Upload
mariacarla
View
213
Download
1
Embed Size (px)
Citation preview
Monitoring Strategies for Adaptive Periodic Control
in Behavior-based Robotic Systems
Ernesto Burattini, Alberto Finzi, Silvia Rossi
Dipartimento di Scienze Fisiche
University of Naples “Federico II”
Naples, Italy
{ernb,finzi,srossi}@na.infn.it
Mariacarla Staffa
Dipartimento di Informatica e Sistemistica
University of Naples “Federico II”
Naples, Italy
Abstract— The main goal of our current research is the designof a behavior-based robotic architecture that has the capabilityof adapting its behaviors to the rate of change of both theenvironment and its internal states reducing the computationalcosts of input processing. Inspired by research on biologicalclocks, we introduced a simple schema theory model wherereleasing mechanisms are combined with adaptive internal clocks.In this paper, we describe the design and development ofa complete robotic architecture implementing this model. Inparticular, we considered a mobile robot domain that simulatesthe navigation behavior of a Catagliphys ant enhanced withsimple visual capabilities.
I. INTRODUCTION
The main goal of our current research is the design of
a behavior-based robotic architecture that has the capability
of adapting its behaviors both to the rate of change of the
environment and to changes of its internal states. This adaptive
behavior requires effective mechanisms for controlling sensors
and effectors with respect to the internal constraints imposed
by the control system and the external environment. The
executive control is to combine different low-level strategies
(such as obstacles avoidance, walls follow, gates crossing,
etc.), with high-level activities (such as achieving a goal,
returning to the nest, etc.), giving them, from time to time,
different priority values both for allocation of resources and
for action selection processes. The low-level activities affect
the safety of the system, and may be achieved by applying
principles of reactive control. However, high-level activities
generally are achieved by processing more complex tasks, and,
therefore, require high computational costs for both the inputs
processing and the acquisition of data from the environment.
In this context, self-regulating mechanisms balancing sensors
elaboration and behavior execution play a crucial role. Indeed,
frequent sensors readings could result in a high computational
charge that degrades the performances of executive and delib-
erative processes, while, adopting large latencies in sensors
sampling is unsafe since the environment may change too
much in between two consecutive readings. In this sense,
any effort to build a cognitive architecture for dealing with
dynamical and flexible behaviors has to deal with an efficient
processing capability of the sensors elaboration.
In a previous paper, Burattini and Rossi proposed the
concept of periodical/timed activations of behaviors [1], [2].
Inspired by internal self-regulation mechanisms in living or-
ganisms [3], they introduced internal clocks to tune and adapt
the frequencies of the behaviors activities. Such mechanisms
can speed up or slow down the period of behaviors activation
and thereby the reading frequency of the sensors according to
both the robot–environment interactions and the interactions
that may arise within the robots itself [4], [5], [6].
In this paper, inspired from this work, we propose a real
robot implementation of a behavior-based control architecture
endowed with periodic releasing mechanisms. In particular, we
investigate the feasibility and effectiveness of the approach and
we explore how the adaptive clocks affect the computational
load for the elaborations of perceptual data, considering each
single behavior and the overall system. The question we have
to answer is twofold. First, is the robot we are building reliable
and safe? This means that the robot should be able to timely
react to the external stimuli avoiding dangerous situations.
Second, does the introduction of adaptive clocks in a single
behavior provide better performances of the robotic system,
in terms of a reduction of computational resources spent to
achieve the tasks?
To test the proposed model we implemented a behavior-
based control system endowed with these regulation mech-
anisms and evaluated different settings. These mechanisms
should provide an interconnecting layer between adaptive
behaviors and high level cognition [7].
