Darwin’s Devices

Chapter 3 ENGINEERING EVOLVABOTS

“IF YOU UNDERSTAND IT, YOU CAN BUILD IT.”[10] THIS IS THE engineers’ secret code. It is so secret, in fact, that I can’t be sure that it’s real. No engineer has ever said this to me, a non-engineer, but I’m guessing that it’s the last question on every licensing exam and is whispered during their secret handshake.

Regardless of whether they say it, it’s definitely how they work. I figured this out for myself, having worked with many of them over the years, building robots. Engineers decide what their device should do—they understand it—and then they build it. Not surprisingly, that attitude drove me crazy—I was working and thinking in the opposite direction. At one company’s early design meeting, in which the hardware, software, and mechanical engineers were pressing me for specifications, I let loose my exasperation: “Let’s build the robot and then see what she can do! If I knew the specifications, then I’d know the answer. What we are doing is testing a hypothesis!” Silence. With eyebrows raised and knowing looks exchanged, the three engineers politely ended the meeting with a collective, “We’ve gotta get back to work.”

This left me sitting with an old friend, Charles, a.k.a. “Chuck” Pell, a chief designer at the company and someone trained as a sculptor and not, I began to appreciate, as an engineer. He interpreted. Chuck explained that engineers never use the word “hypothesis,” were never taught about how you test one, and were, instead, trained to build contraptions that work to do a job. The job that a contraption does, he continued, is defined by the specifications. So most engineers are literally lost without the specifications. You can’t get somewhere without knowing where you are going. This all makes sense, I conceded. But it doesn’t tell someone interested in evolutionary biorobotics the first thing about how to proceed. Damn the code![11]

Or don’t. What I’ve learned since then—thanks to designers like Chuck and his band of merry engineers—is that the code provides a great starting point for any kind of design, even the crazy stuff that we do with evolving robots. In fact, implicitly, the code got us started in Chapter 2, and thanks to it, we now understand more about the game of life, evolution, and how we might go about simulating it. In this chapter we’ll stick to the code and go hunting for an understanding of something more elusive—the first vertebrate. Understanding those first fish-like vertebrates—what they looked like and how they behaved over five hundred million years ago—will help us design and engineer the robotic agents that become the players in our simulation of the game of life.

Designing evolving robots of any kind has a number of important steps. The first and the most important is not something in the engineers’ code, although I think perhaps it ought to be: naming.

I feel it’s my responsibility to point out that if you neglect this first design stage, if you think it’s too silly to spend time on naming your robot, then you’ll regret it. You’ll find that other people will automatically and impulsively toss out names as they encounter and work with even just the idea of the robot. And one of those names—invariably the one that repulses you the most—will stick.

If you follow the examples of roboticists before you, you’ll take one of three approaches to naming. Approach one, eponymism: name your robot after a famous person, preferably someone in robotics or artificial intelligence who is still living and can pay back the favor some day. Honda Motor Corporation took this approach when they named their bipedal spaceman-type robot “Asimo” after the great but late science fiction genius Isaac Asimov, inventor, among other things, of the Three Laws of Robotics.

Approach two, bionymism: name your robot after the animal that inspired it or the job that it does. Michael Triantafyllou at the Massachusetts Institute of Technology created the famous fish-inspired RoboTuna back in the 1980s. Bionymism, when applied to robots, often involves the creation of a portmanteau, the smushing of two words to make a new one. When smushing for your bionymistic robotic purposes, consider the common prefixes “ro-” and “cy-” along with the suffixes “-bot,” “-tron,” “-borg,” and “-droid.”

Approach three, acronymism: name your robot using an acronym that is a random letter string or, heaven forbid, an actual word related to your robot. The military loves nonword letter strings, like VCUUC, which stands for Vorticity Control Unmanned Undersea Vehicle. VCUUC, spoken as “vee-cuhk,” is the serious, naval stage name of RoboTuna. VCUUC is a kind of AUV, spoken as “eh-you-vee,” which stands for Autonomous Underwater Vehicle.

Now we are ready to tackle our “evolving robots.” We call them Evolvabots. We really went out on a limb and smushed, using the functional variant of the bionymistic approach.

Our Evolvabots need to be autonomous agents operating in an evolutionary world, but that’s not all we need them to be—we need them to address our specific hypothesis, such as the relationship between swimming ability and the evolution of the backbone. In order to create those specific Evolvabots and their world, we need to ask and answer a host of mission-critical questions:

* Which animal will we model and why?

* Which features of the animal will the Evolvabots possess and why?

* Which features of the animal’s world will we model and why?

* What is the selection pressure that we apply and why was it chosen?

* How does the Evolvabot and its world, taken together, represent the animal and its world?

