SOME 350 MILLION years after the first amphibians crawled out of the muck, high-tech avatars are retracing those first steps. At USC, a robotic salamander – plunked down in a virtual pond and wired up through an evolutionary computer program – is mimicking the way real salamanders clamber along the ground, glide through the water and easily switch between these two distinct gaits.
“We are moving toward a comprehensive model of a lower vertebrate,” says the robot’s creator, USC computer science postdoctoral fellow Auke Ijspeert. The study may lead not only to breakthroughs in understanding vertebrate locomotion, but to advances in mobility for real-world robots.
The virtual salamander’s nervous system doesn’t contain an explicit program for either of its gaits, says Ijspeert. Nor does it run a program to determine when or how it should switch between gaits. Instead, the pattern of activity across the entire neural network (the robot’s central nervous system) spontaneously produces one or the other gait from different inputs – such as whether the salamander senses water around its body or solid ground under its feet.
This kind of decentralized design philosophy – in which perception leads directly to action without an executive command program – is the hallmark of the biomimetic (“life-imitating”) approach to robotics. As biomimetic technology improves, according to proponents, robots will become more flexible and adaptable.
THOUGH PHYSICALLY unreal, Ijspeert’s salamander is no mere animation. Within its virtual world (rana.usc.edu:8376/~ijspeert/salamander.html), it is a solid entity – contending with gravity, the land’s friction and the water’s viscosity. It must use its “body” to propel itself through different surroundings, and it must evolve these skills using spinal cord circuitry adapted only for swimming.
Ijspeert began by teaching the salamander to swim based on the neural circuitry of a primitive fish. The next step, as in evolution, was to make a small modification. Ijspeert added legs, grafted a few new computing nodes onto the network, simulated thousands of generations of natural selection, and watched a dedicated swimmer evolve into a walker.
“The model we have does indeed produce the two different kinds of gait seen in the salamander,” says Ijspeert. “What’s more, it produces neural activity – output – that is very similar to [electro-muscular] recordings in the real salamander.”
The robot salamander’s journey up the evolutionary ladder has only just begun. The next step will be to give the salamander more sophisticated systems for vision and sensory feedback.
“Amphibians are a good choice for study,” says Michael Arbib, chair of computer science in the USC School of Engineering. “They are simple enough that we can now develop detailed models of their behavior. At the same time, their brains form the ground plan for the more elaborate brains of higher vertebrates.”
LEARN MORE ABOUT IT
Digital Darwinism
AUKE IJSPEERT isn’t exactly playing God, but he’s certainly playing Darwin. The USC researcher started by modeling the neural network of his robotic salamander on the nervous system of real salamanders. Where biological data was incomplete, Ijspeert bridged the gaps with the help of a genetic algorithm – a computer program that evolves over thousands of generations to improve a system’s design.
The algorithm may actually advance biologists’ understanding of how real salamanders walk. Ijspeert is now collaborating with an expert on salamander neurology to see if the virtual salamander can inspire biological experiments of how the real creature moves.
“In the future, evolutionary programming will become part of the technological repertoire,” says USC computer scientist Michael Arbib, who works closely with Ijspeert. “In many cases we will no longer explicitly program a computer or robot, but we will adapt it to its environment. Then, of course, the computer scientist will have to become more like a biologist.”
|
|
