This article first appeared in Personal Computer World magazine, March 1999.
SURELY ANYTHING called "Robo-lobster" can't be serious -- it must be one of those clumsy monsters Godzilla is always fighting, knee-deep in swirling water, the inevitable charred cardboard cityscape smoking in the background. But no -- it is serious. It's one of a new breed of artificial creatures which exhibit unusually complex behaviours. They're called "biobots", and they may be the missing link between biology and artificial life.
Robo-lobster spends most of its time submerged in a tank at Boston's Massachussets Institute of Technology. The brainchild of Dr Thomas Consi, it's a shoebox-sized self-contained robot designed to mimic the ability of lobsters to sense chemicals in the water around them. What makes it unusual is that it's programmed using techniques inspired by how biological systems work. This is "soft computing", where answers don't need to be precise, and where logic is fuzzy.
And robo-lobster now has a cousin: cyber-cricket. Researchers at Edinburgh University's Mobile Robot Group have built a robot which successfully mimics the behaviour of female crickets which fly towards males who sing a mating song. The cyber-cricket doesn't look much like a real cricket. It's actually a three-wheeled laboratory robot called a Khepera, about half the size of a computer mouse. Its "brain" is a Motorola 68000 processor running a neural network.
Neural nets have their origins in attempts to model the real nerve cells found in animal brains. The net comprises a number of interconnected simple processing elements, each of which has a set of inputs, and a set of outputs. Each of an element's inputs has a certain "importance level", or "weight", associated with it, and when the weighted sum of any signals present on its inputs exceeds a certain threshold value, the element "fires" -- just like a real brain-cell -- and sends a signal to its outputs. The outputs then transmit their signal to other elements in the net, which behave in a similar way -- and so on. Eventually, the outermost elements in the net signal the "answer". All this usually happens in a software simulation.
What makes neural nets special is that they are very good at recognising patterns in the data fed into them. This is achieved by an initial "training" period, supplying the net with sample datasets, and repeatedly revising the weights of the inter-element connections to improve the net's response. As Artificial Intelligence pioneer Marvin Minsky puts it, neural nets make their decisions by "weighing the evidence", rather than obeying fixed sequences of program instructions like traditional computers.
The Edinburgh team have fitted tiny microphones to the cyber-cricket, and trained its neural net to respond only to the calling song of males of a particular species of cricket -- Gryllus bimaculatus. Placed in an environment containing random noise and also real crickets chirping their love songs, the cyber-cricket successfully recognises the real crickets, and also trundles toward them. What's amazing is that the cyber-cricket's neural net has only four elements, two for its left side, and two for its right. Each pair of elements is interconnected; their inputs are fed from the robot's ears, and their outputs control the robot's wheels. Part of the neural net's design is based on studies of a real cricket's neural architecture.
Machines like robo-lobster and cyber-cricket are exciting because they demonstrate that multiple behaviours can arise from single mechanisms. For example, it's long been believed that the cricket must use two information processing systems for performing the separate tasks of recognising a male cricket's call, and then locating its position in space in order to move towards it. The cyber-cricket, however, uses its single neural net for both jobs.
But we must be careful about drawing conclusions. Just because a robot equipped with a simple neural net can exhibit the same behaviour as a biological organism, it doesn't follow that the organism works in the same way. It might do; it might not. But what it does show is that unexpectedly complex behaviour can result from very simple processes. It's a pat on the back for the theory of evolution, too, since the neural net "learns" using genetic algorithms, which simulate the copying and mutation processes of DNA.
If the robot researchers are on the right track -- and it looks like they are -- more complex artificial creatures will no doubt soon be scuttling about. Let's hope Godzilla isn't one of them.
Toby Howard teaches at the University of Manchester.