N93-25970

An Autonomous Satellite Architecture Integrating Deliberative

Reasoning and Behavioural Intelligence.

Craig A. Lindley

Department of Computer Science and EngineeringUniversity of New South Wales, Kensington, NSW, Australia

ABSTRACT

This paper describes a method for the designof autonomous spacecraft, based uponbehavioural approaches to intelligent robotics.First, a number of previous spacecraft

automation projects are reviewed. Amethodology for the design of autonomousspacecraft is then presented, drawing uponboth the European Space Agencytechnological centre (ESTEC) automation androbotics methodology and the subsumptionarchitecture for autonomous robots. A layered

competency model for autonomous orbitalspacecraft is proposed. A simple example oflow level competencies and their interaction ispresented in order to illustrate themethodology. Finally, the general principlesadopted for the control hardware design of theAUSTRALIS-1 spacecraft are described. Thissystem will provide an orbital experimentalplatform for spacecraft autonomy studies,supporting the exploration of different logicalcontrol models, different computational

metaphors within the behavioural controlframework, and different mappings from the

logical control model to its physicalimplementation.

Keywords: Spacecraft Control, SpaceRobotics, Artificial Intelligence, Subsumption.

Introduction

AI applications in space systems are becomingmore readily accepted, and constitute a keyenabling technology for ambitious projectssuch as the Space Station Freedom and SpaceExploration Initiative. Current or proposedconstellations of unmanned spacecraft,

particularly in low earth orbit, and multipledeep space missions with longtelecommunication propagation delays, canalso gain substantial benefits from the use ofmore autonomous spacecraft operation.

Teleoperation of industrial space facilities andorbital experimental platforms using highlyautonomous onboard systems may providecrucial competitive advantages m thecommercial and industrial exploitation of

space.

This paper presents an architecture forautonomous spacecraft control that supportsthe integration of behaviour-based approachesto emergent intelligence with numerical andcomputational simulation models, andsymbolic reasoning systems such as expertand knowledge based systems. Firstly, amethodology is proposed for developingautonomous space systems. Using thismethodology, system operational functions arehierarchically decomposed, but functionallevels are not mapped directly ontocomputational models. The operationaldecomposition is used to refine specificationsof layered competencies, based upon a generic

layered competency model. Each competencylevel defines a virtual machine interface from

the point of view of superordinate levels.Hence, the hierarchical decomposition of

system functionality during operationalanalysis does not imply a strict correspondinghierarchical synthesis for design and

implementation, but provides a framework forspecifying system behaviours and resources,and for understanding their interactions.

The realisation or implementation of thefunctionality of successive virtual machinescan be carried out using the most appropriatecomputational paradigm, or a richcombination of paradigms. From this point ofview, a knowledge base or expert system can

be regarded either as a convenient abstractionadopted during the design process to definethe input/output behaviour of behaviouralmodules, or as a resource for use bybehavioural modules much as human

91

https://ntrs.nasa.gov/search.jsp?R=19930016781 2020-04-16T09:18:40+00:00Z

7 "

operators would use expert systems forparticular tasks.

data handling system and onboard computer inthe event of their failure.

A simplified example application of thisapproach is presented. The resulting satellitecontrol architecture is significantly differentfrom previous satellite designs, havingimproved robustness, decreased operatingoverheads, and more autonomous faulttolerance. Current plans are to validate andrefine this approach in a rich simulationenvironment, and eventually to build andoperate a satellite for ongoing in-orbit trialsand experiments.

Precedents

Onboard autonomy is a matter of degree.

Pidgeon et al (1992) describe how a number ofcurrent spacecraft have mechanisms for faultdetection, reporting, and subsequent switchingto component, subsystem, or system fail-safe

The degree of onboard autonomy iscontinuously increasing, and can be expectedto incorporate a wide variety of techniquesfrom AI and intelligent robotics research. Anumber of prototype systems have been builtto investigate onboard expert systemapplications, including DIPOLE, SAGES,APS, SACV, and SICON (see below).

Operational systems may include the CassiniTitan Probe, and many Space Station Freedomapplications. Most of these systems involveonboard architectures comprising a number ofdistinct modular functions. Autonomous

systems of increasing size and complexitytend to have distributed functionality, with thevarious functions running on separate physicalprocessors. Another recurrent theme is thedevolution of autonomous functions to the

lowest possible_ abstraction levels, :modes. Human operators must then diagn0sefaults and initiate appropriate contingencyrecovery procedures. Increasing abstractionlevels in spacecraft command languages havealso been adopted. For example, in normaloperations, the Hipparcos spacecraft iscontrolled by processed commands which aresent to the onboard computer for distributionto other systems, and for possible timetagging. Direct commands are also available,which bypass the onboard computer as abackup in the event of computer failure, andfor more direct access to the controlled

systems. Priority real time commands can alsobe issued, which allow direct switching ofsystems. The ERS-1 spacecraft, which is in a

low polar orbit with limited ground access, hasa similar command macro system, with fourcommand types providing different functionsand levels of authority. The lowest levels ofcommands bypass the onboard computer anddata handling system, again ensuring control ifthose systems fail. The EURECA system,comprising fifteen separate payloads, uses anonboard Master Schedule which contains a list

of time tagged command macros for executionby the onboard data handling system. Thosecommands include rudimentary failureroutines, backed up by safe modes to deal with

command loop failure. Again, directtelecommands can also be used, to bypass the

Tello (1986) describes DIPOLE, a system forsatellite control which is intended to integrate"shallow" heuristic or rule based reasoningwith "deep" model-based reasoning. Theshallow system uses fault-tolerant

mathematical and algorithmic subroutines, andhas the form of real-time expert systems withdata-driven switches for controlling theirperformance. The aim is for the deepreasoning system to take over when theshallow system gets into difficulty, and toallow the shallow system to resume when thedeep reasoning system has resolved theproblem. The DIPOLE architecture addressestwo particular problems for real timedeliberative controllers. Circumspection is the

problem of enumerating all implicit conditionsand assumptions associated with givenknowledge, and the ability to handle situationswhen these are no longer valid. Inferencethrashing is the situation when inferencingcannot produce solutions quickly enough tokeep up with changing circumstances.DIPOLE seeks to address these problems byusing shallow rule based expert systems as

reflex processors in real time, with longerdeliberative processes performed by the deepreasoning system.

