io-insired muIliagenl syslems:

!"##$%&'($ !"*+&,-%&'"*
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

Fic. 9. Local neighbourhood in 3D cubic lattice swarm.

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

Fic. 9. Local neighbourhood in 3D cubic lattice swarm.

Set of local stimulating configurations used to produce these architectures:
the wasp in the central
position of slice z, will put
down
o a brick of type 1 in the
case of configuration
1.x and
o a brick of type 2 in the
cases 2.x.
GIobaI-lo-IocaI robIem
Take an unspecified number of robots and a supply of building
material, give the system a blueprint or other set of specifications
for any arbitrary structure desired, and have a guarantee that
the system will produce that structure without further intervention
CoIIeclive conslruclion: gIobaI-lo-IocaI robIem
Building two-dimensional user-specified structures (Werfel et al, 06)
o robots and blocks are deployed at random into an obstacle-free workspace
o blocks are square and can be attached to each other on all four sides
o a marker indicates the location for the start of construction
o robots can not communicate with one another
o robots fetch blocks from elsewhere in the workspace, bring them to the
growing structure, and travel along its perimeter for some distance before
attaching their block
CoIIeclive conslruclion
How and where robots decide to attach blocks?
! specifying the desired shape
o shape map: lattice-based representation with a coordinate system whose
origin is the marker
o robots can use the structure as a reference to determine their location
! ordering attachment
(A): A block can be attached to 1 or 2, not to 3 or 4
(B): Two separated blocks (sites 1, 5) may not be attached in the same row if all
sites between them are meant to be occupied
(C): Building up a structure by layers, always starting new rows from the end and
extending them clockwise
Conslruclion vilh inerl bIocks
Establishing location of robots with inert blocks:
! identical blocks
o the marker serves as a landmark
o robots follow the perimeter of the
structure until reaching the landmark
! distinct blocks
o all blocks are potential landmarks
o robots maintain a dynamic label map, storing the labels and locations of all
blocks in the structure
! writable blocks
o robots can change the state
of block labels
o blocks store their coordinates
explicitly
Conslruclion vilh communicaling bIocks
Establishing location of blocks with communicating blocks:
! determining whether or not a site is available for block attachment is
shifted from the robots to the structure itself
! blocks connect to each other when connected to the structure
! each block in the structure stores:
o the shape map
o its location in the shared coordinated system
o a state for each of its faces
Block algorithm:
! when a robot is given permission to attach a block, the structure block
sends a message along that row in both directions
! all the blocks in the row update the state of their faces
! the newly attached block obtains from its neighbors the shape map
and its coordinates, and sets the state of its faces
Example of
block algorithm
Conslruclion vilh communicaling bIocks
Conslruclion vilh communicaling bIocks
Types of constraints on block placement:
! geometric: ensure that the structure and its perimeter is of the desired
shape (independent of block type)
! functional: encompass whatever restrictions based on block types
may be dictated by the particular application
o patterned structures: use absolute coordinates
o adaptive structures: use relationships between block types
Conslruclion vilh communicaling bIocks
uiIding adalive slruclures
Types of functional constraints:
! Absolute: no reference to neighboring blocks (e.g., type A may be
attached; B is forbidden).
! Proximity: the types of neighboring blocks are important but their
locations are not (e.g., every A must have at least one B somewhere
within a neighborhood of radius 3; no C can be adjacent to a D).
! Relational: involving both types and locations of neighboring blocks
(e.g., every A must have a B bordering its west edge and a C
bordering its north edge).
square structures built with two block
types and the constraint that no two
yellow blocks be adjacent
patterned structure
with a complex
shape
uiIding adalive slruclures
Sets of constraints can be organized in regions and multiple levels
o Region 1, border of upper-left area: only blue
blocks allowed.
o Region 2, interior of upper-left area: blue and
white blocks both allowed; no white block
may be in the eight neighborhood of
another white block.
o Region 3, elsewhere: blocks must be either
red or white depending on their y-coordinate
Regions defined
by reference to
geometric
shapes and lines
MuIli-robol conslruclion
v2
Experiments on the construction of square structures of varying side length:
MuIli-robol conslruclion
A: The robot initially knows only
that the marker must be
present.
B: Using the line on the floor as a
reference, it picks up a block
from the cache.
C: It can use its RFID reader to
determine its position, and its
camera to follow the perimeter.
Existing blocks are added to
the robots map as it observes
them.
D: The robot reaches an empty
site where a block is desired
E: It maneuvers to attach its block
at that site, dropping it into
place...
F: ...and writes the block's new
coordinates to its tag.
