Page 161 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c004 Final Proof page 147 21.9.2005 9:37am

Evolutionary Robotics and Open-Ended Design Automation 147

a leg, motor, or sensor, or combinations of these. In most cases, the estimation–exploration process
is able to reconstruct a new simulator that captures the actual damage using only 4 to 5 trials on the
target robot, and then use the adapted simulator to evolve compensatory controllers that recover
most of the original functionality. There are numerous applications to this identification and control
process in other fields.

4.5.2 Making Morphological Changes in Hardware

An evolutionary process may require a change of morphology or production of a new physical
morphology altogether. One approach for generating new morphology is to use reconfigurable robots
(Yim et al., 2002). Reconfigurable robots are composed of many modules that can be connected,
disconnected, and rearranged in various topologies to create machines with variable body plans.
Self-reconfigurable robots are able to rearrange their own morphology, and thus adapt in physical
reality. Figure 4.9a shows one example of a self-reconfiguring robot composed of eight identical
cubes (Zykov et al., 2005). Each cube can swivel around its (1,1,1) axis, and connect and disconnect
to other cubes using electromagnets on its faces. Though this robot contains only eight units, it is
conceivable that a future machine will be composed of hundreds and thousands of modules of smaller
scales, allowing much greater control and flexibility in morphological change. As scales decrease,
one may need to switch from classical deterministic reconfiguration processes to stochastic processes
that exploit Brownian motion, not mechanisms in the biological cell. Figure 4.9b shows some large
scale robot prototypes which operate on these stochastic principles (White et al., 2004).
An alternative approach to varying morphology is to produce the entire robot morphology
automatically. For example, the robots shown in Figure 4.4f were produced using rapid prototyp-
ing equipment: These are 3D printers, that deposit material layer by layer to gradually build up a
solid object of arbitrary geometry, as shown in Figure 4.9c. This ‘‘printer,’’ when coupled to an
evolutionary design process, can produce complex geometries that are difficult to produce any other
way, and thus allow the evolutionary search much greater design flexibility. Nevertheless, even
when using such automated fabrication equipment, we needed to manually insert the wires, logic,
batteries, and actuators. What if the printer could print these components too? Future rapid
prototyping systems may allow deposition of multiple integrated materials, such as elastomers,
conductive wires, batteries, and actuators, offering evolution an even larger design space of
integrated structures, actuators and sensors, not unlike biological tissue. Figure 4.9d shows some
of these printed components (Malone and Lipson, 2004).

4.6 THE ECONOMY OF DESIGN AUTOMATION

The examples shown so far are all related to design and fabrication of robotic systems, but the
principles described here are applicable in many other domains. Is there a way to know a ´ priori
where these methods will be successful? Several decades of experience have shown that there are a
number of conditions that suggest such problem domains.

. Known physics. Most evolutionary systems use some form of simulation to determine the conse-
quence of various design choices. Evolutionary algorithms are fruitful when the physics are
understood well enough that simulations are predictive, there is no question about the underlying
physical phenomena, and that simulation can be carried out in reasonable time.
. Well-defined search space. The basic ‘‘atomic’’ building blocks comprising potential solutions are
known, and it is clear how they are allowed to fit together. These two aspects define a search space
in which the evolutionary algorithm can operate. Knowing the building blocks and interfaces does
not imply knowing the solution. It is important to realize that ‘‘building blocks’’ are not necessarily
discrete components — they can be features of a solution or partial solutions.