Ciarlo et al (I987) describe a spacecraft expertsystem prototype study conducted for the

92

European Space Agency. Some of theconclusions of the initial study include:

-highly simplified interfaces typical of

spacecraft modular units reduce integrationand control problems. However, thisseverely limits the information available formonitoring each unit, and the choice ofactions available to correct failure, to the

point of making the advantages of expertsystems questionable when compared tostandard algorithmic or table-drivensoftware.

-it is difficult, and not necessarilyadvantageous, to use an expert system in asatellite designed without this in mind.

Ciarlo and Schilling (1988) report upon workfollowing on from this initial study to consider

an expert system embedded within the CassiniTitan probe, for autonomously managing thedescent of the probe into Titan's atmosphere.The authors note that to keep the complexityand susceptibility of the system to faults aslow as possible, the autonomous systemshould be implemented at the lowest possiblelevel, with capabilities such as component andsensor self testing and redundancy switching.Scientific management, which involvesadaptation to the situation according tocomplex rules, is regarded as an appropriatefunction for implementation as a knowledgebase. Engineering management, involvingFDIR and subsystem control, is regarded as anappropriate function for conventionaltechnology.

The Satellite Autonomy Generic ExpertSystem (SAGES) architecture, developed byRockwell, is based upon the definition of fourintelligent agents, corresponding to phases ofthe mission operation cycle, includingplanning, scheduling, execution, and analysis(providing feedback into the planning phase;Raslavicius et al, 1989). A SAGES prototypehas been developed for a "typical"surveillance satellite.

The Boeing Aerospace Autonomous PowerSystem (APS) testbed has been assembled foruse in developing improved control techniquesfor aerospace electrical power systems (Spierand Liffring, 1989). The main emphasis ofAPS is the development of a programming

environment to properly control theconcurrent execution of multiple autonomous

algorithms coupled with continuous input andoutput data flow. Expert system functions

include fault diagnosis and recovery, andbattery charge control. The expert systems useevent-driven processing within a blackboardenvironment.

The European Space Agency (ESA) StandardGeneric Approach to Spacecraft Autonomyand Automation (SGASAA), is a hierarchical

model which aims to devolve decision-makingto the lowest possible level (Pidgeon et al,1992). To this end, it is a distributed onboard

architecture, with each payload and subsystemhaving a certain degree of "intelligence", inaddition to an Onboard Mission Manager(OBMM) responsible for the control of thespacecraft as a whole. Separate subsystemmanagers are intended to handle their ownfailures and report the results of theirdiagnoses via LAN to the OBMM, along witha proposed recovery action. The OBMM canauthorise the proposed recovery, or block it ifthe failure is caused by a failure elsewhere.SACV (ibid) is an investigation of theSGASAA concept involving theimplementation of a fully autonomousspacecraft based upon EURECA.

SICON, built by LISP Machine Inc, is asimplified prototype system for satelliteintelligent control, concentrating upon theelectrical power system (Leinweber, 1987).The SICON prototype deals with loaddistribution and switching, solar arrayorientation, power system verification andcheckout, fault diagnosis, trend analysis,contingency management, battery charge andreconditioning-cycle optimisation, and fuelcell monitoring and control. Leinweber notesthat there are a number of SSF processes andsubsystems that are amenable to real-timeprocess control, including the electrical power

systeml attitude and orbital control system,environmental and life support system,propulsion system, monitoring of dockedvehicles, manufacturing process control, andground communications and network control.Such real-time onboard applications requireparticular expert system features, includinghigh-speed context-sensitive rule activation,efficient memory recycling, acceptance of

93

interactive commands without suspendingexecution, and communication betweenmultiple expert systemsin order to provideredundancy.

The Space Station Freedom (SSF), willrequireanextensivedataprocessingsupportenvironment. The communications andinformationprocessingbackboneof theSSFisthe Data ManagementSystem (DMS). TheDMS hasthedual role of providinghardwareresourcesandsoftwareserviceswhich supportdataprocessingandcommunicationsneedsofthe system, its elements, and payloads(Erickson, 1987). It also functions as anintegrating entity, providing a commonoperating environment and human-machineinterface for the operation and control oforbitingSSFsystemsandpayloadsbyboththecrew and ground operators. The DMSprovides signal conditioning, and timingsynchronisation of data required forinterpreting time-critical information andresults betweenexpert systems,knowledge-based systems, and robotics elementsdistributed throughout the SSFenvironment.Woods (1992) notes that the DMS may useartificial intelligence techniques for faultdetection,isolation, andrecovery (FDIR) onDMS components. An Integrated SystemsExecutive(ISE) will provideoverall softwareschedulingandcontrol for all other systems,experiments, and elements. Low levelsoftware modules will take care of time

demonstrated more fundamental benefits in

addressing the problems of circumspectionand inference thrashing that have plagueddeliberative robot control systems. There istherefore considerable potential forbehavioural approaches to contribute to theincreased automation of space systems.Toward this end, this paper proposes amethodology for autonomous spacecraftdevelopment which draws from bothbehavioural approaches to mobile robotics andan ESTEC (European Space Agency'stechnological centre) methodology for spaceautomation and robotics. While behaviour-

based robots have been suggested for

planetary surface exploration (Brooks andFlynn, 1989), the behavioural paradigm hasnot previously been used to design orbitalspacecraft. Hence, the proposed methodologyis not fully articulated, but groundwork for amore complete approach is presented. Indeed,it is probably dangerous to suggest anycomprehensive standardised approach untilbehaviour based spacecraft have been welldemonstrated, and the paradigm and itsbenefits in this domain are well understood.

Th_ main technical objectives of a designmethodology include the achievement of fullbidirectional traceability between userrequirements and system solutions selected,the breakdown of complex problems intosuccessively simpler ones, unity of system

architecture, rigorous interfaces between

critical control loops and fault recognition, subsystems, imp.roved communications withDevelopment projects are currently underway end users, precise communication within ain a number of SSF expert system

applications.

A Methodology for Autonomous SpacecraftDevelopment

In the ongoing development of autonomousspacecraft, devolution of decision-making tothe lowest possible level, and themodularisation and distribution of

functionality, are prominent trends. Thesetrends, driven by the particular requirementsof real time autonomous agency, have been

most fully developed in the context of mobilerobotics research, and are captured moststrongly by behavioural approaches toautonomous systems design. Behaviouralapproaches to mobile robotics have also

development team, efficient paralleldevelopment of subsystems, reduced controlcomplexity, greater simplicity of design, andsound data analysis and administration(Elfving and Kirchoff, 1991). The resultingbenefits include a high-quality product which

is easy to maintain and upgrade, better projectcontrol during the design process, reducedtime to completion, and lower cost for system

development.

