Cover page Title: HARPIC, an Hybrid Architecture based on Representations, Perception and Intelligent Control: a way to provide autonomy to robots. First author: Dominique Luzeaux DGA/Centre Technique d'Arcueil 16bis av. Prieur de la C^ote d'Or 94114 Arcueil Cedex, France Second author: Andre Dalgalarrondo. Presenting author: Dominique Luzeaux Corresponding author: Dominique Luzeaux ([email protected]) Keywords: robotics { control architecture { intelligent system.

2

HARPIC, an Hybrid Architecture based on Representations, Perception and Intelligent Control: a way to provide autonomy to robots. Dominique Luzeaux and Andre Dalgalarrondo DGA/Centre Technique d'Arcueil, 16bis av. Prieur de la C^ote d'Or, 94114 Arcueil Cedex, France, [email protected],[email protected], WWW home page: http://www.etca.fr/CTA/gip/Publis/Luzeaux/

Abstract. In this paper we discuss an hybrid architecture, including

reactive and deliberative behaviors, which we have developed to confer autonomy to unmanned robotics systems. Two main features characterize our work: on the one hand the ability for the robot to control its own autonomy, and on the other hand the capacity to evolve and to learn.

1 Introduction As was mentioned recently in a call for participation to a special issue on intelligent systems design, complex intelligent systems are getting to the point where it almost feels as if \someone" is there behind the interface. This impression comes across most strongly in the eld of robotics because these agents are physically embodied, much as humans are. There are several primary components to this phenomenon. First, the system must be capable of action in some reasonably complicated domain: a non-trivial environment within which the system has to evolve, and a rather elaborate task which the system should ful ll. Second, the system must be capable of communicating with other systems and even humans using speci ed possibly language-like modalities, i.e. not a mere succession of binary data, but some degree of symbolic representations. Third, the system should be able to reason about its actions, with the aim of ultimately adapting them. Finally, the system should be able to learn and adapt to changing conditions to some extent, either on the basis of external feedback or relying on its own reasoning capacities. These remarks have guided our research on military unmanned ground robots for search and rescue or scouting missions. In order to bring up some answers and to build robots that could perform well in ill-known environments, we have focused on the robot control architecture, which is the core of the intelligence, as it binds together and manages all the components. In the next sections, we will rst discuss shortly what we understand under autonomy for robots and control architectures. Then we will describe the architecture we propose and give some implementation details and results. The nal sections are dedicated to the possibility for the system rst to control its own autonomy depending on external input and second to learn and adapt.

2

1.1 Autonomy in robotic systems In order to tackle autonomous robots, one has rst to delimit the scope of expected results; this calls for a temptative de nition of autonomy.A rst necessary condition for a system to be called autonomous is to be able to re reactions when faced with external disturbances: this yields a concept obviously parametrized by the nature and diversity of disturbances one can act against. However mere reaction to disturbances cannot be truly accepted as autonomy as it does not encompass longer-term decision abilities. A more demanding de nition includes the ability to change the interaction modes with the environment: this captures the idea that an autonomous organization is not static in its functioning ways and can \adapt". Looking closely at the implications, one sees that such an organization necessarily has to internalize external constraints, which means the ability to integrate knowledge of its own dynamics and representation of the exterior. To sum up, an interaction without explicit representation of both an internal world corresponding to the system and an external world relative to the environment cannot be called autonomous (consider a painting industrial robot with simple contact feedback as a counterexample). Notice that this does not mean the representations have to be entirely dierent (on the contrary, ecient sensorimotor closed loops require an integration of the various representations!). Concluding this paragraph on autonomy, we see that although there are epistemological necessary conditions for autonomy, there is no absolute autonomy: a system can reasonably only be said \more autonomous" than another. In our approach to autonomous systems, we have proceeded in a bottom-up fashion, handling rst the control and perception issues, and looking for adequate representations which could integrate both these issues, leading to sensorimotor i/o behavior [4,12, 13]. Since we are not interested in robot wandering aimlessly within corridors, possibly avoiding scientists strolling through the lab, but have to deal with changing situations, even rapidly changing, ranging from slight modi cation of the environment to its thorough transformation (or at least a transformation of the stored model: e.g. discrepancies between the cartographic memory of the world and the current location, due to stale information or possible destruction of infrastructures), we have incorporated deliberative and decision capacities in the system, that do not run necessarily at the same temporal rate than the i/o behavior execution. In order to scale up to system level, we must then turn to control architectures, i.e. we have to detail how the various functions related to perception and action have to be organized in order for the whole system to ful ll a given objective. To achieve this, we work with hybrid control architectures, integrating a lower level focusing on intelligent control and active perception, and a higher level provided through the mission planning.

