All right so thank you very much. In addition. I'm a research scientist at the Institute at Harvard broadly speaking my research interests are in understanding and designing complex systems in the sense of systems of many independent interacting components where some interesting collective behavior emerges from all of their joint actions that you might not have predicted from looking at the components by themselves and I've looked at problems under that very broad umbrella in a number of research areas the ones I want to talk about today are about collective construction. Program and robots to build things according the user design and I also will talk a little bit at the end about studies of the termites that inspired the approach. So the work that I'm going to be talking about today. Through should be advancing. The Great Work want to talk about today is joint work with Radhika not Paul who is a professor of computer science at Harvard and with Kirsten Peterson who very recently finished her Ph D. and she's heading off for a postdoc in shoot guard. Both of these topics are about construction and I want to take a step back to start with and look at construction in nature. Which is something that. Is almost working something that you see very broadly throughout the animal kingdom so one of the first examples that comes to mind is birds building nests which is something they do actually with a real variety of building materials this example in the bottom left is a bowerbird nest they use a variety of sort of artificial collected materials to help build the nest in the upper right this is a narrow bird that built its nest out of out of soil of mud. And can. More broadly as something that you see insects and other arthropods do this something that you see in fishes it's an activity you see primates and gauging in and actually of course it's an activity you very commonly actually see primates engaging in it's a tremendous human activity red construction is actually a trillion dollar industry in the U.S. alone. With my. You know. OK so. OK So one of the remarkable things I think about the construction industry is the extent to which it doesn't have automation. You know it's it's pretty much all done by skilled human workers using sophisticated machinery. And the. The kinds of problems that are reported with construction are things like low efficiency labor shortages high accident rates very dangerous profession you know these are exactly the kinds of things that automation traditionally helps with so there's a real opportunity for for automation to be able to transform construction the way it's transformed other industries like manufacturing. Now but more than that I think there's a particular opportunity for automation to have an impact in settings where we don't have a system in place for people to build or where we don't want people to be the ones doing the building so for instance if you want to build in disaster areas like to shore up a shaky building after an earthquake if you want to build a. Dam out of sandbags protect against rise in flood waters and these are the kinds of situations where if we have robots do the work could help us keep humans out of harm's way or in settings where we can't easily or safely send people you know. If if we want to build underwater if someday we want to build another planets these are the kind of settings where I think robotics has a chance to really make a particular effect. So there are a variety of reasons why we'd like to introduce robotics into construction. And of course there are a number of ways you think about doing that so one approach might be for instance to build essentially a gigantic three D. printer. Right over the whole building site and print the building you want do that together with a pick in place machine to handle things like plumbing and that's an approach that's been explored for a number of years for instance by bearer of cushion episode U.S.C.. Another approach that people take to to current construction in some situations is occasionally you'll have air cranes you know drive helicopters that lift in parts of a structure depending on the situation and that's a research direction that's been pursued for instance at you pan with struts. Building with these bricks to have these computer controlled helicopters building in some cases very large scale complex structures. And we're just starting to see now construction actually automation making its way into actual construction sites so this is a construction robot that has been developed actually by a construction company. It's semi automated MASON Right it's doing some of the brick laying work here there's a human Mason who's still supervising it telling it where to build and cleaning up the more Directorate. But this is you know this is an exciting step right this is that the first time I think we've really seen true automation in a. An active construction site. So all of these examples are single robots or centralized control it's a more traditional approach to thinking about robotics in. Control. But as I said you know I'm really interested in systems of many independent agents and also because of my employer I'm contractually obligated to be inspired by biology. And so let's take a step and look at. What kinds of groups of animals in nature you see building together so one good example is a group of families of beavers right which build lodges and dams that really transform the environment they live in and here are the groups or maybe a dozen if you want to scale things up a bit. Sociable Weaver birds are an example that these colonies will build nests that really overwhelm the trees that are built in. And here you've got groups maybe on the order of one hundred and so you know I may have actually given away where I was going with this in my title but you know the real champions of construction are insects and in particular termites