Elfving and Kirchoff (1991) present logicalreference models which postulate the essentialfunctions of an automation and robotics

system, to be used for structuring userrequirements and transforming them intodistinct design solutions. This ESTECmethodology is derived from structuring

94

principles of hierarchical decomposition andhierarchical structuring, fundamental 5.properties from disciplines such as controltheory and mechanical engineering, andgeneric principles of structured analysis and 6.structured design. It is characterised by a clearseparation between operational analysis andsystem synthesis, and the application of theprinciple of abstraction in the form of 7.reference model techniques. While the ESTEC

methodology has many desirable features, itdoes not immediately lend itself to the

synthesis of behaviour-based autonomous 8.systems. Behavioural approaches are,however, in need of methodological guidelines

to support a systematic composition ofbehaviours into competencies, to assist in

ensuring that the resulting system design willmeet its overall requirements for a given 9.

application (Brooks, 1990, 1991). It istherefore valuable to draw from both the

behavioural approach and the ESTECmethodology, in order to achieve amethodology combining the benefits providedby both.

Essential Characteristics

Subsumption Methodology

of the

The subsumption architecture was specificallydeveloped to address requirements forautonomous mobile robots (see Brooks, 1986,

1990, and 1991) including multiple, oftenconflicting, goals, multiple sensors,robustness, and extensibility. The approach isbased upon a number of principles which canbe drawn upon and adapted here as principleswhich bear upon the spacecraft autonomy

problem:

1. complex or "intelligent" behaviour can bean emergent phenomenon, arising from theinteractions of a spacecraft with itsenvironment and users.

2. component interfaces should be simplerthan the components that they interconnect.

3. if a module solves an unstable or ill-

conditioned problem, then it is probably nota robust solution.

4. autonomous model-making is important,since idealised models may be inaccurate.

the spacecraft must operate in a three-dimensional world.

relational models can avoid the cumulativeerrors that characterise absolute coordinate

systems.

there is no global internal model, or globalplanning activity with a hierarchical taskstructure.

for robustness, the spacecraft must be able

to perform when one or more of its sensorsfails or malfunctions. Recovery should be

rapid, so built-in self-calibration is requiredat all times.

the spacecraft control problem is

decomposed in terms of layers or levels ofcompetency. Those levels are task-achieving behaviours, and as such areexternal manifestations of the control

system. "Higher" levels correspond withmore specific classes of behaviour.

10. each successive level subsumes as a subset

each earlier level, and provides additionalconstraints on the class of valid behaviours

defined by the earlier levels.

11. successive levels extend competency, butdo not alter the structure of the

implementation of lower levels. A level canreceive data from lower levels in order to

monitor their behaviour, and can outputdata to lower levels in order to modify

behaviours by inhibiting or exciting them.

12. competencies are parallel processes.Hence, lower level competencies can

ensure that spacecraft behaviour is sensible,while higher level competencies take timeto produce more optimal control solutions.

13. there is no central locus of control, either

for the system as a whole, or within anyparticular competency layer. This isessential for robustness.

14. each layer can run on its own processor,and individual layers can also be run overmany loosely coupled processors.

95

15.within eachparticular layer, a traditionaldecompositionof functionalitymaybeused"to someextent".

Theseprinciples conform with the paradigmfor autonomoussystemsdesignwhich Maes(1990a)refersto asthebehavioural approach.Inspired by biological models of autonomy,the behavioural approach has abandoned theolder sense-model-plan-act (SMPA) paradigm

(Brooks, 1990), and in so doing has achievedmany successful demonstrations of greatlyimproved robot performance and robustness.Those demonstrations represent explorationswithin the new paradigm, but can by no meansbe regarded as definitive or fully matured.

It is important to distinguish the logical designof an autonomous control system from itsimplementation. Some explorations of thebehavioural paradigm have concentrated uponsoftware design, with the software being

compiled to run on a single embeddedprocessor (eg. the MIT Squirt robot, Brooks1990). However, while such systems candemonstrate the effectiveness of a control

architecture based upon situatedness andbehavioural interaction, rather than model-based deliberative reasoning, the use of a

single physical processing element creates asingle physical locus of control, and thereforea single point failure mode in the resultingrobot. The behavioural paradigm is a systems

paradigm. As such it should lead to adistributed hardware functionality of the kind

that typifies the more sophisticatedbehavioural robots. Research within the new

paradigm continues at a vigorous pace, and thequestion of how competencies at a conceptuallevel can be achieved as emergent properties

of increasingly parallel and distributedcomputational processes at the implementationlevel has only just begun to be investigated.Continuing advances in parallel anddistributed hardware architectures reinforce

the viability of the behavioural approach, anddemand a radical rethinking of howautonomous systems can be structured.

A number of researchers have adopted thebehavioural approach as a method fordesigning the lower level control functionswithin an autonomous system, and have then

provided one or more "layers" above the

behavioural layers which perform higher leveldeliberative and symbolic computations. Forexample, Steels (1991) proposes a framebased system for "high level" cognitive tasks(such as language use), in which frames aregrounded by sensory inputs and influence thebehaviour of the system in proportion to their"fit" to the current situation. Arkin (1990)describes the Autonomous Robot Architecture

(AURA), a framework for experimenting withthe integration of behavioural approaches withmodel-based reasoning. AuRA allows theadvantages of modularity, incremental design,adaptability, and robustness of the behaviouralapproach to be supplemented by the use ofmodel-based knowledge to configurebehavioural strategies in an efficient form.The AuRA architecture comprises five basic

subsystems: Perception, Cartographic,Planning (both a hierarchical planner and adistributed reactive plan executionsubsystem), Motor (the actuator set interface),and Homeostatic control (monitors internalconditions of the robot for both the higherlevel planning mechanisms and the motorschemas). Flexibility is incorporated into theAuRA system by drawing modularisedbehavioural patterns and sensory strategies

from a library and configuring them to meetthe needs of a particular mission and anyknown environmental constraints. World

models play an important role in thisconfiguration process. Bonasso (1991) alsodescribes an architecture in which a

declarative model is used for reasonin.g aboutplans and controlling the activation ofbehaviours implemented within an underlyingsubsumption layer. Malcolm and Smithers(1990) describe SOMASS, a robot assemblysystem that combines a PROLOG assemblyplanning subsystem with a plan executionsubsystem that handles uncertainty by meansof behavioural modules which accomplish

useful motions of assembly parts. Malcolmand Smithers note that, although the

"cognitive" planning function wasimplemented m a high-level symboliclanguage, and the "subcognitive" executionagent in a low level language, these decisionswere motivated by convenience. Also, thecognitive/subeognitive distinction was itselffound to be a useful construct for the

