Attentional Modulation of Mutually DependentBehaviors

E. Burattini, S. Rossi, A. Finzi, and M. Staffa

Dipartimento di Scienze Fisiche ?,??

University of Naples “Federico II” – Naples, Italy{ernb, srossi, finzi}@na.infn.it, mariacarla.staffa@unina.it

Abstract. In this paper, we investigate simple attentional mechanismssuitable for sensing rate regulation and action coordination in the pres-ence of mutually dependent behaviors. We present our architecture alongwith a case study where a real robotic system is to manage and harmo-nize conflicting tasks. This research focuses on attentional mechanismsfor regulating the frequencies of sensor readings and action activations ina behavior-based robotic system. Such mechanisms are to direct sensorstoward the most salient sources of information and filter the availablesensory data to prevent unnecessary information processing.

1 Introduction

The main goal of our current research is the design of a behavior-based roboticarchitecture that has the capability of adapting behaviors activations both tothe rate of change of the environment and to changes of its internal states. Forthis purpose, attentional mechanisms, balancing sensors elaboration and actionsexecution, can play a crucial role. In particular, attentional processes play twomain roles: direct sensors towards the most salient sources of information; filterthe available sensory data to prevent unnecessary information processing. Asa result of the application of these mechanisms, the robot should react fasterand more effectively to task-related or safety critical stimuli because process-ing resources are focused on relevant stimuli. Attentional mechanisms appliedto autonomous robotic systems have been proposed elsewhere, mainly for visionbased robotics. However, attentional mechanisms are necessary not only to focusthe attention on salient regions of the space, but also to distribute resources andactivities in time [7, 10, 12]. Also in neuroscience researchers started to investi-gate the temporal domain of neural activity (for example neural synchrony [11]),and relate such activity to different cognitive processes such as binding, sensorymotor-coordination, attentional selection and executive functions.? The research leading to these results has received funding from the European

Community’s 7th Framework Programme (FP7/2007-2013) under grant agreementno.216239.

?? The Authors want to thank MariaRosaria Ambrosino for the support provided inthe creation of this work.

In this direction, our working hypothesis is that attentive and adaptive be-haviors can be achieved starting from self-regulated periodic mechanisms, whichmodulate sensory readings and behaviors activations. In previous papers [4, 5],we introduced the Adaptive Innate Releasing Mechanisms (AIRMs) to speed upor slow down the period of behaviors activation and thereby the reading fre-quency of the sensors, according to both the robot-environment interactions andthe interactions that may arise within the robots itself [13].

In the AIRMs, each behavior is endowed with an independent regulationmechanism directly depending on internal and external stimuli, while the mu-tual influence, among the parallel behaviors, is left as a consequence of the overallself-regulating emergent behavior. However, the notion of divided attention [14]suggests that a limited amount of attention is allocated to tasks, when resourcesare shared in multi-task behavior, and attention can be available in graded quan-tity for each task. Indeed, the activations of some behaviors may directly requirethe activation or the inhibition of other behaviors: two behaviors may not beable to activate themselves as frequently as they need without a degrade of per-formances (e.g. cognitive load and interference [8]); otherwise, the activationsof one behavior may directly induce the activation or synchronization of otherbehaviors (e.g. synchrony in attentional selection [11]).

The human behavior provides several examples of tasks that, while appar-ently conflicting, are simultaneously carried on. For example, some research an-alyzed the human behavior while driving and achieving a parallel task, such astalking over a mobile phone [15, 8]. Driving a car is a complex behavior that re-quires the extraction and integration of information from multiple sources. Mostof the information relevant for driving are taken by the view, so every change inthe visual exploration can be significant for a safe driving. For example, in [8] theauthors tried to experimentally assess the effects of cognitive load caused by asecondary task, simultaneously executed. Their results have shown that drivers,under a high cognitive load, execute less saccadic movements consistently withan increase of fixation time and a smaller exploration of the visual field. Theseexperiments show that subjects are able to complete tasks in parallel, but theresources allocated to each task must dynamically adapt themselves to environ-mental conditions and to cognitive and physical capabilities of the subject.