and so this is an example of a termite mound. The tallest on record is over forty feet high it's built out of mud but it's not just a it's not a pile of mud it's actually got all this complicated internal architecture it's got networks of tunnels and it's playing the functional role it's helping to do things like atmosphere of regulation for the colony and these things are built by millions of centimeter long or smaller independent insects and each one is acting on its own it doesn't know what the others are doing it's not like it's getting explicit instructions from the Queen it doesn't know what the overall state of the mound looks like it's just moving around reacting to what it encounters and somehow out of those joint actions you get these incredible structures and so that's you know very inspiring in the sense that it shows us that that can be done you know that's a proof of principle that large numbers of independent agents with local information can build very large scale complex they. Yes And so we said Fantastic So how would we do that right how could we harness that power how would you build and program an artificial termite colony. In order to get them to build for you you know not necessarily one of these termite mounds because we've pretty much got as many from right now as we need but to build anything you ask them for right so that the ultimate goal here is sort of from an engineering standpoint is one day to be able to have robots that take let's say prefabricated square building blocks and use them to build structures that may be very large scale read maybe building ramps and staircases to let them get to places they otherwise couldn't reach so in a sense this is the long term vision. So that this idea of multi robot construction has actually become really an active research area in the last decade or so. And sort of like the there's sort of an ecosystem of different approaches people have taken by analogy to the ecosystem of different building approaches you see in nature and there's been research for instance that's focused on Harvard design that's focused on things like. Cooperation between two robots to help them achieve something that one couldn't do alone. And this is not even to mention the closely related problems. Other areas like program self-assembly information control where in a sense you've got multiple robots that are trying to build a structure out of themselves as opposed to building it out of separate inert material. And all of these kinds of. Research problems. Share a number of challenges you know that are unique to these systems of many independent agents so it's worth asking in fact why you'd want to do things with a swarm why you would want to use a large number of limited robots. Rather than building you know one very very capable one that could do everything itself. And. Again if you know if you look to nature as an example. This one approach is empirically very powerful and if you look at ants for example as sort of the great example of social insects ants make up two percent of known insect species that make it more than fifty percent of insect biomass So there's something that's incredibly successful about this approach if you can get it to work and the kinds of things that swarm system can have as advantages. Are things like if you can break your overall tasks down into a large number of independent subtasks different robots can working on these different things at the same time you can get a larger overall speed up. And it's warm systems tend to be robust for the loss of individual components so if some of the robots break the rest of them can keep working on uninterrupted. Systems can be scalable. Both in terms of so for instance if. If you have a centralized controller that's that maybe not just a single point of failure potentially but it can also be a bottleneck that could be a limit to how much you can keep track of and coordinate at the same time. And as far as control goes the idea that often is that regardless of what else happens in the this dynamic environment let's say you add more robots let's say things change in unexpected ways what the individual robots are doing doesn't need to change they'll just respond to this changing conditions where a centralized controller might need to stop and replan if things change unexpectedly. OK So that's again sort of you know the vision but the same things that make the swarm potentially very powerful also make it very difficult to deal with so again you've got lots of independent agents all going around with. Limited information each doing their own thing somehow you need to make sure they're all working together you know rather than at cross purposes. You may not know in advance how many robots you have that number may change during the course of the task. The order that the robots wind up acting in and the timing of their actions these are all things that you don't know in advance and who every program has to be insensitive to those kinds of unpredictable variations. And then if you're actually talking about building you know large numbers of robots which you know robotics Ideally you want to do. That leads to its own set of sort of unique challenges So for instance the background image in the slide is the kilowatts project also at Harvard and Mike Rubenstein when he was designing and building these robots one of the things that he had to do worry about was. Issues of scalability of operation so like if you normally when you build a robot you put it on and off switch. And if you have a. Welcome send supplies from getting feedback. Feel for a little. OK is a still audible OK. So for instance if it takes you three or four seconds to pick up a robot flip the on switch and put it back down you know you're going to be there for more than an hour just turning on robots right and that's even without worrying about things like how are you going to