II. TIMED ACTIVATION OF BEHAVIORS: RHYTHMIC
RELEASERS
Our aim is to develop a behavior-based control for a
robotic system capable of adapting its emergent behavior to
the surrounding environment and to its internal state. Our
working hypothesis is that adaptive behaviors can be simulated
in the control activity of a robot starting from self-regulating
mechanisms. In particular, our architecture combines Innate
releasing or inhibiting mechanism and biological clocks.
a) Innate releasing or inhibiting mechanisms (IRMs):
Lorentz [8] and Tinbergen [9] identified in many animals an
innate releasing or inhibiting mechanism (IRM) able to control
and coordinate behaviors. An IRM presupposes a specific
stimulus that releases a pattern of actions. For example, a
prey animal may have, as an IRM, the stimulus coming from
2009 Advanced Technologies for Enhanced Quality of Life
978-0-7695-3753-5/09 $25.00 © 2009 IEEE
DOI 10.1109/AT-EQUAL.2009.34
130
Fig. 1. Schema theory representation of IRMs (left) and AIRMs (right). For each behavior we have a schema composed of a coordinated control program(releaser/clock), a perceptual schema and a motor schema. On the left side, the releasing function activates the motor schema; on the right side, the internalclock directly regulates the frequency of the sensor activations enabling/disabling sensors readings.
the view of the predator, which activates the escape behavior.
IRMs were included in the schema representation of behaviors
[10] in the form of releasers, controlling when behaviors
must be activated or deactivated. A releaser is an activation
mechanism that depends on exogenous factors (e.g. presence
of a predator) and/or endogenous factors (e.g. hunger).b) Adaptive Innate Releasing Mechanisms (AIRMs): The
releaser’s function, somehow, recalls the notion of “internal
clock”, already introduced in some approaches [11], [5], [6]
in order to activate motivational states for a robot (for example,
hunger or sleep). In fact, an internal clock, similarly to a
releaser, represents an internal mechanism which regulates
behaviors activations depending on endogenous and/or ex-
ogenous factors. However, there are substantial differences
between IRMs and internal clocks. Indeed, while a releaser
is an instantaneous activation mechanism, the internal clock is
periodical and adaptive.
In [1], [2], the authors inspired by the internal clocks,
connected the concept of IRM to the concept of a periodical
activation of behaviors introducing the Adaptive Innate Releas-
ing Mechanisms (AIRMs). An AIRM is a releasing mechanism
endowed with and adaptive period of the releaser activation.
IRMs and AIRMs are illustrated and compared in Figure 1
deploying a schema theory representation [10]. Here, for each
behavior, we have a schema composed of a Perceptual Schema
(PS), Motor Schema (MS), and an activation mechanism,
either a clock or a releasing function ρ(t). On the left side,
IRM enables/disables data flow σ(t) from PS to MS, the MS
is activated only in the presence of the stimulus, but sensing
data are always processed (i.e. at each machine cycle); in
contrast, on the right side, AIRM directly enables/disables data
flow σr(t) from sensors to PS, therefore when the activation
is disabled sensing data are not processed (sensors readings
minimization). Furthermore, the internal clock regulates the
frequency of the activations, hence the frequency of data
processing (behavior adaptation), using a feedback mechanism
on the processed sensing data σ(t). The AIRM mechanism will
be detailed in the next sections.
III. THE ARCHITECTURE
The robotic system is controlled by a behavior-based ex-
ecutive where each behavior can be described by a schema
theory model [10]. Each behavior of the Robotic System (RS)
is provided with a periodical and adaptive clock that controls
its activations (AIRMs as defined in [1], [2]). In particular, we
assume that (see Figure 1):
• Each clock has a base period (or initial rhythm) called pb,
ranging in the interval [pbmin, pbmax]. The values of pb
and range are experimentally tuned. pb depends on sev-
eral factors such as the sensors precision, environmental
conditions and purposes of the behavior;
• The releasing function ρ(t), is a delta function that, with
period pβ(t), is equal to 1 at each pb machine clocks
and 0 otherwise, and tells the robot when the PS has to
process inputs. This timed releasing function takes data
from a PS and returns an enabling/disabling signal to the
Perceptual Schema itself;
• σr(t) is the actual function representing the data per-
ceived by the sensor at each time step t;• σ(t) is the function that represents data coming from
sensors at each sampling step: σ(t) = f(σr(t))xρ(t),i.e. f(σr(t)) is for processed sensing data and ρ is the
releasing function.
• π(t) is the command sent by the MS to actuators. While
the PS is inactive, no new commands will be sent. This
means that an actuator can keep its output constant until a
new command will come (e.g. the velocity of the wheels
of a robot), or its output will produce an action with its
own duration and then it waits for a new command (e.g.
a gripper that lifts a box).