In this paper, we describe our cognitive/attentive general framework obtainedas an extension of the AIRM architecture [4, 5] that integrates mechanisms formutual influence among attentive behaviors. For this purpose, we introduce sim-ple constraints among the behaviors’ sampling rates. This mutual influence canwork both as an inhibitory or facilitatory process [6]. To assess our framework,inspired by the studies on cognitive distraction [15, 8], we define a case studywhere a real robot is to achieve two conflicting goals. In this context, we com-pare performances of this architecture with respect to non attentive versions ofthe same system. The empirical evaluation shows that the proposed frameworkis capable of harmonizing conflicting goals and distracting activities while main-taining the features of adaptability and reactivity of the previous architecture.

2 Adaptive Periodic Behaviors Modulation

In previous papers [4, 5], we introduced the AIRM (Adaptive Innate ReleasingMechanisms) architecture. In the AIRM framework, the robotic system is con-trolled by a behavior-based executive, where each behavior can be describedby a schema theory model [1]. Each behavior is characterized by a PerceptualSchema (PS), which elaborates sensor data, a Motor Schema (MS), producingthe pattern of motor actions, and a control mechanism, based on a combinationof a clock and a releaser (see Fig.1). The releaser enables/disables the activationof the MS, according to the sensor data. The adaptive clock controls the PS, it isactive with a base period and it enables/disables data flow from sensors to PS.When the activation is disabled, sensor data are not processed (yielding to sen-sory readings reduction). Furthermore, the clock regulates its period p (rangingthe values in the interval [pmin, pmax]), hence the frequency of data processing,using a feedback mechanism.

Fig. 1. Schema theory representation of AIRMs.

3 Attentive Monitoring and Mutual Influence

Our goal is to develop attentional mechanisms supporting a supervisory atten-tional system [12] and providing a kind of divided attention [14], which focusessensory resources and modulates tasks activations taking into account mutualinfluences and constraints among the behaviors.

Attentive Monitoring. Our framework combines the following design principles:

– Behavior-based control: the executive control is obtained from the interactionof a set of parallel behaviors working at different level of abstraction;

– Attention filtering and focusing: Attention modulation mechanisms consistin focusing monitoring and control activities on relevant behaviors and ex-ternal stimuli. For each behavior, the process of changing the rate of sensoryreadings may be interpreted as an increase or decrease of attention towards aparticular aspect of the environment the robotic system is interacting with;The higher is the frequency of the sampling rate associated with one behav-ior, the higher is the resolution at which this is monitored and controlled.

– Internal and external sources of saliency: The sources of salience are behav-ior and task dependent, and these can depend on either internal states orexternal stimuli;

Mutual Influence. In our attentive framework, the attention modulation strate-gies should be suitably regulated not only with respect to the internal or externalsaliency, but also with respect to attentional demand of other behaviors. To ac-count of the problem of mutual influence among attentional behaviors, in thiswork, we propose an extension of the AIRM architecture endowed with explicitconstraints among the internal clocks and regulation mechanisms able to respectthese constraints. The aim is to capture mutual dependencies in terms of interre-lations among the clocks’ sampling rates and then to regulate the clocks’ frequen-cies according to the presence of conflicting or synergetic behaviors. For example,given two mutually exclusive processes, since these are to be interleaved, the as-sociated clock periods should be opportunely changed to allow their alternatedexecution; on the other hand, for two concurrent behaviors, the associated clocksare to be aligned: when the frequency of one clock increases/decreases the otherclock should be accelerated/decelerated and vice versa. However, we want to addthis simple mechanism while maintaining the main features of the AIRM model:the periodic activation of behaviors should provide both a relative decrease inthe computational burden and the ability to monitor the internal/external en-vironment (e.g. the robot reads sensors more often if there is a dangerous orsalient situation and less often in cases of a safe operational situation).

