Simple Neural Networks For Robots Robot Books

Kits and Toys
 Robot Kits
 Stiquito Kit
 Robot Toys
 Solar Kits
 Robot Arms
 Basic Stamp Kits
 Lego MindStorms

 Beginners Books
 Hobby Robots
 Robot Sports
 Robot Minds
 Books for Kids
 Robots at Work
 Advanced Books
 Used Books

More Robotics
 Real Robots
 Robot Motors
 Remote Controls
 Robot Parts
 Robot Tools
 Robot Videos
 Robot News

Courtesy of New Scientist Magazine

If just a handful of neurons is all it takes to make a robot act like an insect, is there less to animal behavior than meets the eye, asks Duncan Graham-Rowe

IN A DARK LABORATORY, a solitary creature shuffles around, searching, lured by the love song of a male she will never meet. She has spent her entire existence in these unnatural surroundings and is doomed never to see the outside world, forced to comply with the whims of researchers. Yet she is unlikely to excite the wrath of animal rights activists nor provoke much sympathy from anyone. Indeed, hers is barely an existence at all since she consists of only a few wheels and a bundle of electronic components. Still, at least it's one less cage to clean out.

The cybercricket's unassuming appearance belies what is a landmark in robotics. It behaves just like a real cricket-and not just outwardly. It simulates the cricket right down to the neurons, and is one of the first attempts to reproduce the pattern of neural signaling found in a living creature. It's also proving to be a challenge to biologists, leaving them open to the charge that they read too much into animal behavior, and see life as altogether more complex than it is.

Decades of research have allowed biologists to theorize that certain behaviors are associated with certain patterns of neural activity. But while they can observe behavior and record neural signaling, to prove that the two are causally linked is extremely difficult. This is where robots are starting to make their mark.

The key to these "biobots" is a better understanding of how natural and artificial neural systems work. Roboticists can now mimic in metal and silicon the natural pathways that link optical and acoustic sensors-eyes and ears-to motors, which act as surrogate muscles. This gives biologists the chance to simulate particular patterns of neural activity and see whether or not they prompt the anticipated behaviors.

The first biobots were very simple, designed to test biological theories. For example, ethologists believe that in order to navigate, rats construct cognitive maps of their surroundings. One popular way to test this theory was to place a rat in an open rectangular box and train it to return to a single location where food had been hidden. Researchers would then place the rat in a different but identical box to see whether it would go to the same spot in search of food. Sure enough, it did. The widely held conclusion from this experiment was that the rat remembered the location of the hidden food by relating it to the geometry of the box. In other words, it constructed a cognitive map.

But Henrik Lund, a computer scientist at Aarhus University in Denmark, and Orazio Miglino, a psychologist at the University of Naples, were not convinced that this conclusion could be drawn from the rat's behavior. They thought ethologists were reading too much into the findings. They wanted to make a robot that could mimic the rat's behavior, without a memory. If the robot had no memory, they reasoned, it could have no map.

Lund and Miglino chose a standard lab robot called a Khepera, a small three-wheeled device that looks like an electronic biscuit. They added two infrared sensors that act like radars, four touch sensors and a pair of motors to power the two drive wheels. The robot's behavior was dictated by an artificial neural network, a cluster of simple mathematical processors, or nodes, that are designed to behave like nerve cells in the brain.

But while rats have millions of neurons, the robot had only 10 nodes. Its sensors connected directly to eight of the nodes, which in turn fed two nodes that controlled the motors. With these basic resources, the researchers then tried to fashion a machine capable of finding a particular spot in a box using a computer program called a genetic algorithm

In an artificial neural network, a node adds together the signals arriving at its inputs, and when the sum reaches a threshold it fires off a signal to other nodes (or to one of the motors). In addition, every input to a node is given a weighting which either increases or reduces the incoming signal. A genetic algorithm treats the nodes' thresholds and weights as though they are genes being passed from one generation to the next.

It creates dozens of different potential robot designs with randomly chosen genes-values for the thresholds and weights-and tests each one in turn to see how well it finds the selected spot. It then allows the best performers to "breed" by swapping genes with other successful robots. The new generation of machines is then tested again. Over many generations, this process allows the best robot for searching out the target to evolve.

Lund and Miglino found that their best robot could locate the food's position as easily as a rat, simply by seeing and feeling its way round the box. "We wanted to . . . look for the minimum conditions to solve this task," says Lund. Their success showed that the open box experiment was not enough on its own to support the theory that rats use cognitive maps. Personally, Lund believes that rats do construct these maps, but he wanted to show that biologists were reading too much into their results.

