Chapter 7 EVOLUTIONARY TREKKERS

REMEMBER: NO MATTER WHERE YOU GO, THERE YOU ARE.[135] That’s one of the most frustrating things about the otherwise wonderful Evolvabots. They visit just a few places in a vast morphospace of evolutionary possibilities (Figure 7.1). As we’ve seen, a population of Tadro3s or PreyRos take but a single path out of a huge number of possible trajectories. Which path they take and how quickly they travel it depends on those three all-encompassing classes of evolutionary mechanism that we introduced in Chapter 2: selection, randomness, and history. Those mechanisms determine what we observe at any time and place in the population’s evolutionary journey: a generation of individuals playing the game of life. Although the game of life is fascinating to watch, sometimes the “there” you are observing is not the “there” where you want to be.

We were curious about the travels of our two different populations of PreyRos. Why had the populations never evolved more than an average of 5.7 vertebrae? Our guess, based in part on the behavior of many populations of Digi-Tad3s, was that the PreyRos had reached an equilibrium or, possibly, an optimum number of vertebrae. But what, if anything, makes 5.7 optimum? Why not 8 or 10 vertebrae? What if PreyRos evolved 12 vertebrae: would the world end?

Why not? What if? Questions about the counterfactual dogs anyone interested in just about anything with a history, including the process of evolution. Whether we are studying the evolution of life-forms,[136] engineering solutions, or artificial intelligence, our curiosity about what didn’t happen[137] or what might have happened motivates much of what we do. A central curiosity-driven question regarding any evolutionary system is: Why did some forms evolve while others haven’t?[138] From this question springs a host of related queries:

* Why didn’t the population evolve along a different path?

* What if the population were to evolve again, from the same starting point? Would it evolve along the same path?[139]

* Why haven’t all imaginable forms evolved?[140]

How to proceed? Curiosity leads us to a classic forbidden-fruit conundrum, presenting us with at least three possible actions. First choice: don’t bite the apple. We could simply let our findings be, content that PreyRo, in two different runs, increased the number of vertebrae under selection for enhanced feeding behavior and predator avoidance. We’ve learned much. Move on to the next study.

FIGURE 7.1. (facing page) There you are. Evolution samples only a small portion of the morphospace, the area of all possible trait combinations. The entire second evolutionary run of PreyRo is shown here, with the points representing the average values of the population for the pair of traits that we found to be evolving in an uncorrelated, mosaic pattern, the span of the caudal fin, b, and the number of vertebrae, N. The thin vertical and horizontal gray lines intersecting each point represent the ranges of each trait for a given generation. The rectangles that those range lines touch represent the population’s total phenotypic footprint, the area that the population has sampled.

Top: The population of PreyRos evolves over a very small region of morphospace in ten generations. The rectangle encloses all individual PreyRos ever evolved.

Bottom: A close-up of the evolution of the population in b-Nmorphospace shows a shrinking phenotypic footprint as selection removes variation, here represented by the ranges, from the population. Loss of variation in traits is a sign of selection.

But no, for our curiosity remains. Remember this inscription that taunts Digory in C. S. Lewis’s The Magician’s Nephew:

Make your choice, adventurous stranger

Strike the bell and bide the danger

Or wonder, till it drives you mad

What would have followed if you had.

Madness, surely, is not something that we desire. If wonder we must, then adventure we take. Strike the bell! Here’s an adventurous choice: employ “directed evolution” and then try to force the system to evolve toward a prescribed goal or along a particular path. Although this approach to engineering might seem new—it has garnered much attention recently because of its success in the synthesis of novel proteins—you could argue that directed evolution is what breeders have been doing for millennia to domesticate the likes of cattle, rice, and corn. It works, but only to a certain point and in certain cases. No matter how much I want flying cows or talking corn, it ain’t gonna happen given the limits genetics and physics impose. Just because you have a target doesn’t mean that you’ll be able to hit it or, for that matter, that it’s even hittable.

We have another choice that helps satisfy our curiosity. We can hit the target by knowing exactly where we want the robot to be in the evolutionary landscape. Unlike chemists trying to engineer enzymes of unknown structure with a targeted function, we, as mentioned in Chapter 3, take the reverse-engineering course, asking about the function of a given structure (rather than seeking a particular function). For example, we could just build a PreyRo with ten vertebrae and see how it functions. This is like dropping a paratrooper behind enemy lines. We can put individual robots at specific points on the evolutionary map and ask them to report back. These robots serve as probes, allowing us to go where no one has gone before. In homage to Star Trek, let’s use eponymism (see Chapter 3) and call these agents Evolutionary Trekkers, or ETs for short.

As we’ve defined them, ETs are not Evolvabots. They don’t evolve. Sorry. Just like the rules for Starfleet officers, the prime directive (General Order 1) applies, and ETs don’t participate in or alter the evolutionary trajectories of the prewarp life-forms that they study or encounter. Even with this limitation, they are a powerful complement to Evolvabots. ETs are the crew members testing the functional waters, so to speak, of different bodies and brains without the historical constraint that Evolvabots drag with them. Keep in mind that ETs cannot test different selective conditions that might drive a system to evolve from one place to another on the adaptive landscape. For this reason, ETs are a separate class of robots, testing hypotheses about the outcomes of evolution rather than the process itself.

MAPPING THE EVOLUTIONARY LANDSCAPE

Before we explore ETs I need to clarify some terms that I’ve tossed around willy-nilly. The term “evolutionary landscape” is also known as a “fitness landscape” or, in the original concept Sewall Wright created, an “adaptive landscape”: I use these terms interchangeably. The metaphor of a landscape gives us a way to conceptualize the hills of fitness heights and the valleys of fitness despair. Fitness is represented by contours lines on a two-dimensional map (Figure 7.2).

You’ve probably noticed that I didn’t show you all three of PreyRo’s evolving traits in Figure 7.2. That’s for the simple reason that maps with more than two traits are difficult to make and interpret visually. For example, for three traits in a three-dimensional surface, you need to be able to rotate the surface so that you can view the selection vectors from different angles. On top of that, literally, you then have to somehow visually code the fitness gradients. Folks more talented than me with visual graphics can manage. But we all fall down when it comes to illustrating maps with more than three dimensions.

Even though adaptive landscapes, as a visual tool, have severe limits, at least for the two characters shown here, they can be very instructive (Figure 7.2). To wit: we had previously decided that the number of vertebrae, N, and the span of the caudal fin, b, evolved independently with respect to each other, a pattern of character interaction that we called mosaic evolution (see Chapter 6). Because this N-b character pair is mosaic and thus uncorrelated, we can’t tell what will happen to one by simply looking at the evolutionary changes of the other. Instead, we make sense of the combined evolutionary history that they share, even if it is uncorrelated, by looking at the population’s evolutionary trajectory and the adaptive peaks in the N-b landscape.

What we see on the map is wild (Figure 7.2). From a bird’s-eye view we see multiple adaptive peaks, a chain of misty fitness mountains running from north to south. In between the peaks we find what looks like valleys and then a whole bunch of white space labeled “terra incognita.” All of the white space on the map, and probably some of the gray hilly parts too, is unknown territory.

