Journal of lntelligent and Robotic Systems 1: 315-342, 1989.

315
9 1989 Kluwer Academic Publishers. Printed in the Netherlands.

Towards Intelligent Autonomous Control Systems:

Architecture and Fundamental Issues
P.J. A N T S A K L I S , K.M. P A S S I N O
Department of Electrical and Computer Engineering, University of Notre Dame, Notre Dame,
IN 46556, U.S.A.

and

S.J. W A N G
Jet Propulsion Laboratory, MS 198-326, 4800 Oak Grove Drive, Pasadena, CA 91109

(Received: 6 July 1988; revised: 12 December 1988)

Abstract. Autonomous control systems are designed to perform well under significant uncertainties in the
system and environment for extended periods of time, and they must be able to compensate for system failures
without external intervention. Intelligent autonomous control systems use techniques from the field of arti-
ficial intelligence to achieve this autonomy. Such control systems evolve from conventional control systems
by adding intelligent components, and their development requires interdisciplinary research. A hierarchical
functional intelligent autonomous control architecture is introduced here and its functions are described
in detail. The fundamental issues in autonomous control system modelling and analysis are discussed.

Key words. Automatic control theory, autonomous control, intelligent control, intelligent systems,
hierarchical systems, hybrid systems, artificial intelligence.

1. Introduction
Autonomous control systems must perform well under significant uncertainties in the
plant and the environment for extended periods of time and they must be able to
compensate for system failures without external intervention. Such a u t o n o m o u s
behavior is a very desirable characteristic of advanced systems. An autonomous con-
troller provides high-level adaptation to changes in the plant and environment. To
achieve autonomy, the methods used for control system design should utilize both (i)
algorithmic-numeric methods, based on the state-of-the-art conventional control,
identification, estimation, and communication theory, and (ii) decision-making-
symbolic methods, such as the ones developed in computer science and specifically in
the field of artificial intelligence (AI). In addition to supervising and tuning the control
algorithms, the autonomous controller must also provide a high degree of tolerance
to failures. To ensure system reliability, failures must first be detected, isolated, and
identified, and subsequently a new control law must be designed if it is deemed
necessary. The autonomous controller must be capable of planning the necessary
sequence of control actions to be taken to accomplish a complicated task. It must be
able to interface to other systems as well as with the operator, and it may need
learning capabilities to enhance its performance while in operation.
316 P.J. ANTSAKLIS ET AL.

Advanced planning, learning, and expert systems, among others, must work
together with conventional control systems in order to achieve autonomy. The need
for quantitative methods to model and analyze the dynamical behavior of such
autonomous systems presents significant challenges well beyond current capabilities.
It is clear that the development of autonomous controllers requires significant inter-
disciplinary research effort as it integrates concepts and methods from areas such as
control, identification, estimation, and communication theory, computer science,
especially artificial intelligence, and operations research.
In this paper, an autonomous controller architecture is introduced and discussed in
detail. For such controllers to become a reality, certain fundamental questions should
be studied and resolved first. These fundamental problems are identified, formulated
and discussed, and future research directions are outlined. Next, the focus of this
paper is established and a detailed description of the results is given.
Autonomous controllers can be used in a variety of systems from manufacturing
to unmanned space, atmospheric, ground, and underwater exploratory vehicles. In
this paper, we develop an autonomous controller architecture for future space
vehicles. Referring to a particular class of control problems has the advantage that the
development addresses relatively well-defined control needs rather than abstract
requirements. Furthermore, the autonomous control of space vehicles is highly
demanding; consequently the developed architecture is general enough to encompass
all related autonomy issues. Future space vehicles must be capable of autonomous
operation to accomplish their missions. Emerging aeromaneuvering vehicles such as
the Aeroassisted Orbital Transfer Vehicle and the Aerospace Plane will be required
to maneuver at high altitudes and hypersonic velocities in a flight regime characterized
by significant uncertainty in atmospheric density and aerodynamic characteristics.
Uncertainty in these parameters may cause significant deviation from the nominal
trajectory, conceivably leading to the loss of the vehicle. Significant time and com-
munication constraints during the atmospheric flight dictate that the vehicles should
perform autonomously for extended periods of time since pilot or ground support
intervention may not be possible. Future space systems, such as manned space
platforms, contain significant flexible structural components. Model uncertainties and
system parameter variations require advanced adaptive control techniques to meet
stability and performance specifications. An autonomous adaptive control system is
needed to deal with gross fundamental and environmental changes in the system. For
space systems these include hardware failures, docking disturbances, payload arti-
culation, and man-motion disturbances.
It should be stressed that all the results presented here apply to any autonomous
control system. In other classes of applications, the architecture, or parts of it, can be
used directly and the same fundamental concepts and characteristics identified here
are valid.
The architecture of autonomous controllers necessary for the operation of
advanced planetary and aeromaneuvering space vehicles is developed here. The
concepts and methods needed to successfully design such an autonomous controller
TOWARDS INTELLIGENTAUTONOMOUS CONTROL SYSTEMS 317
are introduced and discussed. A hierarchical functional autonomous controller archi-
tecture is described in detail; it is designed to ensure the autonomous operation of the
control system and it allows interaction with the pilot/ground station and the systems
on board the autonomous vehicle. Note that a shorter version of these results has
appeared in [4].
Section 2 gives a brief history of the development of control systems to motivate
the necessity for autonomous controllers. The functions, characteristics, and benefits
of autonomous control are outlined. Next, it is explained that plant complexity and
design requirements dictate how sophisticated a controller must be. From this it can
be seen that often it is appropriate to use methods from operations research or AI to
achieve autonomy. Such methods are studied in intelligent control theory. An over-
view of some relevant research literature in the field of intelligent autonomous control
is given together with references that ouline research directions. In Section 3, an
autonomous control functional architecture for future space vehicles is introduced.
The controller is hierarchical, with three levels, the execution level (lowest level), the
coordination level (middle level), and the management and organization level (highest
level). The general characteristics of the overall architecture, including those of the
three levels are explained, and an example to illustrate their functions is given. In
Section 4, fundamental issues and attributes of intelligent autonomous system archi-
tectures are described. An approach to the quantitative, systematic modelling, analy-
sis, and design of autonomous controllers is discussed. It is a 'hybrid' approach since
it is proposed to use both conventional analysis techniques based on difference and
differential equations, together with new techniques for the analysis of systems
described with a symbolic formalism such as finite automata. The more global,
macroscopic view of dynamical systems, taken in the development of autonomous
controllers, suggests the use of a model with a hybrid or nonuniform structure, which
in turn requires the use of a hybrid analysis. Finally, some concluding remarks are
given in Section 5.

2. Conventional and Intelligent Autonomous Control Systems

Autonomous means having the power for self government. Autonomous controllers
have the power and ability for self-governance in the performance of control func-
tions. They are composed of a collection of hardware and software, which can
perform the necessary control functions, without external intervention, over extended
time periods. To achieve autonomy, the controller must be able to perform a number
of functions in addition to the conventional control functions such as tracking and
regulation. These additional functions, which include the ability to tolerate failures,
are discussed later in this section.
There are several degrees of autonomy. A fully autonomous controller should
perhaps have the ability to even perform hardware repair, if one of its components
fails. Note that conventional fixed controllers can be considered to have a low
degree of autonomy since they can only tolerate a restricted class of plant parameter
318 P.J. ANTSAKLIS ET AL.
variations and disturbances. The autonomous controller architecture given in the next
section provides the functions to attain a high level of autonomy. It can interface with
both the crew, ground station and the on-board systems of the space vehicle. A
command by the pilot or the ground station is executed by dividing it into appropriate
subtasks which are then performed by the controller. The controller can deal with
unexpected situations, new control tasks, and failures within limits. To achieve this,
high-level decision-making techniques for reasoning under uncertainty and taking
actions must be utilized. These techniques, if used by humans, are attributed to
intelligent behavior. Hence, one way to achieve autonomy is to utilize high-level
decision-making techniques, 'intelligent' methods, in the autonomous controller.
Autonomy is the objective, and 'intelligent' controllers are one way to achieve it. The
field of artificial intelligence [10, 56] and operations research offer some of the tools
to add the higher level decision making abilities.
Autonomous controllers are evolutionary and not revolutionary. They evolve from
existing controllers in a natural way fueled by actual needs, as it is now discussed.