Inside the animal

Such experiments can be carried out by computer models, but Lund argues that these are no substitute for the real world. "One main advantage of using a real robot in a real experiment is that one is sure to have the same environmental stimuli as in the real animal experiment," he says. Even so, the minimal robot approach is of only limited value because it reveals nothing about what goes on inside the animal. The cybercricket, by contrast, is designed to do just this.

Female crickets home in on potential partners by listening to their song-the sound we hear as chirping. The sound is made of short bursts, or syllables, of a single "carrier" frequency which is specific to different species. This behavior, called phonotaxis, has been studied for decades in the cricket Gryllus bimaculatus, both by watching its behavior and by studying its nervous system. But showing how the two are linked has proved very difficult.

Immobilized or sedated

"It's not impossible, but it's definitely challenging and not many people are in a position to do it," says Barbara Webb, a psychologist at the University of Nottingham and one of the architects of the cybercricket. A living animal must have electrodes placed in its brain to record activity from selected neurons. It is then placed in a spherical treadmill so its movements can be tracked without it going anywhere. "You have to immobilize the animal or sedate it to keep the electrodes in place," explains Webb. "But the question is, can it still behave normally?"

The ideal way to study the insect would be to isolate any neurons thought to be relevant to phonotaxis from the rest of the brain, play some cricket love songs and see what happens. But this just isn't practical. Webb reasoned, however, that with a robot it might work. "I wanted to find an animal system that I could build a complete robot model of. Where the neurons would control the behavior when implemented in the real world," she says.

Webb began building the robot in 1991 while at the University of Edinburgh with John Hallam. Later, they were joined by Lund. As if mimicking the cricket's neural system was not tough enough, the team also had to copy its bizarre auditory system. G. bimaculatus has two ear drums on its forelegs that are connected to two holes in its body via internal tubes (see Diagram). This structure sets up phase differences in front of and behind the ear drums that allow the cricket to tell which direction a sound is coming from. But this works only if the sound is at the carrier frequency of the species.

To simulate the auditory system's performance, Webb and her colleagues used a collection of amplifiers and delay lines, with four microphones positioned on a Khepera robot. A digital signal from these "ears" fed into a neural network. "There was a particular pair of neurons that I copied," says Webb. Neuroscientists had shown that these were crucial to phonotaxis. "I actually tried to tune them so that the model neurons had the same firing patterns as the cricket ones do," she says.

Next, the researchers added extra nodes to see how many more the system needed before it displayed phonotaxis. Biologists had speculated that it would need 20 or so. "But it turns out you only need to add two," says Webb. In her network, the left input node feeds an output node which controls the Khepera's left motor-the same arrangement is repeated on the right side. To complete the network, each input node inhibits any signals to the opposite output node (see Diagram)

This surprising simplicity has been a hallmark of the cybercricket project. The robot was conceived as a way to test the popular theory that phonotaxis needs two neural control systems-one to recognize a male's call and the other to locate it. But when Webb played a call song to the four-node robot, using the same experimental set-up as that used for live crickets, not only did it recognize the sound but also moved towards it. "The two systems are just one," says Webb.

This unexpected bonus raised the question of whether a real cricket's neural system is simpler than anyone had supposed. To investigate this, the team decided to see if the automaton could replicate other life-like behaviors. They played the robot two call songs that it could recognize. Live crickets prefer songs with a faster syllable rate. So did the robot.

For a cricket to choose between two suitors, some biologists argue that crickets must have their preferences somehow encoded in their brains. But the robot showed this was unnecessary. The combination of the right auditory system and simple neural network selects call songs with faster syllable rates automatically.

This suggested to Webb that her simpler neural system is the same as that of a living cricket. Biologists, she says, seem to overcomplicate things, partly because they anthropomorphise. "People look at behavior and assume that there are intentions involved like recognition and goal seeking," she says. "When they look to the neurons they try to find neurons that correspond to those intentions." The problem is that the neurons might not exist, because other animals do not work like humans.

But not all biologists are convinced of the robot's relevance to a real cricket. "There is no problem in studying the behavior that emerges from simple systems," says Gerald Pollack, a neuroethologist at McGill University in Montreal who has worked on crickets. "The problem is [that] most biological systems are not simple," says Pollack, who argues that even if the robot cricket mimicked the animal's behavior perfectly, it still wouldn't constitute evidence that robots and animals are doing it the same way.

Still, as the cybercricket displays more real-life behaviors, Webb feels that her case improves. "You might say one behavior is not very hard to reproduce," says Webb. "But if you carry out a whole series of different experiments and you replicate all of them, then it starts to become a stronger and stronger argument."

