Robots and Human Evolution
Biologist John Long is building robots that swim, eat, flee, and evolve
Hi, this is Steven Cherry for IEEE Spectrum’s “Techwise Conversations.”
Computers are sometimes thought to be electromechanical representations of our minds, and if so, then robots are electromechanical representations of our entire selves—minds plus bodies. We ought, then, to be able to learn about ourselves—as selves, and even as a species—by building and studying robots. But there’s a problem. Most robots are built for very specific purposes, and therefore only do very specific things. They move a part on an assembly line, or weld a joint, or vacuum the floor, or scurry through rubble looking for bodies or bombs.
My guest today is building general robots, with an eye to studying robotic evolution, in order to better understand biological evolution.
John Long is a professor in the biology department at Vassar College and is the director of the Interdisciplinary Robotics Research Laboratory there. He’s also the author of a new book,Darwin’s Devices: What Evolving Robots Can Teach Us About the History of Life and the Future of Technology, published last month by Basic Books.
He joins us by phone from Poughkeepsie, N.Y.
John, welcome to the podcast.
John Long: Well, Steven, great to be here, and I appreciate the opportunity to speak to you.
Steven Cherry: John, maybe let’s just start with an example. Tell us about the Tadro—now this is a fish robot that’s programmed to seek light.
John Long: Yeah, that’s right. It’s a special kind of biorobot, so it’s built to mimic the behavior of fish. So, seeking light is something that fish do because life hangs around light—that’s the basis of where we get photosynthesis, creating in our plants and floating algae, sugar, and everybody wants sugar on some level, or the other animals that eat sugar. So we start with a robot whose job it is to seek out the light and the food that’s there.
Steven Cherry: Now, you call these “biorobots,” and they’re actually embodied robots; they’re not digital simulations. Tell us what they look like.
John Long: That’s right. Digital simulations are something we do do, but these are physically embodied robots, and they are quite literally swimming Tupperware, which does not sound very exciting. But we make them exciting, because the swimming Tupperware has in it a computer, which is doing the calculations of what to do for behavioral output, so it’s the “brain,” if you will, taking sensory input and then converting it into motor output. And so dipping down from the surface of the water is a biomimetic tail, a specially designed soft and flexible structure that has variable number of bones called vertebrae in it, and so we are actually interested in evolutionary processes, in particular what went on 500 million years ago at the origin of the group of animals to which we belong, called vertebrates.
Steven Cherry: Yeah, so let’s talk more about letting robots evolve. There’s a very famous computer game called The Game of Life that was invented by a British mathematician named John Horton Conway back in 1970. I guess that’s sort of the starting point for you here?
John Long: Well, we definitely are after the game of life, and we want those interactions between individuals and individuals and their world to give us the dynamic interplay that we call “behavior.” And behavior then is what matters for evolution. Evolution, we think of as, oh gosh, look, you’ve evolved a tusk, or in our case here with these fish, you’ve evolved these bones in your back that the first vertebrates actually didn’t have. But, in fact, evolution doesn’t see that. Evolution, or natural selection, which is the creative engine of evolution, sees how animals behave. And so we start with that behavior to reconstruct the game of life, which is what happens in front of us in the laboratory when we put these embodied robots in a large 10-foot-diameter tank. We stand back and let them go and do their autonomous behavior, meaning we’re not remote-controlling them. They’re just interacting based on the rules of their body and brain and how those structures are put together.
Steven Cherry: So tell us about how they evolve.
John Long: Well, we know enough about biological evolution now that we can simulate it, so we start with genetics. You need to be able to pass on your traits to your offspring, and so we build an artificial genome, and that we do on the computer, and the genome codes for, in some of our Tadro robots, things like the number of vertebrae, the sensitivity of a special structure called the “lateral line” to oncoming predators, and then the shape of the tail. And the shape of the tail is very important if you think about the tail as a kind of propeller.