* How will we judge if our Evolvabots are a good model of the targeted animal?

These questions are critical because their answers drive years of effort from a group of people, the research team. If you haven’t answered these questions carefully and used them to guide your design effort, then later, when you are done running your experiments and want to get your project published in a scientific journal, you may find your team saddled with a paper that is DOA.

These mission-critical questions hark back to the “why robots?” question of Chapter 1. You have to be able to show that your Evolvabots and the processes that are used to evolve them represent, in some way, biological reality. The important word here is “represent.” To represent is not the same as saying that you have to replicate exactly the actual vertebrate and its actual environment (i.e., you don’t have to make a cat to model a cat). Instead, you have to demonstrate that the decisions you made in designing your Evolvabots were not arbitrary. Time, equipment, money, and expertise will always constrain those decisions. But the knowledge of your target system must also guide those decisions: you have to show that features of your Evolvabot relate to—represent—features of your target.

Representation is a general process that occurs in many different ways. For example, in biology representation occurs between the information to build the animal and the physical manifestation of the animal itself: the genome of an animal represents its phenotype (Figure 3.1). In modeling with Evolvabots, representation occurs between the robot and its biological target: the robot is a representation of the target.

How does one thing represent another thing? This is a fundamental issue in cognitive science, artificial intelligence, and philosophy of the mind.[12] The most straightforward case that I can think of is when one thing is an instance of a category of things. A Tadro is an instance of an Evolvabot. As an instance, a specific Tadro represents the general category of Evolvabots. You can also flip this on its head: the category of Evolvabots represents, by definition, all instances of any kind of Evolvabot, including all the Tadros.

We encounter this kind of categorical representation all the time when we learn. Someone shows us an example of something new to us. Hey, look at this thing called a chocolate donut! Look at it. Smell it. Feel it. Taste it. This particular donut, the donut master tells you, is one example of a whole category of foodstuffs called donuts. The category, “donuts,” includes other chocolate donuts that look and taste very much like this one, chocolate donuts that don’t look like this one (they have sprinkles) but taste similar, and donuts that neither look like this one nor taste like it either. As you can see and taste, the representation of all donuts by a chocolate donut is created in the human mind by linking the instance at hand (or is it at mouth?) with other imagined instances. The “linking” here refers to features of the donut—looks, smell, feel, and taste—that we can morph in our minds in order to create a new imaginary instance of a donut.

So if our minds do the linking between one thing and another, and this linking is the process by which we create representations, then our mind is doing the representing. Other minds, other engines of representation, are thus the judges of our efforts at representing. If no one else thinks that we’ve done a good job building an Evolvabot to represent a vertebrate, then we haven’t. More on judgment later.

To build scientifically useful Evolvabots, we need to use our minds and the minds of others to figure out, explicitly and objectively, how the Evolvabot represents an animal. Bloody obvious, eh? Maybe so. But keep in mind that we (meaning me and other nerds) often get so excited when we start to do cool stuff like build robots that we just start putting parts together, whatever’s at hand, in order to quickly build something that works. Although this can be an exciting way to start designing robots, the implicit intuitions that guide this kind of spontaneous creation can often miss the mark in terms of clearly representing the thing that we meant to represent. So before you get started: stop! Answer the six design questions![13]

We want to model the mother of all vertebrates—literally. We want the ancestor from whom all other vertebrates evolved. The only problem with this desire is that we don’t know exactly who that ancestor was or what exactly she looked like. The origins of vertebrates are shrouded in mystery (soundtrack: key low Celtic whistle). What to do?

This mystery drives crazy anyone who cares about deep evolutionary history: who were the first vertebrates, anyway? This simple question turns out to be controversial because the information that we use keeps being updated and revised. Damn those meddling scientists! We find new fossils, analyze new genes, and come up with different computer methods to reconstruct evolutionary relationships among species.[14]

Some of the newest information about vertebrate evolution when we were trying to answer this question had come from the laboratory of Frédéric Delsuc at Montreal University.[15] Delsuc and his colleagues examined 146 genes in forty species of living animals, using the similarity among the genes to cluster species into related groups. The group that clustered closest to the vertebrates was the tunicates and not a group called lancelets. This result was a surprise because adult lancelets look and behave like zippy little fish whereas some adult tunicates go by the name “sea squirt” because they are little grape-like balls attached to rocks at low tide who squirt water at finger-poking people (Figure 3.2).[16] In technical terms, any two species or groups of species that are more closely related to each other than they are to any other species or group of species are called “sister taxa,” where the term “taxa” is the plural form of “taxon,” which means any group of related organisms.