Each time the behavior is activated, thus enabling sensory
data flow, sensory data are passed in feedback to the in-
ternal clock which updates its period depending on internal
and environmental conditions. The mechanism for period/rate
regulation is called the updating policy and will be detailed in
the following sections.
IV. MONITORING STRATEGIES AND ADAPTATION
The emergent behavior will result from a combination of
the following parameters:
• the initial period pb;
• the range of allowed values for the period [pbmin, pbmax];• the updating policy.
In particular, the combination of these parameters defines a
monitoring strategy. A monitoring strategy is a policy for
scheduling sensing activities and has to balance the cost of
monitoring and the risk of inaccurate and partial information
about the environment.
We will compare our monitoring strategy (adaptive and
periodical) with respect to other relevant monitoring strategies
proposed in literature. In particular, we consider four different
131
cases: a) the robot is not equipped with any internal clock;
b) the robot is equipped with internal clocks and has a priori
knowledge about the task to achieve (for example about the
distance to cover); c) the robot is equipped with clocks, but
does not have any a priori knowledge about the environment;
d) the robot is equipped with internal clocks but not adaptive.
To illustrate these cases we introduce the following example.
Let us consider a robotic system whose purpose is to cover a
certain distance (goTo behavior).
Fig. 2. Monitoring strategies for goTo behavior. Number of activations inthe case of: (a) continuous monitoring; (b) periodic and adaptive monitoringwithout a priori knowledge; (c) periodic and adaptive monitoring with a prioriknowledge; (d) periodic monitoring.
Without an internal clock (a), goTo will be activated at each
control cycle. If the covered distance is constant at each control
cycle, we have that the number of activations n of goTo is
proportional to the distance dist to be covered n: n ∝ dist(see Figure 2). In case (d), if the value of the clock period is
pb, we have the relation: n ∝ distpb
. Cases (b) and (c) fall in the
category of those strategies called of “Interval Reduction” [12],
which are characterized by a variable period for the monitoring
strategy. For cases (b) and (c), it has been demonstrated that
these strategies, asymptotically, are more effective than those
characterized by a constant periodic monitoring [12], [13] in a
wide class of problems. Moreover, in this setting, an interval
reduction strategy has to increase the behavior activations
frequency while approaching the goal. If the robot is equipped
with adaptive clocks and knows a priori the distance to cover,
we might set the initial period pb proportional to half of
the distance to cover and then halve the period of the clock
after each activation of the behavior following the interval
reduction strategy. In this way, the number of activations
would be: n ∝ log2(dist). We assume that this strategy is
the “optimum”, where for optimum, we mean a strategy that
allows to achieve the goal through the minimum number of
activations, without the risk of jeopardizing the correct and
efficient functioning of the robotic system and its safeguard.
Let us now describe how the robot behaves in the case (c),
namely, when it is characterized by adaptive rhythms, but
without a priori knowledge. In this case, the rhythm must
change gradually in accordance with a law that does not
diverge too far from the optimum case. In fact, the robot, even
if not provided with a priori knowledge, can obtain information
from the surrounding environment, thanks to the use of its
sensors and, through these values, it may determine the choice
of the rhythm. For example, the number of activation in this
case may be approximated as: n ∝ log2(dist)+ (distcov)/pb,
where distcov is the distance already covered by the robot.
We can see that the number of activation in the case (b) will
have as an upper bound the case (a), and as a lower bound
the case (c).
Overall, the benefits brought by adaptive and periodical
monitoring strategies are mainly two:
• periodical mechanisms of activation can reduce the num-
ber of activations of the behavior (with respect to the
standard case (a) in which the activations are performed
at every machine cycle), causing a relative decrease in
the computational burden, and improving performance of
the entire system;
• the use of adaptive mechanisms allows us to obtain a
behavior that adapts itself to the specific environmental
conditions (e.g. the robot reads sensors more often if there
is a dangerous situation and less often in cases of a safe
operational situation).
In the following, we will show that, by calibrating appro-
priately the basic rhythms and using the appropriate policies
to update them, we can obtain a significant improvement in
performance compared to an architecture without rhythms.