program them are you going to charge them all so those are the kinds of challenges that you just don't encounter when you're looking at single robot systems. The biggest issue and sort of the most interesting from my perspective is that in these engineering contexts you're trying to design an emergent system and a particular outcome right so again the the hallmarks of complex systems of interacting agents is. And unexpected collective behavior right. In general we don't have a way of looking at a set of low level rules and predicting what their collective result is going to turn out to be often the best you can do is to try it and see what happens and in these engineering cases you know in a sense it's much worse you're demanding a particular final result and somehow you have to come up with a set of rules and have some guarantee that that's going to have the result that you want. And so the first thing that I want to talk about here is the term is Project which is you know some recent work we did to try to address these problems so this is a system of multiple independent climbing robots. Where you can ask the system for a particular result and you have a guarantee that you know that whatever the robots go out and do they'll wind up producing the structure that you asked for. The name is officially an acronym. As a Officially this is the real source of the name is the genus of mountain building termites that inspired the approach and also more about these later. We tried to make this really sort of an end system in the sense that we've got a firm theoretical foundation that lets you guarantee that for any number of robots sort of whatever order they wind up acting in and what situations they want to encountering. You can prove that you'll get the particular final structure that you asked for built in the end. So that's one and the other and we went through and showed actually building this in real life to demonstrate that these kinds of things are are feasible. So I'm going to say more in the next slides about how we did that but just to say a few things to start with we took this approach of trying to keep things simple so each of these robots. Has only three motors and a handful of sensors. One of the things that we decided not to do was we have them we they're not communicating directly through sort of usual communication channels all of their coordination is being handled indirectly by manipulation of the shared environment so one of the robots puts down a block someplace that's a cue that another robot may come along and encounter and that influences what the second robot wind up doing. So that this idea of stigma. Information stored in the environment and because all of the robots are program the same way because they're all following the same conventions that lets you prove certain things about what's going to result if they're all following those shared rules that let you guarantee the outcomes you want. One of the things I think is really important to getting this to work in the end was that we sort of designed everything together so for instance we from the beginning we were really talking about designing the algorithms and the robots at the same time you know to make sure that the things that we were requiring the robots to do in order to achieve the algorithmic results we wanted were things that were actually reasonable to expect to be able to get the robots to do and converse lead to to make sure that the things that. We were putting into the robots were going to things that were in the service of the collective behaviors we wanted. The robots of locks were designed together so each can take advantage of the other so that it's this idea of mechanical intelligence. The idea that if you can design the system so that the physics of the situation takes care of certain things for you then there's that much less that you have to worry about in control so for instance in the concepts of control. You know presumably you're going to want very precise alignment of building material you know if you're building a brick wall. Kind of thing but manipulation of the physical world is something that mobile robots are notoriously bad at or it's a notoriously difficult problem it's a large active research area even for various Fister kid robots you know much less for ones that are trying to keep very simple. And so you know if you take this approach of building of using building blocks that are self aligning. Right so there are matching features in the top and bottom. They slide into place they lock together with magnets and then the robots only have to get a block to about the right place and the mechanical system takes over the rest of the details of find just meant an attachment. And also more about this last point bit later but we took an approach of rather than trying to prevent errors ever from occurring. We let small errors occur and we give the robots in a feedback to to tell when they were doing something wrong and to try again until they got it right and that turned out to be very productive in terms of preventing larger errors that would have been much more difficult for the robots recover from. So when you talk a bit about the hardware part of the system and I'll talk about the programming issues. Under the heading hardware there are a few things the robots need to be able to do they need to be able to get around in the world to have some sense of where they are as they're doing that and to pick up and put down one of these specialized building blocks directly in for themselves and also a little bit about each of those and about the sensing that they use to achieve each of those functions. So to start with to get around the robots have have two motors there's one that drives the left two wheels together when the truck drives the right two wheels so if they run of the same direction it moves were backward they run an abstraction that turns