Here we return to our main problem: adaptive topography can only be mapped if the population has been in that area—to that “there”—and played the game of life. Only when each individual gets a fitness score can we then calculate the population’s selection vector. For PreyRos my colleagues and I determined where the vector pointed (say, to having five vertebrae and a caudal fin span of 22.25 millimeters).[141] The selection vector represents the direction and magnitude of evolutionary change that selection alone would cause.

FIGURE 7.2. (facing page) Mapping the adaptive landscape. Top: We can use two of PreyRo’s evolving traits, span of the caudal fin, b, and number of vertebrae, N, to create the two-dimensional “morphospace” of the evolutionary map. The points represent the population’s average values for the traits at each generation, numbered 1 to 6 for the first evolutionary run. The black arrows represent the actual evolutionary change of the population from generation to generation. The gray arrows are the selection vectors. Each selection vector has a direction and a strength, with strength represented by the length of the arrow. The random evolutionary mechanisms (mutation, mating, and genetic drift) cause the difference between the selection vector (gray arrow) and the evolutionary vector (black arrow).

Bottom: Because the selection vectors point toward a local fitness peak, they can be used to map the adaptive landscape. The points here are the same average values of the population from above (generations 1 to 5). The arrows are the same selection vectors. Adaptive peaks and ridges can be of any shape. The shape and placement of the fitness features that I’ve drawn here are wildly speculative, given that we have only five selection vectors. Terra incognita (any white area) refers to areas that are unknown and therefore unmapped. Note that the lack of selection pressure (short arrow) means that the population is on an adaptive peak.

But selection does not act alone. You can see in the top diagram that those selection vectors don’t predict exactly where the population moves on the landscape. Deviations between the selection vector and the actual evolutionary trajectory, which can be thought of as another vector, are caused by random processes (mutation, mating, drift). What the selection vectors do, though, is point uphill toward the closest fitness maximum or “adaptive peak.”

Selection maps evolutionary terra incognita. With that in mind, look at our population of PreyRos in generation 3 (find the generation number in the top diagram and then look for the corresponding point in the bottom diagram, Figure 7.2). That population sits in what I’ve labeled as a valley, an area of low fitness compared to two or more close-by regions of higher fitness, the adaptive peaks to the north and the south.[142] The selection vector for generation 3 is small and points due south. The small magnitude of the vector means that the population is nearly sitting right on top of an adaptive peak, with just a little bit of climbing in the b-dimension to reach the local summit. The fact that this generation-3 population never summits but instead shifts off of this peak in generation 4 illustrates one of the great ironies of evolution: random factors like mutation and mating can displace a well-situated population, adaptationally speaking.

Evolutionary conditions change drastically from generation 3 to 4. The compass of selection has backed from the south to northeast. This change means that instead of finding a peak by reducing the span of the caudal fin, the population ascends a different peak now by increasing both the span and the number of vertebrae. Traversing to generation 5, the population overshoots the summit, and selection points back from whence the population came.

Mapping the adaptive N-b landscape shows dramatically how a steady selection pressure—rewarding enhanced feeding and predator avoidance—can produce a tortuous evolutionary trajectory. What we still don’t know, however, is what the whole landscape looks like. How extensive is the range of adaptive peaks? Are they peaks or ridges? Do other adaptive peaks exist?

THE LIMITS OF EVOLUTIONARY BIOROBOTICS

We could overcome our evolutionary ignorance using a modified directed-evolution approach. We could plop down populations of Evolvabots somewhere, run evolutionary trials, and let selection map the local terrain. This beam-me-down-Scotty procedure would be instructive, but it would be hit or miss with respect to the whole map. We’d run the risk of doing all that time-intensive evolutionary work only to find ourselves in the middle of an extensive fitness valley.

Given enough time and money, though, you could absolutely use directed evolution to expand your map. Populations of Evolvabots placed evenly and densely throughout the entire N-b landscape would enable you to accomplish what my colleague Chun Wai Liew calls an exhaustive search of the parameter space. This is not realistic with physical models, but it does work fine in digital simulation, assuming your landscape has only a couple dimensions, such as number of vertebrae and tail span, and covers just a little territory. But what if you add in predator-detection threshold, shape of the vertebrae, length of the tail, activity patterns of the muscles, and all of the various neural control mechanisms? By the explosive mathematics of combinatorics, each added dimension, k, expands the possible number of combinations, n, which in our case are different kinds of genotypes or phenotypes, by the following relationship with the number of possible values or conditions, j, within each dimension:


n=j k


Don’t let this cute little equation fool you. It hides a tactical hurricane. Before the storm, the wind is fair. Let’s say that we have only two dimensions—the number of vertebrae, N, and the span of the caudal fin, b—so that k = 2. If we allow both dimensions to have four conditions, j = 4, then the number of different phenotypes would be n =42, or 16. No problem.

But your lumbago should be telling you that a storm is approaching. Hold on tight. Let’s stay with our two dimensions, k = 2, but now make the number of conditions within each dimension a bit more realistic, say j = 14, which is how many vertebral states are possible in PreyRo (zero to thirteen vertebrae). Even though more conditions are possible for the span of the caudal fin (zero to fifty millimeters, in one-millimeter increments), we’ll just say that both have the same j value for the moment. For just our b-N adaptive landscape, that gives us a low-end estimate of n = 142 = 196 different phenotypes. The wind is pickin’ up. Reef the sails!

In PreyRo we have three dimensions, so k = 3. We’ll stay with our conservative estimate of fourteen possible conditions in each dimension. By adding a third dimension, we now have the possibility of n = 143 = 2,744 different phenotypes. But that figure underestimates the number of conditions, because the span of the caudal fin has fifty and the predator-detection threshold has fifty (10 to 60 centimeters in 1-centimeter increments). Let’s choose j = 25 and see what happens: n = 253 = 15,625 different phenotypes. Gale force winds! Batten down the hatches and man the bilge pumps! We are taking on water, matey.

This little exercise makes several things crystal clear. One: nautical metaphors are extremely annoying. Two: it is practically impossible, in the true sense of both words, to build and test all types of even a simple Evolvabot like PreyRo. Three: map makers wanting extensive or exhaustive maps of the adaptive landscape must resort to digital simulation because that is the only way to approach the number of trials needed in hill-searching and hill-climbing experiments. It’s also worth noting that there are ways to avoid having to do an exhaustive brute-force search of an entire landscape; Chun Wai uses different kinds of evolutionary algorithms to balance the demands of finding all of the peaks and doing so in a reasonable amount of time—like weeks instead of years. He has developed a meta-algorithm that decides when to use an exploring routine to search broadly and when to switch to a focusing routine that finds local hills.[143]

Seduced by the phenomenal cosmic power[144] of digital simulations and their handlers, I can’t help but wonder what in the world I was thinking. Using physically embodied robots? Evolving their biomimetic body parts? Enslaving students to work in the robot factory? Believing in autonomous behavior and situated-and-embodied intelligence? John, you dummy! Think of all the time and goodwill that you’ve wasted. Like it or not, digital simulation, clearly, is the way to go.