1.2 Robot control architectures As a complex system collocating sensors, actuators, electronic and mechanical devices, computing resources, a robot has to be provided ways to organize these

3

various heterogeneous components in order to ful ll its prescribed mission, which furthermore may evolve in time. This is all the more important when additional constraints, such as real-time and cost issues { nowadays a major issue for operational systems { are involved. The control architecture deals with these problems and brings answers to the following questions: { how is the system built from its basic components? { how do the parts build up a whole? { how should components be (re)organized to ful ll missions changing in time? By basic components, one has to understand mechanical, electronical and software aspects, sensors, actuators, but also the ways to relate these elements and the interfaces between the various subsystems. For a general overview of existing control architectures, see [1]. The rst architectures historically introduced in mobile robots derive from the sense-plan-act paradigm taken from hard arti cial intelligence, and follow a top-down approach relying on a recursive functional decomposition of the problem into subproblems down to a grain level where an explicit solution to the problem is given. Such architectures have been shown to suer from the symbol-grounding [5], the frame and the brittleness problems [6]. In other words, they manipulate symbols which cannot be related in a constructive way to features of the environment and they have to rely on a model of the environment which has to be complete and rede ned hierarchically in order to cope with the top-down functional decomposition. While this works for static environments, any unplanned situation will have dramatic impact on the robot! [3] In reaction to that approach, bottom-up approaches have been proposed, inspired by biology and ethology. They do not rely on explicit models of the environment but on input-output reactive behaviors, which may be aggregated together to solve a more complex task. One of the most famous bottom-up architectures is Brook's subsumption architecture. However it is generally admitted that both approaches have failed, mainly because of their radical positions: top-down approaches lead to awkward robots unable to cope with any unforeseen change of the environment or the mission, while bottom-up approaches have led to promising animal-like robots which unfortunately could not solve complex problems or missions. Building on their respective advantages (and hopefully not cumulating their drawbacks!), hybrid architectures have been investigated in recent years. They try to have on the one hand a reactive component and on the other hand a decision or planning module; the diculty is of course the interface between these layers and here lies the diversity of the current approaches.

2 HARPIC 2.1 General description We propose an hybrid architecture (cf. gure 1) which consists in four blocks organized around a fth: perception processes, an attention manager, a behavior

4

selector and action processes. The core of the architecture relies on representations. events

attention manager

behavior selection

representations

activation inhibition

behavior

errors sensors perception current action

action selection

actuators

Fig.1. Functional diagram of the HARPIC architecture. Sensors yield data to perception processes which create representations of the environment. Representations are instances of specialized perception models. For instance, for a visual wall-following behavior, the representation can be restricted to the coordinates of the edge detected in the image, that stands for the wall to follow. To every representation are attached the references to the process that created it: date of creation and various data related to the sensor (position, focus...). The representations are stored in a table with xed constant length, so that a round-robin mechanism keeps a given memory depth. Representations are thus snapshots of speci c landmarks in the robot's environment, about which the spatial and the temporal localization are known. The perception processes are activated or inhibited by the attention manager and receive also an information on the current executed behavior. This information is used to foresee and check the consistency of the representation. The attention manager updates representations (on a periodical or on an exceptional basis), supervises the environment (detection of new events) and the algorithms (prediction/feedback control) and guarantees an ecient use of the computing resources. The action selection module chooses the robot's behavior depending on the prede ned goal(s), the current action, the representations and their estimated reliability. Finally, the behaviors control the robot's actuators in closed loop with the associated perception processes. The key ideas of that architecture are:

 The use of sensorimotor behaviors linking internally and externally perceptions and low-level actions: the internal coupling allows to compare a prediction of the