How can it be that a bag of water is the sister taxon to vertebrates? Even though adult tunicates are ugly bags of mostly water,[17] the pre-adult larvae of tunicates look like zippy little fish, sporting a sensor-filled front end and a long tail flexing with undulatory waves that push water backward and, by Newton’s third law, the larva forward. This resemblance of the larval form of tunicates to the adult form of fish has long been recognized. Walter Garstang, working in the first half of the twentieth century, proposed the then-radical idea that because the larvae of some species were more similar to the adults of others, we needed to consider the possibility that evolution might have worked by chopping off the adult stage to create new adult forms. In fact, back in 1928 Garstang proposed the idea that the larvae of ancient tunicates might have provided the basic vertebrate body plan—seventy-eight years before Delsuc’s molecular data suggested the same thing.[18]

Warning: sloppy-thinking watch in effect. Evolutionary intuitions may cloud inferential cognitive processes. Keep in mind that when we look at the living tadpole larvae of sea squirts, we aren’t looking at the ancestor of vertebrates, even if tunicates and vertebrates are sister taxa. Living tunicates have had at least 530 million years of evolution on their own, after they split with vertebrates, to create their own family lineage. Thinking that every living species is an ancestor of another living species is a common fallacy, what I’ll call the fallacy of the living ancestor.

Secondary warning: the fallacy of the living ancestor has a conceptual sibling, the fallacy of the fossil ancestor. “Paleontology is the search for ancestors,” allegedly claimed George Gaylord Simpson, one of the greatest paleontologists and a cofounder of the modern synthesis of evolutionary biology. But he was wrong. (Did I just say that? Forgive me, G. G.!) The chance that you’ll actually find any ancestor is very small for two reasons. First, the fossil record is incomplete. The accidental mudslides and burials that turn animals into fossils capture only a small percentage of the animals that are alive. In addition, we have good evidence that new species are usually created from small, breakaway bands of a main population; the number of these founding individuals just aren’t numerous enough to be reliably fossilized and then found millions of years later. Probability of finding the actual ancestor of any living species: approaching zero.

In the end, with all of this wonderful confusion surrounding the identity of the mother of all vertebrates, the specific vertebrate we chose as our target for designing Evolvabots was the tadpole larva of living tunicates. What finally sold us was that Matt McHenry, working at the time as a PhD student in the laboratory of Professor Mimi Koehl, University of California, Berkeley, had figured out the neural circuitry involved in the swimming behavior of tunicate larvae.

Using careful experiments in which he altered the direction that light hit swimming larvae in a tank, McHenry showed that the tadpole larvae were using a very simple mechanism to orient toward and then away from the light, in a behavior known as positive and negative phototaxis, respectively (see Figure 3.2). The mechanism is called helical klinotaxis (HK) and refers to the fact that many small swimming animals move in helical pathways, as if along the threads of a screw, as they move toward or away from something in their environment, like light or the chemical plume of a food source. Although spiraling along in a helix may seem inefficient (why not just swim in a straight line?), Hugh Crenshaw, working before McHenry in the laboratory of Steven Vogel at Duke University, had shown that it was actually efficient in terms of control. To control your directions in three dimensions, all you, as a small swimmer, need to do is change two variables: your translational (straight) and rotational velocity.

When I saw McHenry present his work on tunicate tadpole larvae at a scientific meeting, I remember nearly shouting out, “Let’s build a tadpole robot!” His mathematical model, which he had worked on with Jim Strother, gave us what we needed to know about the likely neural control of HK in a chordate. Fortunately for us, McHenry and Strother agreed to help transfer their knowledge of HK and tadpole larvae into a robotic form.

“Keep it simple, stupid.” This quote, allegedly from pioneering aerospace engineer Kelly Johnson, is known throughout the design community as the KISS principle. The KISS principle is important at this stage in the design because in the heat of jubilant complexification, it helps keep your feet on the ground and your eyes on the target. KISS forces you to rephrase design question 2: what is the least we can do to fulfill our overall design goal?

For scientists, doing the simplest thing first has a very important philosophical basis: adding complexity to your model requires a combinatorial explosion of decisions, and each decision has an impact on the outcome of your design. And even more importantly, connecting back from your results to the causal elements in your design requires that you understand every element in your design and how every element interacts with all of the other elements. The simpler your design—the more KISS inspired that it is—the better your chances of understanding what the heck you’ve created. This KISS-first approach is one of the guiding principles at Vassar College when we work with students in the Interdisciplinary Robotics Research Laboratory. Undergraduates Adam Lammert and Joseph Schumacher, both cognitive science majors at Vassar, applied the KISS principle when they built the robots that we talk about in this chapter.