Then, a voice. I hear the Ghost of Christmas Past: “You see, John,” he whispers, “if you had not become obsessed with physically embodied robots, your life and the lives of those around you would have been much different. It would’ve been better.”[145]

Yes, I think, it would’ve been different … better. I could’ve explored the entire adaptive landscape of early vertebrates in the wink of an eye using digital simulation. My students and I could’ve moved beyond vertebrae. We might even have explored why fish evolved paired appendages, swim bladders, different body shapes, and the ability to live on land.

“John,” says the ghost, this time louder, “You must change your methodology. There is still time. Join your friends in the land of digital simulation, and you will come to understand why computational biology is de rigueur.” At the unexpected use of French, I turn, expecting to see my Gallic tormenter, hoping to plead for the opportunity to retool, perhaps in a year-long sabbatical at a stylish Parisian university or, if that’s not possible, in an intensive summer course at a marine laboratory on the Mediterranean. But I see no one.

Without meaning to, I say aloud, “I guess you’re right.” Now I hear the ghost smile, as if that’s possible, and I see him, or at least his face, beaming. “Yes, John, I am right. And you are right to change your ways while you still can, for the sake of yourself and for those around you. Now that you have seen what has been and what might be, I take my leave.”

A chilly breeze rustles the Post-its on my bulletin board as I hear the ghost ask a parting question, tossed casually, as if by a long-lost colleague walking away down the hall: “One last thing, John, that always puzzled me: what do robots have to do with biology?” At last recognizing the trick well played, I freeze, caught between my self-loathing for willfully betraying embodied robots and my embarrassment for falling prey to the old Ghost-of-Christmas-Past ploy. Cue Marlon Brando as the voice-over narrator: “Horror … Horror has a face … and you must make a friend of horror.”[146] Never!

Let us unmask the horror. The face revealed: the why-robots question. Aha! We have faced you before, foul query, back in Chapter 1. But we are different now—stronger. We have data. We can wrestle with you once more, emboldened now by our experience and knowledge. And we know much.

We know that the process that we’ve dubbed “evolutionary biorobotics” works. We know that we can design and build autonomous Evolvabots that represent and hence model extinct and living animals. We can let a population of Evolvabots loose in a simplified world, and that population will evolve under the combined effects of history, randomness, and selection. We know that we can use Evolvabots to test hypotheses about the evolution of the traits of early vertebrates. And we know that by virtue of their explicit simplicity, Evolvabots allow us to witness, interpret, and understand puzzling evolutionary patterns.

But is that enough? Wouldn’t we learn the same thing—and learn it faster—from digital robots? No, no, no, whispering ghost, leave us be! The problem and the difference is physical. We don’t simulate the physical world—we live it. Remember Chapter 1? I’ll recap the reasons for building physically embodied robots and add, in italics, what we’ve learned since we first set eyes on this list.

With physically embodied robots built to model animals (what Webb calls biorobots):

* You can’t violate the laws of physics… because the robots are enacting, not modeling, the laws of physics.

* You can build a simplified version of an animal… using the KISS principle, the engineer’s secret code, and Webb’s modeling dimensions as guidelines.

* You can change the size of the animal… to suit the needs of your experiment or match the physical situation of the targeted system.

* You can isolate and change single parts, keeping all else constant… giving you a decent chance to understand the behavioral complexities that even simple agents produce.

* You can reconstruct extinct animals and some of their behaviors… if you know enough about the anatomy and physiology of the targets and the environments in which they lived.

* You can create animal behavior from the interaction of the agent and the world… without needing to code “behavior” into the “brain” because behavior is the dynamic spatiotemporal event that occurs when an autonomous agent operates in an ongoing perception-action feedback loop with its world.

* You can test hypotheses about how animals function in terms of biomechanics, behavior, and evolution… if and only if your embodied robot is carefully designed to represent explicit features of your biological system.

Phew. One thing that we’ve learned for sure: verbosity. More importantly: a biorobot that is embodied and situated is a physically instantiated simulation, a representation of a biological target, a model. But that’s not all. An embodied biorobot is also a physical thing in and of itself. You can’t take that away from me, or the robot. Even if someone, like one of our intellectual predators from Chapter 6, decides that your Evolvabot is a horrible model of an evolving fish, that Evolvabot is still, undeniably, a physical, material entity. Looks like a material entity. Feels like a material entity. Tastes like a material entity. Anyone for a bowl of material entity? Yes, please. I never eat anything but. Yum.

It’s at this stage that folks like me, working with physically embodied robots, like to claim that the digitally simulated robots, the binary things on the computer, aren’t real. My robots are real. It’s those digital simulations that aren’t. I’m okay, you’re a fake.

However, I’m not going to say that, even though I just did (but I didn’t mean it—paradox alert!). What I’m going to say instead, because it’s a more accurate reflection of reality, is that digital simulations do indeed have a physical reality: electrons, within silicon dioxide microcircuits, by virtue of their controlled movements, carry out a series of Boolean logic functions that, in aggregate, represent the manipulation of symbols defined by a human as part of an algorithm. Those electrons interact with a world of other electrons as well as the constraints and channels of their semiconductive silicon environment. The electrons are not spirits in the material world. They have mass, charge, and velocity. They behave in the same way that embodied robots do—governed by the laws of physics—when they interact with the world. So to say that those electrons aren’t real and can’t behave is false.

Then why all the fuss? What’s the difference between an embodied robot as a model simulation and a digital robot as a model simulation? Barbara Webb, the creator of the field of biorobotics (see Chapter 1), makes the distinction between modeling in software and hardware: “The most distinctive feature of the biorobotics approach is the use of hardware to model biological mechanisms.”[147] Webb elaborates: “a more fundamental argument for using physical models is that an essential part of the problem of understanding behaviour is understanding the environmental conditions under which it must be performed.”[148]

Now we are onto something. The difference is not physical simulation versus nonphysical simulation. It’s not materialism versus substance dualism (see Chapter 5). The difference is how we model the behavior. Do we create the behavior by representing the interactions of agent and environment algorithmically, mathematically? Or do we create the behavior by not representing the interactions at all but instead letting them “just happen”? When behavior just happens, we remove a layer of the simulation, a layer of representation that, when present, increases the conceptual distance between the target and the model (Figure 7.3).

Let me put it this way: if behavior is the physical interaction of—or feedback between, if you prefer—a physical agent and a physical world, then that behavior can be modeled in mathematical representations or not modeled. This makes me think that Rodney Brooks was pulling our collective leg when he said, “The world is its own best model.” Here’s the paradox: the world is not a model; it is simply the world itself. We only make the world into a model when we force it to represent something else. This, then, is the Zen of Physically Embodied Biorobots.

FIGURE 7.3. Things-in-the-world and their representations. Each thing-in-the-world can, if carefully designed by a human, represent other things-in-the-world. The great power of software is that it can represent any thing-in-the-world, even representations of things-in-the-world. The same can’t be said for fish: you would never argue, I hope, that fish-in-the-world represent software-in-the-world. But because representation depends on the intent of the human experimenter, folks can and do argue that they can use fish-in-the-world to represent extinct-fish-in-the-world (primary representation, 1°). We built Tadro3s to represent the tadpole larvae of tunicates that, in turn, we selected to represent early chordate ancestors of vertebrates (secondary representation, 2°, of chordate ancestors by Tadro3s). When we created digi-Tad3s as representations of Tadro3, we created an additional layer of representational distance from the target (tertiary representation, 3°, of chordate ancestors by digi-Tad3s).