5

next perception, estimated from the previous perception and the current control, with the perception obtained after application of the control, in order to decide whether the current behavior runs normally or should be changed.  Use of perception processes with the aim of creating local situated representations of the environment. No global model of the environment is used; however less local and higher level representations can be built from the instantaneous local representations.  Quantitative assessment of every representation: every algorithm has associated evaluation metrics which assign to every constructed representation a numerical value that expresses the con dence which can be given to it. This is important, because any processing algorithm has a domain of validity and its internal parameters are best suited for some situations: there is no perfect algorithm that always yields \good" results.  Use of an attention manager: it supervises the executions of the perception processing algorithms independently from the current actions. It takes into account the processing time needed for each perception process, as well as the cost in terms of needed computational resources. It looks also for new events due to the dynamics of the environment, which may signify a new danger or opportunities leading to a change of behavior. It may also re processes in order to check whether sensors function nominally and can receive error signals coming from current perception processes. In practice, for instance with a vision sensor, the attention will focus on the illumination conditions, on the consistency between the movement of the robot and the temporal consistency of the representations, and on error signals sent by perception processes. With this information it is then possible to invalidate representations due to malfunctioning sensors or misused processes.  The behavior selection module chooses the sensorimotor behaviors to be activated or inhibited depending on the prede ned goal, the available representations and the events issued from the attention manager. This module is the highest level of the architecture. It should be noted that the quantitative assessment of the representations plays a key role in the decision process of the behavior selection: on the one hand a representation might be more or less adapted to the current situation, depending for instance on the sensor used or on the conditions of the perception acquisition (e.g. a day camera used during night operating conditions will yield representations to which a lower con dence should be assigned a priori; the same holds also for instance for a detector relying on translational invariance while the robot has such a motion that this assumption is incorrect) ; on the other hand some representations might be more interesting for some behaviors or might provide an improved help to choose between several behaviors (e.g. a wall-following behavior needs information on contours more than velocity vectors, while a tracking behavior has the opposite needs). Therefore every behavior weights also each representation depending on its direct usability, and this weight is combined with the intrinsic assessment of the representation.  The action selection module regroups the lower-level controllers operating on the actuators. It uses valid representations in order to compute the control laws.

6

This modular architecture allows to develop independently the various processes belonging to each of the four entities, before integrating them together. Its originality relies in both the lower-level loop between perception and action, necessary for instance for active vision and any situated approach of perception, and the decoupling of the perception and the action processes, which become behaviors during their execution. The latter avoids the redundancy of common components, saves computational resources when representations are common to several behaviors and limits the con icts when accessing the hardware resources of the robot. Compared to other hybrid architectures (the so-called three-layer approach which develops the three following levels: symbolic level, reactive behaviors, lower-level control), we have tried to focus more on the relations between these three levels in order to take into account the heterogeneous loop aspect characterizing a complex robot. Furthermore, the proposed architecture with both loops between perception and action, one at a lower level, the other one relying on representations, seems a plausible model from a biological point of view [2]: whereas the lower-level sensorimotor closed loop is characteristic of simpler organisms, the cognitive capacities of higher organisms, like mammals, can be explained by another loop relying on a dynamical model of the environment and the organism, running in feedforward mode with a slower time scale. This higher-level loop allows also open loop planning in order to check the validity of some representations and can trigger their update (e.g. the racing pilot who concentrates on the track and the race, when rehearsing in his head before he actually starts the race). Finally, the attention manager is a notion used in biological vision [9], and provides a very ecient supervision concept for arti cial systems, ring batch processes or reacting on discrete events. The asynchronous property of the control architecture is due to that attention manager, and we think this property is a keystone in complex systems, which have to deal with unpredictable environments and limited resources.

