If Charles Darwin were alive, he’d probably be interested in robots. One of the biggest challenges of studying evolution is that extinct species are no longer accessible. For hundreds of years, scientists have tried to answer biological questions about species origin, mutating genes, and evolutionary pressures with only fossils for clues. The idea of using technologically-advanced robots to re-create some of the world’s most biologically basic and oldest creatures sounds like something out of a science fiction novel — but it’s happening right now in, of all places, Poughkeepsie, New York.
In his upcoming book, Darwin’s Devices: What Evolving Robots Can Teach Us About the History of Life and the Future of Technology, John Long, professor of Biology and Cognitive Science at Vassar College and co-founder and Director of Vassar’s Interdisciplinary Robotics Research Laboratory, explains how robots that look and behave like animals can illustrate evolutionary principles. Long explores engineering and robotics to better understand what autonomous robots can learn without human input. Even a basic robot can create elaborate behavior, quickly demonstrating higher levels of intelligence than scientists program it to have. This amazing concept could forever alter the way we think about engineering, warfare, and even ourselves.
I spoke with Long over the phone about “evolvobots,” what Charles Darwin might think about robotic species, and the possibility of a secret arms race for robotic weapons.
Megan Wood: You are a biologist who studies robots. What do robots have to do with biology?
John Long: I think one of the amazing things about robots now is that we’re in a position where we know enough about them that we can actually build robots that are good models of organisms. So if you can build a robot that’s a model of an animal, then you can ask questions about animals using robots. I just got to the point where I couldn’t answer the questions that I wanted to answer with living fish. I had to start this process of modeling, so I tried, and continue to try, to do computer models of swimming fish, but also to do these robotic models of swimming fish. The great thing about a robot is that you know right away if the experiment is going to work or not.
What exactly is an evolving robot?
One of the things to remember is that individuals don’t evolve. What this means is that we have to be careful to talk about the correct units of evolution, which is a whole population. When I talk about evolving robots I’m actually talking about a group of robots that are competing with each other to see who gets to leave the most children in the next generation. An evolving robot is any special kind of robot that’s called a biorobot, which is the kind that’s built to mimic an animal. With biorobots we build many different copies of them and see how they evolve slightly differently, and then judge their behavior in a particular environment. What happens is that from generation to generation, the characteristics of the group changes. That’s the population: the group. So the group of robots actually changes. Over ten generations we look and see which robots in the group have very different characteristics than what we saw in the group ten generations previously (the sort of ultimate parental generation).
Those evolving biorobots we call evolvobots, because everyone needs a stupid robot name.
What can scientists actually learn by building evolvobots?
You can test ideas about the specific events that have caused or that were likely to have caused the origin of unique features that we see in animals — things like the wings on birds. Birds are really living dinosaurs. How did dinosaurs acquire flight? We’ve got some really interesting fossils out of China that suggest how that happened. But one of the things that we miss in the fossil record, and that we don’t have in living animals either, are the intermediate steps that have occurred between one form and another. One of the things we can do with our robots is watch those intermediate forms as they evolve in our population of robots. As soon as you get to see those intermediate forms and watch them behave, then you can understand the evolutionary process that created the living form that came from the ancestors of flying birds. We can go after questions like bird flight or the evolution of being upright and bipedal in humans.
How do you simulate reproduction with robots?
That’s something we do on the computer, and it’s nothing disgusting. We don’t build our embodied robots to reproduce in front of us. What we do is give every robot a genome where the genome is the total sum of the genes that is has. In the case of the kind of robot that we build, called Tadro, which is short for tadpole robot, we wanted to allow three different traits to evolve. The shape of the tail fin, the number of vertebrae bones in the back, and the sensitivity of the robot to a predator it might be avoiding. Each of those three characters has a bunch of gene coding for different qualities, different shapes of tail, different number of bones, and different sensitivity to the robots. That’s all done using computer software.
We then build physically embodied robots, put them together in a tank, and let them behave autonomously, on their own, independently. These are not remote controlled. They have their own nervous systems and they take in sensory input and behave as a result of what that sensory input is. Then we judge them, who is the better feeder and fleer at the same time? We do multiple trials and then take the top behavior fish — kind of like watching the evolutionary Olympics, the bronze, silver, and gold winners are the ones that get to leave their genes in the next generation.
Aren’t robots, or models of any kind, artificial systems that copycat the outward behavior of the biological system? Can you really learn about how an extinct species would have acted in its natural interests by building a robot version of it?