Embracing the KISS principle, we decided to keep our wish list of features short: (1) behavior: helical klinotaxis; (2) sensor: single eyespot; (3) brain: simple processor that turns the light intensity signal from the eyespot into a turning command for the motor; (4) motor: one, used for both driving and turning the tail; (5) body: a simple round bowl; (6) tail: a notochord with a flared caudal fin. Although this list may seem like a long one, keep in mind that some features are as simple as you can get (e.g., single eyespot, bowl for a body) and some features are simply missing (e.g., muscles, other sensors, a mouth).

The design of the first Tadro started in 2003 with Adam, a Vassar undergraduate and cognitive science major. He was interested in robotics, and we talked about taking McHenry and Strother’s neuromuscular model for tunicate tadpole larvae and making a simple robot, relying on an insight from Chuck Pell.

Chuck had been working on three-dimensional helical klinotaxis with Hugh, of Duke University. Hugh, a biomechanist trained for his PhD by Steven Vogel, had made a true breakthrough by figuring out how to measure and mathematically describe the 3-D motion of the single-celled organisms that swim almost exclusively using HK. Later Hugh, as a faculty member at Duke, and Chuck, working with Professor Steve Wainwright out of Duke’s BioDesign Studio, created the first autonomous robot that used an HK algorithm, a small torpedo-shaped vessel. Capable of navigation with a only a single propeller for control and orientation, the robot would become known as Microhunter. For our purposes, Chuck’s insight was that the three-dimensional HK used by the tunicate tadpoles would also work in two dimensions. This meant that we could stay on the surface of the water, avoiding the engineering complexities of moving in three dimensions while keeping our electronics dry. KISS in action.

For all of this work, Tadro1 was not yet an Evolvabot.[19] The transformation from biorobot to Evolvabot was driven by the interests of another cognitive science major at Vassar, Joe Schumacher, who helped endow Tadro1 with a backbone so that we could begin studying backbone biomechanics using robots.

Rob Root, Chun Wai Liew, Tom Koob, and I had tried to fund our research on the biomechanics of backbones straight-up, with no robots. We had seen two of our proposals to the National Science Foundation (NSF) rejected. The third time was a charm, and the change that made the difference—adding robots—came about almost by accident. In the fall of 2003 I worked on a review panel at NSF down in Arlington, Virginia—it was the same panel that had twice rejected our grant. The real power in the room was the program officer, who had the final say about which projects were funded. When a chorus of positivity would arise from the panelists, she’d put down her pen and start asking tough questions. When a break came in the day’s work, I got a chance to ask her a question. Having previously reminded her of my two failed proposals, I went over and, without any preamble, blurted, “What about robots?” She looked up, paused without giving me eye contact, then, looking at me directly, said, “Robots would be good.” That was all I needed to know.

Back at Vassar, Joe and I started scheming. Tadro1 didn’t have a biomimetic notochord yet, but Adam, Tadro1’s departing creator, helped Joe create Tadro2 by giving Tadro1 two important upgrades: (1) a computerized brain (replacing Tadro1’s analog circuitry) and (2) a genetic algorithm that coded for the size of a flapping tail made out of duct tape. Joined in the summer of 2004 by Nick Livingston, Joe quickly created a water world, programmed the digital brain, and set out to design a biomimetic notochord and vertebral column. For the electronics and the new Tadro body, he enlisted help from John Vanderlee and Carl Bertsche, Vassar’s electronics technician and machinist.

By the time our NSF funding started in January of 2005, Joe was already replacing Tadro2’s duct-tape tail with one that had a simple rod serving as a notochord. He used ten-centimeter-long cylindrical erasers as the notochord, plastic clamps as vertebrae, and then put a flared caudal fin on the end. Together we designed the genetic algorithm that would code for the evolvable traits: the length of the axial skeleton and the number of vertebrae. Nick made an important innovation: he wrote a program that allowed Tadro2 to make its tail adjustments for maneuvering using the same motor that flapped the tail. This architecture further reduced the complexity of Tadro2 and made it much more reliable. With our incoming students, whom we called “Fish Fellows,” we quickly realized that notochords made of erasers weren’t making sense because we couldn’t change the stiffness of the erasers’ material. The solution—building the notochord out of a biomaterial whose stiffness we could vary—would come from Tom Koob, as we’ll see on the next pages. With the change to the brain and the tail of Tadro2, we realized that we really had a new critter: Tadro3 (Figure 3.3).

We had three reasons for thinking that Tadro3 was the Evolvabot we were looking for: (1) the brain would make it autonomous, able to behave on its own without a human “in the loop,” without a remote operator acting as the eyes and brains of the operation; (2) the light-seeking behavior would emulate the phototaxis of the tunicate tadpole larva; and (3) the body would also emulate that of the tunicate larva, possessing a propulsive tail with a biomimetic notochord, a backbone whose properties we could vary by degrees, code with an artificial genome, and cause to evolve under the right ecological situation.