SOMASS system, but is not a necessary orconvenient construct for artificial mentality in

96

general.The interface betweenthe cognitiveand subcognitivesystemspresentsa virtualmachinemodel to thecognitivesystem,whichheavily determinesandpermeatesthedesignof thecognitivesystem.Theneedto presentavirtual machine interface to the cognitivesystemresultsin amodularityof behavioursinthe subcognitivelayer,which is not neededinapproaches(eg. subsumption)which do notinclude a cognitive component. Gat (1991)describes ATLANTIS, a system whichintegrates behaviours and deliberativeprocessesin a threelayeredstructure.The firstlayer comprises a behaviour-based system forrobust motion control and low-level

competencies. A deliberative layer provideshigh level reasoning, such as plan generation.These layers are joined together by a

sequencing layer based upon Reactive ActionPackages.

These different explorations of therelationship between deliberative reasoningand behavioural autonomy are not driven by

any well-proven limitations of the behaviouralapproach. As Brooks (1990, 1991) notes, thequestion of how far a behavioural approachwithout the use of deliberative computationscan go in achieving higher levels ofcompetence in autonomous agents is onewhich must be addressed by ongoingempirical investigations. However, an issuearises within the behavioural paradigmregarding the range of computationalmetaphors that can usefully be adopted fordesigning behavioural units. In thesubsumption architecture, Brooks usesaugmented finite state machine (AFSM)models to define primitive behaviouralelements, and AFSMs are further grouped intomore complete behaviours. However, anynumber of computational metaphors canpotentially be used to describe a system with agiven transfer function. The effectiveness ofAFSMs has been demonstrated for some

behaviours, but metaphors of multipleinteracting agents (see Grant, 1992), objects,processes, production systems, etc. mayequally provide convenient metaphors.Similarly, knowledge-base systems, rule bases(AFSMs are defined by rule sets in Brooks'behaviour language), and expert systemsmetaphors may be convenient. The metaphoradopted should be that which most "naturally"

describes the behaviour of a module from the

perspective of the system designer. Adoptingthe behavioural paradigm for the overallcontrol system architecture does not rule outthe adoption of other metaphors for structuringand designing the computational processeswithin a given behaviour or module in order toachieve a desired set of input/outputmappings. The danger in adoptingheterogeneous metaphors within a behaviouralcontrol system is if any metaphor distorts thebehavioural framework by encouraging thecentralisation of behavioural coordination and

control, or the centralisation of data flow. Itcan be argued that this should not occur in thecase of spacecraft control if the control systemis modelled upon manual spacecraft operation,since ground based, manual spacecraft controlinvolves highly distributed, cooperativedecision-making by numerous "agents" ofmixed expertise, generally in a way that ishighly redundant, robust, and adaptive. Thebehavioural paradigm is not violated by

experimentation with alternate metaphors fordeveloping the internal structure ofbehavioural modules within a behaviouralcontrol framework.

Towards Reference Models for the

Behavioural Methodology

The reference models described by Elfving

and Kirchoff are intended to provide cleartraceability from user requirements to designsolutions, which justifies technical decisionsmade and avoids excessively flexible,

expensive, and complex systems with hightechnical risk. The reference models areintended to cover robots, mobile vehicles, and

process control, and reflect the operational useof the system. Elfving and Kirchoff divideA&R capabilities in space into three majorfields:

- external servicing of payloads- servicing of scientific experiments within

pressurised orbiting laboratories- surface mobility and sample acquisition for

planetary exploration

To this can be added the field of current

concern:

- autonomous orbital spacecraft control

97

The startingpoint for thisESTECautomationand robotics (A&R) control designmethodology is an established conceptuallayoutfor theA&R system,typically availableas a result of mission and systemdefinitionstudies.The logical referencemodelscapturetheessentialprinciplesof thismethodology.

The objective of Operational Analysis (orTask Analysis) is to derive the essentialabilities of an A&R system such that missionobjectives are sure to be fulfilled. This is doneby a step-by-step decomposition into levels ofequal importance, from mission objectivesdown to a level of elementary actions. Thisdefines "what has to be done", ideally without

anticipating any solution, but by using initialknowledge about system functional layoutbased upon initial mission and systemdefinition studies. This analysis phase requires

system operational expertise. System Synthesisinvolves the definition of solutions that satisfy

the different operational features required atthe various levels of decomposition, therebyaddressing "how is it to be done?". System

Synthesis involves an aggregation process,from elementary abilities to system

capabilities. Synthesis requires A&Rtechnology expertise.

Given these processes of analysis and

synthesis, the key methodological question isthat of how to achieve traceability between thetwo areas, assuming that the decompositionlogic of analysis must be in agreement withthe synthesis logic, so the inherent structure ofthe analysis process is a virtual systemsolution. This nexus is achieved by the use of

the logical reference models. Three logicalreference models are distinguished:

. Functional Reference Model (FRM):

represents a decomposition of all functionsand information flow and structures, and is

valid for all A&R applications.

, Application Reference Model (ARM):

represents an FRM derivative whichfocuses on individual classes of A&R

applications.

3. Operations Reference Model (ORM)"represents an FRM derivative which

focuses on models of operation andsystematics for man/machine andpreparati0n/uti!isafion allocation.

The objectives of applying reference modeltechniques are:

- guarantee a common 'thinking model' whichis valid for all A&R applications, to enableand ease communication within and between

development teams-provides a generic system information

structure and identifies essential functions

for which application-specific designsolutions should be found

- establishes "rules-of-thumb" and heuristic

design strategies that allow the designer tosystematically derive good and relevantsolutions to common types of problems

- uses formal documentation techniques and

graphic support tools that emphasise thehierarchical functions and information flow

and structures of complex systems

A logical model represents the essentials of asystem, assuming ideal internal technology ofthe system and excluding application-specifictopics. A physical model represents theimplementation of a specific application, andtherefore needs to represent the actualconstraints imposed by the chosen internaltechnology. Hence, only logical models can bereference models that meet the requirement ofbeing valid for a multitude of applicationsand/or implementation technologies.However, a major goal of a designmethodology is to maintain a strong andtraceable link between the logical and physicalmodels of a system in order to achieve designcommonalities and open implementationarchitectures.

