TWO LAYERS ACTION INTEGRATION FOR HRI

Action Integration with Attention Focusing for Interactive Robots

Yasser Mohammad and Toyoaki Nishida

Graduate School of Informatics, Kyoto University, Kyoto, Japan

Keywords:

HRI, Action Selection, Natural Listening.

Abstract:

Behavior architectures are widely used to program interactive robots. In these architectures multiple behaviors

are usually running concurrently so a mechanism for integrating the resulting actuation commands from these

behaviors into actual actuation commands sent to the robot’s motor system must be faced. Different architec-

tures proposed different action integration mechanisms that range from distributed to central integration. In

this paper an analysis of the special requirements that HRI imposes on the action integration system is given.

Based on this analysis a novelle hybrid action integration mechanism that combines distributed attention fo-

cusing with a fast central integration algorithm is presented in the framework of the EICA architecture. The

proposed system was tested in a simulation of a listener robot that aimed to achieve human-like nonverbal

listening behavior in real world interactions. The analysis of the system showed that the proposed mechanism

can generate coherent human-like behavior while being robust against signal correlated noise.

1 INTRODUCTION

Behavioral architectures are widely used to program

interactive robots (Ishiguro et al., 1999), (S Karim,

2006). Behavioral architectures employs a vertical

decomposition of the software design into a set of co-

running behaviors that continuously map the current

state of the environment into commands to the actua-

tors of the robot. A basic problem that must be solved

in any such architecture is how to integrate the results

of multiple running processes into a ﬁnal command

sent to the actuators of the robot. The proposed solu-

tion to this problem (referred to as the action integra-

tion problem hereafter) can be divided into selective

solutions that selects the action of a single behavior at

any point of time to control the robot and combinative

solutions that generates the ﬁnal commands based on

the proposed responses of multiple behaviors and hy-

brid solutions that tries to combine cooperation with

coordination. Another dimension of comparison be-

tween action integration solutions is whether or not a

central integrator is employed. According to this di-

mension available solutions are central or distributed.

The action integration mechanism proposed in this

paper can be classiﬁed as a hybrid two layered archi-

tecture with distributive attention focusing behavior

level selection followed by a fast central combinative

integrator.

The next section formalizes the requirements for

an action integration mechanism suitable for HRI ap-

plications. Section 3 introduces the L

EICA archi-

tecture for which this action integration mechanism

was designed. Section 4 gives the details of the ac-

tion integration mechanism proposed in this paper

and section 5 reports an experiment with a humanoid

robot that used the proposed action integration mech-

anism.The paper is then concluded.

2 REQUIREMENTS

One goal of HRI research is to realize robots that

can use human-like verbal and nonverbal interaction

modalities. To achieve this goal researchers usually

employ behavioral architectures because of their ro-

bustness and fast response. A special property of HRI

applications is that the details of the external behav-

ior of the robot will be interpreted by human users

as intentional signals (Breazeal, 2005), and usually

combining nonverbal messages results on absolutely

different messages. This is why combinative archi-

tectures are not common in HRI research and most

of the HRI speciﬁc architectures employ a selective

Mohammad Y. and Nishida T. (2008).

TWO LAYERS ACTION INTEGRATION FOR HRI - Action Integration with Attention Focusing for Interactive Robots.

In Proceedings of the Fifth International Conference on Informatics in Control, Automation and Robotics - RA, pages 41-48

DOI: 10.5220/0001482600410048

 SciTePress

integration strategy. For example (Ishiguro et al.,

1999) proposed a robotic architecture based on situ-

ated modules and reactive modules in which reactive

modules represent the purely reactive part of the sys-

tem, and situated modules are higher levels modules

programmed in a high-level language to provide spe-

ciﬁc behaviors to the robot. The situated modules are

evaluated serially in an order controlled by the mod-

ule controller. Research in nonverbal communication

in humans reveals a different picture in which mul-

tiple different processes do collaborate to realize the

natural action. For example (Argyle, 2001) showed

that human spatial behavior in close encounters can

be modeled with two interacting processes. It is possi-

ble in the selective framework to implement these two

processes as a single behavior but this goes against

the spirit of behavioral architectures that emphasizes

modularity of behavior (Perez, 2003). This leads to

the ﬁrst requirement for HRI aware action integra-

tion: The action integration mechanism should allow

a continuous range from pure selective to pure com-

binative strategies. In other words the system should

use a hybrid integration strategy. The need to manage

the degree of combinativity based on the current situ-

ation entails the second requirement: The action inte-

gration mechanism should adapt to the environmental

state using timely sensor information as well as the

internal state of the robot. In current systems this re-

quirement is usually implemented by using a higher

level deliberative layer but in many cases the interac-

tion between simple reactive within the action inte-

grator can achieve the same result as will be shown in

the example implementation of this paper.