To understand more about the backbone as a feature we targeted, we need to dig a bit deeper into some of the assumptions we’ve been making about the evolution of notochords and vertebral columns of chordates. As I said in Chapter 1, a species called Haikouichthys ercaicunensis, a small, sporty little fish that lived 530 million years ago (Figure 3.2), appears to have been conducting its own evolutionary experiment on turning a notochord into a vertebral column. Widely spaced bits of cartilage or bone can be seen along its notochord.[21] The proto-vertebrae, as some authors have dubbed them, are too far apart to resemble the tightly packed vertebrae that we see in most other fossils or living species that have vertebrae. However, for all of their differences, the proto-vertebrae of Haikouichthys allow us to infer three important things about the evolution of vertebrae:

* The earliest vertebrate fossils had a backbone that was primarily a notochord, supporting the contention that the notochord is the an cestral state of the vertebrate axial skeleton (no one is surprised by this, by the way, because evolutionary trees have long inferred this pattern, as we’ll see in a minute).

* Vertebrae, even though they appear early in vertebrate evolution, take millions of years to evolve into what we now recognize as a vertebral column. Philippe Janvier, a paleontologist specializing in the earliest fishes, estimates the origin of an internal skeleton of calcified cartilage or bone at about 443 million years ago, about 90 million years after Haikouichthys’ experiments.

* Because the backbone of Haikouichthys does not have the large vertebrae and thin intervertebral joints that we see in living fishes, but just the opposite, we need to be careful to recognize that the two states of the axial skeleton, notochords and vertebral columns, really demarcate the ends of a spectrum of possible axial skeletons. With that in mind, we’d expect to see throughout living and extinct vertebrates variations in the size, shape, and number of vertebrae and intervertebral joints.

Phylogenetic analysis gives us another clue about the polarity of the states, or spectrum of states, of the axial skeleton. The notochord, without any signs of vertebrae, is possessed by both tunicates and lancelets (see Figure 3.2). If, as Delsuc’s tree showed, tunicates are the sister group to vertebrates and lancelets are the sister group to tunicates + vertebrates, then the simplest, most parsimonious explanation is that notochords evolved in the common ancestor of all three groups, well before the vertebrates split off and began to evolve the vertebrae that we think we see in Haikouichthys.

Additional evidence for the notochord evolving first is that it also appears first in the development of living fish, prior to the formation of vertebrae; vertebrae are then built in and around the notochord.[22] Although being first in development isn’t, by itself, evidence for evolutionary polarity, the notochord is a central structure in early embryo development, one that is necessary for the formation of the nervous system and the growth of the embryo. Every vertebrate embryo grows a notochord first and then, if they grow one at all, a vertebral column. This invariant pattern of the notochord guiding the embryonic development of vertebrates and their vertebrae is consistent with the hypothesis that notochords evolved before vertebral columns.

In development and evolution the axial skeleton functions to stiffen the body. As we talked about in Chapter 1, stiffness is the mechanical property that dictates how much a structure changes shape—lengthens, shortens, twists, or bends—in response to having forces applied to it. Put a rubber band on your two index fingers and apply a tensile force to it by increasing the distance between your fingers. The rubber band, at least at first, lengthens easily. Now do the same thing with a shoelace, the ends of which you hold between index finger and thumb. The shoelace does not lengthen much, even if you apply as much force as you can. In engineering terms the shoelace is “stiffer in tension” than the rubber band.

Bending or flexural stiffness of the notochord can be increased by adding vertebrae.[23] Working with Tom Koob and Lena Koob-Emunds at the Mount Desert Island Biological Laboratory in Salsbury Cove, Maine, we analyzed hagfish, a group of eel-like fish that never evolved jaws and retain, as adults, a fifty-centimeter long notochord. After a hagfish died, we removed and bent its notochord to measure the notochord’s flexural stiffness. We then threaded onto the notochord, like pearls on a string, a series of rigid plastic rings that snugly fit the notochord. Sometimes we added just a few rings, widely spaced like the vertebrae of Haikouichthys, and sometimes we added more, leaving less space for bending. The result? More vertebrae created an axial skeleton with increased flexural stiffness. With this in mind, in Tadro3 we allowed bending stiffness itself, rather than number of vertebrae, to be the character that was genetically coded to evolve.

