Content uploaded by Roderich Groß
Author content
All content in this area was uploaded by Roderich Groß on Dec 21, 2015
Content may be subject to copyright.
Auszug aus: Künstliche Intelligenz, Heft 4/2008, ISSN 0933-1875, BöttcherIT Verlag, Bremen, www.kuenstliche-intelligenz.de/order
Self-Assembling Robots
Roderich Groß
Self-assembly is a process by which pre-existing components organize into patterns or structures without human
intervention. Such processes are responsible for the generation of much of the order in nature. This thesis investigates
the use of self-assembly in autonomous mobile robots, and relates the findings to the biological literature.
1 Introduction
One of the grand challenges of robotics is the design of
robots that are adaptive and self-sufficient. This can be cru-
cial for robots exposed to environments that are unstruc-
tured (in space and time) or not easily accessible for a human
operator, such as the inside of a blood vessel, a collapsed
building, the deep sea, or the surface of another planet. Mod-
ular reconfigurable robots are among the most flexible robots
that exist. They are made of one or a few types of compo-
nent modules which can be connected into many distinct
topologies. Therefore, exploring a limited set of modules, it
is possible to set up a robot with context-dependent mor-
phology.
An interesting category of modular robots are self-
reconfigurable robots, which can autonomously transform
between different morphologies [8]. For instance, a self-
reconfigurable robot could adapt its locomotion strategy by
transforming from a snake morphology (which could offer
advantages when navigating through narrow passages) to
a hexapod morphology (which could offer advantages when
navigating uneven terrain) and vice versa. In many of the cur-
rent implementations, self-reconfigurable robots are initially
manually assembled and once assembled, they are incapable
of assimilating additional component modules without ex-
ternal assistance. In our view, this lack of autonomy is a se-
vere limitation to the adaptivity and self-sufficiency of these
robotic systems. In contrast, this thesis focuses on reconfig-
urable robotic systems whose components are capable of
self-assembling autonomously. Thereby, the components can
in principle set up modular robots of arbitrary size, composi-
tion, and function.
2 Methods
Natural self-assembly processes are our primary source of
inspiration [10]. Of particular relevance to our study in au-
tonomous robots are processes involving macroscopic com-
ponents, such as social insects like ants or bees [2, 4]. We
follow the principles of swarm intelligence [3], aiming at sys-
tems that are fault tolerant, robust, and scalable. We consider
robots of “identical” hardware and with decentralized con-
trol. The robots make little use of memory and take actions
on the basis of local information. The control policies (e.g.,
artificial neural networks) are designed in simulation using
evolutionary algorithms [6], and subsequently ported onto a
physical system. Their performance is assessed in a range of
different conditions, and compared with the performance of
reference strategies and with a lower/upper bound perfor-
mance.
3 Contribution
We review half a century of research on the design of systems
displaying self-assembly of macroscopic components. We re-
port on the experience gained in the design of 22 such sys-
tems, exhibiting components ranging from (externally pro-
pelled) passive mechanical parts to (self-propelled) mobile
robots. We present a taxonomy of these systems, and discuss
design principles and functions.
We then focus on systems in which the components that
assemble are (self-propelled) mobile robots. Previous work
in mobile robotics has focused on self-assembly per se, that
is, on the process by which structure forms through inter-
actions of specifically designed robots. Instead, we look at
self-assembly as a mechanism that helps robots to accom-
plish autonomously concrete tasks. In particular, we address
a simple object manipulation task—the group transport of a
heavy object.
In a first study, we simulate robots that have very lim-
ited acting and cognitive abilities. They can neither per-
ceive teammates nor communicate with them directly. Us-
ing an evolutionary algorithm, we train groups of these
robots to accomplish a transport task. The underlying ob-
jective function does neither explicitly reward the robots
for self-assembling, nor does it impose any bias concern-
ing the spatial organization of the robots during task per-
formance. Nevertheless, self-assembly behaviors evolve and
in many cases are the most effective. The “emergence” of
self-assembly is a striking result, confirming that such ca-
pability (as in social insects) can provide adaptive value to
the group. The analysis reveals a variety of proximate mech-
anisms that cause coordinated behavior in groups. Interest-
ingly, some of these mechanisms are also exhibited in groups
of robots that were trained for solitary task performance (bi-
ologists reported that, in some species, individuals show no
difference in behavior when engaged in solitary and group
transport, for example, see [9]). As a result of this, we hy-
pothesize that in some species group transport has evolved
from solitary transport, presumably from situations in which
solitary transporters, without being aware of each other, co-
operatively transported a common load. Further analysis of
our system shows that the transport is relatively ineffective
when the assembled structures are large. This might be a
consequence of the low degree of mobility of our simulated
robots; animals that are subject to similar limitation do not
form large, self-propelled structures either.
Page 1
KI 4/08
61
Dissertation
