Robotics Reference
In-Depth Information
search is under way in the Golem
7
project at Brandeis University, under
the direction of Hod Lipson and Jordan Pollack. Their “parent robots”
consist of a computer running an algorithm that simulates evolution and
produces a design for new robots based on trial and error. This is linked
to a three-dimensional printer that makes small plastic shapes. The off-
spring are small plastic trusses (made of bars) with actuators, propelled
by motors and controlled by neural networks.
8
Artificial neurons are
the building blocks of their control modules (their brains). Bars con-
nected with free joints can potentially form trusses that represent arbi-
trary rigid, flexible and articulated structures. The bars connect to each
other through ball-and-socket joints, neurons can connect to other neu-
rons through synaptic connections, and neurons can connect to bars.
Human intervention is only necessary to attach the motors and connect
the wires—the robots do all the rest of the assembly themselves, includ-
ing telling the humans what to do.
Lipson and Pollack's simple thermoplastic components were employed
in the development of what they call a polymorphic robot, a machine
that can change its shape to suit the job in hand.
9
The Golem robots
had one simple physical task to perform—they had to discover a way to
move.
Starting with a population of 200 “machines” that were comprised
initially of zero bars and zero neurons, Lipson and Pollack simulated
the evolutionary process. The fitness of a machine was determined by
its locomotive ability—how far its centre of mass moved within a fixed
amount of time. At each stage of the process, fitter machines were se-
lected and offspring were created by adding, modifying and removing
building blocks, and these new machines replaced the less fit ones in the
population. This process typically continued for 300 to 600 generations,
with both the body (the shape of the machine) and its brain (control
module) evolving simultaneously.
Typically, several tens of generations passed in the experiments be-
fore the first robotic movement occurred. For example, as a minimum,
a neural network must first assemble and connect to an actuator in or-
der for any motion at all to be possible. Selected robots out of those
with winning performances were then automatically replicated into re-
7
Genetically Organized Life-Like Electro-Mechanics.
8
Shortly before this topic went to press in 2005, Lipson and Pollack announced that they had
also developed a robot whose offspring were exact copies of themselves.
9
See the section “Robots that Change Their Own Shape” in Chapter 8.