in place. Wheels have this sort of pin wheel shape. The name for that. Is. This is a term coined by rhetoric when a Case Western. As a combination of wheels and legs. To have some of the adventures of both the idea goes back longer. The reason for that particular shape was actually empirical so early in the project person started by. Trying a bunch of different configurations and combinations of wheels and the way exe and treads and different sizes and shapes and found imperiously that this particular set of webs was the most effective in terms of certain measures we were interested in. So the result is a robot that's able to move around on structures built out of these blocks and to climb up stairs that are not more than one block high. And here are these notches in the blocks this is an example of the kind of mechanical intelligence I was talking about so the notches one of their functions is their part of the self alignment there's matching features on the bottom part of the block to help them slide and lock together but you can see they're also acting as sort of intermediate stairs to help the robot climb to higher places and they also because of the shape help to straighten out the robot as it climbs if it was rifting off center. So this is good for movement in structure environments and also in less structured environments and so here is the same robot moving through grass through gravel plant material and through the snow. As the robot moves around it you have some sense of where it is. And for that there are these black and white markings we put on the on the blocks on the bottom side of the robot there are these six active infrared sensors which can detect these black and white patterns sort of in extremely poor resolution it's a six pixel camera red and as the robot. Around on the structure these sensors see different patterns of black and white. Which let the robot keep track of its position relative to this physically embodied coordinate system that it's building out of blocks as the structure grows. Together with the sensor that tell it when it's climbing up or down a step it can keep track of its position. Relative to the. Some starting point as it moves over the structure when it's away from the structure it doesn't have that available as a cue. And it has these these ultra sound sensors that it uses you know echolocation two to follow around the perimeter of the structure. So here's the robot walking around the perimeter using sonar. And at some point when it gets around to this point it encounters this stripe on the ground can detect that with the infrared sensors. Follow the stripe in to reach the structure and at that point when it climbs onto the structure. That's a unique landmark It knows where it is that basically the this right guides it to the origin of this committee coordinate system and thereafter it can keep track of its movement by reference to the markings on the blocks until leaves the structure again. So the last thing the robot needs to be able to do is to pick up and put down blocks. And as I said the robot has only three motors and two of them are being used for wheels so there is one motor left for you know picking up and putting down blocks and holding them securely and releasing them so the design that Kirsten came up with for this ripper. Involves these these passive prongs that are normally held in with torch and springs but pressing against the backside of it forces them open and so the motor rotates. The whole assembly down which presses the back end against this wall and forces the gripper open. And so one motor then is enough to to pick up an whole block securely and also to release them when it puts them down. Then there is too tactile sensors on the gripper that tell the robot about. Whether it successfully picked up a block. Or whether it's done something wrong or needs to try again. So all of this together is enough for one of these robots in this video to build actually a pretty large scale structure so this this is sped up with the full processor was twenty four minutes long and winds of building a staircase that's about eighteen times its own volume. Which if you want to measure things that way is technically comparable in speed to the fastest human bricklayers. So something you see here is this point about the the robot is frequently making small errors right so it often takes it several tries either to pick up a block or to realign itself with the grid underneath that it often slips while it's climbing and. Again we took this approach where we said OK let that happen but give it the feedback to detect when that's happening and to take steps to correct it and that wound up working well enough for this robot to be able to build this large structure completely autonomous Lee right the only intervention that we're giving it here is we're reloading this brick cache we're nose to go back and get more material aside from that it's working completely on its own. And this is you know very long string. Of primitive operations primitive operations in the sense of pick up a block turn right from the forward. All of which need to be successful in order to build this large staircase. OK so that's. You know I think that's great from the hardware perspective but when I talk about very long string of primitive operations and then you think about building something larger scale than a ten block staircase or doing something with a large number of robots you know again you don't know in advance how many you you have that brings you very directly to problems of programming the swarm right to get the results that you want. So here's the approach we ultimately went to taking for that. As I said the user should be able to ask for a particular structure that they want the system to build the end. So we have an offline compiler so there's many different ways that this could be built. So we have a step where an offline