Justin Werfel is a senior research scientist at Harvard’s Wyss Institute for Biologically Inspired Engineering, where he works on topics in complex and emergent systems, including swarm robotics, termite behavior, engineered molecular nanosystems, and evolutionary theory. He lead s the Designing Emergence Laboratory, and works closely with a number of other collaborating labs.
In this podcast, Justin Werfel talks about what termites can teach us about using autonomous swarms of robots.
Termites have an amazing ability to create and maintain large, complicated structures with very limited capabilities.
Justin talks about the opportunity to learn from the termites capability to create impressive structures and use that to create structures with autonomous swarms of robots. We get to hear about how the Termes project aims to learn from termites and build on their capabilities to create any desired structure.
We also hear how Justin was drawn to robotics by the balance between theoretical and practical work.
This podcast is part of the Wevolver network. Wevolver is a platform & community providing engineers informative content to help them innovate.
Learn more at Wevolver.com
If you are interested in sponsoring the podcast, you can contact us at richard@wevolver.com
Per Sjoborg: Welcome to the podcast version of Robots in Depth. This interview is with Justin Werfel and is produced in cooperation with Wevolver. Robots in Depth is supported by Aptomica. Visit Aptomica.com to connect. You will find all past episodes and more in RobotsInDepth.com. How did you get started in robotics and because there's always that nucleus that takes your career in a different way.
Justin: That's a funny way to describe it actually when you say takes your career in a different way. I actually got into robotics very indirectly. As an undergraduate I was in physics and for a long time I was really sure that physics was my field. It was the thing I wanted to do that was the path I was going to pursue. There was a lot that was just very appealing about physics from the point of view of understanding the world and this beautiful idea of reductionism that everything in the universe, all the variety we encounter can be explained in terms of four forces and six particles. Then at some point in my undergraduate physics training I found that the experience that I had with encountering modern physics was different from what had originally drawn me to the subject so it was sort of less about things that involved understanding the world in any sort of intuitive way and more about sort of doing math and magical manipulations on a page that wound up having great predictive power but not in any sort of way that felt satisfying to me. Around that time, I read actually a popular press book on complex systems research. This was in the late 90s and I think it was starting to become sort of a field that was popularly discussed beyond the people who had been working on it. This idea of complex systems and of emergence actually is a counterpoint to the reductionism that had been drawing me to physics. It seemed like it actually was much more interesting and in some ways more satisfying if it could be answered. Rather than how does everything that we encounter get boiled down to a few forces and particles the opposite question is actually in a sense much richer is how do we go from just a handful of forces and particles to all of the incredible variety that we encountered. That seemed like in some ways a very, rich, compelling intellectual problem.
When I apply to graduate schools I actually wound up applying in seven or eight different fields because I wasn't sure because there is no or certainly wasn't at the time there's no sort of department of complex systems I could apply to. I applied to places that look like they were working on interesting problems so I applied to a couple physics departments for things like condensed matter research. I applied to a few neuroscience departments. I applied to MIT actually in three different departments for different arresting things and I wrote in my application essays about a couple different things. One of the themes I wrote about for instance for neuroscience programs was the brain. It was consciousness because that's maybe the closest to home amazing, emergent system. We've got this block of 10 to the 12 neurons that let's pretend for the sake of argument that are simple which isn't even a little bit and connect them up in the right way and we get basically magic. To us it's like the experience of consciousness is magic.
Per: Yes I still remember I read Daniel Dennett and DouglasHofstadter and the mind's eye and I got just slightly spooked that when thinking I exist somewhere in there and it flows around.
Justin: This illusion of consciousness and it's like I can just think about something happening in the world and it's happening. This hand is moving. That’s amazing. That is what a thing that I wrote is the neurosciences is like I want to understand consciousness and how that emerges from the interactions of neurons. One of the other things that was a really compelling sort of drive and research question forming was I had read about these termite colonies where you've got millions of termites each sort of going around doing its own thing, reacting to whatever encounters and somehow from all of their actions and interactions you wind up with these massive complex ruptures. How does that work? Actually I was already interested in the opposite question so that shows us that it can be done. So understanding how the termites build the termite mounds is one great question. Another question is how could you create a system, how would you program the individual agents to get a particular emergent result that you wanted.
Per: Instead of a mound you want something else and how do you go.