A Behaviour-Based Functional ReferenceModel

The FRM proposed by Elfving and Kirchoffinvolves a hierarchical ordering of functionsand information with increasing precision anddecreasing planning horizon from top tobottom, where the hierarchical informationinterface layers are:

1. A&R mission: the objectives of end users

98

2. task: clusteredactivities performedon asingle subsystemdefined as the elementsubmittedto motion or actuationby theA&R system

3. action:elementaryactivity for a functionalsubsystem

This decomposition leads to hierarchicalfunctional and information layers, with thecontrolled devices and processes at thebottom,and successivelayersfor action,task,andmissionexecution,planningandcontrol.Stateinformationflows from eachlayerto thenexthigh layer,while commandsandactivityattributesflow downthroughthestructure.

This hierarchical decomposition, with"vertical" divisions between distinctobjectives,tasks,andactionsat eachlevel, isnot compatible with a behaviouralmethodology. To be compatible with abehavioural approach, the followingmodificationsarerequired:

-instead of a decomposition of systemfunctionsin termsof a tripartite structureofmission,tasks,and actions,the systemcanbe decomposed in terms of layeredcompetencies. This overcomes thesomewhatarbitrary parsingof activity intothree levels, providing a more flexiblelayering and abstractionmechanismwhichsubsumes and extends the tripartitestructure. It also has the advantage ofretainingthevisibility of what thesystemisrequired to do, rather than decomposingsystem functions according to how thosefunctions are implemented. Systemrequirementscan be mappeddirectly ontocompetencies,and competenciesthenontotheirimplementation.

- insteadof placing all sensorand actuatorinterfaces at the lowest level of ahierarchical structure, each competencylevel is associateddirectly with a subsetofthe total set of sensors and actuatorsrequired,in addition to having interfacestothe competency levels above and belowitself.

A&R Mission planning and control can beretainedasthe highestgenerallevelof systemcompetency.The lowestlevel is alsogeneral,

and comprisesbasic survival competencies.Intermediatelevelsarevariablein numberandformacrossdifferentA&R applicationclasses.TheresultingverticalFRM structureis thenasshownin figure 1.

Elfving andKirchoff describeforwardcontrol,nominalfeedback,andnon-nominalfeedbackfunctions across all hierarchical levels.Forward control involves activitydecompositionandexecutionplanning,basedon a priori knowledge,and plan execution.Nominal feedback deals with forseenrefinementsand updatesto ensure that thesystemachievestheforwardcontrolexecutiongoals.Non-nominalfeedbackcovers thecaseof systemperformancediverging from theallowable region around the nominalexecutionplan and is realised by meansofmonitoringdiscrepanciesbetweenactual andallowable states,diagnosis of reasons forpossible discrepancies, and generation ofrecoverystrategiesandconstraints.

STATE I _ CO_CYINFO. LEVEL nO B/ECI'IVES

MISSION |

PLcANN1NGoNROL& I LEVEL

" _ DATA ____)

A FLE L "

s_Ic I_SURVIVAL I LEVEL

COMPLiaNCY' J 0

Figure 1. Behavioural FRM.

This model may be used to address individualcompetencies in the vertical decomposition ofsystem functionality. However, nominalfeedback, non-nominal feedback, andfeedforward control functions should not be

regarded with restricted computationalparadigms in mind. It must be regarded as anopen question how the respective functionscan best be modelled for any particular

application or level of problem abstraction. Itis yet another question how the modelledfunctionality should then be implemented. For

example, a logical model of planning may ormay not be the best way of specifying a typeof planning activity that may be required.

99

Moreover,if the planningactivity is definedlogically, the implementationof the plannermaytakequiteadifferent form,suchasthatofa highly distributed and emergent function(eg.gradientfield approaches,asdescribedbyPayton 1990, analogical planning, asdescribedby Steels1990,or emergentgoalarbitration, asdescribedby Maes1990b).In abehavioural system, inputs and outputs forthesefunctionsinclude behaviourmonitoringand control signals, in addition to sensorinputs and actuatoroutputs,dep.endingupontheirrole within thecompetencym whichtheyoccur. In other words, nominal feedback,forward planning,andnon-nominalfeedbackbecomebehaviours within a competency

layer.

These redefinitions result in a very differentFRM, less detailed than that presented by

Elfving and Kirchoff, and as such placingmore importance upon Application ReferenceModels for the different classes of A&R

applications.

Behavioural Application Reference Model

The Application Reference Model (ARM)tailors the functional blocks and theinformation structure of the FRM to thecharacteristics of each class of A&R

application, in order to ease acceptance andunderstanding. Elfving and Kirchoff identifythese classes as:

- motion systems with fixed linkages to theenvironment (eg. robot manipulators)

- mobile systems (eg. moving vehicles)- continuous processes (eg. climate control)- event/sequence control (eg. PLC equipment)

To this can be added the class of current

concem (a subtype of mobile systems):

- autonomous orbital spacecraft

In physical realisations, a combination ofthese classes may be integrated to form theoverall A&R control function.

The ARM leads to a more detailed

decomposition than the FRM. From thebehavioural viewpoint this amounts to

identifying for each class of application:

- the competency layers required- generic requirements for nominal feedback,

non-nominal feedback, and forward control

within each layer- individual behaviours within each

competency

Ongoing research is needed to identifyalternate implementation strategies for variouscompetencies, and to systematically analysethe tradeoffs between different strategies as abasis for rational design decisions.

Proposed ARM for Autonomous OrbitalSpacecraft

Due to very significant domain differences,the competency layers proposed here forautonomous orbital spacecraft bear littleresemblence (other than at the highest level) tothose proposed by Brooks (1986) forbiologically-inspired mobile robots operatingin earth-gravity environments. Some of thesedifferences arise due to the different

operational environment and sensor andactuator sets of orbital spacecraft in

comparison with mobile surface robots. Otherdifferences arise due to a major aspect of

spacecraft functionality which concerns thecollection, distribution, and reception of data.

The definition of layered competencies ismade in terms of decreasing criticality andincreasing autonomy. In this sense, theapproach represents an extension of automatedsafe mode transitions into a more complex setof behaviours, and a set of autonomous

operations which function during normalspacecraft operation in addition toemergencies. Consideration of the layeredstructure shows the non-hierarchical nature of

the control system. Mission critical survivaldecisions, often made at the lowest levels,

must have priority over higher-level decisionmaking during emergency situations.However, the higher levels must be able tomodify resource allocations, timingrelationships, and data flow within the lowerlevels in order to establish priorities betweenactivities within a wide range of variation of

nominal operating modes in order to achievehigher levels of competence in autonomouslymeeting and optimising mission objectives.