compiler marks up the blueprint that the user originally specified with what amounts to traffic rules so these arrows so the color of the each of these sites is just the height of the stack of blocks there. The color of the arrow is showing the direction of the arrow. And so the idea is a robot at any of these sites has different choices about where it's allowed next. So you know a robot here can can come in this way and can travel out in any of these three directions for joyous. So this is a structure a specific set of rules that governs how the robots are allowed to move through the workspace and this is the set of traffic laws that you give to all the robots to make sure they're all you know following the same rules. And as they move through the workspace. They look for a place where they would like to attach the block they're carrying and before they do they have to make sure that it's not going to cause problems if they do so there is a single set of safety checks that the robots perform before attaching a block for any structure. To make sure that this is going to. To lead to correct completion of the structure you asked for the problem is that if robots attached lots sort of carelessly and where they're supposed to go but without any regard to where blocks of already been attached it's very easy for them to get into trouble so let's suppose that robots coming along the structure and it says I know they're supposed to be a block attached there I'm going to turn around and attach it right now. So physically it's capable of doing that right but if it does then it's just built a step that's too high for robots to climb so it's cut off that path. Which is you know a problem for the future Same idea if it decided to attach to its right you know again it can do that but now it's built a cliff that it can't climb down. Or if it went a little bit further ahead you know decided to touch a block here is not going to interfere with anybody's movement but now there's nowhere that somebody can stand in order to attach a block at the site where this one is right now standing. And even if there was a place for someone to stand attaching a block they are now would mean trying to force one directly between two others which is actually very difficult operation mechanically especially again for the limitations we are were put in the robots. So it's to prevent those kinds of situations from happening that these checks are necessary and I'll tell you what they are it won't be obvious why these are a good idea but what should be clear is that these are all local checks that the the. Robot can evaluate the truth of each of these things based on observations of its immediate neighborhood only. So if you're a robot you're standing in a site you're thinking about is this a good place to attach a block right now. You need to think about two things One is how did I get here and the other is where am I going next. So for every site I could have come from according to traffic rules I need to make sure that either the stack of blocks there was at a high right now is at a higher level than the stack where I'm standing. Or the stack that is at the same level as the stack or I'm standing and it's complete you're not supposed to be any more blocks attached to the previous site. One of those two things has to be true for everywhere I could have come from to get here. When I say ever a could come from I mean immediately one step previous. The other thing is where I go next for every site I could go to next the height of the stack of blocks there has to be the same as the stack here. And if those conditions are satisfied that it's OK to attach here I can go ahead and step on to any of those next sites that the traffic rules allow turn around and put down the block where I was just standing otherwise don't cause problems. So this is a local set of checks that one of these robots can perform. That you can show will lead to the structure that you asked for and not get stuck along the way So in brief. These checks erd explicitly constructed to prevent the kinds of problematic situations that I showed a few slides ago and you won't get situations where construction stalls where you could still physically attach more blocks we're just not allowed to because the rules forbid it everywhere. Because you can show that if there were such a situation actually construction is still allowed and possible somewhere else on the structure. And because all of these. Yes these checks and these actions are local because all of the robots are moving through the structure in this sort of directed way according to these flows. The actions that one robot takes are going to be able to have a negative impact on something that someone else is doing somewhere else. So the way that this structure actually gets built can vary very much from from one instance the next so in a sense there's this idea of emergent process that occurs each step that can happen next you know depends on what steps have already taken place and there's many different routes by which the same structure can be built. But the final results every time is guaranteed to be the specific thing that you asked for. OK then just to emphasize this is a reactive approach not a planning one so that the robots reckoned whatever it encounters behind its back anything could be going on. Right it's not seeing something it makes that needs doing and saying OK I'm going to go get a block come back here and making plans to do that particular thing because in this dynamic environment with many other robots as soon as it leaves a site the information had a stale and something else can be happening behind its back so instead it's responding just to the immediate local circumstances that encounters. OK so I'm going to completely change gears now I've been talking about these robots right that were inspired by termites I want to spend a few minutes and