Lecture 29: Self-replicating Xenobots.

Hinzugefügt: 2026-05-12

[00:00:00]Okay, welcome to the final lecture. Um, we are going to finish up our discussion today of evolving physical structure alongside neural control of rigid soft differentiable meta and biological uh robots which will take us to the end of uh your tour through this particular approach to creating autonomous, safe and useful machines.

[00:00:27]uh next week, just as a reminder, by 11:59 p.m. uh next Wednesday, be sure to submit your written report, your written report and oral presentation video to Brightpace and then we'll be all here together again bright and early Thursday morning for oral presentations.

[00:00:48]All the instructions about how to prepare and submit your written report and oral presentation are available through this link and through Brightpace. I went through all of that last time. Any questions?

[00:01:01]Okay. All right. So, we're going to return to lecture 24. We were introducing this idea of trying to automatically uh create autonomous machines out of biological rather than technological components. I walked you through hopefully some familiar territory last time where we saw uh Sam Creman, my former doctoral student who took this class, took HyperNet and adapted it a little bit for use with biological materials. So what we saw last time was hyper neat evolving populations of compositional pattern producing networks or CPPNs which as CPN are want to do paint patterns within a three-dimensional space. In this case the hyper uh the CPN's are painting or proposing a geometry and distribution of two cell types frog heart tissue and frog skin tissue. And as we saw last time, eventually we got to this point where Hypernet was proposing uh this particular solution which Doug Blackston, our micro surgeon at TUS, nodded his head and said, "Yes, I think I can build it." And it turns out that actually he was able to build it under the microscope as we saw last time till eventually we obtained this little creature which does indeed walk along the bottom of a petri dish filled with room temperature freshwater. So you can imagine there's many more videos in which we failed to cross the gap, but at least uh it is possible or was possible.

[00:02:40]Since then we've moved on in a lot of different directions. We're going to focus on one direction today when we get to lecture 25, which is one of the nice things about biological materials is that they tend to want to produce copies of themselves. So if we want just not just autonomous, safe and useful machines, but a lot of them maybe we can convince them once they do whatever it is we want them to do is to produce more of themselves as a side effect. Okay. So we we showed you last time that by 2020 we could actually cross the simto toreal gap which suggested this idea of biological machines was possible. Uh we published this work in uh January 2020 and as you can imagine self-replicating frog bots funded by the US military tended to catch some people's attention and I spent two weeks sitting in my office on the other side of this wall fielding uh interviews with uh uh reporters and we got our 15 minutes of fame or at least xenobots got their 15 minutes of fame. Here's just a snapshot of some of the uh public reaction to this. Uh it was certainly an exercise in public communication. How do you go about communicating to the general public that self-replicating frog bots funded by the US military might have some good uses? So, um, we're not going to spend time going into that, but if you're interested in science communication and the interface between scientists and the public and at least this scientist's attempt to do so, feel free to have a look through the media.

[00:04:15]Uh, we got all sorts of, uh, feedback.

[00:04:18]Uh, we ended up briefly on uh, the late show with Steven Colbear. You can go watch this uh, at your at your leisure.

[00:04:25]I will say it was quite interesting looking at sort of the range of responses that we got. Uh, as you can imagine, the majority of the responses are, "Oh my god, why would you want to create the Terminator?" Turns out most of those responses tended to come from older folks. We also got and still get to this day emails from younger folks.

[00:04:46]U, my [snorts] favorite was from a 15-year-old uh, high school student who said, "How do I become a xeno roboticist?" or how do I train to be a xenoboticist? I said, great question.

[00:04:57]Job doesn't exist yet, but I hope by the time you get to uh grad school, it will.

[00:05:02]So, mixed responses. So, fear, uh interest, you know, amusement, you name it. Okay. Uh the our 15 minutes of fame lasted until about March 2020. So, it lasted about two months until another little creature came along and stole our limelight. But there you go. Okay, that's [snorts] history. Okay, for our purposes at least, now that this was possible, we started to think about, okay, it's obviously not an actual useful safe machine, not something we're going to deploy uh out into the world yet, but what might this technology be possible be capable of as we move forward? So, just for fun, Doug took a bunch of Xenobots, which you're going to see uh in the bottom.

[00:05:52]Took a bunch of Xenobots and put them into a petri dish together. Remember, a Xenobot is about 1 millimeter in diameter. Um, you can tell that the cells that make up the Xenobots are still a little bit adhesive, not as much as they were earlier, but they tend to stick together a little bit. And based on that invivo experiment, so invivo meaning in actual living tissue, Sam, my PhD student, went back and went from real to sim and built a simulation of a whole bunch of bots moving around together. And it's a little difficult to see. Let me see if I can put this on loop for you. If I can remember how. I can't remember how. That's fine.

[00:06:33][snorts] Uh, you might be able to see in vivo. There are some very very small little bits of leftover cells, little small bits of plastic lint that's in the dish and the xenobot swarm in this case are bumping into those small bits of material smaller than them. So Sam went and littered uh the simulator with a whole bunch of small cubes. And not surprisingly, these things that tend to move around pu uh push these blocks around. And we started to get the idea of pushing this material into piles. So very much inspired by Bradenberg vehicles and Roombas. Maybe someday xenobots could be used for a cleanup task.

[00:07:23]I'm going to skip to this video for a moment. this particular video, this invivo bot that you're going to see in the bottom, you can see this more clearly now. It seems to be interested in this small bit of material in the dish.

[00:07:39]If you'll remember all the details we talked about last time about designing and then building these xenobots, what do you think? Is this invivo xenobot actually quote unquote interested in this small bit of material that it bumps into? What's going on here?

[00:08:07]surrounded by >> okay possibly uh if this is even is a clump of cells I don't know whether it's organic or not but if it is organic maybe cells tend to like to be surrounded by other cells so it's possible there I'm going to put scare quotes around all your verbs there interested in in order order to be interested. Of course, this thing would have to be able to sense and react to these bot to these little bits of material in the dish.

[00:08:43]There'd have to be some kind of sense, think, act cycle. Did we build a sense, think, act cycle into our simulated xenobots?

[00:08:53]If you recall, where's the neural controller inside this thing?

