Darwin's Devices

I AM A BIOLOGIST, AND I STUDY ROBOTS. BUT AS SOON AS I started describing my research to other people, it was clear I was in trouble. I was speaking to a longtime friend and colleague about a biology grant I’d just gotten from the National Science Foundation to build robots when he stopped me in my tracks. “What do robots have to do with biology?” he asked. I knew then, with the certainty that only dread can provide, that this was an inescapable question—it was the issue that would come up first from now on, every time one of my students or I presented our strange new work to biologists.

What’s the problem? First and foremost, biologists do not study robots. They work on organisms—living things, their environments, and their evolutionary history. They use machines as tools to ascend a rainforest canopy, as instruments to measure biomechanical properties, as modes of transportation to collect fish from a coral reef. As my friend had stated so succinctly, machines in general, and robots in particular, have nothing to do with biology—from his point of view. Not from mine.

I tried to fight back, blurting out the well-rehearsed line that I’d included in the grant proposal: “We use robots to model extinct vertebrates.” With that being not so much an answer as a statement of intent, I got a raised right eyebrow and then a gentle “Well, I hope it works out for you.” Communication over and out.

I needed a better answer.

No matter how cool robots are (to me), their swaggering presence wasn’t enough to justify their usefulness in biology. And this was a problem not only for me—at least I had a grant—but also for the undergraduate researchers in my lab, who would rather avoid being recognized as having been trained by a known kook. So we talked it over until we hit upon a solution that has proven to work for about half the biologists we encounter.

We decided to equate different models of biological systems: those run on computers and those run on robots. Both are machines, after all, and computers are already used in almost every branch of biology, modeling—among myriad other things—neural networks, predator-prey interactions, virus evolution, and perambulating Tyrannosauruses. In fact, “computational biology” is the hot field right now, the bull’s eye on the what-to-be-working-on-if-you-want-a-job-in-academia dartboard.

Robots—mobile ones, anyway—are essentially self-propelled computers. They are machines that run sets of instructions—their software—and produce an output. Certainly, the outputs seem different. Computers output binary bits that we use to represent numbers, and those numbers, in turn, represent everything from screen colors to mathematical formulae to electronic books. Robots output what we recognize as behavior, but underlying it all are the same bits.

That’s not to say there might not be important distinctions between a robot and a computer. Jeff Staten, a senior engineer at IBM, says a robot “is a computer that’s inside out.” Jeff’s point is that computers today are networked and make decisions based on input from other computers most of the time and humans tapping at a keyboard only some of the time. A robot, although it has a computer inside, makes decisions on its own, with information gathered only through its sensors. The messages the robot receives and sends are physical. The mobile robot, unlike most computers, can be autonomous. What autonomous robots have, which computers don’t, is agency.

Agency is what human observers ascribe to anything, organic or artificial, that appears from an external perspective to act on its own. For those of us working in the world of artificial intelligence and cognitive science, an agent can be an organism or a machine. Humans are agents. My dog, Kooka, is an agent. And so long as there is no unseen human pulling the strings of remote control, robots are agents too.

An autonomous robot is an agent using its own sensory inputs to perceive the world, make decisions about how to move, and, in turn, having those movements affect how it perceives the world. This constant feedback between what an agent perceives and how it moves is what my colleague Ken Livingston and I call a perception-action feedback loop. Multiple perception-action loops in an agent can be operating in parallel, working in combination, fusion, or competition. What we observe the agent doing—moving and interacting with its immediate environment—is what we define as behavior.

An agent’s behavior is its computational output. “Behavior emerges from the agent-environment interaction,” as Livingston is wont to say. And behavior, by this definition, is something that autonomous robots have that computers do not. Oops. Have we just defined ourselves into an identity error that invalidates the logic of our first response? No and yes. No, because autonomous robots have, as part of their agency, embedded computers. Yes, because robots are more than simply computers that move.

Autonomous agency is, ultimately, the answer to my colleague’s question, what do robots have to do with biology? They enable us to build models of how organisms behave. Of course, building models raises another question.