A robot-in-the-world and software-in-the-world can both be built as primary representations of an extinct-fish-in-the-world. In that sense they are equivalent as model simulations. What differs is how they represent behavior. As part of its representation of an extinct-fish-in-the-world, software-in-the-world must represent the physical interactions of the agent and its environment. This is a hidden or implicit level of representation (let’s call it 0°) that increases the conceptual “distance” between the target and the model that represents it.

As physically embodied biorobots, we’ve already established that ETs aren’t Evolvabots: they don’t evolve and, hence, they can’t directly test hypotheses of evolutionary process. However, as suggested by our list of the seven reasons to use embodied robots (page 178), ETs can test hypotheses about how extinct animals functioned and behaved. Because we’ve taken a page from cognitive science and defined behavior as the interaction of an autonomous agent with its environment, testing the behavior of ETs allows us to examine what Robert Brandon, back in Chapter 2, called “function in the ecological situation”—one of the six pieces of physical evidence needed for explaining adaptation.

Thus ETs, as primary representations, inform our evolutionary investigations by testing behavioral hypotheses of extinct or nonexistent animals. Behavior, as we’ve shown with our Evolvabots, is what selection “sees,” the action in the game of life that we judge using the fitness function. However, because behavior doesn’t fossilize, recon structing it with ETs is a great way to remember the past.

REMEMBRANCE OF THINGS PAST

As I promised at the end of the last chapter, I’m not going to use Tadros, at first, to look at what we can learn by using ETs to study the biology of extinct organisms. I’m not going to look at fish nor am I even going to discuss backbones (at least not very much). Instead, our magical mystery tour of lost behaviors in the evolutionary landscape continues with the ET known as Robot Madeleine (Figure 7.4). Madeleine, scallop-shell shaped like the petit madeleine cake, is the first robotic creation named after a French pastry (Figure 7.5). Launched in 2004, Madeleine the robot served, just like Proust’s tea-soaked madeleine the pastry, as the catalyst for explorations into things past,[149] lost vertebrates known as plesiosaurs who, with their four propulsive flippers, were the giant top-level predators of the Jurassic seas over two hundred million years ago.

FIGURE 7.4. Robot Madeleine, a four-flippered Evolutionary Trekker. Madeleine helped launch a new scientific journal, Bioinspiration & Biomimetics, in 2006. Madeleine is designed to represent aquatic tetrapods, descendents of the four-footed vertebrates that evolved on land and then evolved back to the water (see Figure 7.7). By varying the pattern of how she uses her flippers, we can test the hypothesis that four flippers, compared to two, produced swimming behavior with faster top speeds, quicker acceleration, and better braking. The cover image is used with permission of the Institute of Physics. I took the picture of Madeleine while she was going through her first shake-down cruise in the outdoor pool of my friend John Keller.

FIGURE 7.5. A petit madeleine, chocolate, left lateral view, showing its streamlined scallop shape as seen just prior to consumption. Robot Madeleine is the first robot named after a French pastry. We recognize that petite madeleines don’t have flippers and don’t swim. However, the chocolate petit madeleines, in particular, are very tasty, like little moist cakes.

Well, perhaps not quite remembrance: we know of plesiosaurs only because they’ve left us their skeletons as fossils, not because we were around to see them when they still lived. The first plesiosaur was discovered by twenty-two-year-old Mary Anning in 1821 as she scoured the cliffs of Lyme Regis, a West Dorset coastal town on the English Channel. Anning’s sea dragon was clearly a vertebrate—with its many vertebrae forming the great chain of bones along its axis—but was otherwise odd, with four large paddles instead of legs and a girdle of bones instead of gracile ribs amidships.[150] The Reverend Conybeare named it for science in 1824 as Plesiosaurus, from the Greek plesio (= close or near) and saurus (= lizard), and described it as a “comparison with the paddles of the sea turtle will exhibit such fresh analogies as to indicate that in respect of the various forms of animal extremities, the Plesiosaurus holds as it were a middle place between it and the Ichthyosaurus; for we may remark in the first carpal series of the turtle three bones not unlike those of the Plesiosaurus.”[151]

FIGURE 7.6. Plesiosaurus dolichodeirus, cast of fossil. Note the strangeness: a tiny head on a long neck, short body reinforced with robust bone, and, best of all, four large flippers of a form that appears to be shaped for doing the hydrodynamic work of an underwater wing. This cast, about two meters long, is from the Warthin Museum of Geology and Natural History at Vassar College. It was purchased in the nineteenth century by the college and was listed as Item 225 in the Wards Scientific catalogue of 1866. Photo by Rick Jones.

Richard Ellis, in his book Sea Dragons, notes that Conybeare’s scientific description was followed by Dean Buckland’s more famous 1836 construction: “To the head of a lizard, it united the teeth of a crocodile; a neck of enormous length, resembling the body of a serpent; a trunk and tail having the proportions of any ordinary quadruped, the ribs of a chameleon and the paddles of a whale.”

Strangest of all is the fact that these descriptions are not flights of fancy. The darn things look pretty much as described in Figure 7.6.[152]

Although their utter strangeness makes plesiosaurs so compelling, in an evolutionary sense they were pedestrian. Plesiosaurs[153] are just one example of what Carl Zimmer has described as a repeated series of past and ongoing evolutionary experiments in which sea creatures descend from four-legged terrestrial vertebrates known as tetrapods.[154] From land to sea they go. Because of their heritage as terrestrial tetrapods, any vertebrate lineage that has crawled back to the sea, evolutionarily speaking, is called an aquatic tetrapod. Living aquatic tetrapods that you might recognize include whales and dolphins, sea turtles, penguins, otters, and seals and sea lions. And there are more!

Not even including amphibians, it’s simply stunning how many times terrestrial tetrapods have spawned species that have returned to the sea and adapted to aquatic locomotion. Analyzing a beautifully preserved Late Cretaceous mosasaur (mosasaurs are yet another group of giant, extinct aquatic tetrapods), Johan Lindgren, a researcher at Lund University in Sweden, and his colleagues have shown how selection for enhanced swimming performance has apparently and repeatedly built streamlined bodies, vertebral columns reshaped to enhance and regionally control bending stiffness, and caudal fins with increased span.[155] Keep in mind that any pattern of convergent evolution (see Chapter 2) is excellent circumstantial evidence that a strong and steady selection pressure has been at work.

Convergent evolution creates one of those goose-bumps moments for biologists. I mean, how cool is it that dolphins and whales evolved from mammals to look like extinct ichthyosaurs and mosasaurs that were, in turn, independently evolved from reptiles? Not only does convergence provide great evidence for evolution by natural selection, but it also suggests that, in some kinds of situations, only a few options exist for pushing the performance envelope in the game of life.