[00:09:00]There there isn't one. There's just sine waves that we built in that are expanding and compressing the simulated heart tissue which you can see in this particular bot. Sam put uh put a a little yellow cube in front just for fun. But this simulated bot absolutely does not sense that yellow cube and is not reacting to it. So this is actually probably a pretty poor approximation to what's actually happening here.

[00:09:30]We didn't build a sense, think, act loop into the simulated xenobots. So assuming that the invivobot in the bottom part is actually sensing, thinking, and reacting to whatever that little bit of detritus is, how is that possible?

[00:09:53]like maybe the skin or heart cells like have [clears throat] sense within them and it's able to react via those cells.

[00:10:04]>> So absolutely every single cell every single skin and heart cell in the invivo bot has its own sense think act cycle.

[00:10:14]The individual cell is arguably the most complex you know robot or machine that's out there. Every cell is covered in tens of thousands of sensory receptors and it's able to move and deform uh itself in response to those sensory signals. So the the bot that you see in the bottom is really not a bot. It's a swarm of bots that are all kind of stuck together.

[00:10:41]And if they actually are sensing and reacting to this bit of detritus in the dish, that means that those 2,000 cells that are kind of stuck together are coordinating themselves somehow.

[00:10:56]It's a hypothesis unknown at the moment.

[00:10:58]Definitely clear that the cells have their own sense, think act cycle, but are they coordinating enough at the quote unquote bot level to coordinate action in a response to their environment?

[00:11:09]>> It's weird that like just beline straight to that chunk of material because even if it could sense it, I feel like it wouldn't be able to sense it from that far away.

[00:11:18]>> Right? So, typically xenobots, as you may have seen by now, they tend to move in this spiral pattern. Sometimes they move straight, sometimes they don't. Why they quote unquote choose to or not, we still don't know. And it looks like it's making a beline towards that piece of material. I think this is just a coincidence. We just happened to capture some video of this event. It's what's going on when it gets near this thing that's still unclear.

[00:11:49]Yeah. I wonder if it's just something like physical in that something about its gate when it hits it or comes close to it. Um, and there's like bits of pieces or whatever that it like is off balance and just goes.

[00:12:05]>> Excellent. So there's a competing hypothesis which is the the mechanical hypothesis. So the mechanical hypothesis says yes the cells themselves are sensing, thinking and acting, but the bot as a whole is better to think of as just kind of it's not coordinating or the cells are not coordinating. Remember that these things look kind of spherical, but they have a little bit of 3D geometry. So you can almost think of this as like a ship that's sunk to the bottom of the ocean, but it's still able to push itself along the bottom of the ocean. And the hole has a particular 3D shape. And perhaps it's lying on one half of its hole and moving in a bit of a curving path which results in this sort of spiral motion. and then it happens to bump into something on the bottom of the ocean which tips it onto the other half of the hull or some other facet of the bot's 3D geometry which switches the way in which it moves and now it moves like this. So there is a competing hypothesis this thing is either sensing thinking and reacting to the piece of detritus or it's not.

[00:13:17]When we saw this video, I asked my biology colleagues, which of these two hypotheses seems more likely? So, some of you might have come across this idea of AAM's razor.

[00:13:29]What's AAM's razor for those that have heard of it before?

[00:13:33]>> The simplest explanation is usually right.

[00:13:35]>> The simplest explanation.

[00:13:37]>> Exactly. The simplest explanation is usually the right one. Which one here is the simplest explanation?

[00:13:46]physical one.

[00:13:47]>> Maybe maybe the physical one is because it doesn't invoke too much magic, right?

[00:13:53]The cells don't have to coordinate. It's just just being the AAM's razor piece.

[00:13:59]It's just mechanical what you're seeing.

[00:14:01]It's more a result of 3D geometry and simple physics of how these how the heart expands and contracts.

[00:14:09]On the other hand, this is a ball of machines that have their own sense, think, act ability. And each individual cell in this ball has 3.5 billion years of experience going back through its ancestors of dealing with different situations. cells that have made it to the present day. Part of the reason why they're here is because them and their ancestors were very good at staying alive and coordinating with their neighboring cells to do whatever the collective wanted to do. These individual cells are very motivated and have tons of experience of coordinating with their neighbor. So to say it's just a mechanical explanation and to assume that the cells aren't coordinating and cooperating to do whatever it is they're doing here is actually maybe the simplest explanation.

[00:15:04]As a group, we tend to keep going back and forth. Still not quite clear what's going on here. We don't have time to talk about it in this lecture, but we do have a series of papers we've published looking at the cells themselves and how they communicate and whether they communicate. Cells communicate through a lot of different channels. One of the main channels they communicate is using calcium. And we have a lot of data now uh showing that the cells inside one of these bots, you can see calcium flowing between these cells and it is absolutely not random. Whatever it whatever is going on, whatever these cells are communicating with one another, it's definitely not just a whole bunch of people squished together, screaming and yelling and completely uncoordinated.

[00:15:50]There is something going on that is definitely non-random. Whether whatever is going on leads to coordinated sense, think act cycles at the level of the bot as a whole, still unknown. If they are coordinating then you successfully made robots within robots.

[00:16:09]>> Absolutely. Yes. If they are coordinating it has a lot of implications. One of them is that there's bots within bots. The other implication is if we're going to try and build autonomous machines out of biological cells then the machine itself is kind of cooperating in the self assembly self-organizing process. One of the difficult things as you've probably seen now in this course about designing and building and training machines made out of metal and plastics and ceramics is that obviously the metal and plastic and ceramic doesn't get the message. It's inert. It just kind of does what it does and then based on entropy it gradually doesn't do what we want it to do. The machine falls apart. There's perverse instantiation.

[00:16:54]That doesn't seem to be the case when we build with biological materials.

[00:17:00]On the topic of the bots, you're saying that they communicate with calcium?

[00:17:05]>> We're not saying they communicate with calcium yet, but the pattern of calcium that flows between them is definitely not random. That much we know.

[00:17:14]>> If that were the case, would you be able to then put them in maybe a state of surprise to see what happens by putting calcium around them?

[00:17:24]>> Uh, put them in a state of surprise and see what happens. Absolutely. I don't remember if I have this slide here. Let me let me see. Let me jump ahead. Oh, yes, we do. Okay, back to the topic of surprise for a moment.