In this new setting, for each set of clocks p1, ..., pn, we can introduce a rela-tionship R(p1, ..., pn) that specifies the mutual influence. However, in this paper,we mainly focus on the relationships between couple of behaviors. In particular,we consider binary constraints R(pA, pB) like mutex or synchronization con-straints. In this case, the frequencies of the clocks pA and pB , associated withthe two behaviors, depend not only on the salience of the tasks, but also onthe joint frequencies. Examples of these constraints will be provided in the casestudy presented in the following section.

Related Work. The problem of mutual influence among behaviors were tackledin different approaches. For example, in [9] the author presents a homeostaticsystem where couples of behaviors are connected through “successor” or “con-flicter links” to inhibit or activate each other. These links play a role which isanalogous to that of our mutual constraints; however, our regulation mecha-nisms are different because they are based on attentional modulation of clockssampling rates. Moreover, our focus is not on the constraint per se, but on theeffects of constraints on our architecture.

Concurrent tasks interacting with the attentive processes are considered in[16] where a robot architecture integrates active vision and tasks execution.However, mutual influence is not considered while attentive and goal-directedbehaviors are integrated and coordinated using a perceptual memory.

Our attentive sampling can be also related to flexible scheduling for periodictasks in real-time systems. In [3], period modulation is exploited only to keep the

system load balanced. Similar techniques can be incorporated in our framework,however, in our case sampling rate and interaction among behaviors depend notonly on the computational load, but also on saliencies due to environmentalchanges, internal states, and goals.

4 Case Study

Inspired by the studies on cognitive distraction during driving activities [15, 8],we designed a case study with two conflicting goals. In a hallway there are someclusters of green blobs distributed on the left and on the right wall. The robothas the task of running across the hallway in the shortest time possible, whilecounting all the green blobs (see Fig.2(a)(b)). The two tasks conflict on the speedof the robot. In fact, the first task would require a high speed, while the second,in order to effectively count all the blobs, would require a slow one.

Environment. The hallway is straight, without obstacles, 14m long and hasa width of 1.60m (see Fig.2(a)). All along the walls there are 27 green blobsarranged in 3 clusters of 9 blobs each, symmetrically disposed as a 3x3 grid (seeFig.2(b)). The three grids are randomly distributed along the walls.

(a) (b)

Fig. 2. (a) A snapshot of the robot in the environment. (b) A snapshot of the robotfield of view with a superimposed grid to identify different areas.

Behaviors. In order to accomplish the two tasks we implemented three behaviors:RUN, SEARCH and SCAN (see Fig. 3).

SEARCH looks for green blobs on the left and right wall. When the behavioris activated, it causes a random movement of the pan-tilt camera. This behavioris activated every machine cycle until it detects a green area. In this case, theperiod is increased from 1 to 9 machine cycles proportionally to the amount ofgreen color detected in the wall (i.e. the minimum time to allow to the SCANbehavior to identify the 9 blobs composing the cluster) (see Sec. 4.1).

SCAN counts the blobs once a salient area is identified. According to activevision [2] approach, in order to count an object on the wall, the camera has tocenter the object in its field of view, simulating a saccadic movement. For thispurpose, we split the field of view of the camera into nine areas (see Fig.2(b)).

RUN sets the speed of the robot. Differently from the previous behaviors, theeffect of the activation of such behavior continues even when the behavior is off.In fact, after the behavior has sent a command to the motors, the controller ofthe robotic system will keep such speed until a new command will arrive. Thevalue of the speed is in inverse proportion to the value of its period. The rangeof allowed speeds is from 0.01m/s to 0.24m/s.

The system starts with a medium speed, looking for green objects on the wallsof the corridor. Its behavior will change according to the visual percept. Whenthe system detects a green object, the SCAN behavior period decreases, allowingthe robot to slow down its speed and to count the objects it detects. Similarly, ifno green objects are detected, SEARCH and RUN periods become smaller, allowinga more accurate exploration (moving several times right and left the cameralooking for objects), and increasing the system speed in order to reach the endof the corridor as fast as possible.