So we build in a heredity code that allows us to give the instructions of whoever happens to reproduce—and I’ll talk about that in just a second—to that next generation. So we start with those genetics, we make variable robots—so in our Tadro4 world, we had six different kinds of these Tadros in any generation, which were acting as the prey robot. Each one of those robots was built differently. So they had different number of vertebrae—those bones in the back—different sensitivity to predator, and different tail shape. And so what we do is essentially compete those individuals to see among them who are the best at scouring the world that they’re in, in our laboratory for food, and also we introduce a predator, and so you have to be able to stay away from the predator and not become food for someone else. So that’s the basis for the game of life, and it sets up this natural selection, which I like to think of as kind of “evolutionary Olympics,” and you have to then be judged by a faceless, nameless panel of judges that determines who gets to be the parents to leave their genes to the next generation.
Steven Cherry: Yeah, and it has to be across generations, because individuals don’t evolve in their lifetime. Species evolve over time; that’s how evolution works. Now you have a robot that doesn’t evolve. Tell us about Madeleine.
John Long: Yeah, so the Tadros that we have to evolve, we call those “evolve-a-bots,” and they really test the process of evolution—what might have been the evolutionary drivers, the selection pressures, that cause certain events to happen? That other kind of robot that we build, and Madeleine is an example of that, is called an “evolutionary trekker.” Evolutionary trekkers really test the outcomes of evolution, and Madeleine is a four-flippered robot. Some people like to think of her as a robo sea turtle, and she is designed to give us control over her brain and ask the question, What happens when you swim with two flippers? What happens when you swim with four flippers? And the reason this is interesting to me as a biologist, is that 90 million years ago, the seas were ruled by giant creatures called plesiosaurs. Some of them had long necks, like the Loch Ness monster is alleged to have, and they were really the top-level predators, and so they all swam using four flippers. Now, contrast this, fast-forward to now, and when we look into the water, the top predators are whales and dolphins and seals and sea lions. None of them use all four of their flippers, which they brought with them as the history of limbs that they had when they lived on land. None of those four flippers are used at the same time. If you think about a sea lion in your head, they have these nice big front flippers that they use for moving around. So the question is, If it was good enough for plesiosaurs, why isn’t it good enough for the animals that had gone back into the water more recently? And so that’s what we can test with robot Madeleine.
Steven Cherry: So why are we down to two flippers now?
John Long: Well, the great thing about Madeleine and what she’s taught us is that, you know what? There are costs and benefits all the time, and that’s one of the lessons about evolution: There are always trade-offs, right? So if you swim with four flippers, here’s what you get: You get fast acceleration, fast braking, and really good maneuverability. However, once you get up to your top cruising speed, you are just burning twice the amount of energy as if you were using two flippers. You don’t go any faster once you get up to your top cruising speed. And so what does this suggest to us? It suggests that the plesiosaurs 90 million years ago were probably doing more maneuvering, or if they were smaller plesiosaurs growing up, maybe they were sit-and-wait predators that just hung around and then—boom!—hit the cheetah button and accelerated to grab a prey. And what we see in living vertebrates that have gone back into the water using those two flippers is they tend to be cruisers. You know, some of them, like sea lions, are highly maneuverable, but then they’re also having to cruise quite a while to find their fish. So this looks like an explanation that’s based on your behavior—how you want to use your energy. Do you want to cruise, or do you want to sit and wait for your food to come to you?
Steven Cherry: Now, it sounds like that’s the kind of result that would have been harder to find with a computer simulation. Is that a reason for actually embodying these robots?
John Long: Well, that’s right, and one of the things we do do is we do do computer simulations as well, and we have on our team mathematician Rob Root and computer scientist Chun Wai Liew, both at Lafayette College, on that team. And one of the things we find is that you need to create what’s called a “physics engine.” You need to have the equations that actually solve all the things that make any object move in the world. And so it turns out that being in water is probably the most complex kind of physics that you can imagine. You have this dense fluid that you can move through. If you’re an animal, you’ve got a flexible body that has its own nonlinear capacities that have to be modeled, and then you have this very strange motor that we see in animals called muscle. And muscle is not a very linear motor in mathematical terms. It’s quite a nonlinear motor, so it’s actually a whole bunch of complexity in the motor, and the body, and the fluid, and all that stuff has to be solved at the same time. So the physics is actually its own challenge. So when we build embodied robots, we get the physics for free. We don’t have to solve that, and we really put our engineer’s hat on here and say, “You know what, if we can build something and it works, then we understand it.”
Steven Cherry: That makes sense, but you do have to actually build the robot, and some of this seems really complicated. In several of these experiments, you’re basically building backbones, columns of vertebrae, and that seems really complicated.