2.2 Management of perception resources

Behaviors are chained logically and this chaining depends on a succession of events that can be detected by the perception algorithms which are part of these behaviors. For example, an obstacle avoidance behavior follows naturally a road following behavior when the perceptual subset of the avoidance behavior has detected an obstacle. More generally, to every active behavior corresponds a subset of perception algorithms among all perception algorithms of the robot, and their activation must allow the detection of an important event that signi es the modi cation of the behavior of the robot. Among detectable events a distinction will be made between the important events for the security of the robot, the events that have only an immediate in uence on the current behavior, and the events that have no direct in uence or have a very low probability of occurrence. Thus the perception algorithms can be classi ed into three categories. Figure 2 shows an example of such relations between sensorimotor behaviors and perception processes. Since every such perception process can use dierent

7 Perception processes

ll

Sensorimotor behaviors

De

t

a ect

De

te

r ct a

De

te

n ct a

ion

cts bje ark le ark t tac get le o dm andm rge tar l obs a lan obi a n m a ta t a t k t c a t c c c e ck st c e t a e t t r te Te T De Tra De De De ect

ers

oad

wa

int

Follow a wall Follow a road Avoid an obstacle Circumnavigate an obstacle Keep a heading Backtrack Reorient Go back to starting point Estimate own movement Visual search Supervise Stroll Go to a target Stop

: Mandatory

: Consistent

: Inconsistent or useless

Fig. 2. Correspondence between sensorimotor behaviors and perception processes. techniques with dierent parameters, such a table shows how to relate the current goal of the robot (the active behavior) and a partially ordered subset of the perception algorithms. Besides ordering perception processes as a function of their utility, the attention manager has to guarantee a proper reactivity for the robot. To achieve this, we propose to consider the computing cost of every perception algorithm and to allocate to the manager a computing time quota that will be shared between mandatory (directly in uencing security) and consistent perceptions, so that the former can be activated regularly. The detailed steps are presented on gure 3. As illustrated previously, the classi cation of the perception algorithms is given to the attention manager as a look-up table, which is prede ned by the robot designer. This allows to take a priori knowledge into account; besides, the classi cation is a complex task which is not accessible to the attention manager unless it loses its simplicity. However the possibility of modifying dynamically these tables should not be excluded. For instance the duration allocated to the attention loop could depend on the speed of the robot, in order to increase the ring frequency of the perceptions dedicated to security safeguard.

8

Read name of active behavior Fire mandatory perceptions / active behavior Gather consistent perceptions / active behavior Select and activate chosen consistent perceptions Detect new events T = (allocated duration) - (loop duration) no

yes T<0

Wait during T

Reduce duration of next loop

Fig. 3. Algorithm of the attention manager.

2.3 Assessment mechanisms within the control architecture

We have mentioned in a previous section that the representations manipulated by the architecture were quantitatively assessed. We go back on that critical issue in the following paragraphs and discuss also the various quantitative assessment mechanisms introduced at each level in order to make the architecture operational (see gure 4 for a graphical illustration). In our robot control architecture, the action selector executes behaviors, built up from both perception and action processes, whereas the attention manager only selects perception processes. A rst consequence of such an organization is the distinction between perception and action processes. A second consequence is that the perception processes are passive in the sense that they do not control the frequency of their activation. Thus they are not the best choice to assess the temporal consistency of the representations they yield. Indeed such an assessment needs not only a regular activation of the underlying process but also the knowledge of the previous actions and perceptions. Therefore it seems logical for the assessment of the temporal consistency to be done by the action processes. The perception processes will only take care of their instantaneous evaluation with regard to the input data, the obtained computed results and the internal knowledge of the processing algorithms.

9 Attention manager Perception selector

Selector

Comparative evaluation

Long-term evaluation

Events

Proprioceptive and exteroceptive data

Events

Behavior selection

Commands Perception process Perception processing

Instantaneous evaluation

Representations

Commands

Events

Action process Control law

Control signal

Temporal evaluation

Fig.4. Repartition of the assessment mechanisms in the various components of the control architecture.