2.1. DESIGN METHODOLOGY- HISTORY

Conventional control systems are designed using mathematical models of physical
systems. A mathematical model which captures the dynamical behavior of interest is
chosen and then control design techniques are applied, aided by CAD packages, to
design the mathematical model of an appropriate controller. The controller is then
realized via hardware or software and it is used to control the physical system. The
procedure may take several iterations. The mathematical model of the system
must be 'simple enough' so that it can be analyzed with available mathematical
techniques, and 'accurate enough' to describe the important aspects of the relevant
dynamical behavior. It approximates the behavior of a plant in the neighborhood of
an operating point.
The first mathematical model to describe plant behavior for control purposes is
attributed to J.C. Maxwell who in 1868 used differential equations to explain instability
problems encountered with James Watt's flyball governor; the governor was intro-
duced in 1769 to regulate the speed of steam engine vehicles. Control theory made
significant strides in the past 120 years, with the use of frequency domain methods and
Laplace transforms in the 30s and 40s and the introduction of the state space analysis
in the 60s. Optimal control in the 50s and 60s and stochastic, robust and adaptive
control methods in the 60s to today have made it possible to control more accurately
significantly more complex dynamical systems than the original flyball governor.
The control methods and the underlying mathematical theory were developed to
meet the ever-increasing control needs of our technology. The evolution in the control
area was fueled by three major needs:

(i) The need to deal with increasingly complex dynamical systems.

(ii) The need to accomplish increasingly demanding design requirements.
TOWARDS INTELLIGENT AUTONOMOUSCONTROLSYSTEMS 319
(iii) The need to attain these design requirements with less precise advanced
knowledge of the plant and its environment, that is, the need to control under
increased uncertainty.

The need to achieve the demanding control specifications for increasingly complex
dynamical systems has been addressed by using more complex mathematical models
such as nonlinear and stochastic ones, and by developing more sophisticated design
algorithms for say, optimal control. The use of highly complex mathematical models,
however, can seriously inhibit our ability to develop control algorithms. Fortunately,
simpler plant models, for example linear models, can be used in the control design;
this is possible because of the feedback used in control which can tolerate significant
model uncertainties. Controllers can then be designed to meet the specifications
around an operating point, where the linear model is valid and then via a scheduler
a controller emerges which can accomplish the control objectives over the whole
operating range. This is, for example, the method typically used for aircraft flight
control. In autonomous control, we need to significantly increase the operating range
of the plant. We must be able to deal with significant uncertainties in models of
increasingly complex dynamical systems in addition to increasing the validity range
of our control methods. This will involve the use of intelligent decision-making
processes to generate control actions so that a performance level is maintained, even
though there are drastic changes in the operating conditions.
There are needs today that cannot be successfully addressed with the existing
conventional control theory. They mainly pertain to the area of uncertainty. Heuristic
methods may be needed to tune the parameters of an adaptive control law. New
control laws to perform novel control functions should be designed while the system
is in operation. Learning from past experience and planning control actions may be
necessary. Failure detection and identification is needed. These functions have been
performed in the past by human operators. To increase the speed of response, to
relieve the pilot from mundane tasks, and to protect operators from hazards, autonomy
is desired. It should be pointed out that several functions proposed in later sections,
to be part of the autonomous controller, have been performed in the past by separate
systems; examples include fault trees in chemical process control for failure diagnosis
and hazard analysis, and control system design via expert systems.

2.2. FUNCTIONSOF AN AUTONOMOUSCONTROLLER

There are certain functions, characteristics, and behaviors that autonomous systems
should possess [62, 20]. These are outlined below. Some of the important character-
istics of autonomous controllers are that they relieve humans from time con-
suming mundane tasks thus increasing efficiency, enhance reliability since they
monitor health of the system, enhance performance, protect the system from inter-
nalty induced faults, and they have consistent performance in accompIishing complex
tasks.
320 P.J. ANTSAKLlS ET AL.
There are autonomy guidelines and goals that should be followed and sought after
in the development of an autonomous system. Autonomy should reduce pilot/crew/
ground station work load requirements for the performance of routine functions. The
gains due to autonomy would be superficial if the maintenance and operation of the
autonomous controller taxed the operators. Autonomy should enhance the functional
capability of the future space vehicle. Since the autonomous controller will be per-
forming the simpler routine tasks, men will be able to dedicate themselves to even
more complex tasks.
There are certain autonomous system architectural characteristics that should be
sought after in the design process. The autonomous control architecture should be
amenable to evolving future space vehicle needs and updates in the state of the art.
The autonomous control architecture should be functionally hierarchical. Highest
authority lies nearest the pilot, crew, or ground station; for lower level subsystems to
take some actions, they have to clear it with a higher level authority. The system must,
however, be able to have the lowest level subsystems, that are monitoring and
reconfiguring for failures, act autonomously to enhance system safety.
There are also certain operational characteristics of autonomous controllers.
Ground controllers and/or the pilot or crew should have ultimate supervisory over-
ride control of future space vehicle autonomy functions. Autonomous activities
should be highly visible, 'transparent', to the ground controllers and the flight crew
to the maximum extent possible.
Finally, there must be certain features inherent in the autonomous system design.
Autonomous design features should prevent failures that would jeopardize the overall
space vehicle mission goals or safety. These features should enhance crew safety, and
avoid false alarms and unnecessary hardware configuration. This implies that the
controller should have self-test capability. Autonomous design features should also be
tolerant of transient errors, they should not degrade the reliability or operational
lifetime of future space vehicle functional elements, they should include adjustable
fault detection thresholds, avoid irreversible state changes, and provide protection
from erroneous or invalid external commands.

2.3. INTELLIGENTAUTONOMOUSCONTROL
The necessity for a succession of increasingly complex control systems from classical
to adaptive and intelligent control, to meet the ever increasing performance require-
ments on the current and future complex dynamical systems, is described. The basic
elements of intelligent controllers are highlighted and an outline of the relevant
research on intelligent control is given.

2.3.1. Motivation." Sophistication and Complexity in Control

The complexity of a dynamical system model in terms of determinism, nonlinearities,
etc., and the increasingly demanding closed loop system performance requirements,
TOWARDS INTELLIGENT AUTONOMOUS CONTROL SYSTEMS 321
necessitate the use of more complex and sophisticated controllers. For example,
highly nonlinear systems normally require the use of more complex controllers
than low order linear ones when goals beyond stability are to be met. The increase
in uncertainty, which corresponds to the decrease in how well the problem is
structured or how well the control problem is formulated, and the necessity to
allow human intervention in control, also necessitate the use of increasingly sophisti-
cated controllers. Controller complexity and sophistication is then directly pro-
portional to both the complexities of the plant model and of the control design
requirements.
Based on these ideas, [49, 24] suggest a hierarchical ranking of increasing controller
sophistication on the path to intelligent controls. At the lowest level, deterministic
feedback control based on conventional control theory is utilized for simple linear
plants. As plant complexity increases, such controllers will need for instance, state
estimators. When process noise is significant, Kalman filters may be needed. Also, if
it is required to complete a control task in minimum time or with minimum energy,
optimal control techniques are utilized. When there are many quantifiable, stochastic
characteristics in the plant, stochastic control theory is used. If there are signficant
variations of plant parameters, to the extent that linear robust control theory is
inappropriate, adaptive control techniques are employed. For still more complex
plants, self-organizing or learning control may be necessary.
At the highest level in their hierarchical ranking, plant complexity is so high, and
performance specifications so demanding, that intelligent control techniques are used.
The plant is so complex that it is either inappropriate or impossible to describe it with
conventional system models such as differential equations. For instance, even though
it might be possible to accurately describe some system with very complex nonlinear
differential equations, it may be inappropriate if this description makes subsequent
analysis too difficult.
The complexity of the plant model necessary for design depends on both the
complexity of the physical system and on how demanding the design specifications
are. There is a tradeoff between model complexity and our ability to perform analysis
on the system via the model. However, if the performance specifications are not too
demanding, a more abstract model can be utilized, which will make subsequent
analysis simpler. This model intentionally ignores some of the system characteristics,
specifically those that need not be considered in attempting to meet the particular
performance specifications. Often, to obtain an abstract model, high level symbolic
representations are utilized [60, 42, 33, 32, 46]. The choice of the modelling tech-
nique affects most aspects of analysis and design of a controller for the system;
consequently, special control methodologies must be used with the abstract models.
Such methodologies include advanced decision making techniques from the field of
AI, which are used to reason over these representations and decide what control
actions are appropriate to take. Since the AI techniques generally model the human
decision-making processes, about what actions to take next, they can easily provide
for human interface.
322 P.J. ANTSAKLISET AL.
It is perhaps of interest to notice that all controllers in the hierarchy described
above can be considered to be a type of problem solving system. (For an overview of
the theory of problem solving see [56].) This is because there is a desirable goal
behavior and the problem solver generates actions to change an initial undesirable
behavior to the goal. It is our view that problem solving systems can be classified into
two categories, conventional and AI. Several characteristics distinguish these two
classes of problem solving systems. The conventional problem solving system is
numeric-algorithmic, it is somewhat inflexible, it is based on the well developed theory
of algorithms or differential equations, and it can thus be studied using a variety of
methodical modelling, analysis, and design techniques. Classical control systems are
an example of conventional problem solving systems. An AI problem solving system
is a symbolic decision-maker, it is flexible with graceful performance degradation, and
it is based on formalisms which are not well developed; actually there are very few
systematic mathematical modelling, analysis, and design techniques for these systems.
AI expert and planning systems are examples of AI problem solving systems. When
comparing the characteristics of AI and non-AI systems, one can make the following
observations: The decision rate in conventional systems is typically higher than that
of AI systems. The abstractness and generality of the models used in AI systems is
high compared with the fine granularity of models used in conventional systems.
Symbolic representations, rather than numeric, are used in AI systems. High-level
decision-making capabilities similar to those of humans exist in AI systems to a much
greater extent than in conventional systems. The result is that a higher degree of
autonomy exists in AI systems than in conventional ones.
In the hierarchical ranking of increasingly sophisticated controllers described
above, the decision to choose more sophisticated control techniques is made by
studying the control problem using a controller of a certain complexity belonging to
a certain class. When it is determined that the class of controllers being studied (e.g.,
adaptive controllers) is inadequate to meet the required objectives, a more sophisti-
cated class of controllers (e.g. intelligent controllers) is chosen. That is, if it is found
that certain higher level decision-making processes are needed for the adaptive
controller to meet the performance requirements, then these processes can be incor-
porated via the study of intelligent control theory. These intelligent autonomous
controllers are the next level up in sophistication. They are enhanced adaptive
controllers, in the sense that they can adapt to more significant global changes in the
vehicle and its environment than conventional adaptive controllers, while meeting
more stringent performance requirements.
One turns to more sophisticated controllers only if simpler ones cannot meet the
required objectives. Below we list some of the reasons why it is necessary to use
intelligent autonomous control for future space vehicles:

(i) Future space vehicles will be increasingly complex. Some characteristics that
are needed in the model used to design their controller can only be described
by symbolic representation techniques.
TOWARDS INTELLIGENT AUTONOMOUS CONTROL SYSTEMS 323
(ii) Control functions normally performed by the pilot, crew, or ground station
must be incorporated into the controller for autonomous operation.
Therefore, expert personnel's control decisions will have to be automated.
(iii) Human intervention in the control process should be allowed. A facility to
interrupt the autonomous operation of the controller in case of design objec-
tive changes or controller failures should be included.

The need to use intelligent autonomous control stems from the need for an increased
level of autonomy in achieving complex control tasks. In the next section a number
of intelligent control research results which have appeared in the literature are
outlined.

2.3.2. Intelligent Autonomous Control: A Literature Overview

The field of intelligent autonomous control is new. Some of the recent research efforts
have been reported in the Proceedings of the 1985 Workshop on Intelligent Control,
1986 Intelligent Autonomous Systems Conference, the Space Telerobotics Workshop,
and the Proceedings of the 1987 Symposium on Intelligent Control, and a wealth of
useful references can be found there. Research that had a direct influence on our work
is outlined below.
Intelligent controllers are hierarchical and the theory of hierarchical systems is
relevant [36]. This work sets some of the fundamental concepts in intelligent control
such as the need for varying degrees of abstractness in models used at the different
levels in the controller. It also presents a theory of coordination for all subsystems of
the intelligent controller. Coordination issues are also examined in [12]. The work in
[18] extends Mesarovic's work. Fundamentals of intelligent systems such as the
principle of increasing intelligence with decreasing precision, granularity, time scale
density, model abstractness are discussed in [53, 54, 37]; the need for the integration
of techniques from AI, operations research and conventional control theory to
perform intelligent control tasks is also discussed there. The integration of AI and
control theoretic methods is discussed in [13, 22, 11].
In [20] the authors explain how a wide variety of AI techniques will be useful in
enhancing space station autonomy, capability, safety, etc. This project oriented book
points to relevant AI techniques, research areas, and progress in solving the posed
problems. In the Space Telerobotics Workshop, the work of several researchers from
NASA, the Jet Propulsion Laboratory, Ames, Johnson Space Center [55, 21, 30, 6]
and others, e.g., [57], is outlined and specific details of their programs and research
directions are given. In [62] a detailed study of characteristics of autonomous space
systems is given and an architecture for the complete autonomous operation of the
space station is presented; examples are used to illustrate the behavior of the auto-
nomous system.
There has been much work on developing intelligent controllers for robots. A good
overview is given in [51]. The work in [7] describes an initial effort towards a
324 P.J. ANTSAKLIS ET AL.

hierarchical intelligent controller based on AI planning methods. Balaram developed

an architecture for the planning system that incorporates intelligent control funda-
mentals and that is accurately structured for his control task. Other intelligent
controllers that use planning techniques are given in [26, 25, 8, 15, 16, 47]. The vision
problem for intelligent controllers is examined in [17].
The work by Saridis and Valvanis in [49-54], and [63-66] probably represents the
most complete mathematical approach to the analysis of intelligent machines. They
stress a three level hierarchy for intelligent systems with execution, coordination, and
management levels, and the "principle of increasing intelligence with decreasing
precision". They use entropy as an unified quantification of disorder in each of the
three levels in their intelligent system. In an intelligent controller, they choose the
control action that will decrease the entropy in their system.
Other important work in the field of intelligent control is given in [1, 2, 37-40, 29,
19, 67]; a nested hierarchical controller is described in [38, 68]; some similarities
between planning and intelligent control are given in [28]; and an interesting black-
board architecture is studied in [9]. Supervisory control with distributed intelligence
is examined in [69], and heuristic control of real time processes is reported in [48]. The
intelligent restructurable controls problem for aircraft was studied in [61, 58, 43]. The
fault detection and identification problem in an intelligent controller was examined in
[27, 44--45].
There have been numerous studies on the use of expert systems to control various
process. For instance, in [23] the authors use an expert system in the adaptive control
of large space structures. In [5] expert systems have been used in chemical process
control and the term 'expert control' has been coined. There are interesting relation-
ships between the type of problems examined in intelligent autonomous control,
'fuzzy control' [71], and 'automated reasoning' [70]. There exist numerous books and
articles on these topics. As reported in the above literature, several limited versions
of autonomous systems have been constructed. There are many sorts of robots and
vehicles, each achieving a certain level of autonomy. They have been used for
manufacturing [34], underwater exploration [41], and other applications.
A detailed functional architecture for autonomous controllers is the essential first
step in their development. Such an architecture is introduced in the next section. It will
show how to combine intelligent functions in a controller to achieve autonomy. Based
on this architecture, the fundamental concepts and methods that need to be developed
are identified.

3. An Intelligent Autonomous Control Architecture for Future Space

Vehicles
In this section, a fundamental architecture of an autonomous controller for future
space vehicles is introduced and discussed. This hierarchical architecture has three
levels, the execution level, the coordination level, and the management and organiza-
tion level. The functions of each level are described in detail. The architecture exhibits
TOWARDS INTELLIGENT AUTONOMOUS CONTROL SYSTEMS 325
certain characteristics, as discussed below, which have been shown in the literature to
be necessary and desirable in autonomous systems. Based on this architecture we
identify the important fundamental issues and concepts that are needed for an
autonomous control theory.

3.1. ARCHITECTURE OVERVIEW: STRUCTURE AND CHARACTERISTICS

The overall functional architecture for an autonomous controller is given by the

architectural schematic of Figure i. This is a functional architecture rather than a
hardware processing one, therefore it does not specify the arrangement and duties of
the hardware used to implement the functions described. Note that the processing
architecture also depends on the characteristics of the current processing technology;
centralized or distributed processing may be chosen for function implementation
depending on available computer technology.
The architecture in Figure 1 has three levels. At the lowest level, the execution level,
there is the interface to the vehicle and its environment via the sensors and actuators.
At the highest level, the management and organization level, there is the interface to
the pilot and crew, ground station, or onboard systems. The middle level, called the
coordination level, provides the link between the execution level and the management
level. Note that we follow the somewhat standard viewpoint that there are three major
levels in the hierarchy. It must be stressed that the system may have more or fewer
levels. For instance, see the architecture developed in [62]. Some characteristics of the
system which dictate the number of levels are the extent to which the operator can
intervene in the system's operations, the degree of autonomy or level of intelligence
in the various subsystems, the dexterity of the subsystems, and the hierarchical
characteristics of the plant.
The sensors and actuators are implemented mainly with hardware. They are the
connection between the physical system and the controller. Software and perhaps
hardware are used to implement the execution level. Mainly software is used for
both coordination and management levels. Note that the multiple copies of the
different levels reflect the distinct character of the various control functions
necessary to achieve autonomy. For example, there may be one control manager
which directs a number of different adaptive control algorithms to control the
flexible modes of the vehicle via appropriate sensors and actuators. Another control
manager is responsible for the control functions of a robot arm for satellite repair.
The control executive issues commands to the managers and coordinates their
actions.
Note that the autonomous controller is only one of the autonomous systems on the
vehicle. It is responsible for all the functions related to the control of the physical
system and allows for continuous online development of the autonomous controller
and to provide for various phases of mission operations. The tier structure of the
architecture allows us to build on existing advanced control theory. Development
progresses, creating each time, higher-level adaptation and a new system which can
326 P.J. ANTSAKLIS ET AL.
PILOT & CREW / GROUND STATION / ONBOARD SYSTEMS