V. IMPLEMENTATION
Based on the model introduced in the previous sections,
we designed a system whose behavior is mainly guided by
the visual information in a 3D environment, constituted by an
office space, that implements the three monitoring strategies
described here. We used a PIONEER 3DX provided with
a blob camera and sonars. The robot is controlled by a
Player/Stage client [14].
c) Catagliphys domain: We evaluated our approach us-
ing a mobile robot that simulate the navigation behavior of
a Catagliphys ant enhanced with simple visual capabilities.
The robot has the task of searching food in its environ-
ment without any a priori knowledge and then return to its
nest following a straight path, without taking into account
the trajectory followed during the searching of food. The
Catagliphys domain is a good testbed for our purposes: the
domain is interesting from a behavioral point of view and
well analyzed in several ethological field trials [15]; the ant
is provided with internal mechanisms such as guidance and
dead-reckoning; we can associate with this testbed all the
monitoring strategies previously presented (constants, with a
priori knowledge, without a priori knowledge).
d) Behavior-based control: The robot behavior is ob-
tained as the combination of the following primitive behaviors
(see Figure 3): AVOID, WANDER, PATH INTEGRATION,
MOVE TO FOOD, MOVE TO NEST and FIND LANDMARKS,
132
organized in two meta-behaviors called LOOK FOR FOOD and
RETURN TO NEST.
Fig. 3. Control architecture for the mobile robot in the Catagliphys domain.Each behavior receives data from sensors and generates the actions that canbe combined (circled +) or subsumed (circled s).
e) System assessment: To assess the system perfor-
mances, we compared the adaptive control system with respect
to a not adaptive one (i.e. with sensor readings fixed at
every machine cycle). Our aim is to show that a significant
improvement in performance can be obtained by appropriately
tuning the basic periods of clocks and the policies to update
them. Therefore, for each specific behavior we will evaluate
the updating policies, seeking, on one hand, to optimize
performances (i.e. less sensor readings), and, on the other
hand, to enhance the robot safety and the correctness of the
overall system behavior.
f) Behaviors settings: For each behavior, we have to
define the monitoring strategy composed of the 1) updating
policy and the 2) base period, that is empirically defined after
a phase of testing.
1) Updating policies: We start describing each behavior and
the associated updating policy.
The WANDER behavior provides a random search in the
environment. Its output consists only in a random direction
for the robot. Since the activation of this behavior is periodic,
but not adaptive, WANDER can be associated with a constant
clock. That is, the updating policy is the following: pβ(t) =pb = const, where, the clock period pβ remains equal to the
base period pb.
In contrast, the AVOID behavior, responsible for obstacle
avoidance and its output consists in a direction and sets the
navigation velocity for the robot. This behavior is safety
critical and needs an adaptive clock and an associated updating
policy to timely react to dangerous situation. In this case, a
good policy is to change the AVOID clock period according
to the first derivative of the sensory input σ(τ), that represents
the distance from the nearest object, evaluated by the sonars.
More formally, the clock period pβ is updated as follows
pβ(t) = ϕ(σ(t)−σ(t−p′
β)
p′
β
× ρ(t)), where p′β is the period at
the previous behavior activation (initially pβ(t0) = pb), ρ(t)is the releasing function that enables the sensory sampling,
ϕ is a normalizing function mapping the derivative into a set
of permitted values and σ(t − p′β) is the value of σ at the
previous behavior activation. Intuitively, the clock frequency is
adaptive with respect to the environmental changes: the higher
the change the smaller the sensory sampling.
MOVE TO FOOD detects the food and guides the robotic
system towards it, setting its direction. In order to obtain
reliable information from the camera, this behavior forces
the robot to slow down its velocity. By choosing an adaptive
clock period for this behavior (i.e. reducing the number of
activations), we allow the robot to reach the food as soon as
possible. However, we have to set the base period and the
updating policy taking into account that we want to reach
the goal with the minimum docking error. Thus, we have
to balance effectiveness and precision. The idea is to keep
constant the base period pβ(t) until the robot reaches a fixed
distance threshold and then to decrease the period linearly with
the food distance as follows: pβ(t) = ϕ(distance(t) × ρ(t)),where ϕ is a normalizing function, distance(t) is the distance
at time t, ρ(t) is the releasing function that enables sensory
sampling. Here, the behavior performances will depend on the
optimal balance between the clock frequency and distance.