It turns out that, much like most biologists think they shouldn’t be studying machines, many think they shouldn’t be studying models, either. One criticism of models or simulations of any kind, instantiated on either a computer or a robot, is that they are, at best, artificial systems that merely copycat the outward behavior of the biological system. Models, the argument goes, fail as true or accurate representations of the underlying causal phenomena because the underlying functional mechanisms are different from those operating in the system of interest. As Norbert Weiner, the founder of the field of cybernetics, is alleged to have said, “The best model of a cat is a cat.” Although this is not quite like saying that to understand cats, you can only study cats, some people do jump to that conclusion.

With this cats-only criticism in mind, my colleagues and I jump to a different conclusion about models: you have to be very careful when you build them. You have to take care to explain what you are trying to do, how you intend to do it, and how you are discriminating between a bad model and a good one. Bad models won’t be anything like cats, and therefore, they will perhaps tell you something about your ability to make a noncat—but little else. Good models, argues Barbara Webb, a biologist at the University of Edinburgh and one of the founders of the field of biorobotics, are those with explicitly defined goals and goals that are attained. Some models are meant to behave like the targeted system. In those cases a close or perfect behavioral match between your robot and its biological target—if it walks like a cat and meows like a cat—means that you have a good model.

As a real example, Sarah Partan, an animal behaviorist at Hampshire College, wanted to study how squirrels respond to the behavior of other squirrels, so she built a robot squirrel that flicks its tail and adjusts its posture. For Partan, a good squirrel model is able to trick the real squirrels into responding to the robot as if the robot were a squirrel. Happily for Partan, the trick worked.

Behaviorally speaking, Partan’s robotic squirrel is a good model of body posture and tail motion. At the same time, it is obviously a bad model of the neuromuscular mechanisms involved in limb motion. However, to model neuromuscular mechanisms wasn’t her intent. If it was, you’d judge the goodness of her neuromuscular mechanisms model not by how well it elicits tail flicks in other squirrels but rather by how closely it matches the underlying functional mechanisms used in muscles and nerves. If the artificial muscles worked like biological ones, generating peak force at only one length and for short periods of time, then she’d have modeled that mechanism accurately, irrespective of whether or not her robotic squirrel can dupe real squirrels.

Different models, then, serve different goals. Webb enumerates seven: emulating behavior and mechanism (both of which we’ve just seen) as well as abstractness, medium, generality, level, and, particularly important for us, whether or not the model tests a hypothesis—that is, an idea that you have about the biological system.

For Webb, if you are interested in any particular aspect of cats and you have learned as much from real cats as you can, then go ahead and model a cat. I think this would be Weiner’s position as well. But if your goal is to learn more about cats, Webb says, avoid the temptation to build something cool just for the sake of building it. There must be a specific target: you shouldn’t just try to build something cat-like. Her criticism is aimed at two fascinating fields, adaptive behavior and artificial life, in which many workers model invented animals, called animats. Although animats illuminate general operating and cognitive principles, Webb argues that the adaptive behavior and artificial life approaches do little to test specific hypotheses about how real animals work. For her, biologists building animats to test biological hypotheses usually walk away empty handed.

Webb’s critique of adaptive behavior and artificial life gets to the heart of my colleague’s skepticism. He intuited two of her points related to the value of any model to a biologist. First, your model must have a specific biological target. Second, your model must be relevant and must enable testing a hypothesis about the targeted system.

Of course, Webb’s critique also raises some problems for our response to the skeptics—that we should be able to use robots because robots are essentially computers. If computer models aren’t necessarily biologically relevant, then our robots may not be either. So our critique of the critiquer became this: he was asking the wrong question, or at least asking a sufficiently vague question to allow literalists like us to misunderstand. Instead of asking, “Why robots?” our skeptic ought to be asking, “What is the scientific purpose of your model, and why is your model in the form of a physically embodied robot?” But then there’s probably a question you’d like to ask of me: how did a biologist ever find himself sticking up for robots? The answer’s quite simple: it was for the love of fish.