[00:17:41]Okay, here's a xenobot. Remember, millimeter across. And Doug, our microsurgeon, goes in with very, very small clippers and cuts this robot almost in half. And as the text says, over a few hours, this bot starts to stitch itself back together again. This is something that no one in robotics has been able to achieve yet. You take your Roomba and cut it almost in half and you got a dead Roomba indefinitely.

[00:18:14]Another implication of building with biological materials is that the cells have 3.5 billion years of experience dealing with surprise. how they deal with it and what they do. It depends on the circumstances. But in this case, you got a whole bunch of cells that are in a configuration they've never been in before. They've been they're expecting to be a tadpole and then an adult frog.

[00:18:38]They're not. And then we add insult to injury by then cutting them in half. and they seem these cells although they're probably not happy, they're not in their comfort zone, they are still able or they've retained the ability for wound healing. So the cells have brought along a lot of the competencies that they would normally have in a frog or in you.

[00:19:01]If you nick yourself, the wound heals up. That competency carries forward into these new kinds of biological constructs. Also a good thing from a robotics perspective. We have a machine that is safe in the sense that when it's surprised, it does what you hope it would do, which is try and reform back into its normal shape and then continue on with whatever it is we wanted them to do.

[00:19:29]Okay, let me back up again.

[00:19:33]Okay, so we've started to explore collective dynamics. This was um this was another shot that Doug took. We're still looking down through the microscope and you can see there's about a dozen xenobots in the dish here. And he put very, very small glass beads into the dish, millions of them. And you can see just by the way that the xenobots tend to move, they tend to pu push these beads into piles like a Roomba would. If you got enough dust bunnies in your apartment, the Roomba is not going to get them all. It's going to end up bulldozing them into larger piles of dust bunnies. That's more or less what you're seeing here. Probably the biobots are probably not aware of or uh intentionally reacting to these small glass beads.

[00:20:22]Okay, just for fun again, my former PhD student Sam took a little small uh bit of material and put it inside one of the hollow xenobots. Uh you might have remembered from last time there were a hundred evolved designs. Some of them are hollow. So this was us kind of thinking out loud about potential applications.

[00:20:44]This potential application if it ever comes uh bears fruit, it's going to be quite a few years down the road. This is exploring the idea of intelligent delu uh drug delivery. Unfortunately, for a lot of things that can go wrong with you inside your body, it's very difficult at the moment to get medication or therapeutics directly to the problem such as a cancerous tumor. So there is a branch of robotics dedicated to intelligent drug delivery which as you can probably guess by now is a pill that's surrounded by or embodied as a very small robot and you either swallow it or have it injected into you and like a Roomba or like a Braenberg vehicle it's now going to follow a signal in the body to deliver the drug to exactly where it needs to go. Again, as you can imagine, these kinds of robots are still very much in the early prototyping stages, but some robotics labs have moved on to tests in uh rats and pigs at this point. So, we're already on to mammals.

[00:21:52]What's the signal?

[00:21:55]If you want a robot to conduct intelligent drug delivery, what is the what is the taxis behavior you're interested in? Remember our discussion of the the soft robot navigating through a field of debris looking for human survivors? Kind of similar.

[00:22:20]Imagine the robots looking for a cancerous tumor in the body.

[00:22:24]>> Looking for certain cell types that the body would normally deploy to try to handle something like that.

[00:22:31]>> Absolutely right. So follow the immune system. The immune system hopefully is also trying to combat the tumor and it is sending uh it is sending workers to go and try and break up the tumor. You could design a robot to follow that that signal go in the direction of where the immune system is going. What else?

[00:22:53]If you want to create a reliable taxis behavior, it's usually good to rely on more than one signal as we saw when we were imagining our robot looking for human survivors. What else?

[00:23:06]We're digressing a little bit into cancer biology here.

[00:23:14]Even if you don't know much about cancer, imagine this big ugly tumor growing inside the body. It's very complex. It's trying to absorb a lot of resources from the host body.

[00:23:27]>> I also want to follow abnormality in the body train to understand well what the body should look like.

[00:23:34]>> Follow abnormalities in the body.

[00:23:36]Unfortunately, again in cancer part of part of one of the thousand things that makes cancer so difficult is the abnormality is very localized. Right?

[00:23:44]There's there's the tumor itself.

[00:23:46]>> Just follow the resource flow. Follow the resource flow, right? One of the things that cancer is very good at unfortunately is drawing a lot of resources to it. So again, follow the resources or abnormal uh increases in resource flow toward in some direction and now you've got maybe a more reliable signal for moving upstream towards the tumor.

[00:24:11]Tumors also shed cancerous cells. Again, it's part of the insidious nature of cancer. Those individual cells, if they end up getting lodged in other parts of the body become tumors themselves. This is what makes cancer cancer so scary.

[00:24:28]If you can follow those cells that are being shed from the tumor, go in the opposite direction of those individual cells. Okay, this is all still very much theoretical medical research. This is not a practical application yet. But this all sounds good. Even if we could do all of that, even if we could make a reliable robot that is able to swim towards and find a cancerous tumor, there's another problematic aspect of this potential application.

[00:24:59]>> Dropping off >> possibly dropping it off that that's pretty complicated.

[00:25:06]So, if this robot in pill form is swimming through the body and it's not made from biological components, it's made from the technological components you've seen in this course so far.

[00:25:18]What's the problem?

[00:25:23]>> The body absolutely will attack it. One of the things that our bodies hate above all else is small bits of metal and plastic inside the body for obvious reasons. Not a good thing. So one of the out one of the biggest challenges in this field of robotics around intelligent drug delivery is the immune response. How do you sidestep the immune response?

[00:25:49]Hint hint kind of hint hint. It's actually a little tricky. The you said before that the designs with really small holes were the unstable ones or just ones with holes.

[00:26:06]>> The one the ones that had small holes or small concavities were were are currently impossible to build by hand by Doug the microsurgeon. Doesn't mean that will always be the case. In this case, it's a large enough concavity that he was able to construct a robot like this.

[00:26:22]That's not the problem.

[00:26:27]>> Could you put like the cells that the immune system will be responding with attach that to the body and kind of disguise it immune system. So we could we could try and do some camouflage.

[00:26:40]Yeah, [snorts] I told you that these are frog cells.

[00:26:43]How well does your body like frog cells?

