Professional Documents
Culture Documents
Collective Construction
Collective Construction
!"##$%&'($ !"*+&,-%&'"*
Josefa Z. Hernndez, Nik Swoboda
{phernan,nswoboda}@fi.upm.es
Master Universitario en Inteligencia Artificial
Molivalion
! Social insects collectively build nests whose
architectural complexities exceed the
perceptual and cognitive capabilities of single
individuals
! Scientific question: How these insects collectively build the
structures they do?
! Engineering question: How one could design and program a
system in which an artificial swarm collectively builds complex
structures?
AIicalions
! Deal with traditional construction problems: low efficiency,
high accident rates, low quality, and skills shortages
! Facilitate the production of low-cost housing
! Introduce automation in settings difficult or dangerous for
humans to work in, e.g. other planets, underwater.
! Maintain and repair structures during their lifetimes
! Modify structures according to changing circumstances
Hov insecls do il`
! Models of positive and negative feedback have been used to
explain such engineering feats
! Stigmergic communication: the perception of the result of
previous work triggers specific construction behaviors
" quantitative (continuous) stigmergy: the different stimuli that
trigger behavior are quantitatively different (e.g. termite-mound
building)
" qualitative (discrete) stigmergy: stimuli can be classified into
different classes that differ qualitatively (e.g. wasp nest building)
(Brooks, 86): Proposed for robots and extended
into a new view of AI
! It is a hierarchy of task-accomplishing behaviors (situation # action rules)
! Multiple behaviors may fire at the same time
o Layered hierarchy
o Lower layers behaviors inhibit higher level ones
! The resulting systems are, in terms of the amount of computation they do,
extremely simple
! Some of the robots do tasks that would be impressive if they were
accomplished by symbolic AI systems
Svarm Robolics: Iirsl Sles
Steels Mars explorer system (Steels, 90), using the subsumption
architecture achieves near-optimal cooperative performance in
simulated rock gathering on Mars domain:
The objective is to explore a distant planet, and in particular, to
collect samples of a precious rock. The location of the samples is
not known in advance, but it is known that they tend to be
clustered. A number of autonomous vehicles are available that
can drive around the planet collecting samples and later re-enter
a mother ship spacecraft to go back to Earth. There is no detailed
map of the planet, although it is known that the terrain is full of
obstacles hills, valleys, etc.- which prevent the vehicles from
exchanging any communication
Svarm robolics: Mars exIorer syslem
The system uses two mechanisms:
! Gradient field: determine direction to mother-ship via
radio-signal gradient
! Indirect communication: by dropping beacons
(crumbs). Inspired by ant-populations communication
through pheromones.
Svarm robolics: Mars exIorer syslem
! Individual robots behavior:
(1) IF detect an obstacle THEN change direction
(2) IF carrying samples AND at the base THEN drop samples
(3) IF carrying samples and NOT at the base THEN travel up
gradient
(4) IF detect a sample THEN pick sample up
(5) IF true THEN move randomly
! Subsumption hierarchy: (1) < (2) < (3) < (4) < (5)
! The solution is cheap (minimal computing power) and robust
(loss of a single robot will not affect overall system too much)
Svarm robolics: Mars exIorer syslem
: If sample is found, drop crumb trail while
returning to ship (special rocks appear in clusters!). Other robots will
weaken trail on way to samples. If sample cluster is empty $ no trail
reinforcement $ trail dies.
(1) IF detect an obstacle THEN change direction
(2) IF carrying samples AND at the base THEN drop samples
(3) IF carrying samples and NOT at the base THEN drop 2 crumbs
AND travel up gradient
(4) IF detect a sample THEN pick sample up
(6) IF sense crumbs THEN pick up 1 crumb AND travel down
gradient
(5) IF true THEN move randomly
! Subsumption hierarchy: (1) < (2) < (3) < (4) < (6) < (5)
Svarm robolics: Mars exIorer syslem
v1
Svarm conslruclion: hrsl sles
Lunar base construction robots (Brooks et al, 90)
! small bulldozers
! no explicit coordination nor communication
! cooperate to achieve building tasks
o digging out trenches
o stockpiling a supply of loose lunar soil
o covering the habitation units with lunar soil
o .