The attention manager will be the best choice for a comparison of all representations, since it can re at the same time or almost a bunch of perception processes the results of which can subsequently be compared. Such an assessment takes into account dierent representations, or representations of the same type but coming from processes using dierent techniques, which have been generated at approximately the same time. A long-term assessment of the behavior of the perception processes with regard to the execution of the tasks assigned to the robot has also to be performed: the most logical choice to realize this assessment is the action selector since such a mechanism needs knowledge that is usually associated to deliberative processes, and the planning is made at the action level. It should be noted that such a repartition enhances the autonomy and competency of every component of the robot control architecture. It allows an asynchronous scheduling of the assessment tasks, which can be performed at the own rhythm of every module. Another point we wanted to make is not to use a central omniscient assessment mechanism. We have delegated the instantaneous assessment to the perception processes because of two main reasons: rst we think that every representation has to be given a mark by its designer; secondly, if the perception abilities of a robot have to be increased by adding new perception processes, such an addition should imply few or no change on the other components of the robot. The methods

10

used for that assessment will be mostly empirical methods based on appropriate measures on data and/or results of the processing algorithm (see [15] for a detailed exposition of existing measures and references). However such methods do not care about the semantics of the object they evaluate. Therefore we will try to include within the metrics something that not only measures the quality of the result but that states whether the algorithm is adapted or not to the current images. There are no such general metrics and they have to be developed for every algorithm, possibly by a learning procedure during the design phase by working on representative images and controlled disturbances. The measure associated to every perception process consists in a numerical mark between 0 and 1 (for normalization reasons, when representations of dierent types are created and fusion procedures are used) and in a validity threshold, that says whether the representation is within or outside the prede ned domain of validity of the algorithm. The mark can be determined by a normalized dierence between the result and a model or any other ad hoc measure: in the application described later, where a robot will drive on a track or o-roads, we use an algorithm for the detection of the road sides; the mark is based on the dierence with a trapezoidal model of the road (an ideal road with parallel sides has a trapezoidal shape when seen by the camera mounted on the robot due to the perspective). This is only an example and shows that during the design phase the roboticist has to de ne and tune the internal marking mechanism for every perception process. The attention manager will then be in charge of the assessment of all created representations; since he controls the parallel activation of the perception agents, he is able to compare the representations by controlling their temporal and spatial dierence. Such an assessment is based on the marks given to every representation by the perception process which has created it:

{ when several representations are available from dierent algorithms, but of

the same type as the representation used by the current behavior, the attention manager will notify the behavior selector which perception process is best suited to the current behavior by sending an event. Notice that the case of redundant processing is one of the key ideas of the architecture and a raison d'^etre of the attention manager; { for all representations that are complementary to the current representation, the mark should tell whether they are better suited to the robot environment and will directly be used to de ne their priority in the activations made by the attention manager. A representation with a bad mark will lower the priority of the associated perception process.

Thus the instantaneous assessment made by every perception process is crucial for the good performance of the attention manager. As discusses [7], we advocate the use of perception processes which are well-known as to their models and limits, in order to have an ecient global assessment procedure. The measure of temporal consistency done by every action process is based on a dierence measure between the former representation and an estimate computed from older representations: in our application, we use a Kalman lter.

11

2.4 Implementation and experiments Fundamental capacities of our architecture encompass modularity, encapsulation, scalability and parallel execution. For these needs we decided to use a multi-agent formalism that ts naturally our need for encapsulation in independent, asynchronous and heterogeneous modules. The communication between agents is realized by messages. Object oriented language are therefore absolutely suited for programming agents: we chose C++. We use POSIX Threads to obtain parallelism: each agent is represented by a thread in the overall process. We do not aim at promoting multi-agent techniques, for us it is merely an interesting formalism and our architecture could be implemented without them. All the processing is done by a PC bi-PIII 500 MHz running Linux, with a frame grabber card. We did not use any specialized hardware or real-time operating system. The robot used both in indoor and outdoor experiments is a Pioneer AT from RWI equipped with a monochrome camera, with motorized lens, mounted on a pan-tilt unit (cf. gure 5). The robot links to the computer are a radio modem at 9600 bauds and a video link which transmits 768  512 pixel images at a 25 Hz rate. The robot displacement speed is about 400 mm/s.

Fig.5. Experiments with the robot (indoor and outdoor navigation). We have developed perception agents, for lane, obstacle and target detection, and action agents for lane following, obstacle avoiding, target tracking and lens aperture adjusting. The processing time of the perception agents varies between 5 and 30 ms according to their algorithms. Experiments were conducted in indoor and outdoor environments without changing anything to the hardware and software architecture of the robot in between. For these experiments, the robot can choose among two behaviors: wallfollowing and obstacle avoidance. The wall-following behavior can be activated only after the detection of a wall edge and if there are no obstacles detected.