Justin: That was one of the things that I've wrote this these essays about because it was a problem I really wanted to investigate and at that time really nobody was working on that kind of problem. I wound up actually going to MIT and joining a neuroscience lab in computational neuroscience. This probably shows my impatience but after three or four years of studying neuroscience I came to the personal conclusion that science is unlikely to understand consciousness within my lifetime in a way that I feel like is the kind of personally satisfying understanding that I thought I was looking for. Of course great progress has been made in parts of the question. It was actually pretty late in my graduate career. I had wound up sort of going off and working with a few different people on different kinds of topics and one of these at some point I was at this stage which I understand is very typical in extended graduate programs where I was just sick of everything. Just like I absolutely wanted to just write up whatever I had and get out and who cares what comes next. I was talking with one of the PI's I was working with on some other problem and I expressed this kind of general dissatisfaction and he asked me so what would you work on if you could pick any problem what would it be? I said well you know for a long time I really wanted to work on this thing with the as I described the termites and the engineers.
Per: The both way of the problem, how does termites do it and how can we reverse the process and control it to get any result we want.
Justin: I was really focusing on the latter. I didn't want to go out and study termites exactly but I wanted to say like okay, great so that that shows us it can be done. How could we create a system that works like that? He said if you'd said almost anything else I would have advised you to finish, write up your dissertation graduate and try to go a switch and work on that other thing as a post doc but that's something that I think you could actually do in reasonable time for your dissertation. It was in like my fifth year of grad school that I switched to that topic.
Per: Okay brave.
Justin: Yes, well and it worked out. It turned out to be a problem that I was actually able to make real progress on. It became my dissertation and most of it was theory because I'm really interested in theory. I wound up building a sort of proof of concept system as hardware demonstration literally out of things that were lying around the MIT building because in robotics as I found is when I started submitting papers they really want you to have a robot.
Per: It's called robotics.
Justin: There is great theory in robotics but especially if you don't have sort of the background so people know that you know what you're talking about. It’s incredibly helpful to actually have a physical demonstration as a sort of demonstration to say like it's okay, I know what I'm talking about you can take me seriously.
Per: It also keeps you grounded. It keeps us from becoming too far from the subject I think we'd mentioned that in the beginning. You did it. That’s so cool that you have the opportunity at Harvard to absolutely do that.
Justin: But that's exactly why it's so important in robotics to have that because there is great working theory but there's also work in theoretical robotics that would not really be possible to translate to reality.
Per: That robot keeps you grounded and sorts out the good from the bad and all that stuff. So you build this prototype then and what do you do then?
Justin: So I wound up doing this this project at the end of my graduate program actually working with Ron Kanak Paul, who I had known when we were both graduate students at MIT and then this other professor who was the one who convinced me that I should do what actually wanted to do I suggested I go talk to her and see if she wanted to support this project. It worked out brilliantly so we wound starting on this at that time. I graduated. I did one year postdoc with her and I went off and I did a different postdoc on very different kinds of subjects. I was doing some work in evolutionary theory and cancer modelling. Then when the Wyss Institute was getting started up, I worked with Radhika of course and she was trying to recruit me. I've done some work with Don Ingber who was the head of the institute and he suggested I come in. This was what I wound up doing.
I was describing before and how my interest in this problem was to learn about these termites and say this is fantastic. This shows that this can be done that large numbers of independent agents can together build large scale complex things. That’s fantastic. How would we program a bunch of termites do that, a bunch of artificial termites. Building the artificial termites obviously is going to be a huge challenge but that's for someone else to worry about. Not my interest so around the time the Wyss Institute was starting up Kirstin Petersen in Denmark was working on very much the same problem. She had learned about the termites and said this is fantastic, this shows it can be done. How would you build a bunch of termites, a bunch of robots to build the thing you want and I don't really care how you program them. Someone else going to worry about that.
Per: I see a beautiful friendship starting here.
Justin: So her adviser talked to Kasper Stoy, who had done work with Radhika and they sort of you know engineered that Radhika hired Kirstin to come work and she recruited me exactly because that the carrot was we're going to get to do exactly this thing that's really been driving us.
Per: Yes and you've been interested in this project for very long time. I also think that it's been in your mind for a long time and your mind has matured around that issue. I'm quite sure there's something like that going on.
Justin: Maybe. It definitely gets easier to think about the longer you think about it.
Per: Then you're there. You’ve got somebody dealing with the hardware.