Svarm conslruclion
! 15 x 10 x 20 cm.
! Tank-like treads
! A scoop, which can be lifted up and down and tilted back
and forth
! A head which constantly rotates with
various sensors mounted on board:
o beacon sensor
o proximity sensors
o force sensors
o inclinometers
o ..
! Two microprocessors: one for motors and sensors data and the
other for the control programs
Svarm conslruclion
! Global task: selection of a promising area to construct the
lunar base
! Global functionalities:
o Navigation in group in search for a promising area
o Adoption of the next task once such an area have been located
! Envisioned scenario: once the robots are on the lunar surface,
every robot starts executing the same set of rules
Svarm conslruclion
! Behavior rules of every robot:
1. maintain a minimum distance from the robots surrounding it
2. match velocities with the robots in its neighborhood,
3. move towards the perceived center of mass of the robots in its
neighborhood,
4. the velocity of a robot is proportional to the number of big rocks
it perceives in its neighborhood (or the degree of levelness of the
local neighborhood),
5. when a robot hasnt moved much for a while it goes into a new
mode, adopting a new set of behaviors which is appropriate
for the next global task.
flock-like
movement
the flock stops wandering around and starts
working on the next task
Svarm conslruclion
Two main problems:
! Local-to-global: Given a set of rules for a construction
team to follow, what will it end up producing?
! Global-to-local: Given a desired artifact, what rule set will
end up producing it?
! Nest building by wasps:
o attach the future nest to a
substrate with a stalk-like pedicel
o build two cells on either side of the
pedicel
o subsequent cells are added to the
outer circumference of the combs
between two previously
constructed cells
o as more cells are added, parallel
rows of cells are formed
o a row tends to be finished before
initiating a new row
o rows are initiated by the
construction of a centrally located
first cell
LocaI-lo-gIobaI: ModeIIing coIIeclive buiIding
LocaI-lo-gIobaI: ModeIIing coIIeclive buiIding
! (Theraulaz, Bonabeau, 95): Formal model of distributed building
inspired by social wasps
! Lattice swarms model
o stateless agents move randomly on a 3D
cubic lattice
o an agent deposits a brick every time it
finds a stimulating configuration
o a lookup rule table specifies all such
configurations and defines an algorithm
ModeIIing coIIeclive buiIding
nest-like structures
What makes the difference?
! Biological-like architectures require coordinated algorithms
o the desired architecture is
decomposed into a finite
number of building steps
o the local configurations that are
created by a given state and
which trigger building actions,
differ from those created by a
previous or forthcoming building
step (disjoint steps)
o all individuals cooperate in the
current building state at any time
ModeIIing coIIeclive buiIding
ModeIIing coIIeclive buiIding
c. 1nrr:ii:z :Nb r. noN:nr:i 388
stimulating congurations required to grow the
architectures shown in Figs 4(d, e and n). All
architectures except g and h, have been obtained with
coordinated algorithms. The dierence between (g, h)
and all other architectures is striking. One given
coordinated algorithm always converges towards
architectures that possess similar features, despite the
dierent realizations of the random walks performed
by the agents in dierent simulations. Any uncoordi-
nated algorithm seems to diverge: the same algorithm
leads to dierent global architectures in dierent
simulations. This tendency to diverge comes from the
fact that stimulating congurations are not organized
in time and space and many of them overlap, so that
the architecture grows in space without any
coherence. As an illustration of this fact, architectures
d and e result from two successive simulations using
the same coordinated algorithm, and architectures g
and h result from the same non-coordinated
algorithm. Moreover, even in shapes built with
coordinated algorithms, there may be some degree of
variation, which is larger in cases where the number
of dierent choices within one given building state is
large. But the result may also be deterministic in some
cases (that is, all simulations lead to exactly the same
architecture despite the random component in
individual behaviour), through the implicit hand-
shakes and interlocks that are setup at every stage.