4.1 Mutual Influence Rules

The regulation of mutual influence of two clocks, with periods pA and pB , de-pends on the statical and the dynamical priorities between behaviors and therelationship R(pA, pB).

Priorities. Priorities in changing periods depend on the importance of the behav-ior in accomplishing the task and in ensuring the safety of the robot. Behaviorsthat are safety critical have the maximal priority, hence the other behaviors willbe activated consequently. In the case of behaviors with the same priority, thepolicy for updating the value of the period is “the first takes all”, i.e., at eachmachine cycle, the first behavior that changes its period has to notify to otherbehaviors such variation. The other behaviors have to modulate their periodsaccordingly.

Relationship. If two behaviors A and B, with respectively pA and pB periodsand with ranges [pAmin, pAmax] and [pBmin, pBmax], share the same resourcesand are potentially in conflict, we have to define a relationship between thesetwo values. To better understand, we consider what happens in the frequencydomain in which a low pass filter prevents the passage of frequencies belowa particular cutoff frequency. If K is this cutoff frequency (i.e. in some waythe maximum bandwidth available) representing in our case the maximum rateof behavior activation, and fA = 1

pAand fB = 1

pBrespectively represent the

activation frequencies of two conflicting behavior, with the relation: fA+fB ≤ K,we indicate that each frequency will benefit from the breadth bandwidth notused by the other and vice versa. Likewise, if the activation period pA assumesa particular value within its allowed range [pAmin, pAmax], the period pB can

only assume a value within [pBmin, pBmax], limited to the remaining bandwidth.However, if two behaviors need to be executed simultaneously in order to realizea macro behavior, or if their outputs may be summed and are not in conflict,we may assume the following synchrony relationship: |pA − pB | = 0. In ourarchitecture we have that the SCAN behavior (with period pSc) and the SEARCHbehavior (with period pSr) cooperate on the achievement of one of the tasks, butconflicts on the use of the pan/tilt camera. On the contrary, RUN (with periodpR) conflicts with SCAN on tasks. Indeed, the first one has the goal to reachthe end of the corridor as soon as possible, while the second one needs to slowdown as much as possible the speed of robot in order to optimize the countingphase. Finally, the RUN behavior and the SEARCH behavior can cooperate in theachievement of their own task. In fact, both require a high speed. Let α, β, andγ be constants equal to α = pScmax +pSrmin, β = 0 and γ = pScmax +pRmin , inthis test the relationships among the periods of these behaviors can be formalizedas α ≥ pSc+pSr (or α = pSc+pSr if we want a strong dependence), β = |pSr−pR|and γ ≥ pSc + pR (or α = pSc + pSr if we want a strong dependence).

Fig. 3. Control architecture for the mobile robot.

4.2 Results

In order to evaluate the performances of our system, we compared three dif-ferent architectures, each with different behaviors settings, implemented on aPioneer 3DX, equipped with a pan/tilt camera and range sonar (see Fig.2(a)),and defined as follows:

– AIRM: all behaviors are equipped with adaptive clocks;– AIRM v max: adaptive clocks only in SCAN and SEARCH; the speed of RUN

is kept constant at the highest value (0.24m/s);– AIRM v med: adaptive clocks only in SCAN and SEARCH; the speed of RUN

is set to a medium value (0.11m/s);

– SC2SR8 v med: the behavior activation is periodic (for SCAN pSc= 2, for

SEARCH pSR = 8 and for RUN pR = 1), while the speed of the system is keptconstant to a medium value (0.11m/s);

– SC5SR5 v med: the same as the previous case with different periods (forSCAN pSc

= 5, for SEARCH pSR= 5 and for RUN pR = 1);

– Sub v max: the behaviors are active at every machine cycle and they arecoordinated by a subsumption architecture (i.e., SCAN subsumes SEARCH).The speed is equal to 0.24m/s;

– Sub v med: the behaviors are active at every machine cycle as in Sub vmax, but the speed is equal to 0.11m/s.