II
INTERFACE
I CONTROL EXECUTIVE
Management & ~ GENERATION UPPER
Organization ~ (Upper Level) MANAGEMENT
Level ~ MONITORING Decision
~ ASSESSOR Making &
~ (Upper Level) Learning

I
IIa CONTROL MANAGER
Higher ~ MIDDLE
~ (Middle Level) MANAGEMENT
II ~
Coordination ~ Decision
Level ~ (Lower Level) Maldng,
Learning &
lib CONTROL IMP. SUPERVISOR Algorithms
Lower ~
~ *SCHEDULER ~ LOWER
*LEARNING (Low Level) MANAGEMENT

ADAPTIVE C O N T R O L &,
IDENTIFICATION Illl
HI ~ Illl Algorithms in
Execution ~ ID & STATE EST~[I[ Software &
ALGORn'aMS Illl aardwarc
, *IblFO"DISTRIBAUTOR IllI

Hardware

VEHICLE & ENVIROMENT

Fig. 1. Autonomouscontroller architectural schematic.

be operated and tested independently. The autonomous controller performs many of

the functions currently performed by the pilot, crew, or ground station. The pilot
and crew are thus relieved from mundane tasks and some of the ground station
functions are brought aboard the vehicle. In this way the vehicle becomes more
autonomous.
TOWARDS INTELLIGENT AUTONOMOUS CONTROL SYSTEMS 327
3.2, FUNCTIONAL OPERATION

Commands are issued by higher levels to lower levels and response data flows from
lower levels upwards. Parameters of subsystems can be altered by systems one level
above them in the hierarchy. There is a delegation and distribution of tasks from
higher to lower levels and a layered distribution of decision-making authority. At each
level, some preprocessing occurs before information is sent to higher levels. If requested,
data can be passed from the lowest subsystem to the highest, e.g., for display. All
subsystems provide status and health information to higher levels. Human inter-
vention is allowed even at the control implementation supervisor level (IIb).
The specific functions at each level are described in detail in later sections. Here we
present a simple illustrative example to clarify the overall operation of the auto-
nomous controller. Suppose that the pilot desires to repair a satellite. After dialogue
with the control executive via the interface, the task is refined to 'repair satellite using
robot A'. This is arrived at using the capability assessing, performance monitoring,
and planning functions of the control executive. The control executive decides if the
repair is possible, under the current performance level of the system, and in view of
near term planned functions. The control executive, using its planning capabilities,
sends a sequence of subtasks sufficient to achieve the repair to the control manager. This
sequence could be to order robot A to: 'go to satellite at coordinates x y z ' , 'open repair
hatch', 'repair'. The control manager, using its planner, divides say the first subtask,
'go to satellite at coordinates x y z ' , into smaller subtasks: 'go from start to x~yLz]',
then 'maneuver around obstacle', 'move to x 2 Y 2 Z 2 ' , 9 9 9 'arrive at the repair site and
wait'. The other subtasks are divided in a similar manner. This information is passed
to the control implementation supervisor, which recognizes the task, and uses stored
control laws to accomplish the objective. The subtask 'go from start to x~ y~zm', can
for example, be implemented using stored control algorithms to first, proceed forward
10 meters, to the right 15 degrees, etc. These control algorithms are executed in the
controller at the execution level utilizing sensor information; the control actions are
implemented via the actuators.
It is important at this point to discuss the d e x t e r i t y of the controller. The execution
level of a highly dexterous controller is very sophisticated and it can accomplish
complex control tasks. The implementation supervisor can issue commands to the
controller such as 'move 15 centimeters to the right', and 'grip standard, fixed
dimension cylinder', in a dexterous controller, or it can completely dictate each mode
of each joint (in a manipulator) 'move joint ! 15 degrees', then 'move joint 5 3
degrees', etc. in a less dexterous one. The simplicity, and level of abstractness of macro
commands in an autonomous controller depends on its dexterity. The more sophisti-
cated the execution level is, the simpler are the commands that the control imple-
mentation supervisor needs to issue. Notice that a very dexterous robot arm may itself
have a number of autonomous functions. If two such dexterous arms were used to
complete a task which required the coordination of their actions, then the arms would
be considered to be two dexterous actuators and a new supervisory autonomous
328 P.J. ANTSAKLIS ET AL.

controller would be placed on top for the supervision and coordination task. In
general, this can happen recursively, adding more intelligent autonomous controllers
as the lower level tasks, accomplished by autonomous systems, need to be supervised.

3.3. THE E X E C U T I O N LEVEL (llI)

The functional architecture for the execution level of the autonomous controller is
shown in Figure 2. Its main function is to generate, via the use of numeric algorithms,
low level control actions as dictated by the higher levels of the controller, and apply
them to the vehicle. It senses the responses of the vehicle and environment, processes
it to identify parameters, estimates states, or detects vehicle failures, and passes this
information to the higher levels.
The Sensor and Actuator subsystems are depicted in Figure 2. These devices which
physically accomplish the functions for the autonomous controller are at the lowest
level of the architecture. The complexity of these devices depends on the dexterity of
the controller. All sensors which provide information from the vehicle and environ-
ment to any component in the autonomous controller are included here. On the
execution level, the controller will need feedback information about control variables.
The state estimator and parameter identifier also use such outputs for their respective
tasks. The failure detection and identification (FDI) algorithms need these outputs
and those of special failure sensors to enable them to detect failures. To perform
'execution monitoring' for the planning systems at the higher levels, the dynamical
response of the system must be sensed and passed to the planning system so that
it can determine if a plan has failed. The implementation supervisor also needs
sensor information so that it can, for instance, make the smooth transition in the

Conuol
............f......
Implementation Supervisor
Information
Assessor
Control
ImplementationSupervisor FDI lib

Execution
Information
Distributor
Paran~te,
Identifier&
State
Estimator I FDI
Algorithms