talk about actually termites right so the. The inspiration from the termites in a sense is very high level right the robots like the termites are independently controlled they're restricted to local information. They're reacting to the environment then counter they're building things much larger than themselves climbing over the things that they're building in order to get to areas they couldn't otherwise reach. OK but the details of what the robots are doing are very different from the details of what termites do. And there are a number of reasons for that one reason is that they actually have different goals right. Termites are not trying to build a particular mound. You know every termite mound of the same species you can recognize as being a mound of the same species but no two are identical they're building some mound that works for the colony in the place where they're building. And you can use the term system with different sets of rules that I've just been talking about to build structures in a similar kind of way where the result is. Emerges from the building process. Rather than being a particular outcome that you demand. But typically in human construction projects you know you start with a blueprint you give that to a contractor and if they build you something that is kind of close to what you asked for but not really in kind of is bending in the there's going to be hard feelings so for relevance to human construction projects you know we wanted there to be this ability to demand a particular result and get it so that's one reason that the rules the robots are following are different from the termite rules another reason that the details of what the robots are doing are different from what the Germans are doing is that actually there is not that much known about the details of what termites are doing. As social insects go termites are really a neglected research area everybody studies ants you know everybody loves bees termites are not beloved. You know it's actually kind of funny in general to talk about you know. Termite inspired construction robots because the familiar termites around here are the ones that destroy things right their excavators So that's the other thing that makes them inconvenient to study is that if you want to study the ones that actually build things you have to travel halfway across the world. So that's what we did so for the last few years we've had this collaboration with Scott Turner who's a. He's interested in especially in the physiology of the termite mounds and he's a professor at SUNY E.S.F. and he has a research program in Namibia. So we have now gone twice to visit him there and. Twice myself and Kirsten and on the second trip Radhika joined us and also Neil snap who was a post-doc with Roddick at the time is now a professor at Buffalo. And the goal is to try to get some. Some understanding of what program the termites are running. And connect that the things they build. So. You know once you're there you can let's say knock a hole in the mound and stick a camera inside and this is not a very good way to study their behavior partly because it's you know it's very and controlled it's very complicated there's there's a lot going on can't keep track of which Termite is which for very long another reason this is not a good way to study them is this happens right away so instead we set traps and take hundreds of the matter time back to what I will charitably call the lab. And we put them into controlled situations and petri dishes and record their behavior. So one of the. The approaches that we've been taking to these experiments has been to. To take petri dishes where we split. Soil between one half and the other in two different conditions so Scott has identified a number of characteristics of the soil that have an impact on termite response to it so for instance the amount of water in the soil is something that they respond to they tend to excavate from. And from weather areas and deposit material in drier ones there's this idea of a cement pheromone that when they put down the Tiriel there's a chemical in it that evokes building material from others. The physical shape what they encounter has an impact on what they do this is the one that's maybe most closely related to the robot work you know where the robots are looking at the configuration of material around them. Termites have. You know in a high level sense the same kinds of responses and the texture of the soil so it's important their relative composition of sand versus clay. And so what we've tried to do in these experiments is we put down. You know this is this half and half composition of soil which lets us look not just at the termites behavior. On each side as we record in controlled issues that have. Only a single kind of soil in them but this also lets us see what happens when they encounter this transition because in countering differences is a major part of what drives their behavior. So we take video recordings of these things in the first step then and again this is and it was a familiar topic for this audience is to create a tracker so that you don't have to sit there and label by hand the positions and orientations of the termites and every friend. Once an undergraduate has been developing that and then person extended that to to start a sort of semi automated. Mostly automated No they do it does that. The heavier boring lifting automatic labeling of different behaviors and you can start to classify what the termites are doing at different times and places within this dish. And. This is really very much ongoing work we're starting to ramp it up we've got a number of recordings from our second trip that we've been looking at but we're really currently preparing new experiments new design for a trip that we're planning for the spring. But what we have brought back we've started to already see some some interesting results and so