I think it boils down to this: do you think that we can learn anything about an organism or any biological system by creating a model of it? This is a really serious question in all of science. There are some people who say you just have to study the thing itself, whatever you’re interested in. There are other people who say, but the thing itself is either not available to us or it’s so complicated that we can’t study it. So any animal, or any organism — even a single cell organism is far more complicated than our most complex robots. I do think we can [study robot versions] because when we build these robots we pick specific features of the animals we want to study, then we create something that is missing in the fossil records, dynamic behavior — the “what it takes to live” part. We can look at how those biorobots are interacting with each other.
There’s an important caveat here, which is we’ll never know for sure how anything has evolved. The best that I can do with evolutionary biorobotics is to tell what is more likely and what is less likely to have happened. On the other hand, even though I can never know for sure what went on, I can at least circumscribe the possible with the evolutionary robotics approach.
A large part of what you do is reconstructing extinct fish. How are you able to redesign an animal that no living human has ever seen? For example, the Haikouichthys fish that lived over 530 million years ago?
It has to do with what information you bring to bear when you build a model. For the haikouichthys fish, what we have are beautiful 500 million year old fossils from China, but all they show are pieces of anatomy. We look at the anatomy of the fossils, but we also have to look at the physiology of living organisms. Fortunately, some of the nervous control of behavior of fish is highly conserved evolutionarily. In every living species that we see, we see the same kind of neurocircuitry. It’s a pretty good guess when we bet these other early vertebrates probably had it too. We can build a robotic brain using a computer that we think works very much like real living fishes. That’s a guess, but it’s the best guess we’ve got, as opposed to a universe of infinite possibilities.
What do you think Charles Darwin would have to say about evolving robots?
I think he would love it. I think he would look at what we’re doing and in many ways feel vindicated. What Charles Darwin ran into in the late 19th century was resistance to the idea not of evolution itself, which had been talked about for over 100 years, but of natural selection. That was really his contribution to evolutionary theory. The problem was he didn’t have a good theory of genetics, and evolution happens so slowly that the best thing he could do was talk about natural selection that animal and plant breeders did. He started breeding doves, orchids, and pigeons, but people said that’s artificial selection, not natural selection — as though humans aren’t animals, so that was part of the problem at the time.
This is another example that we actually know how to harness evolution. We can put it to work in the case of our robots to actually evolve them over generational time. Not only do we use that mechanism to study biological questions, we can put evolution to work to create new kinds of engineering design as well. That would be totally consistent with the ideas that Darwin was talking about back in 1858.
The U.S. Navy has been funding research on fish and dolphins since 1946. Can evolving robots be used as weapons in war?
Yes. There are drones being operated right now in the Middle East. Military or CIA operations are remote controlled, so there’s a human in the control loop making a decision about where the drone is going to fly or send out a bomb. What people in the military industry (which I’m not in) are talking about is the next logical step, which is to have those drones become autonomous, meaning the human is no longer there telling the drone — or what I think of as a robot — to do. The robot will have a pre-planned mission and will make decisions on its own. It wouldn’t need to wait for instructions for a human to interact on the battlefield.
But that’s exactly what people who write about the ethics of robot war are worried about. They say we have to be very careful when we make our robots in war autonomous because we will be more likely to send them into battle and care less about the human damage they cause on the other side. There are actually proponents for building into these autonomous robots a sense of ethics on how to behave.
The next thing would be to have an autonomous robot in the battlefield that is capable of adapting its behavior. We would call this learning, sort of evolving the software. The next thing after that would be to evolve robot’s hardware on the battlefield. If you allow variations in the shape of the legs or sides of the feet, some robots might do better in the swamp or on sand than the others. With quick turn around time in these robots they could respond to changes in the physical and strategic environment of the battle. There are already over 50 countries that have defense programs that have to do with robots in war. My guess — I don’t have any secret information — is that there’s a secret arms race going on for robotic weapons right now.
What about the future of robotics? Why haven’t robots taken off as much as we predicted they would, in the past?
I think they have taken off. If you have a Ford Focus you can tell your car to be a robot and parallel park. We have autonomous robotic functions that we’re bringing into our consumer lives. We’ve got Google cars that can drive themselves around. We have states in the U.S. that are passing laws allowing autonomous cars to travel. There are vacuum cleaning robots. Auto-pilot on a plane is in essence a robotic artificial intelligence, in fact, planes could land and take off without having a pilot do it. The robots are here, we did it, we brought them into our lives. We built them. They’re all around us.