12 Activation chronogram of the perceptions of the current behavior 8 3 : Obstacle perception through ultrasound 4 : Wall edge perception de bord de mur through gradient 5 : Wall edge perception through mean and variance 6 : Wall edge perception through multifractal measure

7

Index of the active perception

6

5

4

3

2

1

0 0

20

40

60

80

100 Time (s)

120

140

160

180

Fig.6. The Pioneer robot during a wall-following experiment, with the activation chronograms of the perception agents used by the current behavior.

The perception of a wall edge can be done by four perception algorithms: a perception through the ultrasound sensor based on a simple thresholding of the time of ight, and three algorithms based on vision through a video camera. The rst is a gradient extraction and can detect the edge of the wall. The second is a texture segmentation based on the computation of the mean and the variance, followed rst by a thresholding that delimits the boundary between the ground near the robot and other elements of the scene, second by an estimation of the segment representing the wall in the least-square sense. The third algorithm uses a segmentation based on the calculus of a multifractal texture measure [8], and proceeds then as previously; it is much more adapted to strongly textured images but is rather computing time consuming. These various algorithms are a good illustration of the possible situations: they have very dierent computational costs and return very dierent results, and they use several sensors. In the indoor environment experiment, the robot has walked through the lab for about twenty minutes, looking for walls and following them, avoiding potential moving obstacles and running through such a loop inde nitely. If one looks at the chronogram on gure 6, where only part of the experiment has been kept in order to keep some readability, one notices that the wall detection with the gradient algorithm is not used very much, and that the algorithm mainly used is based on the statistical texture segmentation (perception with index 5).

13 Activation chronogram of the perceptions of the current behavior 8 3 : Obstacle perception through ultrasound 4 : Wall edge perception through gradient 5 : Wall edge perception through mean and variance 6 : Wall edge perception through multifractal measure

7

Index of the active perception

6

5

4

3

2

1

0 0

10

20

30

40

50

60

70

80

90

Time (s)

Fig.7. The Pioneer robot during a curb-following experiment, with the activation chronograms of the perception agents used by the current behavior.

In the outdoor experiment, the same wall-following application is used for curb-following without any parameter change. The robot has followed the curb in a satisfactory manner (for an external observer) for several dozens of meters. A look at the chronogram on gure 7 shows immediately (the time scale is slightly dierent and has been expanded for readability) that the perception agents change more frequently. This comes from a higher irregularity of outdoor images compared with the previous indoor corridor. Although there were no obstacles on the trajectory of the robot, the obstacle avoidance algorithm is activated very often: this happens when there is lack of a valid curb edge detection. However this is not annoying since it only re ects the complexity of the environment (recall that the image processing algorithms were not adapted to these speci c images) and shows on the contrary how the system performs in dicult perception conditions in a way consistent with the de ned objectives. The most interesting thing that arises from a comparison of both chronograms is the choice of the algorithms red in both environments. Indeed, one notices that for the outdoor environment, the wall edge perception agent most associated to the wall-following behavior is the multifractal texture segmentation (perception with index 6). This choice of the perception algorithm is not the result of a deliberative process or of an a priori choice of the robot designer, but is an adaptation of the robot to its environment which emerges from the functional organization

