Integrated Control Methodologies for Road Vehicles

TIM GORDON, MARK HOWELL

AND FELIPE BRANDAO *

SUMMARY

This paper considers the scope, methodologies and architectures for the design and development of
interacting control systems in road vehicles. The increasing use of electronic controls leads inevitably to
an increase in overall system complexity. Given the time and economic constraints of the modern
automotive industry, it is not feasible to synthesise and validate the full set of vehicle controls in the form
of a unified and centralized controller. On the other hand a fully decentralized approach to control system
development and operation will induce performance limitations from un-modelled or unexpected
interactions; at worst, such interactions can cause instability and loss of function. There is now increasing
pressure to achieve control coordination whilst maintaining a modular approach to the overall system
design. With this in mind, the paper provides a framework to review current practice in integrated vehicle
control, assesses recent developments in control integration methodologies that are most relevant to the
vehicle application, and formulates an enhanced multi-layer architecture that includes explicit coordination
functionality. Overall emphasis is placed on the role of control system architecture, the resulting flow of
control information and the implications for control system design. An example from handling dynamics is
presented, demonstrating the viability of new and flexible approaches. In conclusion a number of
outstanding research problems are highlighted.

1. INTRODUCTION

Any control system comprises hardware such as sensors, actuators, communications

links, power electronics, switches and micro-processors. Such hardware is becoming
increasingly common on modern motor vehicles, with 30-50 microprocessors not
uncommon on a single vehicle [1]. Hardware is typically distributed around the
vehicle, but of course the physical location tells us relatively little about the design of
the functional control structure, which is the main topic of this paper.
In the past it has been common for separate vehicle functions to be controlled
‘independently’, or rather in parallel. This means that control hardware can be grouped
into discrete subsets, with sensor information and control demands operating in parallel
processes and with no possible ambiguity or conflict over the responses demanded of
the actuators. Of course interactions may occur, when for example brake and rear-steer

*
Address for correspondence: Department of Aeronautical and Automotive Engineering,
Loughborough University, Loughborough, Leicestershire, LE11 3TU, England. Tel +44 1509 227201,
email [Link]@[Link]

1
 T.J. GORDON ET AL.

actuators control the same (yaw) degree of freedom of the vehicle. We shall refer to
this as a parallel or decentralised control architecture; the term architecture will be
taken to mean an abstraction of the pattern (or topology) of how sensor information and
control commands interact between various control sub-systems and components.
Indeed we shall seek to resolve the term architecture into more detail as it becomes
clearer that the control structure operates at a number of levels; for example an
important aspect is the pattern of sub-system functionality (the assignment of control
objectives to sub-systems) designed into the vehicle control system. This makes the
overall control architecture a deeper concept that simply the physical layout of control
components, or even the causal relationship between distributed control processes.
The parallel vehicle control architecture has largely arisen by default, as different
controlled sub-systems are developed and manufactured by independent supplier
companies, or by different groups within a vehicle manufacturer. But there is a very
limited number of relevant degrees of freedom in a typical car: for motion control there
are six rigid-body modes, and even adding the major system states of engine,
transmission, wheel spin and wheel-hop, the number is still only around 20. So as the
number of control functions and actuators increases, it is inevitable that interactions and
performance conflicts arise within any highly parallel architecture. For example, the
use of single-wheel braking to reduce oversteer or understeer will certainly conflict
with the requirement for traction under forward acceleration. The unsatisfactory nature
of this has been recognised for some time (e.g. [2-5]) and yet the overall issue remains
[6]. The foremost concerns are to do with reducing complexity, improving
performance and removing unnecessary and costly duplication of hardware [6].
Coelingh [6] presents an illustration similar to that in Fig. 1, illustrating the way that a
‘Complete Vehicle Control’ structure might limit the growth in overall system
complexity as the number of actuators and sensors increases. As a global structure is
imposed, complexity is reduced, at least in the physical structure of the control system;
however as is also suggested by Fig. 1. the complexity may simply be masked,
preserved intact within the software.

Driver Driver

Actuator Actuator Sensor

Actuator Actuator Sensor

Fig. 1. Unstructured vs. structured control architectures

2
 INTEGRATED CONTROL OF ROAD VEHICLES

Overall, in the development of an Integrated Vehicle Control System (IVCS), the aim is
to combine and supervise all controllable subsystems affecting vehicle dynamic
response [2-4,7-9] to: improve multiple-objective performance from available
actuators, by removing the implicit constraint of a one-to-one link between control
objectives and particular sets of sensors and actuators
• reduce or contain complexity at the number of control components increases
• improve safety and comfort
• reduce or contain system costs, by avoiding unnecessary duplication of
components and by information sharing between sensors.
• improve flexibility
- permitting a modular and distributed design process
- moving towards ‘plug-and-play’ extensibility
- incorporating diagnostics and condition monitoring
- improving system reliability
- allowing fault-tolerant and failure management capability