John Long: It is, and it’s a real fun challenge for us. It’s, in the field of engineering, called “biomimetics,” where you’re trying to mimic biology. And one of the things that animals do that most of our engineering materials don’t do is they make their materials out of flexible, soft, and wet stuff. So as soon as you want to be like an animal, you have to figure out how to engineer in this sloppy, wet environment, and that is really interesting, because we have to do things like start with molecular collagen, build up that collagen, then stabilize it chemically in a form that allows us to keep it well hydrated, have it be very wet and flexible, and then we have to figure out how to control the interface of that sloppy thing with the little hard rigid bones that we want to put in to mimic the vertebrae that are in our back. So that’s its own fun piece of engineering there. And so if you’re interested in biomaterials or biomimetics, that is something that our team has to bring to this effort which—correct—is complicated. And then those biomimetic structures go into the robots themselves, which are programmed to be autonomous, which we borrow from embodied artificial intelligence, the principles, to make these autonomous, freely behaving robots.
Steven Cherry: You mentioned the fourth version of the Tadros before, and I guess you learned something really interesting about the number of vertebrae with that model.
John Long: Yeah, we were testing the idea that the vertebrae—that, ironically, were not in the first vertebrates, even though we’re named after those bones—the vertebrae may have appeared because of selection pressure for enhanced feeding and fleeing. So, making sure you can get a meal and not be somebody else’s meal. And what we did is we started our population in the middle of a number of vertebrae, about five on these biomimetic tails that we have, and that would allow them to “un-evolve,” if you will, or “de-evolve” the number of vertebrae or increase that number. We didn’t want to start with no vertebrae at all, even though that was the ancestral condition, because we wanted to not force the system to actually have more vertebrae. And when we applied this selection pressure over 10 generations, so 180 different trials of these different individuals and populations of robots, what we found is that the number of vertebrae increased under this selection pressure. So what we do here is we can never say, “Oh, we know exactly why evolution occurred,” but we can circumscribe the plausible. And so this selection pressure, for basic stuff, right, feeding and not being eaten, is a sufficient selection pressure, at least in our robots that are built to mimic 500-million-year-old fish. It’s a sufficient selection pressure to drive the evolution of vertebrae. And that was quite different than what we found in the Tadro3 world, where just feeding alone was not sufficient.
Steven Cherry: So one difference in building actual robots is that you can watch them. Is that like going to an aquarium, or is it more like watching flowers grow. What’s it like?
John Long: [laughs] Well, let’s see: These are either very exciting flowers or fairly boring fish. They’re somewhere in the middle. They’re not as exciting as watching the sharks in your local aquarium swim around, but, you know, they’re moving around this 10-foot-diameter tank, and they make a circuit in probably 20 to 30 seconds. What’s actually pretty exciting is when we have the predator in with our population of robots, and the predator really creates chaos in there. Because the predator, who is not evolving by the way, its sole job is to hunt down the prey robots. And so the prey robots have a brain that is modeled on the brains of fish. So we know enough about the neurobiology of fish that we can put it onto a computer, and it does this: The brain of a fish says, “I am going to cruise around and feed, but if I detect a predator, I’m going to stop whatever I’m doing and accelerate as fast as I can.” And that’s exactly what the Tadro prey robots do. They just cruise around—it doesn’t look very exciting—but, uh-oh, predator there, and then they undergo this choreographed kind of escape response, which allows them to accelerate as much as they can.
Steven Cherry: Well, John, fun or not, it’s a fun book, and it’s terrific research, so thanks for doing it and for writing about it and for joining us today.
John Long: Well, it’s been a real honor, Steven, to be with you, so thank you for having me.
Steven Cherry: We’ve been speaking with biologist John Long about the biorobots that he describes in his new book, Darwin’s Devices: What Evolving Robots Can Teach Us About the History of Life and the Future of Technology.
For IEEE Spectrum’s “Techwise Conversations,” I’m Steven Cherry.
Announcer: “Techwise Conversations” is sponsored by National Instruments.
NOTE: Transcripts are created for the convenience of our readers and listeners and may not perfectly match their associated interviews and narratives. The authoritative record of IEEE Spectrum’s audio programming is the audio version.