one that old spent a couple minutes telling you about has to do with this issue of the cement pheromone. It's a topic that pub or do any of us who is a post-doc with Scott has been very interested in. So there's this standard idea. As I mentioned that the termites are adding some pheromone to the soil as they build with it that triggers other termites that encounter a piece of soil put down by one to deposit more soil there and this came from studies in nine hundred fifty nine. By you know one of sort of the fathers of this field here Paul grass a. Who observed that you know one termite would put down the trail another one would come and put it down where the first one had put it and that's it was in that work that he coined the term stigma. The etymology of that is it's work evoked by science. So that's been basically the standard understanding for the last you know fifty five years that this cement pheromone is the thing that's driving that aspect of building behavior. OK but for something to be a pheromone Firman his work was also introduced in one hundred fifty nine is a big year. It's a couple things One is it it has to be a specific chemical So termites. Trails with a trail pheromone kind of the way that ants lay trills and the trail pheromone is well characterized you can write down a formula for it you can synthesize it in a lab you know you can draw your own trail with the artificial pheromone and the termites will respond exactly the same way as they would if another termite laid it down. And in fifty five years nobody has identified a cement firm. So it doesn't mean that there isn't one as I said you know there's been less study in this area than others but you know it's interesting that things like the trail for Mons have been discovered quickly and no one's found a cement one. Also importantly for something to count as a pheromone it has to evoke a particular behavior. And so the idea is that the cement pheromone is supposed to trigger deposition. And in fact when we say if we do one of these split experiments so let's say that half the dish now is fresh clean soil the termites not been building with it is otherwise matched in terms of water content and so on to the other half and the other half is material that they've just been building with and so it should be full of cement pheromone. And then you use this automatic behavior classifier on what are they doing left put them into the dish. So they're certainly doing something different on the two halves right there spending a lot more time on the side that has the putative pheromone Laden soil. But if you look at the breakdown of how they spend their time on each side. Or. The relative distribution of how they're spending their time is actually not significantly different. On the side with you know supposedly the pheromone in the side without So it's not that it's triggering a specific behavior like deposition which is the standard understand understanding for for many decades. It may be that there's. There may not be a cement pheromone there maybe something like a colony odor. You know that is making its way into the soil that termites like the smell of it they prefer spending time on that side it did you know make them feel safe make them feel at home but what they actually do on that side is not matching the standard understanding of what this pheromone The supposed to be doing so that's the kind of thing that we're starting to see as we look at these recordings and you know we're looking forward to to what new things are going to come out of that the next few years. I understand that there's a class time coming up so I'm going to wrap up just before I do I want to say that you know we're always looking forward to graduate students and post-docs who so if you think this kind of thing that I've described sounds of particular interest if you're finishing up a program or if you were advising a student who's especially good while i'm here we should definitely find a chance to talk and. I'll close with that I'm going to run videos of the robots in the top and video within the mound at the bottom and leave those playing will lead to questions thanks thanks. Right. Most definitely chemical So this is this is almost a it's so this is this is a repeat for the audience I guess for the video recorder so what's going on if it's not a pheromone So this is actually something that. Is I think of particular interest to the entomologists right this is not a pheromone in the sense that they think of a pheromone. That fits the definition. Engineers have wound up with very different definitions of very different sense of what stigma means you know what pheromone mean so like in conversations with Scott you know for him Stigler G. specifically means one thing triggers another thing. In robotics people used to mean basically anything where information is being stored in the environment one way or another so in the sense of there is some chemical totally there is some chemical is it a pheromone in this sense of the entomologists maybe not. Yes And actually some of the images that have shown throughout actually the direction has changed at different times in the project. Is Right. There is a difference I can't speak to precisely what it is because the physical things really persons domain but the upshot is. Every tiny detail has an important functional result so that the direction the reason for going from the straight. Spikes spokes rever on the Web's to the curved ones has to do with things like It helps keep them in places they turn. When they were straight they're more likely to fall off the structure. The direction of it has effects in terms of helping with climbing but occasionally getting stuck on things right and so. I'm not sure what the specific problem was that led to them being reversed at different times but whatever direction they wound up in is the most effective. It's a great question