This may seem bass-ackwards, I admit. Why not just build an artificial notochord and then add plastic rings, as the game of life demands, to model the number of vertebrae? Our rationale for evolving bending stiffness as a proxy for vertebrae went as follows. If you evolve only whole vertebrae—they are either present or absent—then your resolution is limited to those stepwise changes. You can’t see what “half” a vertebrae looks like. But do half-vertebrae evolve? Yes, sometimes. In the fossil record for the group of fleshy-finned fishes that were the outgroups to the first land-living tetrapods, we see partial ring vertebrae, little crescents of bone that cup the bottoms of the notochord.[24] At least in this group, it looks like vertebrae form from different pre-existing centers of bone formation, in this case the ribs. Sindre Grotmol and his colleagues at the University of Bergen, Norway, have shown a similar process in the development of living fishes.

Here’s the rub if, like us, you are interested in evolutionary biorobotics: how do you make partial vertebrae? We’ve tried, trust me. My students can tell you many a tale of working on making tails with partial vertebrae and vertebrae of various sizes and shapes. But in almost every case the vertebral column would tear (fracture, strictly speaking) at the interface between the bit of vertebra and the notochord.

Our solution, at the time, was to forget about the vertebrae and make a continuous structure, a biomimetic notochord whose material stiffness we could alter and, in so doing, alter the notochord’s flexural stiffness.[25] We also realized that if you changed the length of a structure, you alter its structural stiffness: for a given flexural stiffness, a longer structure deflects more than a shorter one. Our biomimetic notochord was, in the lingo of material scientists, a hydrogel made, as I said, of collagen, thanks to Tom Koob.

When you take the powdered gelatin and add it to heated water, it dissolves nicely if you stir the pot. As you cool the mixture, the gelatin, now evenly spread throughout the forming solid, makes some chemical bonds between the scattered molecules. Pour the cooling liquid into a mold of some kind and pop it into the refrigerator. In the cold the motion of the collagen fragments slows, allowing even more bonds to form. Presto! You’ve created a solid from a liquid: a molded hydrogel!

For biomimetic hydrogels, we poured the hot gelatin and water mixture into an array of molds that made cylindrical rods about 10 centimeters in length and about 0.5 centimeters in diameter. Once the gelatin had set in the fridge, we pulled the rods out and then did something you wouldn’t do with your dessert: chemically embalm them. Embalming, or what a biochemist would call fixation or, in this case, cross-linking, keeps tissues from degrading and, gulp, spoiling.

For our hydrogels, the mortuarial embalming agent we use is called glutaraldehyde, and it does two things. First, glutaraldehyde allows us to let the biomimetic notochords warm up to room temperature without melting—it keeps the hydrogels solid. Second, glutaraldehyde allows us to control the stiffness of the hydrogel: the more time that the hydrogel spends in the glutaraldehyde solution, the stiffer it becomes as more chemical crosslinks form between collagen molecules. Here, finally, was our method for getting any intermediate flexural stiffness that a genetic call for a partial vertebra might require.

The world or arena that you design for your robots is as important as the robots themselves. Thus, we have in hand one of the design principles for embodied robots expounded by Rohlf Pfeifer and Cristian Scheier: build a robot for a specific ecological niche.[26] In other words, you have to build the agent with a particular world in mind. This is obvious when we think about the difference between a fish-like robot and a dog-like robot: water versus land. But what about a fish-like robot swimming in the nooks and crannies of a coral reef and one swimming in the open ocean? If we use fish as our guides, these robots ought to be very different kinds of agents, the first skilled at precise maneuvering and station-holding and the second skilled at cruising and perhaps navigation.

The world also has other players. For evolutionary biologists, the other players are called “biotic factors” and everything else is “abiotic factors.” For an individual robot, biotic factors are all the other robots and animals with which it might interact. Abiotic factors include the physical and chemical situation in which it’s placed. Together, biotic and abiotic factors make up the ecological niche, here what I’m calling the stage, the modeled world, or the selection environment.

We wanted a world that, like Tadro itself, was a simplification of our best-guess of ancient reality. The ancient world for the first vertebrates was, as far as we can tell, oceanic, near the shore, and full of biotic factors like giant arthropods, trilobites, anemones, and worms with legs.[27] Obviously they all had to eat, and some of them likely competed with the first vertebrates, jostling for position at the donut store, figuratively speaking. KISS demanded we leave most of that cast of characters out.

The simple world we built was a water world, a walled tank 2.5 meters across with a single sun, limited time, and three Tadro3s (Figure 3.4). The sun was a hundred-watt flood light suspended above the surface of the water. Time was limited to three minutes for each trial. Each Tadro3 competed in six different trials, with three robots in each trial. To account for the fact that each Tadro3, even though built to be identical in every way but for their variable tails, may vary in performance, we swapped the biomimetic tails among the three Tadro3s and made sure that all possible combinations of tails and robots were tested. This swapping allowed us to make sure that no particular tail lost the game because it was always stuck with a sluggish robot.