[00:26:48]If it's in your digestive system, that's actually wonderful depending on your point of view. outside the digestive system, not so good.

[00:26:59]If you've got a cancerous tumor inside the digestive system, this alone might be okay, but it's a frog that's going to have drop frogbot that's going to have to drop off some medication on the way through.

[00:27:11]The harder case obviously is where most cancerous tumors are, which is not on the lining of the gut or the intestine.

[00:27:21]How are we going to go after those kinds of tumors? Can we go after those tumors with this kind of machine?

[00:27:30]The answer is at the moment, of course, not at all. But thinking ahead, you mentioned disguising it. Disguising it with what?

[00:27:38]>> Sounds like human cells.

[00:27:41]>> Human cells. So, what do we do? Do we try and wrap a frog bot with human cells?

[00:27:50]Do we use human cells instead of frog cells?

[00:27:53]>> So, we've been talking about xenobots in this course so far. Zeno being short for xenopus leavis, the particular frog that we take these cells from. Xeno in Greek also means like newcomer or stranger.

[00:28:08]So, we thought that that was awesome.

[00:28:10]Um, also goes along with xenomorph if you know what xenomorphs are.

[00:28:15]What we're talking about now are anthrobots.

[00:28:20]And again, we don't have time to go into this in this course to go into this. We have a couple of publications on anthrobots. You can go find them on Google. These are more or less the same things, but built from human cells. I'm also not going to talk in this class about where we get those human cells from. Uh, if you've had your breakfast, you might not want to hear that story.

[00:28:43]But we can collect human cells and build anthrobots, which are human cells. So getting closer to avoiding the immune response to a foreign body that's trying to deliver medicine to a cancerous tumor, which anthro, which human problem.

[00:29:13]>> Absolutely right. Best way to avoid your immune system is to reintroduce your cells into the body. So there is now an active line of research around anthrobots for medical applications many years from now. And the idea would be we would harvest cells from the patient themselves. You wouldn't really be f fooling the immune system at all. you would literally be introducing you back into you. Okay, bit of a digression, but that's something we're looking into.

[00:29:48]I showed you wound healing. Um, in finishing up this first lecture on xenobots, then we'll move on to self-replication.

[00:29:55]Where does this body of research fit into the larger scientific landsc landscape? took us a while to try and figure this out and there's still kind of a lot of argument about where all of this fits. But obviously in this course we've been talking about robotics uh evolutionary robotics which is a very small subfield within the much larger field of robotics.

[00:30:19]Now there is a sister discipline to robotics known as synthetic biology which is this general approach to try and create new kinds of biology, new kinds of organisms that have very different form and function from naturally occurring biology. So it's not quite robotics but it's nearby. So let's dive down into this field of synthetic biology for a moment. At the moment, it's a very new field and not surprisingly, most of the new kinds of biology that have been built within this field have been built or designed by humans.

[00:30:58]Some of these uh human designed uh human designed synthetic constructs are actual organisms. They contain no technological components at all.

[00:31:09]Anybody know of any examples?

[00:31:13]Humans have been trying to create new kinds of biology for a very long time actually. So this category depending on how you view it is actually quite broad.

[00:31:24]How have we been doing it?

[00:31:27]>> Agriculture since like the beginning of farming.

[00:31:30]>> Okay. Agriculture. So 10,000 years back, right? That's one way we design new kinds of organisms.

[00:31:39]Can you go even further back?

[00:31:46]kind of like animals.

[00:31:51]>> Anyone have a dog or cat at home? These are not naturally occurring organisms, right? They've been selected and altered and some designed maybe in scare quotes, but we've definitely been convincing naturally occurring organisms to adopt increasingly different form and function from what's found in nature. So, we've actually been doing synthetic biology for a very long time. You can go millions of years back if we're talking about dogs and cats or 10,000 years back if we're talking about agriculture. And then let's go about 30 or 40 years back.

[00:32:25]We figured out a third way to convince organisms to adopt new forms and function.

[00:32:33]The 30 or 40 years back being the hint.

[00:32:36]How else do we do this?

[00:32:40]genetic >> genetic modification right so GMOs also now sits in this rather broad category so we figured out ways to convince uh only living material to adopt new forms and functions much more recently there's sort of this new category which has uh assumed the name of bio-hybrids which as the name implies are hybrid machines they're made up of part technological components and part biological components. [snorts] Um, what you're looking at here is arguably the first paper reporting this idea. It's go it goes back to 2012. And in this case, they were making a very small jellyfish.

[00:33:21]So, you can see the scale here, 1 millimeter. So, this thing is about this big. Um, it might be difficult to see here, but there's some seethrough material that's been cut in the shape of a little jellyfish. So, this is a little bit of see-through uh silicone. actually not that different from the dragon skin we saw in this class two weeks ago. So there's some technological uh components.

[00:33:45]They threaded into this material some very very narrow gold threads which you can't see in the photographs but are sort of uh visualized here on the cover of Nature Biotechnology. very very gold uh very very thin gold threads and then they cultured on top of this rat muscle cells.

[00:34:07]Sound familiar?

[00:34:09]Why?

[00:34:14]What are the rat muscle cells doing?

[00:34:21]>> Receiving signals from whom? From what?

[00:34:32]Who knows? Wherever they're receiving signals from, you're going to say, and then >> like expand and contract, >> expand and contract. They're on the anterior, the top surface of this little creature. So when they contract, what do they do to the arm, the silicone arms of this jellyfish?

[00:34:57]Now, if all the rat if all of the rat heart if all the rat muscle cells themselves all contracted independently, you probably wouldn't get any swimming.

[00:35:08]So, the gold wires that are placed on top of this thing actually do conduct electricity. They send electricity between the rat muscle cells. And it turns out that just allowing the rat cells to uh communicate with each other electrically is enough to increase the synchrony of the cells so that you get swimming. Another instance of this fact that individual cells they will do whatever they can to try and communicate and coordinate with their neighbors.

[00:35:44]These are what you're seeing over here is still the product of human thinking and human design. There were some very talented biologists and roboticists who came up with this design which on version 191 finally worked and produced this. In this course we've been spending uh this was a more recent updated version of this experiment. Um you'll notice that they're now uh approaching this robot from the front left with a light source. This is obviously a tip of the hat to whom?