When the game moves from land back to water, the aquatic tetrapods basically have two choices to overcome their locomotor roadblock: limbs or body axis. Frank Fish, professor of biology at West Chester University and an expert in the biomechanics of aquatic locomotion, has proposed a functional path for the evolution of aquatic tetrapods in mammals.[156] According to Fish, it’s probably the case that swimming with appendages is the way that the shift back to water starts; this seems likely because almost every terrestrial mammal we see will swim using a variant of the “dog paddle” when they hit the water. Some lineages, such as seals and sea lions, stick with appendages, expanding and oscillating their rear flippers, in the case of seals, or flapping their front flippers, in the case of sea lions. Others—whales and manatees, for example—evolve their developmental programs to stop building appendages, losing the rear limbs and using the body axis, anchored by that pesky vertebral column again, to move the evolutionarily novel flukes up and down.

Pleiosaurs evolved along roughly the same track as seals, sea lions, and sea turtles, sticking with appendages as they readapted to life in the water. Here’s a curious observation, one that makes me wonder and drives me mad with evolutionary might-have-beens: none of the living aquatic tetrapods ever use all four appendages to swim underwater—they use only two. The plesiosaurs, however, appear to have used all four limbs, which were modified into wing-shaped flippers (see again Figure 7.6). If four flippers were good enough for plesiosaurs to rule the seas as the top-level predators in the Mesozoic, why aren’t they good enough now?

Let’s ring the bell and bite the apple! We want to know: why, why, why? Why don’t mammals and sea turtles alive today use all four flippers for propulsion? From a mechanical point of view, it sure seems like using four flippers for propulsion should be better in almost any way imaginable. If you think about each flipper as a propeller, then any agent—animal or robotic—using four flippers instead of two should be able to accelerate more rapidly, reach a faster cruising speed, and brake more quickly. So why wouldn’t they? That, then, was the behavioral mystery within the context of evolutionary paths taken and not taken that we set out to solve with an Evolutionary Trekker.

BUILDING PROPULSIVE FLIPPERS

Here’s where we need Robot Madeleine. We built her as a generalized aquatic tetrapod, with four identical flippers that propel her as she swims underwater. She’s 0.78 meters long stem to stern and weighs twenty kilograms dry, roughly the length and mass of an adult green sea turtle or a small species of plesiosaur, minus the long neck (see Figure 7.6). Each of her flippers, called a Nektor by engineers, has the cross-sectional shape of a wing or, to be precise, a hydrofoil.[157] Each flipper is oscillated around an axis, “in pitch” as the engineers say, by separate motors inside Madeleine’s hull. To avoid building an overly powerful super ’bot, we chose motors that would approximate the power density of vertebrate skeletal muscle, about ten watts per kilogram of body weight. And of course, Maddie has a body shaped like a petit madeleine pastry. From a biological point of view, this pastry shape would be described as bilaterally symmetrical, with a fusiform shape for streamlining.[158]

All of Maddie’s features just mentioned were chosen with mechanistic accuracy in mind. So as convergent evolution seems to do for aquatic tetrapods, we focused on her locomotor behavior and the structures related to generating propulsion. To judge whether we’d done a good job of recreating an aquatic tetrapod, we rely on five of Webb’s criteria for a model robot: biological relevance (criterion 1), behavioral match between the target and the model (criterion 2), mechanistic accuracy (criterion 3), level of structure (criterion 5), and substrate (criterion 7).

Perhaps the biggest complaint we get about Maddie is that she does not represent any species in particular, giving her low concreteness (or, conversely, high abstraction, Webb’s fourth criterion). But … precisely! That’s what we wanted, and Webb’s criteria help us recognize and explain that Robot Madeleine can’t—and shouldn’t even be able to—do it all. In fact, that’s why I named her after a French pastry. I didn’t want to pretend that she was a robotic turtle, for example, as she has come to be named in the popular press. Can I have my pastry and eat it too? If Madeleine is not a robotic turtle, then how can I claim that she is a robotic plesiosaur? I don’t. What I claim is that Maddie uses some of the same propulsive principles that we think both turtles and plesiosaurs use and used. Thus, Maddie’s mechanistic accuracy is high for any aquatic tetrapod that flaps flippers to swim. Thinking about modeling as the process of representing, Maddie’s behavior represents the behavior of turtles and plesiosaurs in the specific sense that she is about their size and swims with flippers.

Another important critique of Madeleine is that her flippers don’t work in exactly the same way as the flippers of a turtle or a plesiosaur. This gets to the issue of accuracy of the model at the level of the propulsive elements. The motion of Maddie’s flippers, or Nektors, as they are known, is unbiological, in no small part because their movements are much simpler than, say, the movements of the front flippers of sea lions.[159] Both Maddie’s flippers and a sea lion’s rotate in pitch during each stroke—but that’s all Maddie’s do. A sea lion’s also roll as the front limb pivots about the joint of the shoulder. While the sea lion is rolling its flipper down and pitching it about its long axis, it is also yawing the flipper about the shoulder joint, moving the flipper’s tip rearward toward its hip. To jump into the lingo of engineering, then, the flipper of a sea lion has three degrees of freedom, whereas Maddie’s has only one. And even three degrees is still too simple! I’ve conveniently neglected to mention that the flippers change shape as they rotate, with joints at the elbow, wrist, and five fingers flexing and extending. Let’s see, if we assume that elbow, wrist, and five finger joints are all simple planar joints, that adds seven degrees of freedom. In sum, each flipper has ten degrees of freedom from ten joints, and each joint has be actuated, controlled, and, as a group, coordinated.

Are you ready to give it a try? You can do it! But before you run off and build a better flipper, keep in mind that by jumping from one to ten degrees of freedom you’d be violating the KISS principle, at your own peril. You’d be doing the opposite of making the simplest device possible. Let’s give that anti-KISS principle a name: Make It Complicated, Einstein, or MICE. With MICE, you’d need to create an internal skeleton with joints strong enough to withstand all of the hydrodynamic loads yet supple enough to bend without requiring too much force. You’d need to figure out how to move all ten of the joints that you’ve created: do you put motors out in the fingers and add bulk and weight, or do you run wires or hydraulic tubes from the inside of Maddie? Then you’d have to cover and embed the skeleton with a flexible, body-like material that maintains shape yet reconfigures as the flipper moves. And, once covered, that limb would need to allow you to get back inside for repair and maintenance.

Let’s say that you took care of all of that mechanical engineering, using the MICE principle, and now you are looking at your handiwork. Now what? You’ve got to build another flipper, for the other side of … an underwater robot. Yikes! You’ve put the robot part on hold while you earned your PhD designing and building the biomimetic flipper. Okay, no big deal. Now you build a robot to anchor the two sea lion flippers and hold the batteries to power the motors. Done? Not quite. Now you need to design and test the software to control and coordinate the twenty motors. Twenty degrees of freedom aren’t too much of a problem because you are just getting the flipper to flap down and back and then recover to the position needed to repeat the motion. Your software controller works great in the lab, in air, on the bench. But then you put the robot in the water. The moment it starts to flap and move, the long flippers undergo drag that tends to push back and bend the flippers. You realize that if the flippers are going to maintain a specific shape and change shape in a way that you specify during the stroke, you’ve got to have sensors at each joint.