Actuators Sensors

Vehicle and Environment

Fig. 2. Execution level.
TOWARDS INTELLIGENT AUTONOMOUS CONTROL SYSTEMS 329
implementation of a newly designed control law. Sensory information is also used in
performance monitoring, capabilities assessing, tuning, scheduling, and display to the
pilot, crew, ground station, or other onboard systems. The actuators are the usual
control actuators (transducers) which translate the outputs of the controller to actions
meaningful to the vehicle. For a highly dexterous controller, a whole manipulator
may be considered to be an 'actuator'.
The main function of the Controller in Figure 2 is to execute the control algorithms
and to issue commands to the actuators. It performs advanced conventional adaptive
control functions. It receives, in real time, all the necessary data (from the information
distributor) to execute the current control algorithm. The information consists of
current output values from the sensors, model parameter estimates and state esti-
mates, as they are generated from the identifier. The adaptation part of the controller
algorithmically interprets the values of the measured plant variables and the estimated
plant parameters and states; and it adjusts, on-line, the coefficients of the control law
which runs in the execution part of the controller. These functions correspond to
conventional adaptive control. The adaptation algorithm can contain information
about the model to be followed, thus implementing 'model following' adaptive
control. Since the model parameters are explicitly estimated and then used in the
control law adaptation, the structure appears to suggest an 'indirect' adaptive control
approach. However, notice that this is not necessarily the case since the model
parameter estimates from the identifier can simply be ignored and the adaptation
algorithm can directly process the information from sensors to directly estimate the
control law coefficients, thus implementing 'direct' adaptive control. If a fixed control
law is used, then the appropriate sensor data are simply fed back to the control law
which is being executed. The sensor data are values of measured variables; these
include states as well.
All possible conventional control functions can be performed via the proposed
achitecture. For fixed control laws, one could envision a loop containing the sensors
providing feedback information, through the information distributor, to the con-
troller; the control actions are performed via the actuators. For adaptive control this
also involves the model parameter identifier. In addition to advanced adapative
control functions, the controller has the following capabilities: The controller allows
intervention from above. It of course allows the introduction of reference signals for
set points and tracking as conventional controllers do. In addition it receives com-
mands: (i) To alter the parameters of adaptation (as determined by the tuner in the
coordination level), and (ii) To switch to different control laws altogether suggested
by the scheduler or the control redesigner.
If the higher levels of the architecture are ignored, the intervention to the controller
can be envisioned as being that of a human operator who adjusts certain parameters
depending on performance, sets the set points, switches to different algorithms
from time to time or to new control laws when failures occur. It returns status infor-
mation to the higher level, such as what particular control law is currently running;
also information about the health of the system (errors in implementation, etc.).
330 P.J. ANTSAKLISET AL.
The controller has access to a variety of stored control laws. The particular location
of the stored programs is not important in this functional architecture. They could be
located in the controller, or in the level above (implementation supervisor). If they are
located above, then one should allow for down loading these programs. Since control
law switching is desirable, transition programs, for smooth control law switching are
necessary. When the scheduler and control redesigner send new control laws to be
implemented they should also attach a program to ensure the smooth transition from
the current to the new control law. The controller consists basically of software plus
hardware as is necessary.
The main function of the information distributor shown in Figure 2, is to distribute
sensor, parameter, and state information where it is needed. Since the control models
and therefore the control, identification, estimation, and FDI algorithms do change,
it is essential to guarantee that the execution level subsystems receive each time the
correct information. Information about the current control models and current
algorithms is provided from above. Using stored information, the distributor
provides the correct sensor information to the controller for control feedback pur-
poses, to the identifier for model parameter identification and state estimation, and
to the FDI for detection and isolation. After perhaps some preprocessing, it also
provides this information to higher levels.
The main function of the adaptive parameter identifier and state estimator shown in
Figure 2 is to execute parameter identification algorithms and state estimation
algorithms and to continuously pass this information to the controller, to the FDI
algorithms, and to higher levels. It receives all appropriate sensor information from
the information distributor. The parameters and the states, the estimates of which are
sought, depend on the particular control model used. Since the control model and the
control law do change, the parameter identifier and state estimator should be able to
switch control models and identification and estimation algorithms. This information
is given from above. It provides the necessary parameter and state estimates to the
controller and to the FDI algorithms via the information distributor. It returns to the
higher level, parameter estimates and state estimates of the current model (via the
information distributor) and information as to the status and health of the system
directly.
The main function of the FDI algorithms shown in Figure 2, is to execute FDI
algorithms for failures detected at the execution level of the autonomous controller.
It receives all appropriate sensor information via the information distributor. This
includes information from sensors specifically located to detect failures at the actuator
level of the control system; it also includes model parameter and state estimates from
the identifier. It has the ability to switch algorithms and plant models. The FDI
algorithms return information to the higher level FDI subsystems.

3.4. C O O R D I N A T I O N LEVEL (IIb)

The functional architecture for coordination level lib is shown in Figure 3. Coordina-
tion level IIb receives commands to perform predetermined specific control tasks from
TOWARDS INTELLIGENTAUTONOMOUS CONTROL SYSTEMS 331
Control Manager FDI Ha

Adaptive
Tuner
Supervisor ~ FDI
*Status Monitoring IIb
*CrisisManagement

Infomiation
Assessor

o o o ~ o ~ o a , woa, o ~ o ~ o o o o o a B o o o o w i

ControUer Information Idcntificr FDI

Distributor Algorithms
Fig. 3. Coordinationlevel Ilb.

the control manager in the level above. It provides the appropriate sequence of
control and identification algorithms to the execution level below. Its ability to deal
with extensive uncertainties is limited.
The main function of the control implementation supervisor shown in Figure 3, is to
carry out the sequence of control actions dictated by the control manager. It can
accomplish predetermined control actions and cope with limited predetermined crisis
situations. The supervisor receives the sequence of control tasks to be accomplished
from the control manager and it has access to a variety of control models, and control,
identification and estimation algorithms. It selects appropriate reference signals for
the controller and it optimizes the subsequences of actions to accomplish the tasks
dictated by the above levels in the best way possible. The supervisor uses the scheduler
to decide what models and algorithms to use in the controller and identifier; it uses
the tuner to decide how to adapt parameters in the algorithms, which are currently
used, and it sends this information to the execution level. It monitors the status of the
system at lib and III, i.e., what algorithms and models are currently used, and the
health of the systems. The supervisor does performance monitoring on lib and III
levels using information provided by the information assessor and FDI lib. It
contains a crisis management facility to deal with certain failures. This includes a
number of methods to maintain performance or to maintain a certain degree of safety
in operations, while degrading performance gracefully. For example, if a failure in an
actuator or sensor is detected, it can switch to an alternative control method using
other actuators or sensors to maintain performance. If performance cannot be main-
tained, it should degrade gracefully, guaranteeing safety (stability). It will take the
necessary steps to maintain stability after a failure is detected and it is isolated
and identified. The control implementation supervisor uses learning to improve the
332 P.J. ANTSAKLISET AL.
implementation of the (predetermined) control forms. It thus improves the speed and
accuracy of tuning with experience, it improves its crisis management and the schedul-
ing of algorithms, and it learns how to more efficiently optimize the overall opera-
tions, as a good human supervisor would do; it also learns completely new control
methods sent from the level above. It informs the control manager about the health
of the system in levels IIb and III, about its status (the progress in performing the
tasks) and it notifies the manager if failures, and unexplained (at that level) per-
formance degradation is occurring.
The main function of the Scheduler shown in Figure 3, is to determine, during the
performance of a specific control function, if certain conditions are met in order to
switch to alternative control laws (and plant models) and to appropriate identifica-
tion, estimation and FDI algorithms. It receives information from the implementation
supervisor as to the control function to be performed, together with information
about the plant models and their validity range, the corresponding control laws, and
the rest of the algorithms. Based on information it receives from the supervisor it
decides when to switch to the proper algorithms and models. The criteria for switching
are predetermined, in perhaps tabular form, and they also depend on information from
environmental sensors. This information is transmitted from the higher level through
the supervisor. Examples will be the schedulers used for docking control. Depending,
for example, on approaching speed and attitude, an appropriate new control law is
selected. Here, the scheduler also selects corresponding plant model when necessary.
The scheduler does not deal with crisis situations.
The main function of the adaptive tuner shown in Figure 3 is to determine, during
the execution of particular algorithms, if specific conditions are met in order to adjust,
tune, certain parameters in the adaptation laws. It receives information from the
implementation supervisor as to the current algorithm being executed, control and
identification algorithms, and also information from the information assessor (via the
supervisor) necessary to decide first if timing is appropriate. Then based on predeter-
mined criteria, it selects the new values for the parameters in the adaptation laws. The
criteria for tuning will be based on excessive output, state, and parameter errors, and
the selection of the new adaptive parameter values will depend on algorithms or
heuristic rules using performance measures and actual past and present inputs
and outputs. In this way parameter tuning of identification and control algorithms
(adaptive, robust, optimal) is accomplished.
The main function of the information assessor shown in Figure 3 is to process and
distribute sensor, state and parameter information to the information distributor
(execution level) and the implementation supervisor. It receives information from the
supervisor as to the current plant model, control, estimation and identification, and
FDI algorithms, and it instructs the information distributor to pass the necessary
sensor information to the controller, identifier, and FDI systems. It receives, from the
identifier via the distributor, information about the current model parameter and
state estimates. After instruction from the supervisor it supplies to the supervisor
processed information such as errors in state and parameter estimation. To do this,
TOWARDS INTELLIGENTAUTONOMOUSCONTROLSYSTEMS 333
it uses sensor data supplied by the distributor and models supplied by the supervisor.
This processed information is used by the tuner, the scheduler, and the control
implementation supervisor for performance monitoring.
The main function of the FDI IIb subsystem shown in Figure 3 is to supervise the
FDI algorithms (execution level) and to detect and identify, using algorithms and
heuristic methods, failures that occurred at the execution level. It passes the informa-
tion about the current models used from the supervisor to the FDI algorithms. It
sends appropriate FDI algorithms to be executed to the lower level. It receives the
outputs of those algorithms. It compares them with additional information from the
supervisor, and it proceeds, after detecting a failure, to isolate and identify it. It
informs the FDI IIa subsystem about the status of the failure and it also informs the
supervisor so that predetermined crisis measures can be taken if necessary and
possible. If the crisis cannot be dealt with at that level, the information is passed to
the FDI IIa, and the designer via the control manager.