PATH INTEGRATION manages the direction and the dis-
tance to return to nest. This behavior uses odometric sensor
information to calculate the robot shift w.r.t. the previous
position. Analogously to WANDER the behavior remains always
active with a constant clock period, until it reaches the food. In
the return process, the behavior is activated only after an abrupt
change of course, e.g. due to the presence of obstacles, which
diverts the trajectory of the robot that has to be recalculated.
Therefore, we can state that pβ(t) = pb × ρ(t).
RETURN TO NEST set the direction of the robot toward
the nest. Like MOVE TO FOOD it requires reliable information
from the camera and so will slow down the robot velocity
while its PS is active. This behavior exploits a priori knowl-
edge (i.e. the distance to the nest) that allows us to set the
base period value proportional to half of the distance from
the nest, pb = distanceToNest(t0/(2k)) and then reduced
accordingly, i.e. pβ(t) = ϕ(distanceToNest(t)×ρ(t)), It has
been shown that, asymptotically, this strategy is very efficient
[12], [13].
2) Setting the base period: Once we have defined the
behaviors’ schemata and the associated updating policies, it
remains to set the base periods pb. These are empirically
defined after a phase of testing. In the following, we detail
the case of MOVE TO FOOD.
As previously mentioned, the updating policy is to keep
constant the clock period until a certain distance threshold
is reached, then it should be linearly reduced. During the
exploration of the environment, the robot can detect the food at
different distances, hence we need different distance threshold.
In order to optimize the performances and simultaneously
133
distance=200cm distance=120cm distance=60cm
activation docking activation docking activation dockingpb time (s) error (cm) time (s) error (cm) time (s) error (cm)
1 20.274 ∓ 0.630 0.178 ∓ 0.018 7.759 ∓ 1.320 0.289 ∓ 9, 25e−01
0.149 ∓ 0.006 0.215 ∓ 1.125e−004
2 7.638 ∓ 0.480 0.28 ∓ 2.37e−004
3.667 ∓ 0.274 0.287 ∓ 3, 11e001
0.194 ∓ 0.003 0.206 ∓ 6.75e−005
4 5.116 ∓ 0.884 0.28 ∓ 1.17e−004
3.046 ∓ 0.145 0.292 ∓ 3, 25e−01
0.167 ∓ 0.004 0.173 ∓ 1.57e−004
8 2.991 ∓ 0.311 0.28 ∓ 2.17e−004
1.392 ∓ 0.335 0.231 ∓ 4, 25e−01
0.237 ∓ 0.004 0.116 ∓ 6.30e−004
16 1.213 ∓ 0.452 0.250 ∓ 0.004 0.649 ∓ 0.059 0.195 ∓ 0.002 0.167 ∓ 0.004 0.020 ∓ 2.83e−004
32 1.292 ∓ 0.363 0.085 ∓ 0.0019 40.45 ∓ 0.053 0.123 ∓ 5, 7e+001
0.199 ∓ 0.003 0.012 ∓ 2.70e−006
Fig. 4. An evaluation of the MOVE TO FOOD behavior, in terms of activation time and docking error, varying the distance and the base period pb.
take care of the system security constraints, we tested the
MOVE TO FOOD behavior performance under different robot–
food distances. The aim here is to define, for each distance
threshold, how to dynamically reset the base period pb (w.r.t.
the distance) before proceeding with its linear reduction.
We tested the behavior under three different initial values of
the robot–food distance and compared the obtained result w.r.t.
all the possible allowed values for the period [1,base period].For each case, we established the period with the best results
in terms of minimum activation time and acceptable docking
error. In Figure 4, we report the average and variance of the
values gathered in 10 runs for each case. For example, looking
at the results, we notice that, if the initial distance from food is
equal to 120cm, the period with the best performances is pβ =8, because it allows to minimize the time of behavior activation
with a very small docking error. Moreover, for each case, a
horizontal line separates safe/unsafe settings: above the line
there are settings where the robot can stop at a safety distance
from the target, hence not incurring in dangerous situation.
Finally, we noticed that if the robot detects, for the first time,
the food at a distance smaller than 60cm the period has to
reach the minimum value immediately. Looking at the docking
error and the total amount of time for the behavior activations
obtained in each test, we can choose the best period/distance
that balances the tradeoff between safety and efficiency.