The Hybrid Coordination approach presented in

(Perez, 2003) is the nearest approach to achieve this

ﬁrst requirement. In this system every two behav-

iors are combined using a Hierarchical Hybrid Co-

ordination Node that has two inputs. The output of

the HHCN is calculated as a nonlinear combination

of its two inputs controlled by the activation levels

of the source behaviors and an integer parameter k

that determines how combinative the HHCN is, where

larger values of k makes the node more selective. The

HHCNs are then arranged in a hierarchical structure

to generate the ﬁnal command for every DoF of the

robot (Perez, 2003). Although experiments with the

navigation of an autonomous underwater robot have

shown that the hybrid coordination architecture can

outperform traditional combinative and selective ar-

chitectures, it still has some limitations in the HRI do-

main. One major limitation of the hybrid coordination

system is its reliance on binary HHCNs which makes

it unsuitable for large numbers of behaviors due to the

exponential growth in the number of HHCNs needed.

Another problem is the choice of the parameter k for

every HHCN. Yet the most difﬁcult problem for this

system is ﬁguring out the correct arrangement of the

behaviors into the HHCN inputs. This leads to the

third requirement: The action integration mechanism

should not depend on global relationships between

behaviors. One of the major problems with this ar-

chitecture is that every behavior must calculate its

own activation level. Although this is easy for be-

haviors like avoid-obstacles or go-to, it is very difﬁ-

cult for interactive processes like attend-to-human be-

cause the achievement of such interactive processes is

not manifested in an easily measurable speciﬁc goal

state that must be achieved or maintained but in the

exact way the overall behavior of the robot is chang-

ing over time. This leads to the fourth requirement:

The action integration mechanism should separate the

calculation of behavior’s inﬂuence from the behavior

computation.

The number of behaviors needed in interac-

tive robots usually is very high compared with au-

tonomously navigating robots if the complexity of

each behavior is kept acceptably low, but most of

those behaviors are usually passive in any speciﬁc

moment based on the interaction situation. This prop-

erty leads to the ﬁfth requirement: The system should

have a built-in attention focusing mechanism. HRI

systems usually work in the real world with high lev-

els of noise but it is required that the robot shows a

form of goal directed behavior. This leads to the sixth

requirement: The action integration system should be

robust against noise and data loss to provide a goal-

directed behavior.

In summary the six requirements HRI imposes on

the action integration system are:

R1 It should allow a continuous range from pure se-

lective to pure combinative strategies

R2 It should adapt to the environmental state utilizing

timely sensor information.

R3 It should not depend on global relationships be-

tween behaviors

R4 It should separate the calculation of behavior’s in-

ﬂuence from the behavior’s computation

R5 It should have built-in attention focusing

R6 It should be robust against noise and data loss.

Table 2 compares the action integration scheme of

some well used behavioral architectures with the pro-

posed system in terms of the six requirements

In this paper an action integration mechanism that

has the potential of meeting these requirements is pre-

sented.

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

Table 1: Comparison of the Action Integration Capabilities of Some Behavioral Architectures in Terms of the Six Require-

ments in section 2.

Architecture Integration Implement. R1 R2 R3 R4 R5 R6

Subsumption (Brooks, 1986) Selective Distributed

√ √ √

ASD (Maes, 1989) Selective Distributed

√ √

PDL (Steels, 1993) Combinative Central

√

Motor Schemas (Arkin, 1993) Combinative Distributed

√ √

Hybrid Coordination (Perez, 2003) Hybrid Distributed

√ √

AVB (Nicolescu and Matari

c, 2002) Selective Distributed

√ √ √

Situated Modules (Ishiguro et al., 1999) Selective Central

√ √ √ √

Proposed System Hybrid Hybrid

√ √ √ √ √ √

Figure 1: The Proposed Action Integration Mechanism.