3.5. COORDINATION LEVELIla

The functional architecture for coordination level IIa is shown in Figure 4. Coordina-
tion level Ila receives commands from the management level which it must determine
how to perform using the designer and planner and considering information from
FDI IIa and the control implementation supervisor. It generates a sequence of control
actions that the control implementation supervisor can recognize and passes them to
it. This coordination level has abilities to deal with significant uncertainties.
The main function of the control manager shown in Figure 4 is to accomplish the
control tasks given by the control executive. It can accomplish predetermined control
actions, using the lower levels, but also it can cope with failures to a large degree. In
general it is equipped to successfully carry out the control tasks under a wide variety
of unanticipated vehicle and environmental conditions. It can also be directed to
prepare for future requirements, building new control laws and contingency methods

Control Executive Control Executive

Designer ~ Control Manager

.~ " I~
[~ Monitoring FDI Ha
[ *CrisisManagement
Planner ~ *Optimization ~

8" ,W W O K , r ~ G ~ ' C

. . . . to;;ol FDI lib

Fig. 4. Coordinationlevel lla.
334 P.J. ANTSAKLISET AL.
using the designer. When a control task is given, it breaks it down into a sequence of
control actions, using the planner, and it passes it to the implementation supervisor.
It receives processed sensor information from the supervisor about the current
positions and information from the above about the goals so that it can plan its
actions. It also passes to the supervisor newly designed algorithms and contingency
plans. It receives, from the implementation supervisor, status and health information
and it passes its status and health information to the executive. It does performance
evaluation and monitoring on II and III levels. For example, it evaluates the per-
formance of a sequence of tasks so that, changes in the next sequence are implemented
if necessary. It also contains a crisis management facility to deal with failures. It is
similar to the one in the implementation supervisor but it deals with higher level
contingencies. It has significant learning abilities to improve its performance. It does
optimization of the system below it using the planned actions, and it suggests new
strategies in algorithm selection to the implementation supervisor.
The main function of the designer shown in Figure 4 is to develop methodologies
to deal with novel situations for which no prior designs have been made. These include
detected failures via the FDI system, in which case new control laws must be designed
on-line, using the models and specifications provided by the control manager. They
also include dealing with new control tasks suggested by the manager or the higher
level. When there are no requirements to develop new methods in real time, the
designer, under direction of the manager, works on developing new methods to build
up the crisis management algorithm inventory, and the inventory of algorithms to
deal with new control tasks needed some time in the future. These algorithms are
passed to the control implementation supervisor at the direction of the control
manager. In this way, the system is enriched and improved, it becomes more experi-
enced, as it can deal with a greater number of contingencies and tasks. The designer
uses decision-making under uncertainty (symbolic based methods) to select design
algorithms. When the designer must react in a very short time to deal with, say a major
failure, it may decide to initially suggest a method which will preserve the system
integrity without meeting all the performance specifications. In the mean time it can
work to produce a full solution to the problem.
The main function of the planner, shown in Figure 4, is to plan the sequence of
control actions, to be given to the control implementation supervisor (lib), in order
to accomplish a higher level control task. If, for example, the control executive orders
the robot to move to a specific location, the planner, based on the current and possible
future robot locations, will devise a sequence of actions to be taken so that tasks will
be accomplished. It will, perhaps at the beginning, suggest a skeleton plan which will
be refined as it is being executed. For example, start moving to the right 15 degrees,
report if passage blocked, etc.
The main function of the FDIIIa system shown in Figure 4 is to detect and identify
failures which occurred at levels lib and III. It also supervises the FDI lib system. It
receives failure information from the execution level (III) via FDI IIb and additional
information from the control manager. It informs the manager about the failure
TOWARDS INTELLIGENTAUTONOMOUSCONTROLSYSTEMS 335
location and its severity, so that measures can be taken, using perhaps the services of
the designer. It directly informs the control executive about the status of the failures
detected at any level since they are very important in capability assessment. It uses
high level decision-making involving heuristics and few algorithms.

3.6. MANAGEMENTAND ORGANIZATIONLEVEL(1)

The functional architecture for the management and organization level (I) is shown
in Figure 5. It interfaces to the pilot, crew, ground station, and other onboard systems
and performs the highest level control functions. It oversees and directs all the
activities at both the coordination and execution levels. It is the most 'intelligent' of
the three levels.
The main function of the control executive shown in Figure 5 is to accomplish high
level control tasks given by the pilot, crew, ground station, or other onboard systems.
Such a task could be: Change orbit to . . . . deploy satellite (open door, turn, etc, then
deploy), repair satellite via robot A (send robot to satellite, open hatch, repair),
retrieve satellite, etc. It performs high level planning. It optimally breaks down the
'macro commands' into simpler commands for the control manager (IIa). It performs
capability assessing of the control system. It receives information about faults from
FDI and about status and health from the control manager and performs high level
performance monitoring. It evaluates the current situation and it predicts what can be
reasonably expected to be accomplished in a certain time. For example, 'docking
procedures under way, estimated docking in 30 seconds'. It provides this information
to the goal generation facility, which by exchanging information, having a dialogue,
with other onboard systems, pilot, or ground station via the interface, generates
attainable realistic goals to be accomplished by the autonomous controller. For
example, in view of the current situation, 'docking can be achieved in 30 seconds but
not in 20 seconds as requested'. These goals are then used by the planner in the control
executive, to plan the necessary steps which lead to their accomplishment. The control

Pilot & Crew'/Ground Station / Onboard Systems

9 C o n t r o l Executive
~ Monitoring ~
~ Monitoring oi.e.ammg Interface
~ Assessing oPlanning
~ Generation

. ". . . ." . . ". . . ". . . ." . . . . . . . " "vol"

Control
Manager rla
Fig. 5. Managementand organizationlevel.
336 P.J. ANTSAKLISET AL.
executive has significant learning abilities. It uses past experience to increase its
efficiency and to improve its capability assessment. It is informed by the control
manager about new capabilities possible, by newly generated control methods. It
suggests preparation for future control tasks. It uses decision making exclusively. It
interprets reports from below and execution commands from above. It can request,
through interface, additional information from the pilot, crew, ground station, or
other onboard systems which may be useful in the control system. This includes
navigation information, future uses of the autonomous controller, etc.
Learning is essential to the development of a true autonomous system. High-level
learning will occur at the management and organization level. At each level of
learning, beginning at coordination level IIb, information is for instance, successively
generalized via induction. The controller may need to learn the model of the plant,
the problem solving strategy, the goals to obtain, and the required performance level.
The main function of the interface, shown in Figure 5, is to provide the liaison,
interface, between the autonomous control system and the pilot and crew/ground
station/other onboard systems. It is an intelligent interface as it allows user friendly
dialogue. It is a language translator, translating language of other systems or the crew
or ground station into a language familiar to the autonomous controller. It displays
data from the control subsystems if requested. It passes the control status to the crew
etc., and desired behavior, and goals, to the control executive.

4. Fundamental Characteristics and Issues

Based on this architecture we identify the important fundamental concepts and
characteristics that are needed for an autonomous control theory. Note that several
of these have been discussed in the literature as outlined in Section 2.3.2. Here, these
characteristics are brought together for completeness. Furthermore, the fundamental
issues which must be addressed for a quantitative theory of intelligent autonomous
control are introduced and discussed.
There is a successive delegation of duties from the higher to lower levels; consequently
the number of distinct tasks increases as we go down the hierarchy. Higher levels are
concerned with slower aspects of the system's behavior and with its larger portions,
or broader aspects. There is then a smaller contextual horizon at lower levels. Also
notice that higher levels are concerned with longer time horizons than lower levels. Due
to the fact that there is the need for high level decision making abilities at the higher
levels in the hierarchy, there is increasing intelligence as one moves from the lower to
the higher levels. This is reflected in the use of fewer conventional numeric-algorithmic
methods at higher levels as well as the use of more symbolic-decision making methods.
This is the "principle of increasing intelligence with decreasing precision" by Saridis.
The decreasing precision is reflected by a decrease in time scale density, decrease in
bandwidth or system rate, and a decrease in the decision (control action) rate. All these
characteristics lead to a decrease in granularity of models used, or equivalently, to an
increase in model abstractness. Model granularity also depends on the dexterity of the
TOWARDS INTELLIGENT AUTONOMOUSCONTROLSYSTEMS 337
autonomous controller as discussed in Sections 3.2 and 3.3. Next we discuss an
approach which, in our opinion, is especially suitable for an analytical study of
intelligent autonomous control systems.
The quantitative, systematic techniques for modelling, analysis, and design of
control systems are of central and utmost practical importance in conventional
control theory. Similar techniques for intelligent autonomous controllers do not exist.
This is of course because of their novelty, but for the most part, it is due to the 'hybrid'
structure (nonuniform, nonhomogeneous nature) of the dynamical systems under
consideration. The systems are hybrid since in order to examine autonomy issues, a
more global, macroscopic view of a dynamical system must be taken than in conven-
tional control theory. Modelling techniques for intelligent autonomous systems must
be able to support this macroscopic view of the dynamical system, hence it is necessary
to represent both numeric and symbolic information (see discussion in Section 2). We
need modelling methods that can gather all information necessary for analysis and
design. For example, we need to model the dynamical system to be controlled (e.g.,
a space platform), failures that might occur in the system, the conventional adaptive
controller, and the high level decision making processes at the management and
organization level of the intelligent autonomous controller (e.g., an AI planning
system performing actions that were once the responsibility of the ground station).
The nonuniform components of the intelligent controller all take part in the genera-
tion of the low level control inputs to the dynamical system, therefore they all must
be considered in a complete analysis. For an extended discussion on the modelling of
hybrid systems consult [72].
It is our viewpoint that conventional modelling, analysis, and design methods
should be used whenevever they are applicable. For instance, they should be used at
the execution level of many autonomous controllers. We propose to augment and
enhance existing theories rather than develop a completely new theory for the hybrid
systems described above; we wish to build upon existing, well understood and proven
conventional methods. The symbolic/numeric interface is a very important issue;
consequently it should be included in any analysis. There is a need for systematically
generating less detailed, more abstract models from differential/difference equation
models to be used in higher levels of the autonomous controller (coordination level).
There is also a need for systematically extracting the necessary information from
lower level symbolic models to generate higher level symbolic models to be used in the
hierarchy where appropriate. Tools for the analysis of this information extraction also
need to be developed. Research in this area is underway. In this way conventional
analysis can be used in conjunction with the developed analysis methods to obtain an
overall quantitative, systematic analysis paradigm for intelligent autonomous control
systems. In short, we propose to use hybrid modelling, analysis, and design techniques
for nonuniform systems. This approach is not unlike the approaches used in the study
of any complex phenomena by the scientific and engineering communities.
A practical but very important issue is the simulation of hybrid systems. This
requires simulation of both conventional differential equations and symbolic decision
338 P.J. ANTSAKLIS ET AL.