In the water world, sun represents food. The first food, for almost all life, is the glucose made by plants that harvest the energy from light. In the ocean most critters follow the light to find the food. That’s why most of the sea creatures live in the upper reaches, the shallower depths, because that’s where the light is. And where there is light, there are algae and diatoms, the “primary producers” making the stuff, their fronds and bodies, that feeds the mobile, self-propelled critters like fish.

The world sets the stage for what matters in the game of life—surviving and reproducing better than other agents in your population: outwit, outplay, out-reproduce.[28]

Honestly, no one has any idea what the selection pressures were that drove the evolution of the early vertebrates. As we discussed in Chapter 2, it’s difficult enough to understand what is going on when you have live animals right in front of you. Even then you have to know the evolutionary fitness of different individuals and link their many phenotypic differences to differences in how the individuals behave and interact with their world over their entire lifetimes. This is a daunting task under the best of conditions with live animals. For extinct animals, all we can do is make what seem to be reasonable guesses, or what Brandon called “how-possibly” explanations.

So how do we make reasonable guesses—hypotheses—about selection pressures that drove early vertebrate evolution? We use our understanding of how specific traits function in living animals, assume the same thing was happening long ago, and then conjecture that variations in that trait had functional consequences for the individuals possessing those variations that could help or hinder them in the game of life. We then imagine the conditions of the world, with its physical characteristics and other autonomous organic agents, under which that help or hindrance would matter the most for survival and reproduction. That particular condition of the world is our “selection environment,” to use Brandon’s term from Chapter 2, and the “selection pressure” that we are talking about here is the particular type of interaction between the selection environment and the individuals that most affect survival and reproduction. For example, many people think that avoiding predators is the selection pressure that drives the evolution of fish coloration, body shape, and swimming performance.[29]

We focus our educated guessing on the axial skeleton and the evolutionary change from notochords to vertebrae. As explained above, Tadro3s are built to evolve the structural stiffness of their tails as a proxy for the presence of vertebrae. With tail stiffness in mind, we first think about the mechanical function of the tail: what does it do and how does it do it?

The primary mechanical function of the chordate tail, the section of the body behind the gut and including the terminal caudal fin, appears to be propulsion. No surprise. Tunicate tadpoles, sharks, and bony fish all undulate their tails to create thrust. Undulation is making waves, traveling waves of body flexion that start near the head and move toward the caudal fin. By altering the shape and speed of those undulatory waves, fish alter their swimming speed, turn, and brake.

The structural stiffness of the tail controls, in part, the shape and speed of the undulatory waves. If you’ve ever tuned the string of a guitar or violin, you know that if you tighten the string, it will vibrate faster when plucked, creating a higher pitch. By tightening the string, you’ve stiffened it, and it is a well-known principle of engineering that an elastic structure like a string or steel bridge will tend to vibrate at a particular frequency, called the natural frequency, that is determined by the structure’s stiffness, mass, and ability to dissipate energy. Thus, stiffer tails ought to vibrate faster than flexible tails.

So how might have the stiffness of tails evolved? What might have been the selection pressure? Here we connect the dots. If increased tail stiffness makes or allows the undulatory waves to travel faster, then increased tail stiffness increases the speed at which fish can swim. If increased swimming speed helps fish find food, then increased tail stiffness increases the amount of food that a fish can eat. And finally, if finding and eating more food increases a fish’s chances of survival and its reproductive success, then increased tail stiffness was selected to improve the ability to forage and feed.

We put the feeding selection pressure into action by coming up with a “fitness function,” which is a fancy name for a numerical formula for judging how well each individual does relative to the other individuals in the population. Because foraging and feeding involve detecting the presence of the food, traveling to the food source, and then staying and eating, we reasoned that a number of behaviors should be rewarded at the same time. First, the ability to detect the food could be measured as the time it took a Tadro3 to reach the source; more points are given for shorter times. Second, the ability to get to the food quickly could be measured by the average speed at which the Tadro3 traveled; more points are given for higher speeds. Third, staying and eating could be measured by the average distance of each Tadro3 from the food source; more points are given for a shorter distance. Fourth and finally, the sloppiness of swimmers that waste energy, and thus food, could be measured by their average amount of body wobble; more points are given for smaller amounts of wobble.

Our earliest Chordate ancestors were probably little fish-like swimmers, with notochords for an axial skeleton in their body and tail and with at least a single eyespot. They could probably detect and navigate relative to light gradients in the sea. This simplified and hypothetical ancestor is inferred from what we know about living chordates and Cambrian vertebrates (see Figure 3.2), development of living vertebrates, and the evolutionary relationships among chordates that we reconstruct using phenotypes and genomic data.