3 L

EICA

EICA is a behavioral hybrid architecture designed for

HRI applications (Mohammad and Nishida, 2007).

The basic behavioral element of EICA is the inten-

tion. Every intention implements a simple well de-

ﬁned reactive capability of the robot. Intentions are

more like motor schema of (Arkin, 1993) than com-

plete behaviors. Every intention has three attributes:

Attentionality a real number (0 1) that speciﬁes the

relative speed at which the intention is running.

Actionability a real number (0 ∞) specifying the ac-

tivation level of the intention. A zero or negative

activation level prevents the intention from execu-

tion. This attribute controls the inﬂuence of this

intention on other running components.

Intentionality a real number (0 ∞) set by the dis-

tributed action integration layer and used by the

central integrator to combine the actions of vari-

ous intentions.

4 THE PROPOSED SYSTEM

Fig. 1 shows the block diagram of this hybrid system.

The following subsections will detail the two layers

of the system.

4.1 Behavior Level Integration

The goal of this layer of the action integration mech-

anism is to provide timely values for the actionabil-

ity, attentionality, and intentionality for various inten-

tions in the system. The ﬁrst two of those parameters

will determine how frequent every intention will be

allowed to send actions to the central action integra-

tor, and the third will be used by the action integrator,

along with other parameters, to integrate the actions

proposed by all the running intentions as will be de-

tailed in the next section.

This layer is implemented as a set of processes

connected together and to the intentions through effect

channels.

Every process runs with a speed directly propor-

tional to its attentionality level as long as its action-

ability is positive. The processes are connected to-

gether through effect channels in a network. Every

effect channel has a set of n inputs that use continu-

ous signaling and a single output that is continuously

calculated from those inputs according to the opera-

tion attribute of the effect channel. The example pre-

sented in this paper uses the Avg operation deﬁned as:

y =

∑

i=1

)

∑

i=1

)

, where x

is an input port, a

is the ac-

tionability of the object connected to port i and y is

the output of the effect channel. Every process or in-

tention has a single effect channel connected to its at-

tentionality attribute and another effect channel con-

nected to its actionability attribute, and effect chan-

nels can be arranged into hierarchical structures like

the HHCNs in (Perez, 2003) but rather than carrying

action and inﬂuence information, the effect channels

carry only the inﬂuence information between differ-

ent processes of the system.

TWO LAYERS ACTION INTEGRATION FOR HRI - Action Integration with Attention Focusing for Interactive Robots

4.2 Action Level Integration

This layer consists of a central fast action integrator

that fuses the actions generated from various inten-

tions based on the intentionality of their sources, the

mutuality and priority assigned to them, and general

parameters of the robot. The actionability of the ac-

tion integrator is set to a ﬁxed positive value to ensure

that it is always running. The action integrator is an

active entity and this means that its attentionality is

changeable at run time to adjust the responsiveness of

the robot. The action integrator periodically checks its

internal master command object and whenever ﬁnds

some action stored in it, the executer is invoked to ex-

ecute this action on the physical actuators of the robot.

Algorithm 1 shows the register-action function re-

sponsible of updating the internal command object

based on actions sent by various intentions of the

robot. The algorithm ﬁrst ignores any actions gen-

erated from intentions below a speciﬁc system wide

threshold τ

act

. The function then calculates the to-

tal priority of the action based on the intentionality

of its source, and its source assigned priority. Based

on the mutuality assigned to every degree of freedom

(DoF) of the action, difference between the total pri-

ority of the proposed action and the currently assigned

priority of the internal command, the system decides

whether to combine the action with the internal com-

mand, override the stored command, or ignore the

proposed action. Intentions can use the immediate at-

tribute of the action object to force the action integra-

tor to issue a command to the executer immediately

after combining the current action.

4.3 Requirement Achievement

1. The system can achieve a continuous range of

combinative to selective behavior based on the

values of the actionability and attentionality of the

intentions and the mutuality assigned to the ac-

tions. On one extreme τ

act

can be set to the same

value of the intention with the highest intention

which will lead to a purely selective behavior. On

the other hand τ

act

and action mutuality can be

set to zero while keeping the intentionality of all

intentions equal which will lead to a purely com-

binative behavior.