Some of the structured architectures of Fig. 4 are
reminiscent of natural wasp nests. Among these nests
the presence of plateaus is observed in Fig. 4(b, df, i,)
and the dierent levels of the nests are linked together
with either a straight axis Fig. 4(b, i) or with a set of
pedicels Fig. 4(df). Others possess an external
envelope: nests k and n are shown with a portion of
the front envelope cut away so as to allow for a
visualization of the nests interior, and correspond
to nests found in the genera Chartergus. Figures
4(c, g, h, j, l, m,) are examples of architectures not
found in nature, but only (g) and (h) look
structureless. As an example of how the particular
geometry of the discrete space may inuence global
patterns, we also present two coherent architectures
obtained from coordinated algorithms using hexago-
nal elementary building blocks, that is, with a 30
rotational invariance around the z-axis Fig. 4(o and
p). We see that this particular geometry allows
round shapes to be easily created, which is not the
case for cubic bricks.
In order to understand what a coordinated
algorithm is, we now focus on an example depicted in
Fig. 5. This gure represents the successive steps of
the construction of a nest that resembles nests built
by Epipona. The transition between two successive
building steps is shown to depend on a given number
of local congurations stimulating the deposit of a
brick. Once all the bricks in the current step have
been deposited, the building process goes to the next
step. Steps 15 correspond to the enlargement of the
top of the nest, including the rst sessile comb of cells
(step 3). Step 6 represents the construction and
lengthening of the external envelope, from which
parallel combs will be built (steps 7 and 8). These
steps determine the distinction between this nest,
where the entrance and access holes at the dierent
levels lie in the periphery of the comb, from the
(a) (b) (e) (f)
(c) (d) (g) (h)
Fic.. 4(ah)
c. 1nrn:ii:z :Nb r. noN:nr:i 388
stimulating congurations required to grow the
architectures shown in Figs 4(d, e and n). All
architectures except g and h, have been obtained with
coordinated algorithms. The dierence between (g, h)
and all other architectures is striking. One given
coordinated algorithm always converges towards
architectures that possess similar features, despite the
dierent realizations of the random walks performed
by the agents in dierent simulations. Any uncoordi-
nated algorithm seems to diverge: the same algorithm
leads to dierent global architectures in dierent
simulations. This tendency to diverge comes from the
fact that stimulating congurations are not organized
in time and space and many of them overlap, so that
the architecture grows in space without any
coherence. As an illustration of this fact, architectures
d and e result from two successive simulations using
the same coordinated algorithm, and architectures g
and h result from the same non-coordinated
algorithm. Moreover, even in shapes built with
coordinated algorithms, there may be some degree of
variation, which is larger in cases where the number
of dierent choices within one given building state is
large. But the result may also be deterministic in some
cases (that is, all simulations lead to exactly the same
architecture despite the random component in
individual behaviour), through the implicit hand-
shakes and interlocks that are setup at every stage.
Some of the structured architectures of Fig. 4 are
reminiscent of natural wasp nests. Among these nests
the presence of plateaus is observed in Fig. 4(b, df, i,)
and the dierent levels of the nests are linked together
with either a straight axis Fig. 4(b, i) or with a set of
pedicels Fig. 4(df). Others possess an external
envelope: nests k and n are shown with a portion of
the front envelope cut away so as to allow for a
visualization of the nests interior, and correspond
to nests found in the genera Chartergus. Figures
4(c, g, h, j, l, m,) are examples of architectures not
found in nature, but only (g) and (h) look
structureless. As an example of how the particular
geometry of the discrete space may inuence global
patterns, we also present two coherent architectures
obtained from coordinated algorithms using hexago-
nal elementary building blocks, that is, with a 30
rotational invariance around the z-axis Fig. 4(o and
p). We see that this particular geometry allows
round shapes to be easily created, which is not the
case for cubic bricks.
In order to understand what a coordinated
algorithm is, we now focus on an example depicted in
Fig. 5. This gure represents the successive steps of
the construction of a nest that resembles nests built
by Epipona. The transition between two successive
building steps is shown to depend on a given number
of local congurations stimulating the deposit of a
brick. Once all the bricks in the current step have
been deposited, the building process goes to the next
step. Steps 15 correspond to the enlargement of the
top of the nest, including the rst sessile comb of cells
(step 3). Step 6 represents the construction and
lengthening of the external envelope, from which
parallel combs will be built (steps 7 and 8). These
steps determine the distinction between this nest,
where the entrance and access holes at the dierent
levels lie in the periphery of the comb, from the
(a) (b) (e) (f)
(c) (d) (g) (h)
Fic.. 4(ah)
coiirc1ivr niiibiNc n. soci:i iNsrc1s 397
tries to apply the rst set of rules; if it does so
unsuccessfully for an amount of time Tthis implies
some kind of memoryit switches to the next set
of rules, etc) and there are many other possible
renements.