FOR THE LOVE OF FISH

I’ve loved fish since I was a child, when I first saw Jacques Cousteau’s underwater world on television. I followed the wake of swimming fish and other aquatic vertebrates through college, graduate school, and, now, into a career. And there was a lot to learn in the flesh. With Steve Wainwright of Duke University and Mark Westneat of the Field Museum, I’ve scuba dived to videotape the propulsive oscillation of two hundred–pound blue marlin. With Mark, Melina Hale of the University of Chicago, and Matt McHenry of the University of California, I’ve outfitted rainbow trout, bowfin, longnose gar, and African bichir with tiny instruments to measure muscle function during escape maneuvers. With Wyatt Korff of the Janelia Farm Research Campus, I’ve used high-speed video to investigate how Wyatt’s trained Amazonian arawana can propel themselves out of the water to catch food in midair. With Lena Koob-Emunds and Tom Koob, I’ve worked at the Mount Desert Biological Laboratory to study the biomechanics of slimy pink hagfish so that we can learn about swimming without a vertebral column. Finally, with Marianne Porter of the University of California, I’ve measured the mechanical properties of the backbones of dogfish sharks to see how skeletons transmit force.

I love real fish, and more than twenty years’ work has shown me much about their shape and structure, how they move, and how they evolved. But real fish only reveal some of their secrets. As with any science, we are limited by what we can and cannot observe and measure. Sometimes we lack an instrument or a technology. At other times we lack the right fish. Take, for example, the giant blue marlin. They die in captivity, so studying them in a lab was impossible. So we went blue-water diving not because we wanted to (it’s very expensive and dangerous work) but because we had no other choice. We had to film blue marlin from a distance and leave some of our questions unanswered.

Well, not entirely. We could’ve studied marlin in other ways, but other questions would’ve been left unanswered. Say you want to know what’s going on inside a blue marlin when it swims. You could, as Barbara Block of Stanford University did, build a team of engineers and physiologists to design tiny instruments that can be implanted quickly in a marlin that you’ve brought alongside a boat using hook and line. These instrument tags carry their own computer, power pack, and broadcast system, collecting data and sending signals back to a ship or satellite. Block can measure the marlin’s body temperature, muscle activity, speed, and depth as it moves freely about its oceanic cabin.

But for all Block’s approach reveals about physiology, the method reveals nothing about the biomechanics of marlin backbones—and that’s what I wanted to study. The backbone, or vertebral column, runs from an animal’s head to tail, and its presence is one of the signatures of the vertebrates, a group of animals to which amphibians, birds, fish, mammals, and reptiles—some fifty-eight thousand species—all belong. In a fish the backbone prevents the body from shortening while also allowing it to bend, and it gives the whole body important mechanical features, such as the ability to store and release energy elastically like a spring. My pursuit of this question is what would ultimately send me headlong into the world of artificial intelligence and robots.

FROM THE FIELD TO THE LAB

I first met Block—and the blue marlin—back in 1986, when she, as a newly minted PhD from Knut Schmidt-Neilsen’s lab at Duke University, convinced me, a newbie PhD wannabe in Steve Wainwright’s lab, to work on the biomechanics of marlin vertebral columns in the laboratory. Under the guise of buying me a cup of coffee at the Ninth Street Bakery, Block pulled me out of the lab my first day so she could expound the virtues of the marlin.

Of all the fish, she explained in the car, marlin are the best, the fastest, biggest, coolest predators in the sea. “Think tuna are fast?” she asked rhetorically as we pulled into the parking lot. “Well, marlin eat tuna!” As we walked across the street to the bakery, she went for the kill. “Have you seen the vertebral column of a marlin?” she asked, sounding like a minister in the First Church of Poseidon. I knew the proper response: “No, I have not seen the vertebral column of a marlin. What does it look like?”

She introduced me to the mysteries of the marlin’s backbone. “It’s not like a bunch of little bones linked together, like pearls on string, that you see in regular bony fish,” she said. “The vertebral column of a marlin looks like a piece of wood, a long pine board, a one-by-six, with bones overlapping, bones welded together with collagenous connective tissue to form a single, giant spring.” She paused for effect. “And this spring works to store and release energy, the energy that powers the high speeds and spectacular leaps of marlin.”