2. The distributed layer of the system can use the

sensor information and internal state stored in the

processes to manipulate the tau

act

parameter as

well as the intentionality of various intentions to

control the level of combinative behavior based on

timely information from the environment.

Algorithm 1 : Register Action Algorithm.

function REGISTER-ACTION(Action a, Inten-

tion s)

if s.actionability < τ

act

∨s.intentionality <

int

then

exit

end if

c ← current combined command

p ← a.priority + max priority ×

s.intentionality

for every DoF i in the a do

combined ← true

if p < c.priority ∧s 6= c.source then

c.source ← s

c.do f (i) ← a.do f (i)

c.priority(i) ← p

c.hasAction(i) ← true

end if

if c.source 6= null then

c.source ← null

c.priority(i) ←max (p, c. priority (i))

else

c.source ← s

c.priority(i) ← p

end if

if a.mutual = true ∨c.source = s then

c.source ← s

c.do f (i) ← a.do f (i)

c.priority(i) ← p

else

c.do f (i) ←

p×a.do f (i)+c.priority×c.do f (i)

p+c.priority

end if

if combined = true then

return false

end if

c.actionable ←true

if a.notCombinableWithLower then

c.stopCombiningLower ←true

end if

if a.immediate then

execute a

end if

end for

end function

3. The proposed action integration mechanism does

not depend on any global relation between the

intentions or the processes and adding new pro-

cesses or intentions will only involve deciding the

local effect, and data channels related to this new

active component.

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

4. The inﬂuence of any intention on the ﬁnal behav-

ior of the robot is controlled based on its intention-

ality which is in turn controlled by the processes

of the behavior integration layer rather than the

intention itself. This provides the required sep-

aration between the calculation of inﬂuence and

the computation of the basic behaviors.

5. Attention focusing is implemented in the lowest

level of implementation in the proposed system

and is controlled by the attentionality of various

processes and intentions. This calculation is sepa-

rated from the inﬂuence calculation through the

actionability parameters. Most existing robotic

architectures do not separate these two aspects of

behavior control. By separating the actionability

from the attentionality and allowing actionabil-

ity to have a continuous range, EICA enables a

form of attention focusing that is usually unavail-

able to behavioral systems. This separation allows

the robot to select the active processes depending

on the general context (by setting the actionability

value) while still being able to assign the compu-

tation power according to the exact environmental

and internal condition (by setting the attentional-

ity value). The fact that the actionability is vari-

able allows the system to use it to change the pos-

sible inﬂuence of various processes (through the

operators of the effect channels) based on the cur-

rent situation.

6. The central action integrator acts as a low pass ﬁl-

ter that reduces the effects of noise by combining

the actions sent by various intentions and accumu-

lating them for a period that is controlled by the

attentionality of the action integrator itself. This

provides a simple means of noise rejection. A

more subtle advantage of the proposed system in

relation to goal-directed behavior is the distribu-

tive nature of the behavior integration layer which

allows, for well designed robots, the emergence

of complex external behavior from simple internal

processes. The other option to achieve this com-

plex external behavior was to map this complexity

directly to the intentions themselves which would

have complicated the design too much.

4.4 Limitations

Although the previous subsection has shown that the

proposed architecture can theoretically achieve the six

requirement of section 2 the system still has some lim-

itations. One of the major problems with the proposed

system is that it is sometimes difﬁcult design the

behavior level integration layer because of its asyn-

chronous distributed nature and it is also difﬁcult to

learn the parameters needed to control the timing of

the operation of various processes in it. Although

careful division of the task into a set of intentions

and task-speciﬁc integration processes can alleviate

this problem, a general guideline to this process is

needed to simplify the design of EICA robots. This

limitation can be overcame by restricting the number

of processes active at any moment to only one process

but this will lead to over-complication in the design of

the intentions. A better solution to this problem is a

direction for future research.