[00:36:18]Bradenberg, right? It's always Bradenberg vehicles. Okay. So, in this case, these are optogenetically controlled heart muscle cells. So, sorry, uh I misspoke. Not rat muscle, rat heart muscle. In this case, these individual cells have been genetically modified. So, we've uh stolen from the bag of ideas over here.

[00:36:42]These rat heart muscle cells have been genetically modified to contract whenever they uh whenever they're bathed in light. That's the opto part. And they've been genetically modified.

[00:36:55]Optogenetics. You introduce some genes that alter how rot uh rat heart muscle cells normally behave.

[00:37:05]You can see that they're flashing light from the front left of this ray. You can see the gold uh wires here.

[00:37:12]What happens to the heart muscle cell, rat heart muscle cells that are on the left side of this quote unquote stingray?

[00:37:35]They've been genetically modified to react to light. React how, do you think?

[00:37:46]the ones on the left side that contract.

[00:37:49]>> They're gonna they're gonna they're going to contract. They're going to contract. Uh sorry, I forgot to mention one other detail. The lights are flashing at a high frequency. So, whenever there's a burst of light, all of the heart cells on this side suddenly contract. They don't even need to coordinate anymore because they're all being bathed in light and they all contract which causes the little left wing of the stingray to flap. What do you think happens to all the rat heart cells that are on the right side of the creature at this moment in time?

[00:38:27]Remember the Bradenberg vehicle just physics at this point. How much light is falling on the right side of the ray?

[00:38:39]Not much. Or at least less than is falling on the left side of the ray at this moment in time. So what happens?

[00:38:46]>> Flaps less.

[00:38:48]>> It flaps less. So are we in the presence of vehicle 2 A or 2B?

[00:38:54]The coward or the lover?

[00:39:00]Coward.

[00:39:00]>> The coward. Absolutely right. So, the cells on the left are contracting more than the cells on the right because they're receiving less light or at least contracting faster, which causes the left wing to flap more strongly than the right wing. And away it goes. And you can go and look this up and you can see this stingray either swimming away from the light or in some cases they got a little bit more clever with the optogenetics here and they were able to make the lover as well that turns towards the light. Again roboticists being roboticists which is taking these new materials and figuring out how to create simple machines that have different kinds of taxis behavior.

[00:39:41]Okay, still a very new field works in progress. no practical applications yet.

[00:39:47]Okay.

[00:39:49]All right. Um I'm going to skip over the video. You can watch this at your leisure. We're focused in this course on the other half of this field which is still very new. Arguably xenobots and anthrobots are the only members of this other part of the taxonomy for now which is computer-designed or computationally designed or AI designed synthetic biological constructs. We bring in an AI and ask it to figure out how to combine biological and technological components.

[00:40:23]No one's attempted that yet, but I'll bet you either this year or next year there's going to be an AI generated version of the stingray. Somebody's going to figure out how to teach an AI how to put together biological and technological components to create a biohybrid. What you've seen in the case of the xenobots is AI designed machines that are made up of only biological components. And moreover, these biological uh these biological machines are non-GMOed.

[00:40:55]So when we get over here, we're now dealing with ways to alter the form and function of living materials without affecting their genes. Our ancestors a million years ago when they were selecting for dogs and cats, they didn't know they were doing it, but al obviously they were altering the gene pool of those species at the time. They were doing genetic engineering just very slowly and they weren't aware they were doing it. Over here we've got xenobots and possibly anthrobots that are non-GMO. Okay. [snorts] Um, along the way, this particular category, some people refer to xenobots as reconfigurable organisms because the AI is reconfiguring biological materials.

[00:41:43]Um, we also call these CDOS's, so computer-designed organisms, which is a nod or a reminder that CDOS's are not necessarily GMOs. If you break up if you break open any of the cells in a xenobot, it looks identical to wild type frog. Okay, just a little bit of taxonomy for where all this work fits in.

[00:42:06]Any questions about that before we switch to self-replicating xenobots for the last half hour?

[00:42:13]No. Okay, we've already uh we've already had a lecture on self-replication just to remember before we get into the sci-fi and all the rest all the other aspects of self-replication. There is a very good economic argument to be made for creating self-replicating machines assuming we can keep them under control which is uh which is exponential utility.

[00:42:39]It's this economic argument. Utility is the fancy word in economics for how useful is something? This thing has much more utility than this thing. So most people are willing to pay more money for this thing than for this thing. That's utility.

[00:42:56]Economically, we would like to create large numbers of things that have as much utility as possible. Exponential utility is can we create something that's useful that's utile or has utility for people. And as a side effect, that useful thing also creates more copies of itself. The moment something that has a utility of X produces a copy of itself. Now together those two things have a utility of 2x.

[00:43:27]If one of those things creates two copies and those two copies create four copies and those four copies produce eight copies. If you measure utility over time as this self-replication process rolls onward into the future you get an exponential amount of X. the exponential amount of utility. Our technological civilization at the moment is very good at building very large numbers of utile objects, things that people seem to be wanting to spend money on. But that rate is not exponential.

[00:44:01]It's rising like this. With self-replicating machines, even if they were very inefficient to build, took a long time for a self-replicating machine to make a copy of itself. It's possible that if you could figure out how to produce economically useful machines that people want to buy, and they're also self-replicative as a side effect, you could economically compete with General Motors or Apple in the future. I wish you good luck, but at least it's possible. That's how the theory goes.

[00:44:33]Okay. Part that's part of the reason why we're interested in self-replication. So again, just as a reminder, at the moment we're very good at this, which is building factories or building things that make things. The problem is the reason why we don't get exponential utility is that the thing that makes things makes other things. Car factories make cars. Phone factories make phones and so on. What we want to do is create useful machines which among other things make copies of themselves, which make copies of themselves, and which might, as a side effect, make things people actually want to use like cars.

[00:45:14]Okay. All right. Now, for the sci-fi, of course, Hollywood's had a good time with this. NASA's very interested in this. uh if we want to build a self-sustaining uh colony on the moon or Mars, we are going to probably need robots and we're probably going to need a lot of them.

[00:45:31]But sending a huge number of machines to a moon or another planet is extremely expensive. Uh if the rocket blows up, we you lose all that stuff that's inside.