VI. EMPIRICAL RESULTS
In order to assess the system performances, we compared
the system behavior when endowed with adaptive clocks with
respect to a not adaptive periodic version (e.g. activations at the
machine clock). In particular, for each behavior, we considered
the number of activations and the total amount of time spent in
behavior execution. In Figure 5, we show some results from
the analysis of the RETURN TO NEST behavior, where for
each case we plot the results of one trial. We note that the
number activations of this behavior, with an adaptive clock, is
proportional to log2(dist) whereas, in the case of non adaptive
clocks, these increase linearly with time. Furthermore, we note
a significant reduction of the execution time of the behavior
endowed with the adaptive clock.
If we consider the overall system we observe a considerable
advantage in performances in the case of adaptive clocks.
This is illustrated in Figure 6 where we compare the per-
formances in term of activations for different behaviors (i.e.
with different adaptation strategies) considering the overall
8 16 32 64 1280
20
40
60
80
100
120
140
base periodexecution tim
e (
s)
without Internal Clock
with Internal Clock
with Internal Clock without Internal Clocknumber of execution number of execution
pb activations time (s) activations time (s)
8 3 1 8 816 4 5 16 1732 5 12 32 3464 6 25 64 69
128 7 50 128 132
Fig. 5. A comparison between the execution of the RETURN TO NEST
behavior with or without the adaptive clock changing the base period pb.
system. We notice that the adaptive monitoring strategy of
RETURN TO NEST (i.e., with a priori knowledge) produces
better results in terms of number of activations.
VII. CONCLUSIONS AND FUTURE WORKS
In this paper, we explored the feasibility and the effec-
tiveness of a behavior-based robotic architecture where the
behavior activities are modulated by periodic and adaptive
releasers. In particular, inspired by internal self-regulation
mechanisms in living organisms, we introduced internal clocks
to tune and adapt the frequencies of the sensors and activations
associated with the system behaviors. Such mechanisms can
speed up or slow down the period of behaviors activation
and thereby the reading frequency of the sensors according
to both the robot-environment interactions and the interaction
that may arise within the robots itself. To test the model
we implemented a behavior-based control system endowed
with these mechanisms and evaluated different monitoring
strategies. In particular, we developed an application which
simulates the behavior of the Catagliphys ant.
134
Behavior based robotic usually resolves conflicts by deploy-
ing a subsumption architecture or by implementing some con-
trol mechanisms in order to switch between tasks, selecting the
action [16], [17]. For example, in [18] the authors presented a
schema theoretic model for a praying mantis which behaviors
are driven by motivational variables such as fear, hunger and
sex-drive. In this approach, the action selection module selects
only the motivational variable with the highest value. In our
approach, the modulation of behavior is not controlled by an
on/off switch for changing the task. Indeed, the global behavior
emerges from the each single behavior through a rhythmic
controller modulated by a monitoring strategy.
Other authors dealt with this kind of problems. For example,
in [11] the authors presented a parallel architecture focused on
the concept of activity level of each schema which determines
the priority of its thread of execution. A more active perceptual
schema can process the visual input more quickly and a more
active motor schema can send more commands to the motor
controller. However, while in our approach such effects are
obtained through rhythmic activation of behaviors, in [11]
the variables are elaborated through a fuzzy based command
fusion mechanism.
In this research, we performed a preliminary assessment of
an architecture in which the activities of each behavior are
triggered by a clock. In our experiments, we observed that the
number of activations of each behavior decreases strongly, i.e.
the computational overload is lowered. Furthermore, our tests
have also shown interesting results about the synchronization
and the scheduling of the behaviors. In fact, since each
behavior is endowed with its own clock, whose period changes
are based on external and internal conditions, we have that, in
some circumstances, a behavior activated more frequently than
others with the consequence that its influence on the emergent
behavior is stronger than other behaviors. We observed that our
architecture is only partially reactive since we store at each
time the values of the clocks for each behavior and sensor
data. In fact, the mechanism of adaptive clock provides both a
quick reaction to the perceived stimuli and takes into account
the rate of changing in the external environment and internal
states. In other words, our system refers, in some way, to the
history of the rhythms fluctuations caused by the changes of
the environment and of the internal state. This means that in a
pure reactive system we have the same action in presence of
the same stimuli while in our system the action depends also
on previous stimuli.