5 EXPERIMENT

The ability to use human-like nonverbal listening be-

havior is an advantage for humanoid robots that coex-

ist with humans in the same social space and is com-

plex enough to test some of the proposed system’s

features. (Kanda et al., 2007) implemented a robot

that tries to use natural human like body language

while listening to a human giving it road directions

based on the situated modules architecture. The road

guidance task is simpliﬁed by the fact that there are

no other objects of interest in the scene except the hu-

man. In this work we try to build a general listener

robot that can generate natural nonverbal behavior in

an explanation scenario involving unknown number

of objects that can also be moving. As a minimal de-

sign, only the head of the robot was controlled during

this experiment. This decision was based on the hy-

pothesis accepted by many researchers in the nonver-

bal human interaction community that gaze direction

is one of the most important nonverbal behaviors in-

volved in realizing natural listening in human-human

close encounters (Argyle, 2001). This example is in-

tended as a guide for designing systems that can uti-

lize the proposed integration strategy and a proof of

its applicability for HRI.

5.1 Procedure

The evaluation data was collected as follows:

1. Six different explanation scenarios were collected

in which a person is explaining the procedure

of operating a hypothetical machine that involves

pressing three different buttons, rotating a knob,

and noticing results in an LCD screen in front

of a Robovie II robot while pretending that the

robot is listening to the explanation. The data was

collected using the PhaseSpace Motion Digitizer

system (PhaseSpace, 2007) by utilizing 18 LED

markers attached to various parts of the speaker’s

body. The data was logged 460 times per second.

TWO LAYERS ACTION INTEGRATION FOR HRI - Action Integration with Attention Focusing for Interactive Robots

2. The logged data were used as the input to a robot

simulator that implemented the proposed system.

The behavior of the robot’s head was compared

with known human-human behavior in terms of

mutual gaze and gaze toward instructor.

3. For every scenario 20 new synthetic scenarios

were generated by utilizing 20 different levels of

noise. The error level is deﬁned as the percentage

of the mean value of the noise term to the mean of

the raw signal. The behavior of the simulator was

analyzed for every one of the resulting 120 sce-

narios and compared to the original performance.

5.2 Design

Four reactive intentions were designed that encapsu-

late the possible interaction actions that the robot can

generate, namely, looking around , following the hu-

man face , following the salient object in the envi-

ronment, and looking at the same place the human

is looking at. Each one of those intention is imple-

mented as a simple state machine (Mohammad et al.,

2007). The sufﬁciency of those intentions was based

on the fact that in the current scenario the robot sim-

ply has no other place to look, and the necessity was

conﬁrmed empirically by the fact that the three be-

havioral processes needed to adjust the intentionality

of all of these intentions.

The analysis of natural listening requirements

showed the need of three behavioral processes. Two

processes to generate an approach-escape mechanism

controlling looking toward the human operator which

is inspired by the Approach-Avoidance mechanism

suggested in (Argyle, 2001) in managing spatial dis-

tance in natural human-human situations. These pro-

cesses were named Look-At-Human, and Be-Polite.

A third process was needed to control the realization

of the mutual attention behavior. This process was

called Mutual-Intention. The details refer to (Moham-

mad et al., 2007). A brief description of them is given

here:

1. Look-At-Human: This process is responsible of

generating an attractive virtual force that pulls the

robot’s head to the location of the human face.

2. Be-Polite: This process works against the Look-

At-Human process generating a repulsive virtual

force that pulls the robot’s head away from the lo-

cation of the human face depending on the period

of attending to the human.

3. Mutual-Attention: This process tries to pull the

robot’s head toward direction to which the human

is looking.

Table 2: Comparison between the Simulated and Natural

Behavior.

Item Statistic Simulation H-H

value

Mutual Gaze Mean 31.5% 30%

Std.Dev. 1.94% –

Gaze Toward Mean 77.87% 75%

Instructor Std.Dev. 3.04% –

Mutual Mean 53.12% unknown

Attention Std.Dev. 4.66% –

Five perception processes were needed to imple-

ment the aforementioned behavioral processes and in-

tentions. The details can be found in (Mohammad

et al., 2007).

5.3 Results

Some of the results of numerical simulations of the

listening behavior of the robot are given in Table 2.

The table shows the average time of performing two

basic interactive behaviors obtained from the simu-

lated robot in comparison to the known average val-

ues measured in human-human interaction situations.

The average times in the human-human case are re-

ported from (Argyle, 2001). As the table shows the

behavior of the robot is similar to the known aver-

age behavior in the human-human case for both mu-

tual gaze and gaze toward instructor behaviors and

the standard deviation in both cases is less than 7%

of the mean value which predicts robust operation in

real world situations.