[00:45:43]So NASA for a very long time, arguably going all the way back to the beginning of NASA, has kept in the back of its mind this idea of why don't we just send a few smart machines that are able to find the materials in situ on or under the surface of wherever we're sending them that can mine those resources and use them to build copies of themselves.

[00:46:07]If we could get there, suddenly the prospect of building in space or building on the surface of planetary bodies becomes much easier. We just have to loft less stuff into space. Okay, so again, it's an idea that's been around for a long time.

[00:46:24]At the end of the first Xenobots paper back in 2020 when Doug showed us this picture, this kind of led us to this idea of, well, what if we replace these very small red glass beads with just more frog skin and heart cells? What would happen? Doug's answer was, "I have absolutely no idea what would happen, but I do know that they would, the bots would start to push these cells into larger and larger piles." And us roboticists at that point knew enough biology to say, "But the cells are going to be adhesive at that point." So, they're gonna uh probably stick to each other more than they're going to stick to the bots that are pushing them because the bots are older and the cells that make up the bots are less adhesive at this point. So, you've got a big ball of wadded up tape and that tape has gotten very dirty. It's not very adhesive anymore and it starts pushing smaller wadded up pieces of tape and those smaller pieces are going to start to stick together. So, you can start to see how the idea of self-replication started to come up. So, we asked Doug, what would happen if you had a big pile of skin and heart? Would it actually start to move around? And again, Doug's answer was maybe yes, maybe no. But for various biological reasons, the heart cells themselves were kind of problematic. So Doug came back the next week and said, "Let's do the following.

[00:47:58]Let's actually try this. But in this case, we're going to reduce the number of cell types from two to one. We're going to get rid of the heart cells. And Doug gave us a crash course in frog skin biology, which is that if you put enough frog skin cells together, the cells that are on the surface tend to grow these very small subcellular hairs called psyia. And these psyia beat against the surrounding air or the surrounding water. So in a normal frog when it grows into an adult frog, it's got these very, very small hairs all over its surface.

[00:48:36]And these hairs are used to kind of slow off bacteria and algae and other stuff that's bad for the frog.

[00:48:45]You also have psyia not on your skin but in your esophagus and in your nose cavities and the cells the psyia there are also trying to get rid of stuff.

[00:48:57]Mine are working overtime because of all the pollen count out there at the moment. But you've got psyia, frog have cyia. Doug hypothesized that some of these piles, if they were big enough, would grow these psyia on their surface.

[00:49:11]And if [snorts] there are enough of these psyia, they could actually start beating against the room temperature freshwater. And it's possible that some of these piles might start to move. So, the movement now or the motors of these new xenobots you're about to see, they're not going to move because there's contractile heart uh muscle cells inside. They're going to start to swim along the bottom of the petri dish because they've grown a whole bunch of these subcellular psyia on their surface. So far so good.

[00:49:45]Okay, off we go. All right. So, as you can imagine, this actually worked. And we reported this in a paper a year after the first paper. [snorts] Here's how this works. We take a whole bunch of skin xenobots. Now, no heart. We put them together in a petri dish and then we sprinkle a whole bunch of dissociated cells into the dish. You can see some of these xenobots already moving and acting like Roombas. They're just passively pushing a bunch of these cells uh together. And after about 24 hours, some of these piles have gotten pretty big and they've started to sphericalize.

[00:50:26]So, it looks like the cells in some of these piles are starting to coordinate with their neighbors, at least to the point of starting to pull on their neighbors. And if you have a whole bunch of cells in a pile that all pull inward, you end up getting uh you get end up getting some spheres.

[00:50:45]Uh at this point, uh Doug went in and harvest harvested all of these sphericalized piles, put them in another petri dish, left them there for about five days to mature into adult xenobots.

[00:51:02]And over those five days, some of these bigger piles did grow cilia and some of these bigger piles started to move. So now we've got baby xenobots or child xenobots. We have one round of self-replication.

[00:51:20]The the P's here represent parents, parents in scare quotes, and O represent the offspring here, the bigger piles that have formed. These offspring are able to move around. You can guess how this goes. We took the piles that moved, put them together into a third petri dish, sprinkled some more dissociated cells into the dish, and around and around and around we went. There's a little bit of decision tree logic in panel E here. You can see one minus alpha and alpha over here. So if uh alpha is representing the fraction of uh offspring bots that are actually moving.

[00:52:05]So alpha starts to drop the number of offspring that actually are motile or move goes down pretty quickly and at that point this self-replicative process runs out of gas.

[00:52:18]Sometimes there are moving piles and we keep this process going. So what Doug did was to make uh what we call these days basil xenobots.

[00:52:29][snorts] Basil xenobot means kind of base. This is just Doug taking a whole bunch of skin cells from the animal cap of a frog egg, the part of the egg that's going to develop into skin. No AI design here.

[00:52:47]This is just wild type or basil or normal xenobots. And it turns out that uh if you do that with about a dozen of these basil xenobots, you get one generation. You get two or three offspring. They move around a little bit. They never produce any offspring.

[00:53:07]So we go to the halt state. We get one round of self-replication.

[00:53:14]Not very exciting. So then Sam and I came in and said maybe we can get Hypernate to design bots.

[00:53:24]And the fitness function now is not going to the right. The fitness function is self-replication fidelity. Meaning how many times can you go round and round in this cycle? Yeah.

[00:53:39]Okay. Very last evolutionary robotics experiment you're going to see in this class. Here we go. Okay. Oh, just for fun. So, this is a snapshot of an AI designed Xenobot. It's not spherical.

[00:53:52]It's got this Pac-Man shaped. It's got this Pac-Man shape. I think we talked about this Xenobot uh a few weeks back.

[00:54:01]In retrospect, obviously, if you design something like a Pac-Man or a C-shape, it's got this concavity at the front, and it's basically just a shovel. It just tends to produce larger piles.

[00:54:14]Larger piles tend to grow more psyia.

[00:54:17]The more psyia the offspring bot has, the more it's going to move around and the more likely it's going to push yet more dissociated cells into piles.

[00:54:30]All right. So, where did this C-shape come from? You're going to watch uh we're watching snapshots from one hyper neat run. So, we've got one CPN that's painting 3D geometry. It's not even painting. It's not even painting. I'm never going to be able to figure this out. There it is. There it is.

