Robot Kits
Stiquito Kit
BattleKits
Robot Toys
Solar Kits
Robot Arms
Robosapien
Basic Stamp Kits
Lego MindStorms
Beginners
Books
Hobby Robots
Robot Sports
Electronics
Mechanics
Robot Minds
Books for Kids
Robots at Work
Microcontrollers
Advanced Books
Used Books
Real Robots
Robot Motors
Remote Controls
Robot Parts
Robot Tools
Magazines
Robot Videos
Robot News
RoboLinks
Contact
|
Courtesy of New Scientist Magazine
By Clive Davidson
THE SETTING IS SIMPLE: a rectangular
room with its walls painted black except for one white square and one white triangle. In
the center of the room, a video camera hangs from a shaft connected to an overhead gantry.
Below it is a plastic disc connected to a circle of eight, touch-sensitive bumpers. Motors
on the gantry begin to hum, the camera is driven forwards, backwards and from side to side
as it explores the room like an animated Cyclops. After a few moments to orient itself, it
advances steadily towards the triangle and then comes to a stop. The experiment has been a
success.
This "robot" is the brainchild of a three-man team at the Evolutionary Robotics
Group at the University of Sussex. The simple task it has successfully completed - finding
the white triangle in a landscape populated only by one other feature - does not look like
much, as the team members freely confess. But as they explain, "the significant
factor is how we got the robot to solve the problem".
Better bred
The Sussex team used a new approach which might eventually succeed where others have
failed in building autonomous robots that can move around complex environments. Today, the
robot can manage to move around only its blackwalled room: in the future, the goal is a
robot that can handle the ever-changing, real-world environments of offices, warehouses
and hospitals.
The Sussex robot was not programmed to distinguish between a triangle and a square. Its
primitive nervous system was not even wired up so it would pick out geometric objects.
Instead, the robot's brain was made to "evolve". It was "bred" from a
population of robots with randomly designed motor control and vision systems. After being
made to perform in the experimental environment, the "fittest" individuals in
the population - in this case, the ones which came closest to stopping in front of the
triangle - were allowed to "survive", "breed" and "spawn"
the next generation. From a starting population of 30 robots, it took just 30 generations
of artificial evolution over two days before the successful triangle-seeking individual
appeared.
This method of creating intelligent robots differs greatly from the way robot-builders
have proceeded in the past. The traditional artificial intelligence (Al) approach has
attempted to program a robot with sets of rules and knowledge about the world in which it
lives. This places a tremendous burden on the programmer to reduce everything that the
robot needs to know into a set of nonambiguous, utterly precise statement.
A more recent approach has used neural networks - computer systems modeled loosely on the
human brain. The programmer does not write a series of instructions for the network to
carry out. Instead, the machine learns by trial and error to carry out the correct actions
when given certain inputs. During the learning process, the properties of the network are
gradually adjusted until a particular input produces the right output. This
"connectionist" approach has been successful in creating machines that can solve
certain specific problems and has the advantage that the programmer does not have to
specify exactly how a problem is to be solved.
But the problem with both these approaches is that they are difficult to apply to a wide
range of situations. Each set of rules or learned connections relate only to specific
problems. Building robots that can deal with the complexities of the ever-changing, real
world is thus very difficult. As the Sussex team puts it, "If a robot is interesting,
then it's probably too difficult to design." That is why they have turned to
"evolving" their software.
Inman Harvey, Phil Husbands and Dave Cliff are using genetic algorithms as the basic tool
for building their robots. This technique first invented by researcher John Holland in the
1960s while he was working at the University of Michigan, loosely applies the principles
of Darwinian evolution to computer code. This approach sees strings of code which specify
certain structures or behaviors - for example, the way visual sensors on the robot's body
connect to one another - as analogous to chromosomes.
At the start of the process, a number of these chromosomes are generated randomly. The
robots that carry them will "behave" in a variety of different ways. A
"fitness test" is then used to evaluate how well they solve a particular
problem, in this case finding the way to the white triangle in a black-walled room.