In Figure 4(a), we summarize the results collected during the tests, consideringthe number of counted blobs and the time spent to complete the task. For eachsetting, we performed 10 tests. The AIRM architecture performed well in termsof number of blobs counted. In fact, the AIRM implementation counts an averageof 17.8 blobs. In the case of the AIRM architecture with adaptive clocks onlyfor SCAN and SEARCH, the speed of the RUN behavior is kept constant duringthe tests. In these two cases, the number of counted up blobs is smaller thanthe AIRM case. However, for the AIRM v max, time performances are better,while in the case of medium speed the average time (127.7s) is comparable withthe AIRM case (123s). Another important thing to highlight is that, while theaverage speeds in the medium case and in the AIRM v med case are comparable,the number of counted blobs is better in the AIRM case. This is because thesystem will adapt itself to the surrounding environment speeding up or slowingdown, taking advantages of empty areas to accelerate, while decelerating whenit perceives blobs to count.

In the cases of periodic (not adaptive) activation of behaviors (SC5SR5 andSC2SR8), the performances with respect to the number of counted blobs areworst than in the AIRM case. The case SC5SR5 presents the worst results interms of counted blobs. We experienced a little improvement in the case ofmore frequent activation of SCAN (SC2SR8). However, we have to highlight thatthe periodic activation of behaviors in the case of SC2SR8 determines a highernumber of activation of perceptual schemas (i.e. wasting more resources) withrespect to the AIRM case, elaborating camera data even during the explorationof empty areas.

In the last set of tests, we evaluated the performances of a subsumption ar-chitecture (Sub). In this implementation SCAN subsumes SEARCH. The speed ofRUN is kept constant at 0.24m/s and 0.11m/s. The performances of Sub v medresult to be better with respect to the other cases except for the AIRM thatperforms the best. Indeed, the subsumption architecture resolves potential con-flicts on resources (i.e. the pan/tilt of the camera) while, without an arbitratormodule, such conflicts may reduce performances. However, in this case, similarlyto the periodic activation of behaviors, we have a higher number of activations ofthe SCAN perceptual schema that elaborates camera data at each machine cycle.These results make us foresee that in the case of a higher elaboration load anadaptive architecture may significantly improve performances.

In Figure 4(c), we plotted the cost/benefit (time/counted blobs) evaluation.Also from this point of view, the AIRM implementation performs better than theothers. However, this plot shows that the AIRM case presents a greater standarddeviation in the time performances. A high standard deviation implies a highvariability of the test results. This variability is caused by the adaptability ofthe system with respect to the environment and, consequently, to the changesof the system speed.

Finally, in Figure 4(b) we evaluated the error on the number of counted blobsfor units of time. This error is evaluated as (nB − nCB)/t, where nCB is thenumber of counted blobs, nB is the total number of blobs in the environmentand t is the time spent to accomplish the task.

(a)

n of blobs time (s)

AIRM 17.8± 3.22 123± 11.25

AIRM v med 7.6± 2.76 127.7± 1.64AIRM v max 5.9± 1.45 60.4± 1.84

SC5SR5 v med 4.6± 1.07 127.5± 1.96SC2SR8 v med 9.9± 2.08 128.7± 3.37

Sub v med 13.8± 2.49 127.4± 2.01Sub v max 7.4± 1.17 61.7± 2.83

(b)

error

AIRM 0.07

AIRM v med 0.15AIRM v max 0.35

SC5SR5 v med 0.18SC2SR8 v med 0.13

Sub v med 0.10Sub v max 0.32

(c) AIRM sub v max sub v med AIRM v max Sc2Sr8 v med AIRM v med Sc5Sr5 v med0

5

10

15

20

25

30

c\bvariance n.blobvariance time

Fig. 4. (a) Performances and standard deviations in term of number of counted blobsand time spent to accomplish the task. (b) Error on the number of counted blobs forunits of time. (c) Plot of costs/benefits of the tests. C/b is evaluated as time/countedblobs.