[00:54:55]Okay, we've got hyper population of CPN. And each CPPN now only paints 3D geometry because there's only one cell type, skin. And unfortunately skin is shown as magenta here. It's a little confusing.

[00:55:11]Here is um an initial random CPN from the initial hypernate population. And this CPN basically paints a sphere.

[00:55:23]Okay. We we take 12 of those spheres, put them in a simulated petri dish with simulated frog skin cells. Again, a little confusing. Shown in green. You can see that the parent xenobots do indeed, just by chance, like Roombas, push some of these dissociated cells into piles. And if you look carefully, you'll notice some of these piles are sphericalizing a little bit. They're pulling together into a tighter pile. So Sam took VoxCAD, the physics engine underlying all of this, and put in a little bit of biological detail. He put a little bit of logic into the green cubes, which is if the green cube comes into contact with another one, connect to it with a beam. Remember, VoxCad always simulates voxels as connected by neighboring uh connected by beams. These [snorts] beams are active. They're actively compressing. They're pulling the neighboring xenobot. Xenobot, the neighboring skin cell towards themselves.

[00:56:35]Okay, a lot of detail, but again, very familiar components. We got hypernet, we've got CPNs, we have voxad under the hood.

[00:56:46]Fitness function is how many children, grandchildren, greatg grandandchildren do you produce?

[00:56:54]In this example, uh, in this example, none of these 10 or so, one, two, three, four, five, six, seven, eight, nine, these nine parent xenobots, none of them managed to create a child. What is a child here? A child is a pile that gets big enough to grow cyia. We had absolutely no idea what big enough is.

[00:57:18]So Doug looked down the microscope and he watched all the moving piles and he watched all the piles that didn't move and he figured it out. It's about 3/4 of a millimeter.

[00:57:30]So we took that back into the simulator and in this case this particular pile here, the biggest pile is 74% of the way to being big enough.

[00:57:44]We assigned the CPN that produced these Xenobots. We gave it uh a fitness score of 0.74.

[00:57:54]The the C that CBPN produced a pile that was 74% of the way to being big enough to move. So, or another another way of saying this is the parents built 74 of one child.

[00:58:10]So far so good. Throwing a lot of biological detail at you. Remember that the fitness function is self-replication, but we don't even have self-replication yet. So, we're giving these CPNs like partial points for even getting to the point where self-replication is possible. We've seen that mechanism, that algorithmic mechanism many times in this course.

[00:58:34]What's it called?

[00:58:36]Scaffolding, right? So, we built in a little scaffolding here, which is build big piles.

[00:58:45]After uh a few generations uh the CPN started to paint donuts which do make a big enough pile at least one pile that becomes motile. It's big enough. [snorts] So this is going to loop after I tell it to.

[00:59:10]So we got nine parents that produce one child. So this particular CPBN gets one fitness point and then it gets an additional 0.51 points for what reason?

[00:59:26]One point for producing one child and an additional 0.51 points for point 51% of the way to creating a grandchild. So, we'll watch the nine parents again.

[00:59:44]Okay, here go the nine parents.

[00:59:48]They're going to collectively build one child.

[00:59:52]There it is.

[00:59:55]This child produces a pile that shouldn't move. I think I've got the wrong video here. So this is nine parents producing one child which in turn produces one grandchild which should be two points. And I didn't see how big the pile that the grandchild produces but it's going to get that CPN is going to get a a partial point for the grandchild producing uh something approaching a great grandchild.

[01:00:24]Okay. So sorry there's a little bit of mismatch between this CPN and uh this particular behavior here. Everybody get the idea?

[01:00:34]Why do the donuts produce a child and the spheres don't?

[01:00:44][snorts] It took me like two months to figure this out. It's not so obvious, but maybe some of you can get there faster than me.

[01:00:56]>> They're bigger. Uh, are they bigger? I think they're the same size.

[01:01:03]>> The only advantage here has to come from the from the hole in the middle. But how?

[01:01:15]>> Uh, they don't move any quicker.

[01:01:19]Well, it's like a soft body, right? So, it probably compresses maybe.

[01:01:24]>> All right. Good. Now, my ego is bruised.

[01:01:26]You figured it out. Exactly. The hole in the middle means that this thing is just softer. So when the front edge collides with some stuff, the front edge tends to bend inward more than the solid sphere does, which means you now have indirectly a little plow on the front, which tends to push more stuff together, which triggered the creation of a motile.

[01:01:56]Doug uh Doug took this doughut design and said, "I can make this." He went and made it and sure enough it produced one child. Okay, just a note about our little uh our little cartoon over here.

[01:02:09]So, I keep saying parents in quotation because we're we're obviously a very far away we're very far away from traditional ways in which organisms get together and produce copies of themselves. There's nine of them moving around, but in this particular case, there were two of them that got together and produced a pile that was big enough to produce one child. So there were nine prospective parents. Two of them got together and produced a child. When we published this paper, the title of this paper was kinematic self-replication.

[01:02:44]The two parents here are obviously not getting together the way that mammal parents tend to get together. This is a very different way of combining your behaviors to produce an offspring. And that combination is mostly kinematic. It has to do with how you push and pull material push and pull material around and how your own body deforms. So we called it kinematic self-replication.

[01:03:08]Kinematic meaning having to do with movement.

[01:03:12]Turns out that at least so far, no one's been able to identify a species that produces offspring in this manner. It's theoretically possible. Doesn't seem to happen. So, the biologists, our biology colleagues were kind of excited because we'd also sort of unexpectedly uh discovered a new form of self-replication that biological organisms can perform under the right conditions.

[01:03:38]Okay.

[01:03:39]uh after many more generations. So this is showing you a phoggenetic tree. This picture is showing you um an initial random CPN, the best CPP, the best CPN from generation one, but a mediocre CPN from generation one produced a better CPPN in generation two. Another CPN in generation three, four, five, six, seven, eight. So here in the eighth generation there is a CPN that is producing uh a CPN that produces Pac-Man-shaped uh creatures that produce three children which in turn produce one grandchild and I think that's it. So again my apologies these numbers don't match up with what you're seeing over here. So what is the what is Hyperne doing in retrospect?