making processes. Normally, numeric-algorithmic processing is done with languages

like FORTRAN and symbolic decision making can be implemented with LISP or
PROLOG. Sometimes the two types of processing are done on computers with quite
different architectures. There is then the problem of combining symbolic and numeric
processing on one computer. If the computing is done on separate computers, the
communication link normally presents a serious bottleneck. Combining A! and
conventional numeric processing is currently being addressed by many researchers
and some promising results have been reported in [31] and [14].
It was pointed out in Section 2 that complex control problems required a controller
sophistication that involved the use of AI methodologies. It is interesting to observe
the following [35]: Although there are characteristics which separate intelligent from
nonintelligent systems, as intelligent systems evolve, the distinction becomes less clear.
Systems which were originally considered intelligent evolve to gain more character of
what are considered to be non-intelligent, numeric-algorithmic systems. An example
is a route planner. Although there are AI route planning systems, as problems like
route planning become better understood, more conventional numeric-algorithmic
solutions are developed. The AI methods which are used in intelligent systems, help
us to understand complex problems so we can organize and synthesize new approaches
to problem solving, in addition to being problem solving techniques themselves. AI
techniques can be viewed as research vehicles for solving very complex problems. As
the problem solution develops, purely algorithmic approaches, which have desirable
implementation characteristics, substitute AI techniques and play a greater role in the
solution of the problem. It is for this reason that we concentrate on achieving
autonomy and not on whether the underlying system can be considered 'intelligent'.

5. Concluding Remarks
A hierarchical functional autonomous controller architecture was introduced. In
particular, the architecture for the control of future space vehicles was described in
detail; it was designed to ensure the autonomous operation of the control system and
it allowed interaction with the pilot and crew/ground station, and the systems on
board the autonomous vehicle. The fundamental issues in autonomous control system
modelling and analysis were discussed. It was proposed to utilize a hybrid approach
to modelling and analysis of autonomous systems. This will incorporate conventional
control methods based on differential equations and new techniques for the analysis
of systems described with a symbolic formalism. In this way, the well developed
theory of conventional control can be fully utilized. It should be stressed that auto-
nomy is the design requirement and intelligent control methods appear, at present, to
offer some of the necessary tools to achieve autonomy. A conventional approach may
evolve and replace some or all of the 'intelligent' functions. Note that this paper is
based on the results presented in [3].
It was shown that in addition to conventional controllers, the autonomous control
system incorporates planning, learning, and FDI. An initial study of the FDI problem
TOWARDS INTELLIGENT AUTONOMOUS CONTROL SYSTEMS 339
incorporating both conventional and AI FDI techniques was reported in [45]. Fur-
thermore, AI planning systems were modelled and analyzed in a Petri net framework
in [46].
It must be stressed that the results presented here apply to any autonomous control
system. For other applications, the architecture, or parts of it, and the ideas discussed
here are valid. For instance, to achieve a certain level of autonomy for a particular
application one may modify the functional architecture by removing the management
and organization level. In this case, the limited version of the autonomous controller
would not provide for a user interface, goal generation, high level learning, etc. In
general, modifying the controller for certain applications entails the removal of
portions of the functional architecture which limits the attainable degree of auto-
nomy. Hence, to use the above results for a different application one must decide what
level of autonomy is needed and then include in the autonomous controller architec-
ture those components necessary to achieve it.

Acknowledgement
This work was partially supported by the Jet Propulsion Laboratory, Pasadena,
California.