These algorithms are extracted from a hierarchy of
elementary behaviors developed by Moore (private
communication) to describe dierent levels of growth
complexity. Growth complexity is dierent from
recognition complexity, embodies e.g. in Chomskys
grammars which tell how dicult it is to recognize
that a particular pattern belongs to a given set of
patterns: growth complexity should be a measure of
how dicult it is to grow a pattern. As we deal with
models, such a measure can only exist with respect to
a particular model class. The problem is then to nd
the most appropriate model class to describe a given
phenomenon [there are many ways of generating a
pattern: see e.g. Meinhardt (1982), Murray (1989),
Held (1992) or Sachs (1994)]. While it seems rather
dicult to evaluate the complexity of continuous
growth models, such as reaction-diusion equations,
it becomes possible to do so with formal models
relying on discrete automata. Of course, such discrete
models may not give a reliable image of how a pattern
is actually formed in many cases. Our experience
suggests, however, that our model can be applicable,
despite all artifacts, to nest building in wasps owing
to the apparent if-then/yes-no nature of their
behaviours, which can be described, in a rough
approximation, by discrete automaton rules. Discrete
automata may not describe many other building
behaviours, especially if quantitative responses (e.g.
to chemical gradients) are intrinsically important.
The neighbourhood of each agent comprises the
26 rst cells surrounding the cell it occupies. We
represent this neighbourhood with three slices along
the z direction (see Figure 9).
Below we give the rules (i.e. the whole set of local
stimulating congurations) used to produce the
architectures in Fig. 4(d, e and n). When the wasp
occupies the central position of slice z (marked ), i.e.
when there is no brick in that cell, it will put down
a brick of type 1 in the case of conguration 1.x and
a brick of type 2 in the cases 2.x. We give each local
conguration without taking into account the
symmetries due to rotations and reections. In the
case of architecture 4(n) (Chartergus) we distinguish
the same succession of sub-shapes as in Fig. 5.
:rcni1rc1irrs 4(b, r)
z+1 z z1
(1.1)
2 2 2
2 2 2
2 2 2
0 0 0
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.1.)
0 0 0
0 1 0
0 0 0
0 0 0
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.2)
0 1 0
0 0 0
0 0 0
0 2 0
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.3)
0 1 0
0 0 0
0 0 0
2 2 2
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.4)
1 0 0
0 0 0
0 0 0
2 2 0
2 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.5)
0 0 0
0 0 0
0 0 0
2 2 2
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.6)
0 0 0
0 0 0
0 0 0
2 2 2
2 0
2 0 0
0 0 0
0 0 0
0 0 0
(2.7)
0 0 0
0 0 0
0 0 0
2 2 0
2 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.8)
0 0 0
0 0 0
0 0 0
2 2 2
2 2
0 0 0
0 0 0
0 0 0
0 0 0
2 2 2
2 2 2
2 2 2
0 0 0
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.1.)
0 0 0
0 1 0
0 0 0
0 0 0
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.2)
0 1 0
0 0 0
0 0 0
0 2 0
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.3)
0 1 0
0 0 0
0 0 0
2 2 2
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.4)
1 0 0
0 0 0
0 0 0
2 2 0
2 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.5)
0 0 0
0 0 0
0 0 0
2 2 2
0 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.6)
0 0 0
0 0 0
0 0 0
2 2 2
2 0
2 0 0
0 0 0
0 0 0
0 0 0
(2.7)
0 0 0
0 0 0
0 0 0
2 2 0
2 0
0 0 0
0 0 0
0 0 0
0 0 0
(2.8)
0 0 0
0 0 0
0 0 0
2 2 2
2 2
0 0 0
0 0 0
0 0 0
0 0 0