Noel Sharkey is engaged by a take on the intriguing overlap between biology and robotics.
Darwin's Devices: What Evolving Robots Can Teach Us About the History of Life and the Future of Technology
A book on robotics by a marine biologist sounds a bit fishy, but Darwin's Devices is anything but. John Long takes us on a journey through the wonderful, oceanic world of research on the evolution of the vertebrae of extinct species.
Long's work is innovative because of his use — and strong defence — of modelling with physically embodied robots, rather than the usual software simulations of computational biology. He is also unusual, as a biologist, for the way that he exploits artificial evolutionary methods called genetic algorithms to test evolutionary hypotheses.
Long's accounts of justifying his research to sceptical fellow biologists contain both triumphs and difficulties. He relates how, when a colleague asked what robots have to do with biology, he replied that they are used to model extinct species — and how he realized, “with dread”, that the question would haunt his research. This book is his answer.
The opening chapters lay a solid foundation for the use of robots to model biological theories, underlining why testing models in the physical rather than the simulated world is important. In the rest of the book, Long describes a series of his experiments with the robots he dubs evolvabots. Each leads on to the next, adding only what is necessary to get to the next level: a virtue in any scientific model.
Long's chatty style made me laugh out loud at times. But beneath the levity lie robust and sometimes powerful arguments about biomimetics. He moves from his childhood love of fish to his graduate work on the biomechanics of marlin vertebral columns. He gives us an accessible run-through of evolution, then moves on to the design issues of engineering his evolvabots, including key decisions about which features of the animal to include. His guiding principle follows Albert Einstein's famous dictum to be as simple as possible, but no simpler.
“Who needs a brain when you have a smart body?”
The brain alone, Long shows, is not sufficient to explain behaviour. As he asks, “Who needs a brain when you have a smart body?” He argues that brainless robot bodies moving through the real world can exhibit seemingly complex cognitive behaviour with little computation. Few have demonstrated this point better than biological cyberneticist Valentino Braitenberg, in his book Vehicles: Experiments in Synthetic Psychology (MIT Press, 1984), and Long makes ample use of Braitenberg's core ideas. Robotics researchers such as Rolf Pfeifer, with his morphological computation, have carved out this territory, but Long offers new twists — such as questioning where the brain begins and ends.
By chapter six, sleeves rolled up, we are at the nub of the science. It is time for the test results of Long's key evolutionary hypothesis: “selection for enhanced feeding performance and predator avoidance would increase the number of vertebrae” for robot models of extinct fish. Long is in his stride here and gets down to the technicalities, telling us much about the real evolution of sea creatures.
I was intrigued to see how a biologist might use genetic algorithms to test specific hypotheses about bodily evolution and its impact on behaviour. Such algorithms have been used widely in biologically inspired robotics to create artificial gene strings that determine how robots behave; they allow researchers to run through thousands of machine generations in hours. For each generation, the best robots are selected using a mathematical fitness function rather than natural selection, and their genes are incorporated to make the next generation and simulate breeding.
But in Long's experiments, the evolving artificial genes do not directly change the behaviour of the robot fish. Instead, they add vertebrae to the robot bodies, indirectly altering the behaviour of the robots to make them better at feeding and avoiding predators. This drives home Long's point that behaviour is created in the interaction between the physically embodied robot and the world.
There are minor problems with Long's discussion of genetic algorithms. He seems to think that they tell him more about evolution than may be the case. At one point, he writes: “I think Darwin, a keen observer, would have loved watching our robots evolve.” But although Darwin would have been intrigued, I doubt he would have gained much scientifically from pure observation of these robots. Unlike natural selection, a genetic algorithm has an experimenter-designed fitness function, which demands a God-like decision about what features are important in choosing the fittest artificial gene strings.
Often, if a genetic-algorithm experiment does not work as hoped, we can make minor modifications to the fitness function and selection procedure until we get the desired results. So refuting a hypothesis is not always meaningful, because results can be dependent on the experimenter's design. A positive outcome, however, has different implications: if a genetic algorithm produces a simple model or mechanism that fits current data from a real creature, we have a scientific hit.
Aside from that, and a final chapter that tries to cram in too much — it includes 20 pages on the military uses of robotic fish, in which the arguments seem rushed and a little naive — this is a sound and hard-hitting work. It is also an insider's view of the scientific world: an honest account of the cut and thrust of academic ambition. Because Long questions his own methods and motivations throughout, his book lacks the pomposity of many texts that motivate through weak appeals to scientific method.
I read this book for the robotics, but I learned more about fish and the evolution of their bodies. Darwin's Devices represents a step forward in biomimetics. And, cleverly hidden among the discussions and the humour, gems of scientific philosophy shine.