References
1. Albus, J., et al., Theory and practice of intelligent control, Proc. 23rd IEEE COMPCON, pp. 19-39
(1981).
2. Albus, J.S., et al., Hierarchical control of intelligent machines applied to space station telerobotics,
Proc. Space Telerobotics Workshop, pp. 155-166 (1988).
3. Antsaklis, P.J. and Passino, K.M., Autonomous control systems: Architecture and concepts for future
space vehicles, Final report, Jet Propulsion Laboratory Contract, Oct. 1987.
4. Antsaklis, P.J., Passino, K.M., and Wang, S.J., Autonomous control systems: Architecture and
fundamental issues, Proc. Amer. Control Conference, Atlanta, pp. 602-607 (1988).
5. Astrom, K.J., et al., Expert control, Automatica 22, 277-286 (1986).
6. Atkinson, D.J., Telerobot task planning and reasoning: Introduction to JPL AI research, Proc. Space
Telerobotics Workshop, pp. 339-350 (1988).
7. Balaram, J., et aL, Run time control architecture for the JPL telerobot, Proc. Space Telerobotics
Workshop, pp. 211-222, (1987).
8. Bhatt, R., et al., A real time pilot for an autonomous robot, Proc. IEEE Internat. Syrup. Intelligent
Control, pp. 135-139 (1987).
9. Blank, G., Responsive system control using register vector grammar, Proc. 1EEE Internat. Symp.
Intelligent Control, pp. 461-466 (1987).
10. Charniak, E. and McDermott, D., Introduction to Artificial Intelligence, Addison Wesley, Reading,
Mass. (1985).
1I. Crosscope, J. and Bonnell, R., An integrated intelligent controller employing both conceptual and
procedural knowledge, Proc. IEEE Internat. Symp. Intelligent Control, pp. 416-422 (1987).
12. Cruz, J.B. and Stubberud, A.R., Knowledge based approach to multiple control coordination in
complex systems, Proc. 1EEE lnternat. Syrup. Intelligent Control, pp. 50-53 (1987).
13. DeJong, K., Intelligent control: Integrating AI and control theory, Proc. IEEE Trends and Applications
1983, pp. 158-161 (1983).
14. Despain, A.M. and Patt, Y.N., Aquarius - a high performance computing system for sysmbolic/
numeric applications, Proc. COMPCON S'85, pp. 376-382 (1985).
340 P.J. ANTSAKLIS ET AL.
15. Dudziak, M.J., et aL, IVC: An intelligent vehicle controller with real-time strategic replanning, Proc.
IEEE lnternat. Syrup. Intelligent Control pp. 145-I52 (1987).
16. Dudziak, M.J., SOLON: An autonomous vehicle mission planner, Proc. Space Telerobotics Workshop,
pp. 289-302 (1987).
17. Farsaie, A., et al., Intelligent controllers for an autonomous system, Proc. 1EEE Internat. Symp.
Intelligent Control, pp. 154--158 (1987).
18. Findeisen, W., et al., Control and Coordination in Hierarchical Systems, Wiley, New York (1980).
19. Fiorio, G., Integration of multi-hierarchy control architectures for complex systems, Proc. IEEE
Internat. Syrup. Intelligent Control pp. 71-79 (1987).
20. Firschein, O., et al., Artificial Intelligence for Space Station Automation, Noyes, New Jersey (1986).
21. Freidland, P. and Lum, H., Building intelligent systems: Artificial intelligence research at NASA Ames
Research Center, Proc. Space Telerobotics Workshop, pp. 19-26 (1988).
22. Fu, K.S., Learning control systems and intelligent control systems: An intersection of artificial intelli-
gence and automatic control, IEEE Trans. Automatic Control pp. 70-72 (1971).
23. Gartrell, C.F., et aL, The use of expert systems for adaptive control of large space stuctures, Proc.
AIAA Guidance Navigation and Control Conf., pp. 376-385 (1985).
24. Gevarter, W.B., Artificial Intelligence, Noyes, NJ (1984).
25. Graglia, P. and Meystel, A., Planning minimum time trajectory in the traversability space of a robot,
Proc. IEEE lnternat. Symp. Intelligent Control pp. 82-87 (1987).
26. Guha, A. and Dudziak, M., Knowledge based controllers for autonomous system, Proc. IEEE Work-
shop Intelligent Control pp. 134-138 (1985).
27. Handelman, D.A. and Stengel, R.F., Combining qualitative and quantitative reasoning in aircraft
failure diagnosis, AIAA Guidance Navigation and Control Conf., pp. 366-375 (1985).
28. Hawker, J. and Nagel, R., World models in intelligent control systems, Proc. IEEE lnternat. Syrup.
Intelligent Control pp. 482-488 (1987).
29. Hodgson, J., Structures for intelligent systems, Proc. IEEE Internat. Syrup. Intelligent Control
pp. 348-353 (1987).
30. Jenkins, L., Space telerobotic systems: Applications and concepts, Proc. Space Telerobotics Workshop,
pp. 29-34 (1988).
31. Kitzmiller, C.T. and Kowalik, J.S., Coupling symbolic and numeric computing in KB systems, AI
Magazine 18, 85-90 (1987).
32. Knight, J.F. and Passino, K.M., Decidability for Temporal Logic in Control Theory, Proc. 25th
Allerton Conf., pp. 335-344 (1987).
33. Krogh, B., Controlled Petri nets and maximally permissive feedback logic, Proc. 25th Allerton Conf.,
pp. 317-326 (1987).
34. Lutz, P., Autonomous mobile robots in industrial production environment, in L.O. Hertzgerger and
F.C.A. Groen (eds.) Intelligent Autonomous Systems, North-Holland, NY (1987) (Proc. of an Interna-
tional Conference in Amsterdam, Dec. 1986.)
35. Mendel, J. and Zapalac, J., The application of techniques of artificial intelligence to control system
design, in Advances in Control Systems, C.T. Leondes (ed.), Academic Press, NY (1968).
36. Mesarovic, M., Macko, D. and Takahara, Y., Theory of Hierarchical, Multilevel, Systems, Academic
Press, NY (1970).
37. Meystel, A., Intelligent control: Issues and perspectives, Proc. IEEE Workshop Intelligent Control
pp. 1-15 (1985).
38. Meystel, A., Nested hierarchical controller for intelligent mobile autonomous system, in L.O. Hertz-
gerger and F.C.A. Groen (eds.), Intelligent Autonomous Systems, North Holland, NY (1987) (Proc. of
an International Conference in Amsterdam, Dec. 1986.)
39. Meystel, A., Planning/control architectures for master dependent autonomous systems with non-
homogeneous knowledge representation, Proc. IEEE Internat. Symp. Intelligent Control, pp. 31-41
(1987).
40. Meystel, A., Nested hierarchical controller with partial autonomy, Proc. Space Telerobotics Workshop,
pp. 251-270 (1988).
41. Moizer, A. and Pagurek, B., An onboard navigation system for autonomous underwater vehicles, in
L.O. Hertzgerger and F.C.A. Groen (eds.), Intelligent Autonomous Systems, North Holland, NY
(1987) (Proc. of an International Conference in Amsterdam, Dec. 1986.)
TOWARDS INTELLIGENT AUTONOMOUS CONTROL SYSTEMS 341
42. Ostroff, J.S., Real time computer control of discrete systems modelled by extended state machines: A
temporal logic approach, PhD dissertation, Report No. 8618, Dept. of Elect. Eng., University of
Toronto, Jan. 1987.
43. Passino, K.M., Restructurable controls and artificial intelligence, McDonnell Aircraft Internal
Report, IR-0392, April 1986.
44. Passino, K.M. and Antsaklis, P.J., Restructurable controls study: An artificial intelligence approach
to the fault detection and identification problem, Final Report, McDonnell Douglas Contract, Oct.
1986.
45. Passino, K.M. and Antsaklis, P.J., Fault detection and identification in an intelligent restructurable
controller, Journal of Intelligent and Robotic Systems 1, 145-161 (1988).
46. Passino, K.M. and Antsaklis, P.J., Artificial intelligence planning problems in a Petri net framework,
Proc. Amer. Control Conference, pp. 626-631 (1988).
47. Pearson, G., Mission planning for autonomous systems, Proc. Space Telerobotics Workshop, pp. 303-
306 (1988).
48. Raulefs, P. and Thorndyke, P.W., An architecture for heuristic control of real time processes, Proc.
Space Telerobotics Workshop, pp. 139-148 (1988).
49. Saridis, G.N., Toward the realization of intelligent controls, Proc. 1EEE 67, I 115-1133 (1979).
50. Saridis, G., Intelligent controls for advanced automated processes, Proc. Automated Decision Making
and Problem Solving Conf., NASA CP-2180, May 1980.
51. Saridis, G.N., Intelligent robot control, 1EEE Trans. Automatic Control, AC-28, 547-556 (1983).
52. Saridis, G.N., Foundations of the theory of intelligent controls, Proc. IEEE Workshop Intelligent
Control, pp. 23-28 (1985).
53. Saridis, G.N., Knowledge implementation: structures of intelligent control systems, Proc. lEEElnter-
nat. Symp. Intelligent Control, pp. 9-17 (1987).
54. Saridis, G.N. and Valavanis, K.P., Software and hardware for intelligent robots, Proc. Space
Telerobotics Workshop, pp. 241-250 (1988).
55. Schenker, P.S., Program objeclives and technology outreach, Proc. Space Telerobotics Workshop,
pp. 3-18 (1988).
56. Shapiro, S.C. (ed.), Encyclopedia of Artificial Intelligence, Wiley, NY (1987).
57. Stark, L., Telerobotics: Research needs for evolving space stations, Proc. Space Telerobotics Work-
shop, pp. 91-94 (1988).
58. Stengel, R.F., AI theory and reconfigurable flight control systems, Princeton University Report
1664-MAE, June 1984 (see also 1665 by D. Handelman).
59. Stephanou, H.E., An evidential framework for intelligent control, Proc. 1EEE Workshop Intelligent
Control, pp. 118-123 (1985).
60. Thistle, J.G. and Wonham, W.M., Control problems in a temporal logic framework, Int. J. Control
44, 943-976 (1986).
61. Trankle, T.L., Sheu, P. and Rabin, U.H., Expert system architecture for control system design, Proc.
Amer. Control Conference, pp. 1163-1169 (1986).
62. Turner, P.R., et al., Autonomous systems: Architecture and implementation, Jet Propulsion Labora-
tories, Report No. JPL D-1656, August 1984.
63. Valavanis, K.P., A mathematical formulation for the analytical design of intelligent machines. PhD
Dissertation, Electrical and Computer Engineering Dept., Rensselaer Polytechnic Institute, Troy NY,
Nov. 1986,
64. Valvanis, K.P. and Saridis, G.N., Architectural models for intelligent machines, Proc. 251h Conf.
Decision and Control, Athens, Greece (1986).
65. Valavanis, K.P. and Saridis, G.N., Information theoretic modelling of intelligent robotic systems, Part
I: The organization level, Proc. 261h Conf. Decision and Control, Los Angeles, pp. 619-626 (1987).
66. Valavanis, K.P. and Saridis, G.N., Information theoretic modelling of intelligent robotic systems,
Part II: The coordination and execution levels, Proc. 261h Conf. Decision and Control, Los Angeles,
pp. 627-633 (1987).
67. Villa, A., Hybrid knowledge based/analytical control of uncertain systems, Proc. 1EEE lnternat. Syrup.
Intelligent Control, pp~ 59-70 (1987).
68. Waldon, S., et al., Updating and organizing world knowledge for an autonomous control syslem, Proc.
1EEE lnternat. Syrup. lntelligenl Control pp. 423-430 (I987).
342 P.J. ANTSAKLIS ET AL.
69. Wolfe, W.J. and Raney, S.D., Distributed intelligence for supervisory control, Proc. Space Telero-
botics Workshop, pp. 139-148 (1988).
70. Wos, L., Automated Reasoning: 33 Basic Research Problems, Prentice Hall, NJ (1988).
71. Zadeh, L.A., Fuzzy logic, Computer, April 1988, 83-93.
72. Zeigler, B.P., Knowledge representation from Newton to Minsky and beyond, Journal of Applied
Artificial Intelligence 1, 87-107 (1987).