agriculture rather than construction so we've. We have tried do a little bit of work with a Boston company called Harvest automation. Which actually has a very similar approach to that problem looking at that they have been looking at is in ornamental horticulture so the plants that you go to you know Wal-Mart the river and you buy potted plants they're actually like gigantic farms might not be quite the right word they're grown in pots there are there put out in a field you know that gets sprayed with water and pesticide and so on and as the plant grows it starts taking up more space. So initially you want the plants to be as packed as close together as you can. Because it's in a better use of space and things are spraying with and so on but as they grow if they're too close it winds up looking bad for the plant. Throughout the plant winds of growing in a way that doesn't look as good and won't be worth as much on the market so you actually have to keep rearranging these pots you know hundreds and thousands of pots during the whole time the plants are growing and this is a this isn't thing it's done by hand it's a terrible thing that nobody wants to do it's one of these you know perfect for Abbas kinds of jobs in the villages system to do this kind of thing and it's it's almost nice morphic problem to the term is tough because I mean they're not climbing but they're. They're building you know configurations of material moving things from one place to another. So that's. It's limited in agriculture and I think you're. Maybe thinking about things like you know really large scale crops which. Is a. Industrial Agriculture has become a thing that's where you have sort of like a gigantic centralized machines you know there's a there's a real. Pressure toward crops that can be grown in sort of you know very regular rows and so on and that's not in some ways a natural way to grow a lot of crops and and. I suppose the answer questions we haven't worked on that but yeah this is it's. Another area where these kinds of robots could potentially have a really strong impact you know if you can go through and do sort of one on one spray actually just a thing that's needed just in the place that's needed you know do local monitoring rather than trying to apply an intervention to a whole area. That's the reason for the so the question is about collisions. So the the traffic laws that I was showing are all sort of one way signs right from between any two sites there's a one directional arrow and so there's. There's never a head on collision between two robots. The closest there can be is two robots that want to enter the same space from different areas. And. Or let's suppose you know those robots walking around the perimeter and then one of them has finished building and it wants to reenter the parade walking around the perimeter to reenter. So the only kinds of collisions you get are merged collisions in that case and the the resolution essentially is you know you have some local negotiation as to who's going to go first and in general if you see someone in front of you they get priority. Right so potentially a centralized approach can be more efficient. The decentralized one may wind up with a lot of movement that strictly unnecessary you know if you look at an ant colony you see a lot of insects running around in ways that don't look like they are to the purpose necessarily. You know a certain amount of movement winds up being wasted because. The individuals have stale information about what might or might not need to be done somewhere they haven't seen in a while. If you can do things with you know central control and global knowledge about the state of the whole system that's one of the ways in which that kind of system can have an advantage in terms of that efficiency. You know so we're not actually necessarily planning on. So the termite research is actually you know worse than that because worse than science right now we want to understand how was it the termites are. What are they doing how does what they're doing connect to what they build. We don't necessarily have the. The goal of building systems that use exactly the same approach. And also draw distinction between bio inspiration and biomimicry and so. You know by mimicry as you see something in real life you send them to build something that's just like that and bio inspiration is you see something real life and you say that's great I want something that you know send me up in some direction or. Do you remember this life or. So if. So we wouldn't use chemical pheromones in that kind of setting because you know chemistry is great for biology to use. It's really hard for robots to use. What we would do instead is something like putting our IF idea tags on each of the blocks right and then the robots can use those to write and read information. Again specific to the location where the the information is put and and that's something that we've explored actually in previous work kind of there's this continuum of. Putting more or less abilities into the building material versus requiring the robots you more or less and if you go further down that path actually at some point you get to. So that these blocks are passive right and then you can make them a little bit more capable and they're all distinct and they can. Carry some information and then you can make them all right all right so robots can put any information they want in a pretty place and you go a little further and you put a processor in every block right and now the robots are only there to carry blocks around the blocks can tell the robots where they're allowed to attach things and that helps in some cases and then you take things one step further in program self-assembly the robots are the blocks. Right and so depending on the application and the expense and so on different things call for different combinations but that's. OK. Thank you.