[01:04:34]What it's doing is delaying the loss of self-replication. We've seen that basil xenobots, if you put them together in the right way, they produce one offspring. In this particular run of hypernate, we had a CPN that can create parents which create three children which create one child. Our record so far is five rounds of self replication.

[01:04:58]So we we have the CPNS have found designs that produce great great great grandchildren and then those can no longer self-replicate.

[01:05:09]So when we wrote this up, we were ready for the media this time because now we had selfreplicating frog bots funded by the US military. What a surprise. We got even more media attention this time around. We said, "Don't worry, they're never going to get out of control because they always lose self-replication. They don't produce more and more. They actually produce less and less of themselves over time.

[01:05:34]We're fine." The reporter said, "Yeah, that sounds great, but what if they escape from the petri dish? What's going to happen?" And we kind of lo laughed them off and said, "Don't worry. These things, they can never disappear from the dish." The next day, the grad student who was working on the anthrobot said, "Yeah, I counted them yesterday.

[01:05:50]There were 14 of them. I counted them today. There's 13.

[01:05:54]The day after we'd said they'll never escape from the lab, one of them escaped from the lab. If you've seen it, please let us know. We're still looking.

[01:06:02]Okay.

[01:06:07]Okay. So, like before, um, what what I just walked you through was one of many runs of Hypernate. We did 100 independent runs. We got these 100 designs which gave us a much better idea of why these particular designs were working.

[01:06:29]They're all kind of variations on a theme.

[01:06:37]I think we mentioned convergent evolution in this class at some point which is mother nature and in this case hibernate independently discovers things that work and it turns out that there's only one or a few ways in which things can work. So mother nature keeps independently discovering the same thing over and over again. For example, eyes and ears have been independently discovered in different species many times by mother nature. What is Hypernet independently discovering over and over again here?

[01:07:10]>> Pac-Man. Pac-Man and donuts work pretty well because both end up giving a concavity on the anterior or the front-facing uh edge.

[01:07:21]Okay. All right. So, just backing up for a moment. Um again, this idea of self-replication has been around for a long time. It was proposed theoretically by John Vonoyman in the 1940s. And the way John vonoman explained it was you have a machine and that machine uh has a copier. It has a part of it that's able to make copies of things. One of those things that copier can do is make a copy of uh all of this stuff according to these instructions. And the final thing that B this copier can do because it's a universal copier, it can copy not just the parts of itself following this instruction, it can also copy the instructions themselves and put that instruction into the new machine. That's Vonoyman's uh replicator, which interestingly was proposed right around the time that genes were being uh discovered.

[01:08:21]For fun, uh, we put together this little simulation. So, self-replicating frog bots are cool, but could they actually ever do uh, uh, exponential utility? Um, we have another colleague who works on flexible electronics. Remember the jello robots? If you want to put electronics inside a jello robot, you want the electronics themselves to be soft. How do you assemble soft electronics inside a soft robot? I don't know. So, for fun, we took our self-replicating robots and we had them self assemble soft electronics. So, what are you looking at here?

[01:09:06]You're looking in uh the blue pieces are very short strips of flexible wire.

[01:09:13]You'll notice that there is yellow traveling along the little blue strips.

[01:09:18]These are little bits of those blue strips that have attached to a very very small power source.

[01:09:25]There is electricity flowing out and some of these soft bits of wire end up connecting a b a very small power source to a very very small LED, a very small light, which is the little yellow stuff you see in the corner. Here's a little light down here. And this light has not been attached by a wire to a battery. So this light has not been turned on. So what are you watching? You're watching a swarm of xenobots which are about a millimeter across self assembling some soft electronics.

[01:10:02]And we know that those electronics are being assembled correctly when some of these lights start to turn on. The lights are being connected to batteries.

[01:10:12]The xenobots know nothing about soft electronics. They're just doing what they normally do, which is move around like Roombas.

[01:10:21]We're assuming that the number of lights that gets turned on is our utility. It's the thing that we want. We want we're putting out a bunch of soft electronics into soft robots and we want that soft electronics to be assembled such that lights go on.

[01:10:39]The green cubes that you see as usual are independ skin cells. So there's a lot going on in this video. What you might notice is you have parent xenobots that are utile.

[01:10:55]They're doing some useful stuff. And as a side effect, they're also building copies of themselves.

[01:11:03]You'll see some of these green piles turn pink in a moment. And I think some of them are going to start to move.

[01:11:10]Yeah, there they go.

[01:11:12]We're going to take those offspring xenobots out of this dish and put them into another dish with more nonassseembled soft electronics. And we want those offspring in the new dish to assemble that electronics as well. And we're going to count the number of lights that gets turned on.

[01:11:34]Turns on if it turns out if we keep doing this, we got this picture down here, which is going to be the very last figure we describe in this class.

[01:11:43]In the center of this tree down here is the original Pac-Man swarm. And it's very difficult to see, but there's a red dot here. There was at least one light that was turned on. That swarm did a little bit of useful work for us. It produced some offspring and those offspring got distributed into two other petri dishes. It's hard to see, but there's two edges coming out of the center here. So, we've got petri dish 2, petri dish three, and both of those two edges are red, meaning a [clears throat] light went on in petri dish two, and a light went on in petri dish three. So we took the xenobots out of dish two and three and put them in dish four, five, six and seven. One dish, two dishes, four dishes. As we move outwards in concentric circles in this figure, we see more and more red. There are more and more lights being turned on in more and more petri dishes. This is obviously all sim and this is us running a little bit ahead ahead of what can be done in real. But this was us wanting to demonstrate that at least in theory these kinds of self-replicating machines can produce incre exponentially increasing numbers of themselves and an exponentially increasing amount of useful work on behalf of us humans.

[01:13:14]Okay, you have now reached the edge of what I know about evolutionary robotics.

[01:13:19]Uh I appreciate you coming along for the ride on this uh tour through one particular branch of how to create useful and safe machines. Uh it's been my honor and privilege to uh take you on this tour. I wish you all the best with your remaining exams in the one minute that we have left. Remember, you have one last quiz due tonight, and then we will all be back here together at 7:30 in the morning on Thursday. Make sure you submit your written report and your oral presentation to Brightpace by 11:59 p.m. Wednesday night. I will be here with caffeine and sugar for you, and the TA and I are looking forward to seeing uh your final projects. Okay, thanks very much. See you next week. [snorts]