There is a close link between flexible plug-and-play design, and fault-tolerant

control system operation; given the availability of suitable diagnostics, condition
monitoring, and adaptability in the control algorithms, should a sensor or actuator fail,
the failed component may be de-activated and reconfiguration take place to minimise
degradation, or at least provide a safe and predictable failure mode for the vehicle. In
this way, it should be possible to generate redundancy in a distributed way at the system
level, which is cheaper and more effective than simply duplicating key components.
One particular algorithmic device to achieve such behaviour, based on formal optimal
control, has been presented by one of the authors [10] but there appears to be scope for
significant future developments in this area – see Section 6 below.
Further aspects of IVCS flexibility are described in the discussion of integrated
control for helicopter dynamics by Wills [Link]. [11]. In particular, this paper highlights
the importance of interoperability and openness. The former is a relatively low-level
matter dealing with the need for effective communication between different
microprocessors (e.g. using a CAN bus). This is an essential aspect of any modular
approach to system design - rigorous interfacing standards must be applied for
anything to actually work! On the other hand, openness refers to the need for software,
algorithms and their resources to be sufficiently open to allow for systematic
integration. This is a key point, especially for automotive applications, since in the
foreseeable future, proprietary control algorithms will not be freely made available; but
this need not prevent an open architecture operating between IVCS modules.
Flexibility also encompasses the need for a modular approach to design and
manufacture. Vehicles are not designed and built in isolation; if a manufacturer sells
five different vehicles, each with five optional control sub-systems that may or not be
implemented, and each employs alternative hardware from two different suppliers,
there are 5 × 35 =1215 possible combinations, each potentially requiring a unique
variant of an overall vehicle control algorithm; while these numbers are hypothetical,

3
 T.J. GORDON ET AL.

the problem is not. If the controlling software is not modular, design and development
becomes unmanageable.
These issues provide the motivation for this paper: to clarify the conceptual basis
and future challenges of integrated vehicle control. To achieve this, we review the
current state of the art in Integrated Vehicle Control Systems, with a critical emphasis
on functional control structures and operational architectures. The focus is on the area
of vehicle motion control, with particular emphasis on chassis systems. The aim is to
highlight and demonstrate the major options available, and provide a degree of
synthesis of ideas – hopefully to move towards a clear and consistent approach to
identify suitable IVCS architectures, that not only encompass reported work to date, but
also suggests new possibilities for the future.
As a self-imposed limitation, the important areas of integrated sensor fusion, state
estimation and information management will not be considered in any detail; it will
mostly be assumed that the various vehicle or sub-system controllers have direct access
to sufficient sensor or state information for control decisions to be made. This stretches
the normal use of the well-known ‘Separation Principle’ whereby control and
estimation functions may be developed independently. It is recognised however that
issues of fault tolerance and overall system complexity cannot be fully addressed
without taking account of the sensor and information aspects. The concept of the
vehicle as a subsystem interacting with ‘intelligent’ external roadway and traffic
systems is also not considered in any detail (see the review in [12] for more details).
However it is clear that IVCS ‘expandability’ is an important issue when such external
interactions are introduced. For example [13] the external control of vehicle speed can
be beneficial in reducing congestion, and the automation of speed control for traffic
management, platooning etc. becomes technically and economically very feasible once
sufficient vehicles are fitted with drive-by-wire systems; the global control architecture
must be compatible with that of the individual vehicles, and the various (local and
global) fault/failure-mode behaviours may become a critically important
implementation issue for the future.
In the next section we review the state of the art of IVCS application areas, in
particular to understand where vehicle control integration offers significant
performance benefits, and also which general control algorithms have been used to
support such integration. In Section 3 some basic architectures are considered, as are
the levels at which these operate. Section 4 deals with multi-layered ‘behaviour based’
architectures in more detail, while Section 5 looks more deeply at the coordination
mechanisms available for control integration. This is followed by an illustrative
simulation study for handling control, using layered design and combined
fuzzy/subsumption coordination. Section 6 looks at the scope for extending the
capabilities of coordinators so that they are more systematically linked to the
performance criteria of the associated sub-system controllers. Finally Section 7 brings
together the main conclusions of the paper, and highlights a number of areas for future
research.