Maybe this is why animals have proprioception, the internal positional sensory system that allows you to touch your nose with your index finger with your eyes closed. Thinking with your MICE hat on, you curse yourself for not making the flippers complicated enough from the get-go! You rip open the flippers and put potentiometers at each joint, and wire those twenty sensors to the computer controlling the motion of the flipper. You build a new software module that takes the sensor input as feedback to the motors. Now, when you tell the flipper to have a certain shape and position, you are certain it will do so.

Back at the pool, you put your system, now dubbed SeaLioTron, into the water and are overjoyed to see it swim slowly with a symmetrical and controlled motion of the flippers. Great. Now let’s get SeaLioTron to turn. You’d thought about that aspect of control, but because you were working on the flippers, you saved that problem for later. When you revisit the videotapes that you’ve got of real sea lions, you notice that they have tremendous turning agility because they can bend, using their bodies as big rudders in the water. Because having a flexible body was not part of your project and raises all kinds of issues with the internal payload of batteries, motors, and computers, you either (a) call the SeaLioTron project a success and consider it done, (b) figure out how to turn SeaLioTron with flexible flippers and a rigid body, or (c) start a whole new project to build a flexible body-as-rudder.

Although I’ve made up this MICE counterexample, the idea that someone might tackle SeaLioTron is not complete fantasy. You would certainly encounter the problems that I’ve outlined and you’d probably, along the way, generate some really clever solutions that I can’t fathom. And all that difficulty might be worth it: the SeaLioTron would easily produce a handful of PhD projects and, likely, a bunch of cool patents for flexible, actuated propulsors and multijoint neural controllers. You might even get NSF to fund the project. So what am I doing? I bring up SeaLioTron as an example of the MICE principle to make the following point about Maddie’s one-degree-of-freedom Nektors: yes, Maddie’s flippers are inaccurate as models of sea lion or sea turtle or plesiosaur flippers, but at the time that Maddie was built (2003 to 2004), they were the most bio-realistic flappers around (generating thrust by flapping with a flexible foil), and they are relatively easy to actuate and control. With the KISS principle, you do the simple stuff first. And the simple stuff turns out to be plenty complicated.

TWO FLIPPERS OR FOUR?

So back to the big question: why do all living aquatic tetrapods favor two powered flippers over four, when four seems to be the key to better performance? Even though ETs make answering the question easier than if we tried to explore the morphospace of flippers and the likely evolutionary processes, we still need to proceed with the extreme prejudice that a physically embodied robot will produce results that we failed to anticipate.

Let’s first define our “simple,” two-dimensional morphospace. The traits that we could vary in Madeleine were the number of flippers used and the pattern of flipper use. In the first set of experiments, however, we only changed and investigated the pattern of use in what we might call the neural-control space (Figure 7.7). Although that simplifies matters, the patterns are still awfully complex, a result of the fact that each of Maddie’s flippers is independently controlled.

When one flipper reaches its most downward position, another flipper may be reaching its uppermost position. The difference in the time that two flippers take to reach the same position is called “phase.” If two flippers are in phase, what we’ll label as 0 degrees out of 360, then they are flapping together, perfectly synchronizing their swimming. If two flippers are out of phase, the easiest pattern to see is a phase of 180 degrees, like drumming a steady beat by alternating your left and right hands striking the table top. Unfortunately, there are many other ways to be out of phase, so then the combinatorial world of our experiment explodes. If you think about testing every ten degrees of phase, a crude resolution for this dimension that contains 360 degrees, that gives you thirty-six different conditions to go along with four flippers or—ouch!—more than a million total combinations (Figure 7.7, bottom). Cruel irony. Even with just two dimensions, number of flippers and the phase between them, we can’t exhaustively explore this “simple” neural control space.

FIGURE 7.7. Two flippers or four? (facing page) Terrestrial tetrapods—mammals, reptiles, and birds—have repeatedly spawned lineages that returned to the sea. These aquatic tetrapods evolved in different ways, improving their submerged swimming performance over generational time by shifting from a terrestrial pattern of back-and-forth limb movement to an up-and-down or side-to-side motion. Those changes in motion are associated with a change from drag-based paddling to lift-based flapping. Living flappers, like sea lions of the genus Zalophusor the seals of the genus Phoca, use only two flippers for propulsion. In contrast, extinct flappers like short-necked plesiosaurs of the genus Kronosaurusor long-necked plesiosaurs of the genus Plesiosaurus have four nearly identical flippers that appear, from their wing-like shape and anatomical connections to the body, to have been used in lift-based propulsion. How swimming performance is connected to the motion of the flippers quickly becomes complicated, with millions of possible flipper patterns in four-flippered swimmers. This figure is inspired by the research of Frank Fish.

FIGURE 7.8. Experiments with the Evolutionary Trekker, Madeleine. To test the hypothesis that Robot Madeleine should swim faster and with greater acceleration using four flippers rather than two, we ran her through a series of experiments. In this example Madeleine is using her two rear flippers to start from a stop, swim as fast as she can, and then stop as quickly as she can. These images were taken from underwater video that had been analyzed to show Madeleine’s position (point on her bow traced frame by frame to create the path) over the whole experiment. The snorkeler in the water (that’s me, ahem) makes sure that Madeleine is stationary and located at a depth of two meters before the topside experimenters start the experiment.

In the face of this daunting complexity we sought the refuge of the KISS principle once again. In terms of number of flippers, we had three conditions to test: (1) two front flippers, (2) two rear flippers, and (3) all four flippers. Within each of these conditions, we varied the phase in the following ways. With two flippers, they flapped either in-phase (0 degrees) or 180 degrees out of phase. Simple. With four flippers, we borrowed a page from Frank Fish’s evolutionary model and used four patterns of limb movements, called gaits, that are seen in terrestrial tetrapods: (1) pronk (all four in phase), (2) gallop (front in phase, rear in phase, front 180 degrees out of phase with rear), (3) trot (left front in phase with right rear, right front in phase with left rear, those diagonal pairings 180 degrees out of phase with each other), and (4) pace (left in phase, right in phase, left 180 degrees out of phase with the right).

We tested Madeleine using all eight gaits in the diving well at Vassar’s pool.[160] Using an underwater video camera, we recorded Madeleine’s movements as we accelerated her from rest, raced her to top speed, and then had her break as quickly as she could (Figure 7.8). While she was doing all of this, we measured her accelerations and energy consumption using an on-board three-axis accelerometer and electrical power monitor.[161]

The result? Two flippers are just as good as four in terms of top cruising speed (Figure 7.9). This was not what we’d expected. Another good theory dashed upon the rocky shore of empirical investigation! But this wreck is enlightening. Four flippers, in addition to not moving Madeleine along at any faster top speed, also takes twice the electrical power of two flippers. Two flippers good, four flippers bad—right? Not so fast. Four flippers are really handy if you want to accelerate quickly from rest, and using four gets you off the mark 1.4 times faster than using two flippers.