100

For example, critical docking operationscannot havetheir power supply interrupted.Ensuringthat this doesnot occurcaninvolvepowerconditioning schedulesthat comeintooperation well in advanceof the dockingactivity, and therefore must be plannedbyhigher competency levels, and thoseplansthenconformedto or usedby the behavioursof the lower levels.

A proposedcompetencymodel is showninfigure 2. The Maintain Thermal Balancecompetencyis placedat level 0 as the mostfundamental precondition of spacecraftsurvival. That is, any significant over- orunder-heating can permanently damageordestroythe spacecraft.However, temperaturevariations tend to be gradual, and goodthermal design can ensurethat the normalrange of temperature variations for thespacecraft is limited to a 20° range. Thegreatest danger is from more localisedtemperature fluctuations (especiallyaccumulationof heat),possiblydue to faults,which can be dealt with by temperaturedependent inhibition or excitation of heatgeneratingprocesses.

9 Plan and Execute Mission Based Upon Objectives

8 Rendezvous and Dock ( I Land I Take Off)

7 Plan and Schedule Data Acquisition andTransmission

6 Plan and Execute AOCS Maneuvers

5 Acquire, Condition, Downlink Instrument Data

4 Maintain Auitude Stability and Orbit

3 Acquire, Condition, and Downlink EngineeringData

2 Maintain Telecommand Override

1 Maintain Battery Condition

0 Maintain Thermal Balance

Figure 2. Competency Layers for OrbitalSpacecraft.

The Maintain Battery Condition competencymaintains a battery charge/discharge cycle thatwill ensure long battery lifetime. This must be

adaptive in the face of variations in batteryand solar cell characteristics over the course of

their lifetime, and must accommodateindividual cell failures and variable solar

irradiation cycles.

Levels 0 and 1 ensure spacecraft survivability,while successively higher layers are concernedwith spacecraft accessibility, performance, andincreasing independence from the ground. TheMaintain Telecommand Override (level 2)

competency is provided to ensure that, so longas the survival of the spacecraft is notthreatened, the behavioural control system canbe overidden by direct ground control. It isnot, however, envisaged that this should severthe layered structure of the behaviouralsystem. Rather, it provides an orthogonalaccess mechanism by which it is possible to

inject data into the behavioural control system,and should have the capacity to overrideinternal interconnections between spacecraftbehaviours. Hence this level ensures command

access to the spacecraft.

Acquistion, Conditioning and Downlink ofEngineering Telemetry data (level 3) ensuresaccess to data which is critical for diagnosing

and understanding the status of the spacecraftfrom the ground. Maintainance of AttitudeStability and Orbit (level 4) are secondarypriorities after maintaining the engineeringtelecommunications link, since the link is vital

for understanding the state of the spacecraft,

and initiating telecommand override if

necessary.

The Acquisition, Conditioning, andTransmission of Instrument Data (level 5) canbe ensured as a level above maintaining basic

survival, two way telecommunications, andstability. Stability control (level 4) is anoperational precedent for Planning andExecution of attitude and orbital control

(AOCS) Maneuvers (level 6), and theacquisition and transmission of instrumentdata must also be accounted for in the AOCS

maneuver planning process. AOCSManeuvering competency is a precondition forautonomous Planning and Scheduling of DataAcquisition and Transmission (level 7).

Rendezvous and Dock (level 8) is a high level

competency. Although this is already

101

automated at a low level in systems fordocking Salyut and Progress vehicles with the

MIR space station, in an integrated,behavioural control system it occupies a highlevel in order to automate all of the planningand resource allocation activities associated

with the basic rendezvous and dockingprocedure. Landing and Take Offcompetencies are mentioned at this level ascompetencies of orbital spacecraft which alsoland on and take off from planetary bodies.

Plan and Execute Mission Based UponObjectives (level 9) allows the highest level ofcommand abstraction in specifying behaviourdesired of the spacecraft.

Qperations Reference Model

The FRM and ARM represent generalfunctions necessary to decompose a globaltask into process variables and the relatedinformation architecture. The OperationsReference Model (ORM) addresses a different

question, that of how the overall A&R systemis operated. The aim is to help to structure thedefinition of operating modes and achieve aclear and common understanding of what ismeant. The ORM can also be used as a

support tool for a systematic and consistentdraft specification of man/machine taskallocations, and allocations of which tasks are

to be performed on the ground and which onthe spacecraft.

Virtual machine models are useful as a basis

for increasing command abstraction. Hence,on the basis of the ORM, separate virtualmachine models can be used for telecommand

generation and subsequent injection of datainto each competency level. This provides abasis for telecommand language specificationat several abstraction levels.

From Analysis to Design Synthesi_

The primary sources for the operationalanalysis are the global spacecraft missionobjectives, the characteristics of the process tobe handled, and the layout of the spacecraftdevices to handle this process. The applicationof the reference models means that the

decomposed activity hierarchy of theoperational analysis can be used to definefunctional and performance specifications for

system competencies. Hence the hierarchical

structures defined during operational analysisare used to develop specifications for the non-hierarchical layered competency model, andthis competency model provides the basis forsystem design synthesis, as shown on Figure3.

OPERATIONAL ANALYSIS

A&R Mission Objectives

requirements Task

& attnl_tes HP Action_ s_k_.,_

' COMPETENCY ' SYSTEMt i, ANALYSIS , SYNTHESIS• i!

, A_RMISSIONIi IA_RM_SS_NIPL_d_NING _ PLANNING I&CONTROLI : I _ CONTROL

state • control

B_,c I:1 8_c I, SURVWALI_ SURVWALI: :OMPErENCYp. ICO_ETENCYI' LEVEL 0 • LEVEL 0

Figure 3. Sequence from analysis to synthesis.

System synthesis involves firstly finding theintermediate detailed specification of systemcompetencies based upon a generalApplication Reference Model and theoperational analysis, and then finding theappropriate design solution for thecompetency layers of the controller, as shownin the right hand side of Figure 3. This resultsin system solutions that are valid for classes ofmission scenarios, in that activities to be

performed become more generic towards thelower decomposition levels, supporting a'standard set' of competencies which canaccommodate evolving or changing missiontask requirements without a critical impactupon the design.