Code-bearing chromosomes from the fittest pair of robots are brought together and run
through programs which simulate the recombination of genetic elements that takes place
during reproduction by real organisms. The resulting pool of "offspring" is put
through the fitness test once again - and so on for generation after generation.
Cyclops' Eye
The Sussex group did not, of course, start right at the beginning with the electronic
equivalent of the primordial slime. They began with a species of robot which can move and
see in a rudimentary way. Their robot Cyclops doesn't even have wheels or legs - it
consists of just a video camera eye suspended from a gantry over the scene of the
experiment and connected to a computer. Stepper motors allow the gantry to move
horizontally and vertically and move the eye from place to place. Around the eye is a set
of eight bumpers. These tell the robot if it collides with an object and where the object
is.
This "virtual robot", hanging from the ceiling may seem far from the kind of
robot that could move around on its own. But it isn't really: it is just simpler to work
with in the laboratory. If the team wanted to take the robot out into the real world, it
would be easy enough to use the commands that control the gantry motors to control motors
on real wheels.
The process of evolutionary selection is carried out on the controller - the computer that
processes information from the sensors and controls the motors which move the robot about.
Input from the video camera is first processed to simulate a group of simple
photoreceptors: that is, the video image, which covers a wide field, is divided into a set
of smaller patches. Each patch simulates a simple eye that just looks in one direction and
detects light or dark. The actual number and position of these simple eyes varies from
individual to individual.
Once the virtual robot was built and the laboratory setting was complete - a job that took
over a year - evolution could take place at high speed. In just two days, the robot
controllers were run through 30 generations. During that short time the robot brain
evolved to the point where it could easily guide the robot to the white triangle.
The Sussex team were pleased with such rapid success, particularly because other
researchers had thought the task might prove impossible. Before they pulled it off, Rodney
Brooks, a roboticist at the Massachusetts Institute of Technology had dismissed the
notion, believing that it would be too difficult to test the results of artificial
evolution on physical robots. "The number of trials needed to test individuals
precludes using physical robots for testing the bulk of the control programs produced for
them by genetic means," he wrote in a paper given at the First European Conference on
Artificial Life in Paris in 1991. But now Brooks says of the Sussex researchers:
"They found ways around all of the things I said were too hard." He is
incorporating some of the British ideas into Cog, Brooks' attempt to build a childlike
humanoid.
The Sussex team were also agreeably surprised when they turned to study the design of the
robot which emerged after evolution had done its work. Its control system consists of
computer code which created 11 artificial neurons: three for visual input, five connection
neurons, and three for motor outputs. The design of its visual system was even more
unexpected. Two of the robot's visual sensors were positioned in the center of the robot's
field of view, largely overlapping. The third was placed higher up the robot, but more or
less directly above the first two. The robot identified the triangle by seeking an area
where its two lower overlapping sensors detected brightness - the lower part of the
triangle - while its upper sensor detected only the darkness of the walls outside the
tapering shape of the triangle.
The real challenge now is to see whether the robots can evolve considerably farther. If
you compare the visual task they solved with computers that can find and read car
numberplates from motorway video camera images, or measure and count particles in
microscope images, they are still at a primitive stage.
The team is currently moving the robots one step up the evolutionary ladder by getting
them to find the same triangle when the scene is lit by spotlights, flashing off and on
like disco lights. If the robot can evolve to cope with this changing world, then it may
move on to bigger things. Step by step, it could reach the stage where it could deal with
the more complex shifting scenes inside an office, warehouse or hospital. New skills will
have to be added by the experimenters along the way if the robot is to be autonomous. It
will need a memory, for example. That will enable it to find its way back to wherever it
can recharge its batteries.
After a few million generations of robotic evolution, who knows, the lumbering little
robot of today might be a handy hospital porter - or even Robocop.
Subscribe to New Scientist
|
|