What we have here is called a performance trade-off. If cruising is your game, then you should use two flippers. If accelerating is important, use four. Trade-offs like this are the stuff of evolution, as the selection environment fiddles with what works best in a population at a given time and place. Change how schools of fish are distributed, for example, and maybe you have to cruise more to find your prey.

FIGURE 7.9. Madeleine’s behavior with two and four flippers. To our surprise, Madeleine reaches the same top cruising speed using only two flippers. To maintain the same cruising speed with four flippers, Madeleine has to use twice the electrical power. Four flippers offer an advantage when Madeleine accelerates from a start. The points represent the means for three trials of each of four gaits. The error bars are the standard errors of the mean. Statistical tests back up all claims of a performance metric being the same or different.

Let’s get back to our motivating question: why do we see living aquatic tetrapods using only two flippers for propulsion? Based on the results from Robot Madeleine, my empirically educated guess is that living species tend to do a lot of cruising, relying most heavily on that aspect of their swimming behavior in the game of life. For example, green turtles, when they aren’t asleep on the ocean floor, are cruising around, moving among beds of sea grass, which they visit for their vegetarian meals. Penguins cruise rapidly from shore to fishing grounds, where they dive and maneuver through schools of prey. Seals and sea lions, too, eat fish, although sea lions, in particular, are no-table for their rapid turns and maneuvers, as we talked about with SeaLioTron. Sea lions may be an exception to the two-flipper-cruiser rule that we are sketching out. Frank Fish, Jenifer Hurle, and Dan Costa have measured centripetal accelerations of up to five times that of gravity in the California sea lion, Zalophus californianus.[162] Although those accelerations are angular rather than linear like those that we measured in Robot Madeleine, they highlight the fact that the sea lions don’t need top cruising speeds but instead rapid accelerations to capture quick and elusive fish.

What about our extinct four-flippered plesiosaurs? From the physical evidence Robot Madeleine produced, we can circumscribe the likely scenarios. For small plesiosaurs the size of Madeleine, I can imagine them feeding as sit-and-wait ambush predators: hang in the water until something tasty swims close by and then—bam! Hit the accelerator and grab some lunch. Although this ambush behavior might work for smaller plesiosaurs, I don’t think that it’s possible for the giant short-necked pliosaurs like Kronosaurus or Predator X.[163] The trouble is that at ten or fifteen meters long, they are simply too massive to accelerate quickly. For the same reason that you never see a tractor-trailer beat a sports car off the line when the light turns green, you would never see Kronosaurus waiting and then lunging at a fish traveling by. Neither truck nor pliosaur can generate the mechanical power needed to launch their massive bodies quickly. When starting from a stop, their performance is constrained by (1) the amount of power that either internal combustion or skeletal muscle can produce and (2) their massiveness.

So what’s a poor giant sea monster to do with four flippers? For the Predator X documentary that aired on the History Channel, I did some very crude calculations. The team of paleontologists that discovered Predator X, led by Dr. Jorn Hurum, estimates that Predator X was fifteen meters long. Using data that are available on the length and mass of great whales, I estimate that Predator X had a mass of about thirty-nine thousand kilograms, or thirty-nine metric tonnes. If Predator X accelerated at Madeleine’s peak acceleration from rest, about 0.085 m s–2 (see Figure 7.9), it would move 8.5 centimeters in one second, a tiny fraction of its 1,500-centimeter total length. If you happened upon Predator X sitting still in the water, you’d have nothing to worry about, unless you swam right into its mouth!

If, however, you ran into Predator X while she was already moving, you might be in trouble. I’m guessing that Predator X cruised around and used its large flippers, like those on a humpback whale, to maneuver, redirecting its forward-cruising momentum into a feeding lunge of the type seen in blue whales, accelerating to one or two m s–2 to hit speeds perhaps as high as two to three m s–1.[164] But what’s Predator X predating? Unlike humpback or blue whales, which use their baleen plates to filter whole schools of small fish or krill out of sea water, Predator X, with its large teeth, was probably grabbing onto other large and relatively sluggish animals. The long-necked plesiosaurs may have been this short-necked plesiosaur’s target.

Here’s the best part: if Predator X cruised around looking for an unwary plesiosaur, then perhaps it only needed to use two of its flippers at any one time. Why waste the energy needed to flap all four if Robot Madeleine tells us that you won’t swim any faster for the additional effort? Imagine if you could run a marathon on your legs or your arms. You could race until your legs were sore and then switch to your arms. This is straight out of the Department of Crazy Ideas. But based on what we now know about the physics of four-flippered swimming, switching between front and back propulsive systems makes sense.

For the truth police: we can never know for sure how plesiosaurs swam. Because they are extinct, their behavior is lost. No matter how accurate we make Robot Madeleine’s limb anatomy, no matter how big or small we scale her, no matter how exhaustively we search the neural control space, the very best we can do using an ET is to talk about what is more or less possible. ETs help us circumscribe the plausible; what guides our judgment is the physical reality of behavior, that interaction of an embodied agent and its physical environment.

Acting as an ET, Robot Madeleine journeyed into the neural control space of four-flippered aquatic tetrapods to show us what swimming behavior looks like if you swim with two or four flippers. We compared benefits—speed and acceleration—and costs—power consumption. The differences in Madeleine’s swimming behavior and energy use indicate that a trade-off could exist between high cruising speed and rapid acceleration. If you want both, you are going to pay for it in terms of the food you need to eat to make the energy that your behavior requires. Although our curiosity is not completely satisfied, we haven’t gone mad—or have we?

BUILDING ROBOT MADELEINE

It certainly took a kind of collective madness to build a self-propelled biorobot like Madeleine—madness, induced by a shared vision and then a bunch of people with the know-how, the time, and the money to get the job done.[165]

Robot Madeleine was custom built for Vassar’s Interdisciplinary Robotics Research Laboratory (IRRL) in about a year, from 2003 to 2004, by engineers at Nekton Research, LLC, in Durham, North Carolina. Working as Nekton’s vice president of Science and Technology was Chuck Pell, cofounder with Steve Wainwright of the BioDesign Studio at Duke University in 1990. Under Chuck and Steve, BioDesign Studio had produced the early prototypes of bioinspired devices that would lead to innovations like Nektors and Transphibians, a class of unmanned underwater vehicles of which Madeleine was the first.

Having created a toy chest full of commercializable ideas, in 1994 Steve and Chuck teamed with businessmen Gordon Caudle and Jeff Bourne to form Nekton Technologies Inc., an independent start-up company that became Nekton Research, LLC, in 2000. Nekton, which acquired multiple government contracts under the guidance of president and CEO Rick Vosburgh, was acquired by iRobot Incorporated in 2008. As a result, Madeleine’s four-flippered commercial descendants are now called iRobot Transphibians. Transphibians can be used for mine clearing, surveillance, and reconnaissance in physically challenging shallow and wave-swept marine environments.[166] You can see why Brett Hobson, Nekton’s first ocean engineer and one of the patent holders, calls the Transphibian “Madeleine on steroids.”[167]

Our shared vision for Madeleine began with Chuck sharing his vision. Before Nekton, during his time running the BioDesign Studio, Chuck drew a Kronosaurus on a napkin and started talking to me about building a life-size, swimming pliosaur. No big deal—a life-size and self-propelled pliosaur. As crazy as this sounds in the retelling, I didn’t think that Chuck was nuts, at least not because of that idea. You have to understand that by 1992 Chuck had already demonstrated that he could use a Nektor to propel a surfboard. I first saw him do this at the Duke Marine Laboratory. He sat on the board, legs out straight on the board, and then cranked a large metal lever back and forth between his legs. Because you couldn’t see the underwater Nektor attached to the shaft he was wiggling, his rapid exertions looked for all the world like he was trying furiously to tighten a bolt with a large socket wrench.