14

of the robot control architecture. It validates experimentally the advantage of the attention mechanism in its role of detecting new events and comparing the relevance of several perceptions. Of course, a critic could arise concerning the intrinsic stability of the behavior selector: one could fear a new behavior would be chosen every time step. In order to avoid this, we have established a dichotomy between behavior selection and action selection. A behavior has thus a dierent time scale than the ring of an action or the data acquisition process; a consequence is a temporal smoothing. Besides, the behavior selector uses a knowledge base which de nes which behaviors might follow a given behavior. This reduces the combinatorics and introduces some determinism in the behavior scheduling. One should note that in order not to have too much determinism, we allow the attention manager to interrupt the current behavior whenever it is necessary, depending on unforeseen detected events. Our dierent experiments up to now have shown that this balancing mechanism works well. Such intrinsic instability within the architecture should not be judged too severely, since we introduced intentionally asynchronism and non-determinism in order to be able to cope with unforeseen and changing situations: solving the frame problem and the brittleness problem cannot be done without introducing some structural instability, but one should take care to provide some means to control that unstability as best as possible. To conclude this experimental illustration, let us underline that if other perception agents are added, that use the same image processing algorithms but with ne parameter tuning speci cally adapted to the environment (for instance by giving a wall edge representation a mark greater than the validity threshold only if the edge was perfectly detected), the behavior changes correspond exactly to the passage of the doors in the corridor. This hints at higher-level perceptive abilities, where some kind of introspection could interpret the behavior changes and deduce a categorization of the environment.

3 Computational intelligence and controlled autonomy We have seen in the experiments how a straightforward scheduling of contextual modes is allowed through a parallel execution of a wide range of algorithms and a choice of the \best action", depending on the evaluation of the most adequate representation given a current situation and behavior. For instance, for an outdoor robot, there is no need to tell explicitly the robot when it drives on structured roads and when it is o-road, since the right representation computed by the various vision algorithms should be selected by the navigation behavior. In our experiments within indoor environments with competing behaviors such as wall-following, obstacle avoidance, navigation with given heading, we have observed a natural transition from wall-following to navigation with given heading and back to wall-following when a door occurred. The same occurred in structured outdoor environments, where curbs or road marks can suddenly disappear, either because of a crossroad or more often because the reference structure is partially damaged. Whereas usual roboticists

15

would then ask for more robust perception algorithms, we have preferred a design with enhanced intelligence that does not require perfect perception or action to work. As a consequence of what has been discussed up to now, a most interesting property is the fact that the robot can control its own autonomy by relying on external features, such as perceptual landmarks that can trigger special events or behaviors once they have been detected by the perception processes, transformed into representations and interpreted by the attention manager. A typical straightforward application consists in an autonomous robot looking for trac signs along the road, that limit its freedom of decision. A similar example is an autonomous robot, to which instructions are given for instance through sign language or even speech (speech processing algorithms are then needed among the perception processes), which yield either information on the environment or provide orders to be followed, like modi cation of the current behavior. Such ability to switch from autonomous to semi-autonomous modes is from our point of view a true sign of computational intelligence: while autonomy is often conceived for an arti cial system as not needing external rules to choose its own behavior, we think that being able to follow suddenly rules while you were evolving freely just relying on yourself is a proof of intelligent design!... and a prerequisite for robots to be accepted among humans. More elaborate control modes of the robot's own autonomy obviously can be conceived which do not necessarily rely on a priori semantics to be discovered by the perception processes. Such less explicit control of autonomy is exerted in a multirobot scenario [18], where we have implemented several robots with the previous architecture after having provided it with an additional communication process which allows the exchange of individual representations among dierent robots upon request. The autonomy of every robot is then implicitly limited by the autonomy of its partners, as the completion of the task is evaluated at the group's level and what is actually developed is the multirobot group's autonomy.

4 Computational intelligence and learning In previous work [11,12,16], we have presented methodologies that allow to learn the sensorimotor lower-level loop. They proceed in two phases: identi cation of a qualitative model of the dynamical system and inversion of that model to deduce a fuzzy-like control law. Such research has shown that furthermore higher-level representations could emerge relative to the topology of the environment or the sequencing of behaviors [14,19]. It is interesting to see where and how similar learning mechanisms can be introduced within the control architectures. The lower reactive level which binds the perception and the action selection modules have already been dealt with as was recalled previously. Of course it is always possible to improve the already available methods: for instance, learning control laws up to now does not include any explicit prediction, as there is no quantitative analysis or comparison of the

16

actual eect of the input with respect to the predicted eect. If this was achieved, it would be possible to include corrective terms within the learned control law (as is done in predictive feedforward control) and to have reasoning mechanisms on the consistency of the identi ed qualitative model relatively to the observed eects (techniques on qualitative model veri cation from [10] could be used for that purpose). But it would also be instructive to add learning mechanisms at the representation level, especially concerning their update: as in biological systems, forgetting or short-term and long-term memory mechanisms could be implemented. Concerning the introduction of learning mechanisms within action selection algorithms, the literature is rather rich [17], and most techniques described in the reference could be implemented advantageously in our work. On the contrary, introducing learning at the attention manager level is an open problem: it could be done in order to interpret either the awaited events as a function of the current behavior and the potential behaviors that could follow it, or unexpected events that can trigger emergency answers. Another direction could be to have adaptive sequencing of the behaviors.