But she recognizes that studying behavior alone can take her only so far. The robot allows her to make predictions about what neuroscientists should find, and ideally she wants someone to find out if her predictions are correct. If the firing patterns and behaviors are identical in creature and Khepera, can there still be doubt that they work in the same way?

Whatever the outcome of this debate, many scientists agree that, at the very least, robots are valuable research tools. This is the attitude taken at the University of Zürich in Switzerland, by a team made up of roboticists and zoologists. Their target is ant navigation.

Ant bearings

The desert ant Cataglyphis forages up to a couple of hundred meters from its nest and returns home almost unerringly in a straight line. An essential part of this ability, say biologists, stems from the insect's skill in using polarized light as a compass. As light from the Sun is scattered by molecules in the atmosphere, it is partially polarized. This effect creates a pattern in the sky that is symmetrical about the solar meridian, a line that cuts through the Sun and its zenith point (see Diagram)

Cataglyphis has receptors in its compound eye that are sensitive to different orientations of polarized light, and researchers believe that these feed three neurons-called POL neurons. Each neuron gives a peak output for a different orientation, measured as the angle between the direction in which the light is polarized and the axis of the ant's body. Yet precisely how the ant makes use of these signals is still unknown, says Dimitrios Lambrinos, one of the team's roboticists.

One theory to explain the ant's abilities, called the scanning model, is based on just one POL neuron. It suggests that when an insect goes foraging, it first rotates through 360 degrees until it locates the meridian. The effect is the same as looking through two polarized lenses and rotating one until the light coming through both is brightest. Once the ant has this fix, it heads off in its chosen direction.

To test such theories, the team built Sahabot 2, a six-wheeled "ant" with three sets of polarized light sensors, to simulate those in Cataglyphis. For the POL neurons, the researchers took a "black box" approach. They didn't want to simulate the actual firing of neurons, as Webb had done, they merely wanted to reproduce the same inputs and outputs to and from the neural system. This job fell to three amplifiers, each one tuned-just as in real ants-to give a maximum output at different angles between the direction of polarization and the robot's axis.

This year, Lambrinos and his colleagues let the robot loose in Tunisia, where Cataglyphis lives. In all the experiments, the cyberant behaved very similarly to real ants, travelling more than 100 meters and returning to its start point with an error of less than 60 centimeters.

But the robot performed worst when programmed to follow the scanning model. In the noisy environment of the desert, each amplifier generated only a flattened peak, which made finding the maximum point-the ant's reference direction-a relatively tough task. But what, asked the researchers, if they combined the outputs of all three amplifiers? By subtracting the output of two amplifiers from the third, the team created a sharply defined peak that the robot sensed with greater accuracy. When the robot trundled off with this new program onboard, its error rate fell.

It was a revelation, says Lambrinos. Previously, no one understood why there were three POL neurons, since the scanning model needed only one. It took a test with a robot in the same environment as living ants to provide a better model.

To go further, the team wants to follow Webb's lead down to the level of individual neurons-to study the firing patterns in the tiny ant brain and to recreate them in silicon. This may be easier for the Swiss team than for Webb because it is multidisciplinary. "One of the reasons this is so successful is that we are working together," says Lambrinos. "If we don't agree with each other then we know other biologists won't agree with us either."

Robot modeling is fast becoming a valuable tool for biologists to test theories and create new ones. And it will continue to grow in popularity, according to Webb, not least because of the time savings. "It takes a lot more time to find out what one neuron does than to build one," she says.

And copying nature may hold lessons not just for biologists, says Webb. The cricket project suggests that computer scientists could get far more out of their neural networks if they tried. "Someone making a standard neural network would never bother making a network with only four neurons because they would assume that it simply wouldnt be enough," she says. But as the lovelorn cybercricket found, sometimes four is all you need.

Subscribe to New Scientist



Advertise your product on

Beginners Books  |  Hobby Robots  |  Robot Sports  |  Electronics  |  Mechanics  |  Robot Minds  |  Robot Fiction
Books for Kids  |  Robots at Work  |  Mars Robotics  |  Advanced Books  |  Recommended  |  Roboxers  |  Robot Kits
Solar Kits  |  Robot Arms   |  Robosapien  |  Basic Stamp  |  BioHazard  |  Robot Toys  |  Muscle Wires  |  Lego Mindstorms
Real Robots  |  Robot Motors  |  Robot Tools  |  Microcontrollers  |  Used Books  |  Robot Parts  |  Magazines  |  Holdem
Robot Videos  |  Robot News  |  RoboLinks  |  Contact