Always the enthusiastic, hands-on teacher, Chuck made all of us who were jeering him from shore give his swimming machine a try. Soon we were all hooked on flapping flexible foil propulsion, and we took turns wiggling the lever and zooming around the docks. The demonstration of Chuck’s crazy contraption was exactly the kind of thing that got the business-minded among us fired up. Gordon started thinking about building a company that would make quiet, low-speed trolling motors for bass fishing. Steve saw a science toy that could be used in the sink or bathtub to demonstrate how fish make waves of body bends in order to swim. In fact, this science toy, which Chuck dubbed “The Twiddlefish,” launched the toy company TwidCo, Inc., and Gordon and Jeff soon had little clown fish and sharks on the shelves of museum stores across the country.

But it wasn’t until Nekton was created in 1994 that Nektors were put to work propelling craft. Brett, renowned in the ocean engineering community for his work on high-performance submersibles like Deep Flight, was Nekton’s first employee. Arraying four Nektors around the circular belly of an ellipsoidal submersible, Brett and Eric Tytell built PilotFish, which quickly achieved the world record for underwater maneuverability, showing that Nektors had response rates twenty times faster than conventional thrusters, could provide thrust in any direction, and allowed PilotFish to use all available degrees of freedom.

By the time I approached Chuck about building a four-flippered robot in 2002, Nekton had left Nektors behind, according to Brett, leaving PilotFish on the shelf. Even though they were interested in exploring other body and flipper geometries, they didn’t have the funding. So we got it from the National Science Foundation in an equipment grant that I wrote with my colleagues at Vassar, Ken Livingston, Tom Ellman, Luke Hunsberger, and Bradley Richards.

Money in hand, we began the design of Madeleine in earnest. Chuck already had Maddie’s conceptual design drawn up when I came to work with the team at Nekton. Brett and Chuck were joined by engineers Robert Hughes, Ryan Moody, and physicist Mathieu Kemp. The Robot Madeleine project was important for the group because, said Brett, “[Madeleine] was the first piece of equipment that Nekton produced and delivered and somebody used.”

What made Maddie useful as an ET was not just her glorious Nektors but also the way in which the Nektors were programmed to operate. This programming was Mathieu’s masterstroke. Not only did he program the motor controllers, he also figured out how to have Maddie’s on-board computer interface with all the sensors I wanted. This gave Maddie all of the capacities that Nick Livingston, Joe Schumacher, and I needed to conduct our flipper experiments at Vassar. In addition, Mathieu programmed Maddie to be autonomous and employ a two-layer subsumption hierarchy in 2004. This was the first time, to our knowledge, that an autonomous underwater robot had used Rodney Brooks’s architecture (see Chapter 5). Although we never published Mathieu’s autonomous design, proof of Maddie operating autonomously can be seen on the Australian Broadcasting Company’s science and technology program, Beyond Tomorrow.[168]

In the same way that Chuck’s swimming contraption was a great catalyst for what might come from Nektor-based propulsion, Maddie turned out be a great spokesperson for Nekton and what they could accomplish. “Video of Madeleine performing,” reminisced Chuck, “electrified government sponsors who previously turned us away and who now said ‘yes’ to money for an even more powerful vehicle.” For demos to the Navy brass, Brett and Mathieu would borrow Maddie, give her special fins, reprogram her, and turn her into an amphibious surf-zone explorer (Figure 7.10). As a Transphibian, “[Maddie] was targeted for high-energy environments,” Chuck explained, “where bottom crawlers can deal with terrain and pelagic vehicles can’t deal with surge. So this tactically important zone had no vehicle that could thrive in it.” That’s no longer the case, thanks to the crawling and swimming Transphibian.

OTHER EVOLUTIONARY TREKKERS

Robot Madeleine is not the world’s first and only Evolutionary Trekker. Josh de Leeuw, a former student who has spent time, as an employee, working with and rebuilding Maddie, pointed out that in 1997 Tony Prescott, professor of cognitive neuroscience at Sheffield University, built a wheeled, autonomous robot to test ideas about the locomotor behavior of Cambrian invertebrates. With the simple intelligence to detect a track that it had made, Prescott’s robot showed one possible neural mechanism that the ancient invertebrates may have used to create the foraging trails preserved as complex spiral patterns known as trace fossils.

To explore the neural control mechanisms involved in the evolution of terrestrial locomotion from aquatic vertebrates, Auke Ijspeert, associate professor and head of the BioRobotics Laboratory at École Polytechnique Fédérale de Lausanne, and his colleagues built a swimming and walking salamander robot in 2007.[169] The robot was programmed to walk and swim using neural chains of central pattern generators, which produce rhythmic activation of local body and limb motors. With a simple linear increase in the stimulation of the artificial nervous system, the robot transitioned smoothly from land to water, switching from standing to traveling body waves. Because the behavioral match and mechanistic accuracy of the robotic salamander are both high in terms of locomotion and nervous activation, respectively, this ET provides a plausible model for the evolution of the neural control of walking in vertebrates.

FIGURE 7.10. Robot Madeleine as the first Transphibian, an amphibious surf-zone vehicle. Maddie uses large flippers that rotate unidirectionally on land, like wheels, to move from the beach into the water. Once in the water, she switches to fin-flapping mode. Transphibians are commercially available from iRobot, Inc., which purchased Nekton Research, LLC, in 2008. Photo is courtesy of Brett Hobson.

After all of our efforts to understand the evolution of early vertebrates and vertebrae, you won’t be surprised to know that we have an ET project underway at Vassar to help us out. To explore the morphospace of just the number of vertebrae, we’ve built MARMT (Mobile Autonomous Robot for Mechanical Testing), a swimming robot into which we can put propulsive tails with biomimetic vertebral columns. MARMT is another group project, with Jon Hirokawa, Sonia Roberts, Nicole Krenitsky, Carina Frias, Josh de Leeuw, and Marianne Porter all making important contributions to this Evolutionary Trekker. Rather than letting evolution create our variations, we simply vary the number of vertebrae from zero to eleven, keeping everything else the same, including the tail span. We then program MARMT to either swim steadily, using a variety of tail-beat frequencies, or to escape. MARMT accelerates faster and swims more rapidly with the stiffer tail that higher numbers of vertebrae produce.[170]

With Evolutionary Trekkers we learn how behavior varies in regions of morphospace (1) no longer occupied by living species, (2) never occupied by species, and/or (3) not visited by evolving robots. There you go, and there you are!

Загрузка...