5 Conclusion We have discussed robot control architectures and proposed an architecture which aims at lling the gap between reactive behavioral and deliberative decision systems, while keeping a close eye on the dynamic management of all the resources available to the robot. We have also shown how it was possible to propose various autonomy modes, since the autonomy of the robot can be easily controlled by feeding the environment the landmarks that, after interpretation, provide selective behaviors. Finally, we have discussed the introduction of learning techniques at the various levels of the architecture, and we have argued how such learning mechanisms, if successfully implemented, could yield not only adaptive, but also evolutive control architectures. All of this would then achieve a degree of autonomy, about which we may only dream for now...

References 1. R.C. Arkin. Behavior-based robotics. A Bradford Book, The MIT Press, 1998. 2. A. Berthoz. Le sens du mouvement. E ditions Odile Jacob, 1997. 3. M.A. Boden, editor. The philosophy of arti cial life. Oxford Readings in Philosophy, 1996. 4. A. Dalgalarrondo and D. Luzeaux. Rule-based incremental control within an active vision framework. In 4th Int. Conf. on Control, Automation, Robotics and Vision, Westin Stamford, Singapore, 1996. 5. S. Harnad. The symbol grounding problem. Physica D, 42:335{346, 1990. 6. P.J. Hayes. The frame problem and related problems in arti cial intelligence. In J. Allen, J. Hendler, and A. Tate, editors, Readings in Planning, pages 588{595. Morgan Kaufmann Publishers, Inc., 1990.

17 7. I. Horswill. Specialization of perceptual processes. Ph.D. Dissertation,MIT, Cambridge, MA, 1993. 8. L. Kam. Approximation multifractale guidee par la reconnaissance. Doctorat de l'Universite Paris XI, Orsay, 2000. 9. S.M. Kosslyn. Image and Brain. A Bradford Book, The MIT Press, 1992. 10. B. Kuipers. Qualitative Reasoning: Modeling and Simulation with incomplete knowledge. The MIT Press, 1994. 11. D. Luzeaux. Let's learn to control a system! In IEEE International Conference on Systems Man Cybernetics, Le Touquet, France, 1993. 12. D. Luzeaux. Learning knowledge-based systems and control of complex systems. In 15th IMACS World Congress, Berlin, Germany, 1997. 13. D. Luzeaux. Mixed ltering and intelligent control for target tracking with mobile sensors. In 29th Southeastern Symposium on System Theory, Cookeville, TN, USA, 1997. 14. D. Luzeaux. Catastrophes as a way to build up knowledge for learning robots. In 16th IMACS World Congress, Lausanne, Switzerland, 2000. 15. D. Luzeaux and A. Dalgalarrondo. Assessment of image processing algorithms as the keystone of autonomous robot control architectures. In J. Blanc-Talon and D. Popescu, editors, Imaging and Vision Systems: Theory, Assessment and Applications. NOVA Science Books, New York, 2001. 16. D. Luzeaux and B. Zavidovique. Process control and machine learning: rule-based incremental control. IEEE Transactions on Automatic Control, 39(6), 1994. 17. P. Pirjanian. Behavior coordination mechanisms: state-of-the-art. Technical report, University of Southern California, TR IRIS-99-375, 1999. 18. P. Sellem, E. Amram, and D. Luzeaux. Open multi-agent architecture extended to distributed autonomous robotic systems. In SPIE Aerosense'00, Conference on Unmanned Ground Vehicle Technology II, Orlando, FL, USA, 2000. 19. O. Sigaud and D. Luzeaux. Learning hierarchical controllers for situated agents. In Proceedings of the 14th International Congress on Cybernetics, Bruxelles, Belgium, 1995.