Control System Synthesis

Elfving and Kirchoff (ibid) suggest that cleartraceability of solution decisions made duringdesign synthesis can be achieved by an inverseof the design process: use is made of a set of

possible technical solutions which originatesfrom existing fundamental engineeringtheories and which can be systematicallyordered into solution trees by applyinghierarchical structuring principles. While thisis a desirable goal, it depends upon a wellunderstood synthesis process. For abehavioural approach to orbital spacecraftdesign, this is premature. Since the spacecraft

102

=

i

Ii

i_i

r.

control architecture presented here is novel, it

represents at least one traversal of a morecomprehensive solution tree, yet to be fullyarticulated. There are also major ongoing

research questions regarding the relationshipbetween the logical definition of systemcompetencies and the implementation of thelogical design (as discussed above).

Control System Example

The version of the subsumption architecture

adopted for current purposes allows a moduleinput tO be suppressed by an output fromanother module, and the input signal is

replaced by the suppressing signal (followingBrooks, 1986). Limited aspects of control wiltbe considered here to illustrate the principlesinvolved. Behavioural interactions to achieve

level 0 and level 2 competencies within a very

simple model will be considered. Sensorinputs for this example comprise a set oftemperature sensors. Actuators will comprise atelemetry transmitter and a receiver, each withcontrollable power levels (telemetry dataflows will not be considered).

The level 0 competency maintains the thermalbalance of the spacecraft. Temperature sensorsdistributed throughout the spacecraft andassociated with each controllable component

and subsystem are the primary inputs to thiscompetency. The hypothetical spacecraft is avery simple design, and the only activetemperature control available is by increasingor decreasing the power consumption, andhence heat dissipation, of controllable units.To achieve adaptivity and maximumrobustness, the sensor inputs are mapped intoan analogical representation of the

temperature of the spacecraft (Steels, 1990,describes analogical maps). The analogicalrepresentation stores a model of each heatgenerating component of the system,conductive paths, uncontrollable heat sourcesand sinks (eg., the sun and "dark" space,respectively), and the spatial interrelationshipsbetween these elements. The resultinganalogical representation resembles a low-resolution finite-element model. For each

controllable temperature source in the model,there is a separate behaviour which computesa power level signal as a function of thecurrent and previous temperatures of the

particular element, its immediate spatialneighbours, the heat transfer characteristics oftheir interconnections, and an approximate

measure of the specific heat of the component.This method deals very robustly with thefailure of any given component to respond todirect temperature control, by calculating

power (and therefore temperature) levels as afunction of the average temperature of the

component and its neighbours, therebyallowing indirect control of components whichare not directly accessible by the use oftemperature flow relationships. Each controlbehaviour could be implemented by a separateprocessor, and the method of obtainingtemperature control for inaccessible elementsalso deals with the control of elements whose

associated control processors have

temperaturep_._ I CENERATE Isensors k ¢ " [ LOCAL MAPS[

map _lr _P' map

ICESERATERXI IOENERATErXIIPOWER LEVEL I IPOWER LEVEL I

vel _Level

Figure 4. Example level 0 competency.

malfunctioned or failed. The resultingbehavioural model is shown m Figure 4.The analogical representation is an internalstate model used by the Generate Local Mapsbehaviour. This behaviour may actuallycomprise a separate local map generatorassociated with each controllable unit, so the

analogical model is a virtual construct. Eachlocal map is a model of the immediate (ie.directly connected) thermal environment ofthe associated component, which is used bythe power level generation behaviour of thatcomponent to calculate a power level which isthen sent to the component.

The next level of competency considered isthe level 2 Maintain Telecommand Override

competency. For the current highly simplified

103

example, this is easy to create using a singlebehavioural module. The MaintainTelecommand Override module monitors the

receiver power level, and suppresses the level0 control signal if it falls below the minimumlevel required to operate the receiver fortelecommand override. The minimum powerlevel is then injected into the receiver input inthe form of the suppressing signal from theMaintain Telecommand Override module. The

structure giving this competency is illustratedin Figure 5.

[ temperature[_.l_[ GE/,mRATE I

sensors 1¢" " [ LOCALMAPS I

RX Local[ [TX Loea/map _ _ map

[GENERATE RX[ [GENERATE TX[

10OWERLEVELI IPOWER LEVELI

P°wRXer __vweTXcl rI OVER IDEI

Figure 5. Example level 2 competency.

The level 0 competency can still deal withovertemperatures in the region of the receiver,not by reducing its power level below theminimum required for telecommand override,but if necessary by powering downneighbouring units. The structure andfunctionality of the level 0 competencyensures that this will happen automatically,without explicit control by the level 2competency or by ground operators

In a more complete example, this minimumpower level required to MaintainTelecommand Override is a function of

spacecraft orientation, orbital height, positionin relation to ground control stations, the statesof a number of alternate transmitter systemmodules, etc., so the Maintain Telecommand

Override competency emerges from a muchmore complex interaction between a numberof modules.

Conclusion: The AUSTRALIS-I SpacecraftProject and Ongoing Research

AUSTRALIS-1 is a microsatetlite projectcurrently under development by a number ofAustralian universities. Ongoing research withthe spacecraft control architecture describedhere will be conducted in relation to the

AUSTRALIS-1 project. A comprehensivecontrol simulator for the satellite based uponthe described architecture is currently underdevelopment. The outcome of simulationstudies will be a detailed control architecture

design, with well understood tradeoff studiesof different control system configurations.Another major ongoing area of research willconsider the relationship between the logicalcontrol structure and the physicalcomputational architecture upon which it isimplemented. It is particularly desirable todefine a hardware architecture compatiblewith many possible mappings from the logicalcontrol model to its implementationalstructure. To this end, the hardware designapproach will follow the following sequence:

l, a basic set of modules will be defined

representing subsystems and/orcomponents required to achieve desiredspacecraft competencies in relation to thespecific requirements of the AUSTRALIS-1 mission. Module definitions will include

the definition of outputs from the modules(which are equivalent to sensor values orstatus signals), inputs to the modules(equivalent to actuator commands), thedifferent operational modes of eachmodule, the association of different sensor

outputs and actuator inputs with differentoperational modes, and the definition ofthe conditions under which transitions

between the different operational modeswill occur (both in response to commandinputs, to sensor values, and to internalstates). This set of modules will bereferred to as substrate modules, and

constitute the lowest level description ofthe machine to be controlled.

. a generic module interface unit will bedefined at the physical and protocol levels.The connection between any two modulesshould be asynchronous, bidirectional, and

104

configurablebetween1- and n-bit binarynumbers.

. each substrate module defined in step 1will have a standard microprocessor-basedinterface unit associated with it.