v3
MuIli-robol 3D conslruclion
! 3D structures with communicating blocks (Werfel, Nagpal, 08)
! Extends the 2D assumptions on blocks and robots to 3D
structures: weightless environment, robots move freely in any
direction, blocks can communicate with physically attached
neighbors, ..
! Problem decomposed in two parts:
o where blocks need to go?
o how they get there?
3D conslruclion
Where to put blocks?
(A) A block could be attached at site A but not at site B
(B) A 3D version of the separation rule violated
(C) Even though this partial structure can occur without violating the
separation rule, attaching further blocks at any of the twelve sites
marked * would violate that restriction
3D conslruclion
Blocks rules - Where to put blocks?
Attachment allowed iff the answer to these three questions is YES:
1. Does the structure design specify this site is to be occupied?
2. For each of the three axes passing through this site, is the separation
rule satisfied?
3. For each of the three planes passing through this site, is a block here
being attached contiguously to the locally contiguous group?
Site A violates the first check
Site B passes checks 1 and 3 but fails 2
Site C passes checks 1 and 2 but fails 3
Desired structure
3D conslruclion vilhoul gravily
Robots rules - How to find allowed attachment sites?
! Random movement: Movement in any available direction along the
structure surface.
! Systematic search: Robots circle the structure at a fixed height; if
they return to a previously visited site they move up or down one
level and continue with the next layer.
! Gradient-following: Robots receive explicit directions from the
blocks of the structure. Information includes the number of steps
away from the closest allowed attachment site, as well as the local
direction to travel to reach it.
3D conslruclion vilhoul gravily
successive
snapshots during
the process of
construction by
ten robots, using
gradient-based
movement
Comaring aroaches
S: Steps taken along surface
M1: Messages passed from blocks to robots
M2: Messages passed between blocks
v
3D conslruclion vilh gravily
! TERMES (Werfel et al., 11), a multi-robot construction system
inspired by the building activities of termites.
! The system takes as input a high-level representation of a
desired structure, and provides rules for robots to build that
structure.
MuIli-robol conslruclion: TIRMIS
Robots model:
! Robots can move freely in the plane, follow the perimeter of the
structure in progress, and recognize the marker block when found.
! Movement on the structure:
o can attach a block at an adjacent site at the same level as itself,
o can climb onto a block one step higher,
o can descend to one one step lower,
o can move from one to another at the same level.
! Robots can keep track of
their movement and
location relative to the
structure, use no global
information, and rely on
stigmergy for coordination.
v4
MuIli-robol conslruclion: TIRMIS
! Structures must be representable as a 3D occupancy grid
! Intermediate configurations to avoid:
! Dangers of indiscriminate attachment:
TIRMIS: singIe-hIe alh
! Admissible structures: Those for which a single-file path to be followed by
all robots exists.
! An offline compiler Is supposed to deliver a structpath: the single-file
path plus the desired stack height at each site
! The structpath avoids conflicts when multiple robots encounter:
the farther along the sequence of [fetch block]-[return to structure]-
[follow perimeter]-[follow structpath] a robot is, the higher its priority
Side view of a linear
structure in which blocks
will be added in
the order shown as robots
enter from the left and
leave to the right
v5
! More sophisticated structures can be built by allowing a path to split
and rejoin itself
! A robot reaching a split point picks a branch at random
! Two robots approaching a merge point from different branches will
establish right-of way
TIRMIS: branching and merging alhs
! Additional structures can be built if robots can remove blocks as well
as adding them
! Auxiliary staircases: built in adjacent pairs forming a two-lane
highway
! Two structpaths are generated: one for additive construction and
other for staircase removal
TIRMIS: lemorary slaircases
References
! Brooks R.A., Maes P., Mataric M. and More G. Lunar Base Construction Robots, IEEE
International Workshop on Intelligent Robots and Systems, 1990.
! Steels L., Cooperation Between Distributed Agents through Self-Organisation. In
Demazeau, Y. and Mller, J.P., editors, Decentralized AI First European Workshop on
Modelling Autonomous Agents in Multi-Agent Worlds (MAAMAW-89)
! Bonabeau, E. and Theraulaz, G.. Modelling the Collective Building of Complex
Architectures in Social Insects with Lattice Swarms. Journal of Theoretical Biology, vol. 177,
issue 4, 1995.
! Werfel J., Bar-Yam Y., Rus D. and Nagpal R. Distributed Construction by Mobile Robots with
Enhanced Building Blocks, IEEE International Conference on Robotics and Automation,
2006.
! Werfel J. and Nagpal R. Three-Dimensional Construction with Mobile Robots and Modular
Blocks. International Journal of Robotics Research, vol. 27, 3-4, 2008.
! Werfel J., Petersen K., and Nagpal R., Distributed Multi-Robot Algorithms for the TERMES 3D
Collective Construction System, Workshop on Reconfigurable Modular Robotics, at IEEE/RSJ
International Conference on Intelligent Robots and Systems, 2011.