Justin: That's when the termites project started in 2009 when we all started at the Wyss. The Wyss in fact started in 2009. We started working on exactly the thing that we published four plus years later. It just took a while to get there.
Per: I would presume so. It is a very complicated problem and since you also have the hardware and you're in robotics. You have this robot. It actually has to perform the tasks that your algorithms, it has to be verifiable.
Justin: To me that's where the bulk of the challenge is. The hardware to me was really the most amazing part of the project, the thing where the greatest amount of effort and sort of requirement went into. Kirstin is just brilliant. She managed to make this work. Maybe that she has a different perspective because she's used to working on robots and less so about algorithms but for my thinking like the algorithm part was straightforward. It’s in theory. Everything works in theory but getting something to work in real life is amazing.
Per: She does the hardware and there is just beautiful partnership that way.
Justin: But at the same time this is actually a really true partnership. All three of us we are co-authors on the paper really contributed to every aspect of the project and that's part of what made it possible to get all the way from beginning to end yeah was were talking very deeply from the beginning about but the right design decisions to make it work about what are the reasonable things that are attainable for the robots to do and what are the things they need to be able to do in order to you know accomplish something reasonable all together and as things changed in real life and as experiments showed that something was more or less feasible we'd sometimes change our thinking but was we were all really working on all aspects of that and that's I think a big part of why it wound up being successful both with the strong theoretical foundation but through the real-life realization.
Per: All of you bringing that essential component and fitting together and then doing it as a team I think that is where any great work come from and every great science certainly comes from. I mean we see so many Nobel laureates having one or two or three people on the same ticket so to say and just comes in there. The minds support each other and then you said also Kirstin had talked about the termite problem ahead of time so you were all kind on the same page already.
Justin: We had each done sort of different kinds of work on this subject beforehand.
Per: Again it gets easier to think about the longer you do it. Could you tell us any key points in the termites project that you think that were revelations that really said wow we really took it to the next level?
Justin: Nothing is really coming to mind. It’s a good question. I mean my experience when I started working on this topic completely in theory the biggest surprise in a sense was that it wound up tractable. In a sense it didn't have to be. Like it could have been that I sat down to try to understand a literally incomprehensible system. It could have turned out to be a thing that no human could make progress on. As it turned out it was very tractable like it was it was straightforward to sort of start in one place and increase in complexity and build up this sort of algorithm framework to describe the building of collective structures. When we started the TERMES project in a sense like that wasn't a surprise. We basically knew that we were going to be able to get there certainly from the algorithmic standpoint. Kirstin was very confident that from the hardware standpoint she can get there and so it was sort of just a question of how to get there and where exactly were we trying to get.
Per: What is the minimum requirements for the I mean you're not going to go out and build the real house for someone. That’s for another project but really the technical demonstrator has to be the most simple one that's good enough.
Justin: When you're really starting a new project from scratch that's a much bigger risk in some way. You really don't know whether this thing that no one has tried to do before is actually going to wind up being doable.
Per: The theoretical framework is proven that given enough blocks or building blocks a group of these autonomous agents will complete the structure as predicted. I mean that's a fundamental thing we can all rely on.
Justin: The idea is that if you're the user, if you're the person who wants this group of robots to build you something you can give them a blueprint of what you want them to build but no instructions about how to go about building it. Just show them the final result that you want and give them a supply of building material and walk away. They’ll do the rest themselves.
Per: I mean for the listener, for the viewer we might say that this doesn't depend on the fact that nothing goes wrong or anything like that.
Justin: There's a lot that's going to be very unpredictable. You may not know how many robots you have. The number might change during the process. You might lose some. Maybe things are going too slowly, bring in some more robots to join the team. You don't know what order they're going to wind up doing things in. They might encounter things unpredictably. You don't know what the exact timing of the reactions is going to be so this kind of thing kind of makes it inappropriate if not intractable to try to use a traditional planning approach when you've got, instead you're trying to take the swarm approach where you're going to sort of throw a lot of self-controlled agents at the problem and let them handle the details.
Per: Even if we have one robot working when the project ends and at least as many building blocks as required the project will end. Of course those two things has to be fulfilled, we have to have at least one working robot and at least the minimum amount of components we need.