Although the average time of showing the three

evaluation behaviors are similar to the human-human

values as shown in Table 2, this similarity is not

enough for completely judging human like natural be-

havior and the dynamical aspects of the interaction

must be taken into account before drawing ﬁnal con-

clusions. This method of analysis, although limited,

was selected because of two reasons. First the be-

havior of the speaker in this experiment is not to-

tally natural because the robot did not respond at real

time, which affects the dynamics of the interaction

but it was hypothesized that the effect on the aver-

ages used for evaluation is much less. Second there is

no available data about the detailed dynamics of the

behavior of the listener and the speaker in the human-

human case during explanation scenarios. In near fu-

ture a wide scale human-human experiment will be

conducted by the authors to collect such data for more

accurate evaluation.

Fig. 2 shows the effect of increasing the error level

on the percentage of time mutual gaze and gaze to-

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics

Figure 2: Effect on the error level on the behavior of the

robot.

ward instructor behaviors were recognized in the sim-

ulation. As expected the amount of time spent on

these interactive behaviors decreases with increased

error level although this decrease is not linear but

can be well approximated with a quadratic function

as regression analysis revealed. This means that the

performance degrades gracefully with the increased

noise level even for signal correlated noise.

5.4 Discussion

One of the main purposes of having robotic architec-

tures is to make it easier for the programmer to divide

the required task into smaller computational compo-

nents that can be implemented directly. The pro-

posed action integration mechanism helps in achiev-

ing a natural division of the problem by the following

simple procedure. First the task is analyzed to ﬁnd the

basic competencies that the robot must possess in or-

der to achieve this task. Those competencies are not

complex behaviors like attend-to-human but ﬁner be-

haviors like look-right, follow-face, etc. Those com-

petencies are then mapped to the intentions of the sys-

tem. Each one of these intentions should be carefully

engineered and tested before adding any more com-

ponents to the system

The next step in the design process is to design

the behavior level integration part of the action inte-

gration. To do that, the task is analyzed to ﬁnd the

underlying processes that control the required behav-

ior. Those processes are then implemented. The most

difﬁcult part of the whole process is to ﬁnd the correct

parameters of those processes to achieve the required

external behavior. Currently this parameter choice is

done using trial-and-error but it will be more effective

to use machine learning techniques to learn those pa-

rameters from the interactions either ofﬂine or online.

The current architecture supports run-time adaptation

of those parameters, and this feature will be exploited

in the future to implement learning of the behavioral

integration layer. Those behavioral steps are added

incrementally and the relative timing between them is

adjusted according to the required behavior.

This simple design procedure is made possible be-

cause of the separation between the basic behavioral

components (intentions) and the behavior level inte-

gration layer (processes).

It is informative to compare this procedure with

the procedure suggested in (Ishiguro et al., 1999)

for the situated modules architecture. Every situ-

ated module should have a list of preconditions that

is always checked by the module controller which

chooses that module that is most suitable to the sit-

uation. The problem with this arrangement for HRI

applications is that the evaluation of the preconditions

can be very time consuming or even very difﬁcult to

decide in the ﬁrst place. Let’s consider the situated

module look-at-human that should be activated dur-

ing the interaction enough to make the speaker feel

comfortable but not too much. How can the designer

ﬁnd all the rule to select all the occasions in which

the behavior is to be invoked? and how can this list

be updated for reuse in other applications? The main

problem is that it is too difﬁcult to ﬁnd the precondi-

tions for a behavior as simple as look-at-human and

the only solutions are either to complicate the behav-

iors used (increase the granularity) or to complicate

the module controller (may be by using deliberation).

In the proposed system this problem does not exist be-

cause the behavior itself (the intention) is coded with-

out any need to think about its preconditions. The be-

havior level integration processes are then designed

based on the global view of the task and not the re-

quirements of each intention which means that this

set of preconditions need not be built at any point in

the design process. This allows intentions to be thin-

ner than the situated modules without the need of a

higher deliberative layer.

A widely accepted deﬁnition of intention is:

”a choice with commitment” (Cohen and Levesque,

TWO LAYERS ACTION INTEGRATION FOR HRI - Action Integration with Attention Focusing for Interactive Robots

Figure 3: The evolution of intentionality of the four basic

intention used in the listener robot.