Using this information, we chose as our specific biological target the tadpole larva of living tunicates. Even though no living species is the ancestor to another living species, we still felt confident enough in the similarities between the behavior of tunicate tadpole larvae and the ancient, extinct chordates to use the larva as our model for designing Tadro3. The Tadro3 uses the same neural algorithm that we think the tadpoles use, both have a single eyespot, and both continuously undulate a tail with a notochord. To flap the tail, the Tadro3 uses a single motor instead of a series of muscle cells distributed along the tail. Both Tadro3s and tadpoles turn by adjusting the angle at which the tail meets the body.

We did simplify things, however, by having Tadro3 swim just on the surface of the water rather than underneath and by building the Tadro3 on a scale easy for us humans to manipulate. Tadro3 is about twenty-five centimeters from head to tail; a tadpole larva, however, is just a few millimeters long. We also simplified, as I described above, the physical environment: Tadro3 lives in a small circular pool rather than the ocean. Feeding behavior is also simplified, as Tadro3 merely needs to navigate up the light gradient cast by the single light over the pool. Likewise, although tadpoles encounter many other animals in a twelve-to twenty-four-hour dispersal period, during which most tadpoles die, our robots encountered no other agents (just other Tadro3s), “dispersed” for just three minutes, and could not “die.”

We think these simplifications were justified, but we must always be on guard against simplifications that are not. If we ever cannot justify them to other people, then we have failed in our primary goal: to test an evolutionary hypothesis of animals using robots as model simulations.

Last comes the justification. Barbara Webb delineates seven dimensions that can be used to help describe and characterize biorobotic models: (1) biological relevance, (2) match between the behavior of the biological target and the robot model, (3) accuracy of the model in using the same functional mechanisms as the target, (4) how concrete the model is in terms of mimicking features of the target, (5) the level in the target’s structural hierarchy at which the model focuses, (6) the specificity of the model in terms of the number of elements targeted, and (7) the substrate from which the model is built, either digital or physical.[30]

For Webb, the key dimensions for biorobotic models are biological relevance and substrate. You’ve got to have a robotic system that allows you to test a hypothesis about your target or it isn’t relevant; the test of the hypothesis can be if the robotic system works as expected or, in the case of Evolutionary Biorobotics, if the evolutionary trajectory of the system is as expected. Moreover, she argues that the substrate ought to be physical rather than digital, for all the reasons outlined in Chapter 1.

For Evolutionary Biorobotics models, we include behavioral match and functional accuracy. For example, we want an individual Tadro3 to behave like a tunicate tadpole in terms of (1) using a tail that generates thrust by undulating, (2) navigating up a light gradient, and (3) being part of an interacting and evolving population. That’s the behavioral match, with a three-layer nested hierarchy of the behavior of the organ, the individual, and the population.

We also want Tadro3 to use the same functional mechanisms that tunicate tadpoles do, such as the same kind of undulatory wiggle of the tail and the same neural wiring and sensorimotor loops that we’ve understood and engineered above. We also want the evolutionary mechanisms that we talked about in Chapter 2—selection, mutation, random mating, and genetic drift—to be what drives the evolution of the population of Tadro3s. That’s the accuracy of the functional mechanisms, with a three-layer hierarchy of propulsive mechanism, sensory-neural-motor mechanism, and evolutionary mechanisms.

In sum, relevance, substrate, match, and accuracy are our primary goals for designing, engineering, and running Tadro3s as a model simulation of the earliest fish-like vertebrates. We’ll judge how close we come to meeting those goals when we look at the Tadro3s playing the game of life in their water world in Chapter 4.

Throughout this design process we’ve employed the KISS principle and simplified our Evolvabot and its world whenever we could. We’ve even argued in the scientific literature that the original Tadro1, which was built to swim and behave like a tunicate tadpole, is the simplest possible autonomous navigator because it possesses a single sensor and a single motor control output (tail offset to turn). Tadro3 has that same basic hardware and neural architecture but has the biomimetic tail coded to evolve. Simple, simple.

But as we’ve seen, even with a simple robot like Tadro3, you need to understand a ton of stuff about the animal target. Think: the engineers’ code. Think: specifications. We had to figure out how evolution works (Chapter 2), make an educated guess about the hypothetical chordate ancestor of vertebrates, find a reasonable living proxy for that ancestor (tunicate tadpole larva), understand the swimming behavior of the tadpole, infer the neural control system of the tadpole, measure the mechanical function of a tail with a notochord, and divine a likely selection pressure on early vertebrates. Phew (pant, pant, out of breath …)! This is the kind of understanding that allows us to follow the secret engineers’ code and build a population of evolving robots.

But were we successful?