Justin: You could push that further. I mean there are ways in which things can fail that can really cause problems and could prevent completion for instance. I mean there are there are mistakes that let's suppose that are certainly possible that can be made that requires certain amount of work to detect and to correct. As long as the mistakes are being detected and corrected faster than they're being made we're still going to keep making progress toward the goal but certainly if you're making mistakes all the time and not actually and spending all your time fixing things naturally introducing errors faster than you're solving them you're also not going to get to the goal.
Per: That's very interesting. It’s like the failure of transmitting data. I mean eventually the Skype signal breaks down and it's not understandable as speech anymore. We know very well our edge cases and they're very, there are true edge cases right and they are also reasonable in the fact that if we don't provide enough building material of course the building is not going to complete. I mean that's not something that should come as a surprise to us.
Justin: But then the idea with the robot programming in the TERMES system is that each robot doesn't know anything about the global state of the system. It doesn't know how many other robots there are. It doesn't know what the others are doing. It doesn't know how much has already been built. It has its own sensing. Some robots actually are limited to their own on-board sensing. It’s actually very limited. They have a handful of sensors. They can only tell what's happening really right around themselves and based on that feedback, based on what they encountered they decide what actions to take. Then the trick is in designing their behaviors in that all of these together taking those actions wind up producing the thing you wanted. As they move through the work space they don't know what's already been built. They have no dynamic knowledge about the state of the system. In some sense you can imagine a system where they do keep track of what they've already seen but in a swarm system where there's really a huge number of other agents and also active, also changing the world keeping that information may not actually be useful.
Per: No because it's going to be out of date and of course also sharing the information between millions of actors is not also not possible. This is very interesting. I mean the project is now published in Science Ci. The big splash.
Justin: February 14th this year. Good Valentine’s Day.
Per: A good Valentine’s Day hopefully from more than one perspective but that's a great day in any day. It’s also great for robotic science that we're out there and really participating at that level of the science debate.
Justin: It's really exciting. There’s been a number of papers like that more recently.
Per: I mean I talked to a lot of people and they say robotics is coming into its teens. It's not a toddler anymore and it's not an adult but I think that you see we've one or two more papers coming into these prestigious journals that wasn't the case five or ten years ago.
Justin: I mean there have been robotics papers and science over time but they've become more common recently.
Per: I think that that's also the feeling I get. I don't know if you agree with me that there is a sign, there's a maturity and a stability in our common knowledge within robotics now. There is something to lean on. You don't have to do everything from scratch all the time. I think that that also will make it progress faster and I think we're going to see more and more papers coming into these journals over time because I mean it's simply easier now when you don't have to start from scratch all the time. That’s really interesting. Say that we have somebody out there that is viewing this and they're saying oh cool, I can have a termite with my house. Where is the project now because it's closed now for you right?
Justin: The TERMES project as such is closed. We wrote up the results of that project. It was in a science paper. Kirstin graduated. She’s in Europe now. We miss her deeply.
Per: We'd be very happy to have her.
Justin: The TERMES project is wrapped up. Collective construction work we're still working on and in different directions. Even at that same time Radhika was working with the postdoc Nils Napp, he's now a professor at Buffalo on building with amorphous materials.
Per: I've seen the work. It’s based on a kind of expanding foam or things that are covered in glue. He used matchsticks to construct and bridges and stuff, really interesting in that perspective. What I think is TERMES compared to that is that the TERMES project is it is reusable all the time because the blocks doesn't really stick to each other and I really like the reusability. If you have one structure you could instruct these agents to continuously modify your structure to whatever you need.
Justin: I think that's great for the experimental standpoint if you're trying to build one set of blocks and reuse them to build many things. I'm not sure if that's what you'd actually want from a construction standpoint because I mean typically we build a building. We built it once. It’s finished we don't want to change it after that. Is that only because we can't change it after that or would we actually want to if it were easy enough to take things off and put them back on.
Per: I had some conversations with architects about this. This is like the termites for architects. They have really been coming back to this problem time and time again of modular buildings. We hear Vern talks about it in 130 years ago or something like that of us renewing our cities every year because we tore all the buildings down and build new ones. Architects have come to these problems throughout the ages of trying to make the buildings adapt to the needs of the users. Do you continue to work on this form of construction or is this not taking over the work? Where are you going from now on so to speak?