1990). This deﬁnition emphasizes that for a behavior

to be perceived as intentional it must possess a form

of behavioral inertia that makes the robot stick with

its behavior until its goal is achieved or the situational

changes result on the adoption of a new goal. Care-

ful design of the behavior level integration processes

can achieve this goal in the proposed architecture. For

example Fig. 3 shows the evolution of intentionality

of the aforementioned four basic reactive intentions

in one case annotated with the external behavior as

perceived by the interacting human. As shown in the

ﬁgure the effects of the ﬂuctuations of the input sig-

nals ,although mapped to the intentionality of various

intentions, do not affect the ﬁnal behavior directly be-

cause of the existence of the central action level in-

tegrator that averages those ﬂuctuations and provides

this needed behavioral inertia to make the ﬁnal be-

havior look intentional for the human.

6 CONCLUSIONS

This paper analyzed the required properties for ac-

tion integration mechanisms suitable for interactive

robots, and presented the design of a new two lay-

ers hybrid action integration mechanism that utilizes

both distributive integration in the behavior integra-

tion layer and a central action level integrator that can

achieve the aforementioned requirement. The paper

also presented a simple example of a listener robot

implemented based on the proposed mechanism.

REFERENCES

Argyle, M. (2001). Bodily Communication. Routledge;

New Ed edition.

Arkin, R. C. (1993). Biological neural networks in inver-

tebrate neuroethology and robotics, chapter Modeling

neural function at the schema level: Implications and

results for robotic control, page 17. Academic Press.

Breazeal, C. Kidd, C. T. A. H. G. B. M. (2005). Effects

of nonverbal communication on efﬁciency and robust-

ness in human-robot teamwork. In 2005 IEEE/RSJ In-

ternational Conference on Intelligent Robots and Sys-

tems, 2005. (IROS 2005), pages 708–713.

Brooks, R. A. (1986). A robust layered control system for a

mobile robot. IEEE Journal of Robotics and Automa-

tion, 2(1):14–23.

Cohen, P. and Levesque, H. (1990). Intention is choice with

commitment. Artiﬁcial Intelligence, 42:213–261.

Ishiguro, H., Kanda, T., Kimoto, K., and Ishida, T. (1999).

A robot architecture based on situated modules. In

IEEE/RSJ Conference on Intelligent Robots and Sys-

tems 1999 (IROS 1999), volume 3, pages 1617–1624.

IEEE.

Kanda, T., Kamasima, M., Imai, M., Ono, T., Sakamoto,

D., Ishiguro, H., and Anzai, Y. (2007). A humanoid

robot that pretends to listen to route guidance from a

human. Autonomous Robots, 22(1):87–100.

Maes, P. (1989). How to do the right thing. Connection

Science, 1(3):291–323.

Mohammad, Y. F. O. and Nishida, T. (2007). A new, hri

inspired, view of intention. In AAAI-07 Workshop

on Human Implications of Human-Robot Interactions,

pages 21–27.

Mohammad, Y. F. O., Ohya, T., Hiramatsu, T., Sumi, Y., and

Nishida, T. (2007). Embodiment of knowledge into

the interaction and physical domains using robots. In

International Conference on Control, Automation and

Systems, pages 737–744.

Nicolescu, M. N. and Matari

c, M. J. (2002). A hierar-

chical architecture for behavior-based robots. In AA-

MAS ’02: Proceedings of the ﬁrst international joint

conference on Autonomous agents and multiagent sys-

tems, pages 227–233, New York, NY, USA. ACM.

Perez, M. C. (2003). A Proposal of a Behavior-based Con-

trol Architecture with Reinforcement Learning for an

Autonomous Underwater Robot. PhD thesis, Univer-

sity of Girona.

PhaseSpace (2007). http://www.phasespace.com/.

S Karim, L Sonenberg, A. T. (2006). A hybrid architecture

combining reactive plan execution and reactive learn-

ing. In 9th Biennial Paciﬁc Rim International Confer-

ence on Artiﬁcial Intelligence (PRICAI).

Steels, L. (1993). The artifcial route to artifcial intel-

ligence. Building situated embodied agents, chapter

Building agents with autonomous behaviour systems.

Lawrence Erlbaum Associates, New Haven.

ICINCO 2008 - International Conference on Informatics in Control, Automation and Robotics