a global vision, and we all believe in the same thing in my lab. That's how I form my group.35:37There is definitely very strong cohesion, but within that, there has to be space for ideas coming from everywhere because that's where you get the best ideas. So you're saying it's like a heterogeneous group of people who are in there's no centralized possible. I mean, I talk to everybody. It's like there is, it's it's like it's not centralized in the way like I'm not controlling in that way, but I definitely have a lot of people. So it's a it's a hybrid approach. I have a yeah, definitely hybrid. I have a communication channel to everybody in my team, but they all talk to each other as well a lot.36:10Are there like specific breakthroughs? I like assume physical AI is one that have happened over, you know, or or I I guess would have been eight years ago that really allowed you to accelerate this idea. Absolutely. You know, when I started my faculty position at Cambridge, it was in robotics, at least for multi-agent system or multi-robot systems, nobody really was using AI. And I had worked in various departments before coming to Cambridge, and I was joining a computer science department. I'm like, well, the talent I'm gonna be getting, those are gonna be computer scientists.36:43So I'm gonna lean into the computer science side of things so I can really kind of soak up that talent. And I'm gonna really start to embed machine learning into the multi-robot community. And I was amongst the first to use data-driven methods within the multi-robot community. We were amongst the first to do differentiable communication, so learning communication. We were amongst the first to embed a lot of data-driven heuristics into multi-agent search algorithms and all that stuff.37:14I really leaned into that data-driven component from the moment I joined the computer science department. And it was a good choice because I didn't have a lot of controls talent coming in, not being in the controls department or an engineering department. So it was great for team building as well, and people knew how to talk to each other. Now I have people coming from all sorts of different backgrounds and disciplines. I have comms experts, I have controls experts, I have mathematicians, I have experimentalists.37:44Yeah. So very diverse thought. Yeah, I think we're at about time right now, so we we can end on this. But I I believe one of the questions that I asked you on stage, like unfortunately, my brain has been wired given my jobs, you know, my time via TechCrunch and doing this podcast to think a little bit commercially when it comes to robotics. I I talk to a lot of startups. And I look at something like you know, the heterogeneity, the cross-embodiment, you know, the the reduction of bottlenecks, the the orchestration, and it occurs to me that that there are probably a lot of commercial bodies that are really interested in the work you're doing.38:26Yes, we we we talk quite a bit to startups. I've been working with a few also bigger companies on some of the things we did. I got a couple of early grants from some of those bigger companies who work on really difficult logistics problems, and they were really interested in some of the early work we did. I think at the time I was almost ahead of what was commercially kind of deployable because of, you know, we can see it works empirically and it works really well, but we need to know a little bit more about okay, in which cases does it not work?38:56What are the edge cases? Because obviously, if you're running in a factory of with lots of workers and you know, there are thousands of robots, it's not just about it works 95% of the time, it has to work 99.9% of the times. If not, you know, you have to understand what's going on. Especially because I know you talk about safety too, especially as like people are pushing to remove the cage. Yes. And I think that's ultimately that's a really interesting frontier in robotics right now, really having robots that we can trust to the extent that it's fluid and we're, you know, just feeling feeling good around the robots.39:29You know, there was this really interesting moment this morning. I would I so I'm here, I'm here at the conference with my three-week-old baby, and I was like, oh, we're gonna take a picture next to the humanoid. And I stepped up really close to the humanoid. My husband was watching, and the the the company, the operators of the humanoid, they were super happy I was there taking a picture. And my husband then came up to me and said, You really trusted that that robot wouldn't suddenly, you know, lash out and you know, hurt the baby. I'm like, Yeah, kind of I did, but you're right, you know, we don't know.40:00So that was a really interesting moment, you know, reckoning for me that actually, you know, I trusted that the company had their hands on the red button. Yeah, I'm gonna say that that would be the worst possible PR moment would be to hurt a baby at a big show like this. But yeah. But you know, on the safety side of things, we've been thinking about that quite seriously in my lab. And we have some new work that's so part of it is already published, came out at in Rio this year. and some more is coming out the door really soon.40:31We've been working on understanding what it takes for people to, you know, for anybody to be able to trust a robot. And we've actually built on top of watermarking technology. I don't know if you're familiar with that. So we've been thinking a lot about, okay, so if here you can watermark digital content, why can't you watermark a robot's behavior? So that anybody that can somehow like observe a robot with, say, a camera or some sort of digital device that can take a recording of the robot's movements, and that, you know, say, you know, for example, your mobile phone can then immediately tell you, oh yeah, that robot's policy, it's been certified.41:07And we know who created this behavior. And because it's you know, from a trusted body, you can now trust that robot. So that's something that we've been working towards. And I think that's I'm I'm really super excited about that. Well, Professor Pro Rock, thank you so much for taking the time. Thank you so much for having me. Thanks again to Professor ProRock and the organizers at the Davos Tech Summit. If you're enjoying the show, our show, that is, you can help us out by liking and subscribing and checking out our newsletter over at automated.fm.41:39Thank you for tuning in. We will be back soon with another episode of Automated.