Justin: Lately I've been working with a structural engineer Paul Kassabian and we're interested in looking at taking this approach to, I sort of want to say more like the realistic real-world physical situations but here's the setting we're looking at. One issue is that we are going to go back three steps. The TERMES work is inspired by termites. It’s inspired at a high level. Like the termites the robots are acting independently. They’re building large structures, committed themselves, climbing over the partial structures in progress. They only have local information. They don't know what the others are doing. You know that kind of inspiration but the details of what the termites are doing not very different from the details what the robots are doing and there are a number of reasons for that. One of the reasons is because their goals are very different. Termites are not trying to build a specific mound. Every mound looks different. They’re all recognizably mounds from the same species but as you say it's being remodelled all the time. It depends on the environment that it's built in. You see a lot of mounds built around trees. There’s a giant tree sticking out of the mound and yet there's a functional mound around the tree whereas with human construction projects typically if you're the person who wants a building you start with a blueprint. You know exactly what you want built. You give that to the contractor and say please build me this. If they build you not that but something kind of similar and then it looks like it might have come from the same family. One of the ideas with the TERMES project was for this kind of thing to be relevant to human construction we wanted to have the capability of building completely predefined things.
Now we also demonstrated that you know of course you can use the same system to build things that are not completely predefined. This is actually something that I think it's sort of a minor point in the paper but it's easy to overlook completely that we gave some fairly trivial algorithms for here's how to use the same system and build something where the whole form emerges from the construction process and it's not actually predictable in advance. As I say understandable that wasn't the main contribution but that's more what termites are doing. Again going back to you as the person who wants a building built you don't just go to the contractor with the blueprints. Actually that's the second step, the first step is you go to an architect. You say I've got this site which I haven’t really looked at and I have some idea of what I want built there.
Per: I need three bedrooms, a kitchen, bathroom and the basics they know.
Justin: You give me a blueprint and the architect does that.
Per: They organize it where's the sun coming in, where is the neighbors, where's the road. They look at the site where you want have your garden and they take all these very hard decisions that again when you thought about them for a long time like the architects have it's easier but for a novice it's very hard because there's so many of them and they all, I mean if you move the patio you've got to move the bathroom and then you move the bedrooms and suddenly the patio is in the wrong place again and you just start all over.
Justin: So this work that I've been doing with Paul, the idea is to automate some of that as well, to let the structure that's being built come out of the function that you want to specify but not the precise plan.
Per: So it's taking it a step further.
Justin: And to take advantage of elements that already exist in the environment. That is something that actually Radhika and I have been talking about forever. That is Nils' work that kind of making use of the environment to build something that performs a function. The thing that I'm doing with Paul, we're looking at building with a framework of scaffolding struts and robots that can measure local forces on what they're building. The idea is for them to you know again be sort of taking local measurements, local decisions and deciding based on the forces here is this going to be strong enough to let me keep building higher, shore it up and to build things that perform a function like if there's a gap in the environment that you need to bridge from one side the other build some bridge.
Per: If the ground is uneven flatten it out so that the next level can be.
Justin: Although if you're building out of scaffolding, tubes you're probably not that's a more sort of open strut base kind of a plan so less with the solid as you were saying. That’s something we've been working on a little bit. We’ve been we've had three very good undergraduates over the summer and we've looked at different aspects of this problem. That’s a direction that that we're trying to take this collective construction work in the future.
Per: Very interesting and I think that the theoretical work and the practical work you've done where you show us that what you're original problem was can we add input in the reverse version of the termites project and say I want this and could we actually get it out the other end both from a theoretical point of view and from a practical point of view it's in science because it most like most certainly deserves to be there. It is a fundamental thing that so many will be able to rely on when they go on and take these in directions we're just not going to know about today. I'm sure that the construction industry of all infrastructure living accommodations bridges we talked about everything that is built by human hand which is a huge amount of our world around us and in the US a trillion dollar market or more this is going to be this is going to be a very significant part of the new version of that. I am quite certain of it and that's a major contribution. I'm very grateful for you taking your trying to share these result with us and for participating in the interview. Thank you very much.
Justin: Thank you so much for inviting me.
Per: I hope you liked this episode of the podcast version of Robots in Depth. This episode is produced together with Wevolver. Wevolver is a platform and community providing engineers informative content to help them innovate. It is how engineers stay cutting edge. Aptomica is the founding sponsor for Robots in Depth. Aptomica runs anything in modular robotics. Dream, rent, build. Visit Aptomica.com to connect. I am your host Per Sjöbor. Until the next episode thank you for listening.
END OF TRANSCRIPT