. in addition to the substrate modules, based

upon an estimate of overall bandwidthrequirements to achieve all levels ofcompetency (or a specified subset thereof),a number of additional processing nodeswill be specified, each with a standardinterface, and a generic processor definedin order to implement each additionalnode. (The number and capacity ofadditional nodes required will be explored

by simulation of the architecture.)

5. all behaviours will be integrated into a fullyconnected control signal network via theirstandard interfaces.

. each standard processor will in additionhave a dual-redundant connection to a

behavioural override processor (BOP).The BOP will be able to download transfer

functions to each processor in the controlnetwork, to allow downloading of software

implementing different logical controlarchitectures. The BOP will also be able to

drive each behaviour directly, or bypassthe behavioural control network

altogether.

A significant difference between spacecraftand the mobile robots typically used toexplore the behavioural paradigm is theexplicit and substantial role of spacecraft asdata acquisition, conditioning, andtransmission systems. Options foraccommodating this function includedistributed routing of data via the behaviouralcontrol network, or use of a more conventionaldata communications topology (eg. a bus,

ring, or star network). The approach adoptedwill be to provide for any of thesemechanisms:

7. a high speed asynchronouscommunications link will be part of thestandard processor interface, with

multiplexed connections both to other

processors in the network and to the BOP

processor.

This control architecture will provide ahardware platform that conforms with thebehavioural paradigm, but allows forconsiderable experimentation with differentmethods of mapping the logical controlstructure onto the hardware structure. In

particular, the ability to download moduletransfer functions from the BOP will support

experimentation with:

- dynamic control system reconfiguration- adaptivity and learning of module transfer

functions, network topology, andbehaviours; ie. all of the control structureabove the substrate level

-the distribution and form of multiple

computational paradigms within thebehavioural framework

In-orbit experiments will comprise a series ofempirical evaluations aimed at refining andevaluating different detailed logical controlmodels, the development of a softwarearchitecture and mappings onto the hardwarearchitecture for each logical model, andmethods of achieving adaptivity androbustness within these models, at increasinglevels of behavioural competence. The resultsof the project will be valuable in defining andunderstanding techniques for increasing theautonomy of space systems in many diverse

applications.

References

Arkin R. C., 1990, "Integrating Behavioural,Perceptual, and World Knowledge in ReactiveNavigation", in Maes, P. DesigningAutonomous Agents: Theory and Practicefrom Biology to Engineering and Back,MIT/Elsevier, 105-122.

Bonasso R. P., 1991, "Integrating ReactionPlans and Layered Competencies ThroughSynchronous Control", 12th InternationalJoint Conference on Artificial Intelligence,Morgan Kaufman, 1225 - 1231.

Brooks R. A., March 1986, "A Robust

Layered Control System For A Mobile

105

Robot", IEEE Journal of Robotics andAutomation, 2(1), 14-23.

Brooks R. A., 1990, "Elephants Don't PlayChess", in Maes, P. (ed) Designing

Autonomous Agents: Theory and Practicefrom Biology to Engineering and Back,MIT/Elsevier, 3-15.

Brooks R. A., 1991, "Intelligence WithoutReason", 12th International Joint Conference

on Artificial Intelligence, Morgan Kaufman,569 - 595.

Brooks R. A. and Flynn A. M., Oct i989,"Fast, Cheap, and Out of Control: A RobotInvasion of the Solar System", JBIS, 42(10),478-485.

Ciarlo A., Donzelli P., Katzenbelsser R. andMoiler B. A., 1987, "Satellite On-Board

Applications of Expert Systems", ESAJournal, 11, 31-44.

Ciarlo A., and Schilling K., 1988,

"Application of Expert System Techniques tothe Cassini Titan Probe", ESA Journal, 12,

337-351.

Elfving A. and Kirchoff U., 1991, "DesignMethodology for Space Automation andRobotics Systems", ESA Journal, 15, 149-164.

Erickson J. D., 1987, "Intelligent Systems andRobotics for an Evolutionary Space Station",JBIS, 40, 471-482.

Gat E., 1991, Reliable Goal-Directed ReactiveControl of Autonomous Mobile Robots, PhDThesis, Virginia Polytechnic Institute andState University, revised version.

Grant T. J., 1991/1992, "A review of Multi-

Agent Systems techniques, with application toColumbus User Support Organisation",FGCS7, 413-437.

Leinweber D., Spring 1987, "Expert Systems

in Space", IEEE Expert, 26-36.

Maes P., 1990a, "Guest Editorial", in Maes, P.Designing Autonomous Agents': Theory andPractice from Biology to Engineering andBack, MIT/Elsevier, 1-2.

Maes P., 1990b, "Situated Agents Can HaveGoals", in Maes, P. Designing AutonomousAgents: Theory and Practice from Biology toEngineering and Back, MIT/Elsevier, 49-70.

Malcolm C. and Smithers T., 1990, "Symbol

Grounding via a Hybrid Architecture in anAutonomous Assembly System", in Maes, P.Designing Autonomous Agents: Theory andPractice from Biology to Engineering andBack, MIT/Elsevier, 123-144.

Payton D. W., 1990, "Internalised Plans: ARepresentation for Action Resources", inMaes, P. Designing Autonomous Agents:Theory and Practice from Biology toEngineering and Back, MIT/Elsevier, 89-104.

Pidgeon A., Howard G. and Seaton B., 1992,"Operational Aspects of SpacecraftAutonomy", JBIS, 45, 87-92.

Raslavicius L., Gathmann T. P., and Barry J.M., 1989, "An Artificial IntelligenceFramework for Satellite Autonomy", SecondInternational Conference on Industrial and

Engineering Applications of ArtificialIntelligence and Expert Systems, 2, 536-543.

Spier R. J. and Liffring M. E., Nov, 1989,"Real-Time Expert Systems for AdvancedPower Control", IEEE AES Magazine, 33-38.

Steels L., "Exploiting AnalogicalRepresentations", in Maes, P. DesigningAutonomous Agents: Theory and Practicefrom Biology to Engineering and Back,MIT/Elsevier, 71-88, 1990.

Steels L., 1991, "Emergent Frame Recognitionand its Use in Artificial Creatures", 12th

International Joint Conference on ArtificialIntelligence, Morgan Kaufman, 1219 - 1224.

Tello E. R., 1986, "Dipole: An Integrated AIArchitecture Suitable for Space PRogramApplications", Expert Systems in GovernmentSymposium, IEEE, 168-180.

Woods D., April 1992, "Space StationFreedom: Embedding AI", AI Expert, 33-39.

106