5 Conclusion

In this paper, we investigated simple attentional mechanisms for coordinatingcompetitive and cooperative behaviors in a behavior-based robotic system. Theresults show that the AIRM mechanisms are effective in adapting the frequencyof behaviors activations according to the particular circumstances, increment-ing or decreasing the attention toward salient aspects of the robot environmentor the internal state. We compared our architecture w.r.t. different architec-tures not endowed with attentional mechanisms. In summary, we observe that

the proposed architecture performs better than the others in terms of: num-ber of detected blobs (effectiveness); tradeoff between time and counted blobs(cost/benefit); error of detection (precision); less activations of the perceptualschema (efficiency). Basically, the system can modulate the activation frequencieson the basis of the available resources and external conditions. Indeed, by usingthe adaptive clocks, the number of behaviors activations substantially decreasescompared to the case in which the control system enables the robot’s behaviorsat each machine cycle, and this results in a substantial gain in performances.

References

1. Arbib, M.A.: Schema theory. In: The handbook of brain theory and neural net-works, pp. 830–834. MIT Press (1998)

2. Ballard, D.H.: Animate vision. Artif. Intell. 48(1), 57–86 (1991)3. Beccari, G., Caselli, S., Zanichelli, F.: A technique for adaptive scheduling of soft

real-time tasks. Real-Time Syst. 30(3), 187–215 (2005)4. Burattini, E., Rossi, S.: Periodic adaptive activation of behaviors in robotic system.

Int. J. Pattern Recognition and Artificial Intelligence 22(5), 987–999 (2008)5. Burattini, E., Rossi, S.: Periodic activations of behaviours and emotional adapta-

tion in behaviour-based robotics. Connection Science 22(2), (in press) (2010)6. Cooper, R., Shallice, T.: Contention scheduling and the control of routine activities.

Cognitive Neuropsychology 17, 297–338 (2000)7. Coull, J., Nobre, A.: Where and when to pay attention: The neural systems for

directing attention to spatial locations and to time intervals as revealed by bothpet and fmri. Journal of Neuroscience 18(18), 7426–7435 (1998)

8. Harbluk, J.L., Noy, Y.I., Eizenmann, M.: Impact of cognitive distraction on drivervisual behavior and vehicle control. Tech. rep., 81st annual meeting of the Trans-portation Research Board, Washington, DC (January 2002)

9. Maes, P.: A bottom-up mechanism for behavior selection in an artificial creature.In: Proc. of the first intern. conf. on simulation of adaptive behavior on Fromanimals to animats. pp. 238–246. MIT Press, Cambridge, MA, USA (1990)

10. Miniussi, C., Wilding, E., Coull, J., Nobre, A.: Orienting attention in time: Mod-ulation of brain potentials. Brain (122), 1507–1518 (1999)

11. Niebur, E., Hsiao, S.S., Johnson, K.O.: Synchrony: a neuronal mechanism for at-tentional selection? Curr Opin Neurobiol 12(2), 190–194 (April 2002)

12. Norman, D.A., Shallice, T.: Attention to action: willed and automatic control ofbehaviour. In: Davidson, R.J., Schwartz, G.E., D., S. (eds.) Consciousness andSelf-regulation: Advances in Research. vol. IV, pp. 1–18. Plenum Press, New York,NY (1986)

13. Parisi, D.: Internal robotics. Connect. Sci. 16(4), 325–338 (2004)14. Pashler, H.E.: The Psychology of Attention, chap. Divided Attention, pp. 101–166.

MIT Press, MA (1998)15. Patten, C.J.D., Kircher, A., Ostlund, J., Nilsson, L.: Using mobile telephones:

cognitive workload and attention resource allocation. Accid Anal Prev 36(3), 341–350 (May 2004)

16. Wasson, G., Kortenkamp, D., Huber, E.: Integrating active perception with anautonomous robot architecture. Rob. Autom. Syst. 26, 325–331 (1999)