In future works, we plan to investigate the introduction of
learning mechanism to select the proper rhythms for each
behavior. Moreover, in this work we assumed a setting where
each behavior modulates its own rhythm of activation, how-
ever the behaviors interconnection can be enhanced, in this
direction we plan to investigate how the adaptive periodical
activations of behaviors may influence and constrain each
other.
0
50
100
150
200
250
300
350
AVOID MOVE−TO−GOAL RETURN−TO−NEST
Nu
mb
er
of
Be
ha
vio
r A
ctiva
tio
ns
with Internal Clock
without Internal Clock
Fig. 6. Number of behaviors activations with or without adaptive clocks.
ACKNOWLEDGMENT
The research leading to these results has received funding
from the European Community’s Seventh Framework Pro-
gramme (FP7/2007-2013) under grant agreement no.216239.
REFERENCES
[1] E. Burattini and S. Rossi, “A robotic architecture with innate releasingmechanism,” in International Symposium on Brain, Vision and Artificial
Intelligence, ser. LNCS, vol. 4729. Springer, 2007, pp. 576–585.[2] ——, “Periodic adaptive activation of behaviors in robotic system,” Int.
J. Pattern Recognition and Artificial Intelligence, vol. 22, no. 5, pp.987–999, 2008.
[3] W. L. Koukkari and R. B. Sothern, Introducing Biological Rhythms.Springer-Verlag, 2006.
[4] D. Parisi, “Internal robotics.” Connect. Sci., vol. 16, no. 4, pp. 325–338,2004.
[5] A. Stoytchev and R. C. Arkin, “Combining deliberation, reactivity, andmotivation in the context of a behavior-based robot architecture,” inProc. of IEEE International Symposium on Computational Intelligence
in Robotics and Automation, 2001, pp. 290–295.[6] ——, “Incorporating motivation in a hybrid robot architecture,” JACIII,
vol. 8, no. 3, pp. 269–274, 2004.[7] A. Clark and R. Grush, “Towards a cognitive robotics,” Adapt. Behav.,
vol. 7, no. 1, pp. 5–16, 1999.[8] K. Lorenz, King solomon’s ring. Penguin, 1991.[9] N. Tinbergen, “The study of instinct,” Oxford University press, 1951.
[10] M. A. Arbib, “Schema theory,” in The handbook of brain theory and
neural networks. MIT Press, 1998, pp. 830–834.[11] G. Pezzulo and G. Calvi, “A schema based model of the praying mantis,”
in SAB – International Conference on Simulation of Adaptive Behavior,ser. LNCS, vol. 4095. Springer, 2006, pp. 211–223.
[12] M. S. Atkin and P. R. Cohen, “Monitoring strategies for embeddedagents: Experiments and analysis,” Adaptive Behavior, vol. 4, no. 2,pp. 125–172, 1995.
[13] S. J. Ceci and U. Bronfenbrenner, “Don’t forget to take the cupcakesout of the oven: Prospective memory, strategic time-monitoring, andcontext,” Child Development, vol. 56, 1985.
[14] B. Gerkey, R. Vaughan, and A. Howard, “The player/stage project: Toolsfor multi-robot and distributed sensor systems,” in Proc. of ICAR 2003,2003, pp. 317–323.
[15] R.Wehner and S.Wehner, “Path integration in desert ants:approaching along-standing puzzle in insect navigation,” Monit. Zool. Ital., vol. 20,pp. 309–331, 1986.
[16] R. A. Brooks, “Intelligence without reason,” in Proc. of IJCAI, Sidney,Australia, 1991, pp. 569–595.
[17] M. J. Mataric, “Behavior-based robotics as a tool for synthesis of arti-ficial behavior and analysis of natural behavior,” in Trends in Cognitive
Science, 1998, pp. 82–87.[18] R. C. Arkin, K. Ali, A. Weitzenfeld, and F. Cervantes-Perez, “Behavioral
models of the praying mantis as a basis for robotic behavior,” Robotics
and Autonomous Systems, vol. 32, no. 1, pp. 39–60, 2000.
135