I shuffled forward in the line, unable to muster words. Only images came to mind: marlin leaping and spinning above white caps, and terrified tuna, swimming for their lives but unable to avoid the explosive charges of the spring-loaded marlin. Block waited for a moment, paid for our coffee and muffins, and guided me to a table. Signaling with her hand for me to eat something, she gave me a chance to return from my reverie. Then she said, knowing the answer, “So. Are you in?” I gushed, “Absolutely!”

Giddiness gave way to the not harsh but practical realities of scientific research. Working in Wainwright’s lab, I spent the next five years chasing after the elusive blue marlin, literally and figuratively. I wanted to measure their vertebral column’s mechanical properties, features like stiffness—related to how much the vertebral column would resist the magnitude of bending and how much spring energy it would store—and energy loss—related to how much the vertebral column would resist the speed of bending and how much energy would be lost as heat. If stiffness is large compared to energy loss, the backbone would be a spring; if the stiffness is relatively small compared to energy loss, the backbone would work as a brake. If I could measure stiffness and energy loss of the vertebral column over a range of motions and speeds, I knew that I could have some idea of what Wainwright calls “mechanical design”—in this case how the mechanical properties of the vertebral column allow it to operate as the blue marlin swims or leaps.

Unable to buy an off-the-shelf marlin-testing machine (they don’t exist), I had to design, build, and calibrate a customized vertebral column bender. My DIY guru for this challenge was Steven Vogel, also at Duke, who helped me brainstorm designs and taught me the difference between a DC brushless motor and a servo one. Once I had a working bending machine in place, Block and Wainwright helped get me and my machine out to the big island of Hawaii and the Pacific Gamefish Research Foundation.

On the Kona side of the island deepwater blue marlin are caught by recreational fishers literally in sight of the steep-sloped volcanic beaches, where, hat in hand, I would beg at the local fish houses for the castaway vertebral columns. Once I had one, I was unable to sleep until I had put each individual motion segment, consisting of two vertebrae and the intervening joint, through a series of mechanical tests. I’d bend the segments with varying frequency and amplitude, just like the marlin would have done as it hit the turbo button to pursue a tuna. To get a sense of what parts of the bone and joint structure helped cause changes in stiffness and energy loss along the column, I also measured the size and shape of each joint and the adjoining vertebrae. Lather, rinse, repeat. After several weeks I had tested the vertebral columns from six different marlin ranging in length from four to seven feet and weighing from thirty-six to more than two hundred pounds.

THE MECHANICAL DESIGN OF THE MARLIN’S BACKBONE

Back at Duke I began running the raw data through the Newtonian equations of motion that govern the relation between the bending motions the machine imposed on each joint and the bending torque each joint developed in resistance to that imposed motion. Looking over the range of joint positions, bending frequencies, and amplitude, I began seeing some very interesting patterns. The biggest surprise was that the tail, which looks from the video we took of marlin swimming to be the most flexible part of the body, actually has the stiffest part of the vertebral column. Talking with Wainwright, we realized that this was a counterintuitive result only because we were thinking of a jointed column as a series of bony blocks and frictionless hinges. If instead the joints—the hinges—were very stiff because of all of the overlapping bits of bone that Block had talked up, then the joints themselves appeared to be capable of storing energy as they bend.

But were those same joints able to release that spring energy as they unbent? This is where the energy loss came into play, and the marlin played a trick on us again. With simple ideas of springs in our heads, we had been thinking that as the marlin swam faster, increasing the frequency of their tail beats, their vertebral column would become even more spring-like, storing and releasing more elastic energy to match the power that the faster speeds demanded. We expected stiffness to increase and the energy loss to decrease. Just the opposite occurred.

To make sense of these surprises in the biological context of the swimming marlin, we put this information about mechanical properties into a mental, conceptual model of what we thought might be going on inside the marlin. Our guess was that as marlin increased swimming speed, the vertebral column would be adjusting its mechanical behavior, switching gradually from a spring to a spring with a brake. This spring-and-brake mechanism is exactly how the shock absorbers in your car work, with the spring resisting the initial bump, giving way gently, and then returning the wheel to its place on the road. At the same time, the brake, or what we call a dashpot in a shock absorber, uses fluid to dampen the spring’s motion, keeping the spring from bouncing the car vertically after that first bump.