4
 INTEGRATED CONTROL OF ROAD VEHICLES

2. INTEGRATED VEHICLE CONTROL APPLICATIONS

The fundamental objective of vehicle motion control is to direct the vehicle, according
to driver demands, in the longitudinal and lateral directions whilst maintaining
acceptable variations in bounce, pitch, roll and vehicle slip angle. Although body roll is
intimately linked with handling performance, it is largely a ‘means to an end’ (driver
feedback, bump-steer etc.) rather than a primary goal of vehicle dynamics. The aim is
to provide safe, predictable and responsive motion control, whilst maintaining
acceptable occupant ride comfort. NVH, energy consumption and similar issues are
clearly important, but again play a secondary role in the current discussion. In
principle, suspension control is also secondary - the driver has no direct influence over
this sub-system; however, in practice, its direct influence on vertical tyre loads and the
vertical, pitch and roll degrees of freedom clearly give it a very important role, and it is
normally treated on an equal footing to braking and yaw-moment control for example.
Wallentowitz [4] considered cross-compatibility of sensors and processor operations
in current and potential future vehicles. In this broad review, the case has been made
for a number of integration improvements for chassis-specific functions such as ABS,
stability and traction control under acceleration, active rear-axle kinematics, all-wheel
drive and lateral dynamic stability, as well as active ride dynamics. The paper makes
the case based largely on sensor-sharing and sensor redundancy issues, but also
describes the significant problem of ‘ownership and distribution of IPR’ mentioned in
Section 1.
In another paper on the ‘basic principles’ of integrated control, Kiencke [5] gives a
general overview on how the introduction of Control Area Networking (CAN) provides
convenient scope for the integration of previously independent systems. The work
reinforces the above idea that it is fundamentally the small number of vehicle rigid-
body degrees of freedom compared to available actuators that drives the performance
advantages of IVCS; Kiencke focuses on the combined use of brakes and rear-steer to
augment the driver’s front-steer input in controlling the yaw (and implicitly sideslip)
degree of freedom. A sliding mode algorithm, which involves the real-time estimation
of tyre friction, is described for the integrated control.
To better understand the overall potential performance benefits of integrated motion
control of the vehicle, the basic task can be represented via a ‘g-g diagram’ of vehicle
mass centre lateral and longitudinal accelerations. The driver uses accelerator, brakes
and steering to maintain control of the magnitude and direction of the mass-centre
velocity vector; changes in this vector are then the g-g diagram accelerations. In terms
of the ‘ideal function’ of motion control, the driver is limited only by the friction
constraints of the tyres - plus of course in-plane aerodynamic forces when these become
significant at high speeds. The vehicle controls are expected to provide the driver with
predictable authority over these accelerations, within the physical constraints of the
vehicle ‘friction circle’, and subject to perceived customer acceptability of the
frequency and amplitude dependence of the vehicle responses. Also, given that friction

5
 T.J. GORDON ET AL.

limits change with speed and road surface condition etc., these vehicle control systems
are required to provide adequate feedback of such changes - for example via steering
torque. The concept in Fig. 2 was presented by Tanaka et al [3] as a schematic to
indicate the domain of operation of some typical vehicle control systems, and the areas
where system integration are likely to be beneficial. Although the diagram should not
be taken too literally, it clearly underlines the conclusion from the Toyota Soarer that
integrated control can enlarge and ‘smooth out’ the dynamic response domain of the
vehicle.

Acceleration
TRC + Active 4WS
Tyre Traction Limit

With Integrated Control

Lateral Acceleration

Without Integrated Control

Active Suspension + Active 4WS

ABS + Active 4WS
Deceleration

Fig. 2. Integrated Control and the g-g Diagram

The Toyota Soarer employed combined control of hydro-pneumatic suspension, four

wheel steer and active ABS / traction control [3,14]. Making use of classical control
techniques, with three term (PID) feedback for each subsystem, control integration was
based on logical mode detection (for example at preset thresholds of longitudinal and
lateral accelerations) to regulate controller gains at a supervisory level. The published
results epitomise the degree of control improvement that can be achieved, even without
any recourse to any particular formal or multivariable control algorithms.
In a later paper from Toyota, Hirano et al [7] advocated a similar ethos,
implementing a four wheel steer/ four wheel drive (4WS/4WD) controller via
feedforward and feedback compensators designed using multivariable H∞ methods;
these are suitably inter-related between each subsystem, and adapt for saturating tyre
forces. The authors report improved vehicle stability on slippery surfaces and an
improved steering response in both simulation and experiment. The experimental
vehicle employed Local Area Network (LAN) communications to link a central control
unit to a distributed set of five local control units for 4WS (for the rear steer actuator)
4WD (for actuating drive torque distribution), ABS, engine control and electronic
throttle. Though only 4WS/4WD were explicitly integrated within the control

6
 INTEGRATED CONTROL OF ROAD VEHICLES

algorithm, this paper indicates a physical architecture that is relatively fault-tolerant: if

there is an interruption or failure of the central controller or the LAN, the local control
units can continue to provide a safe basic level of functionality. Note however that this
architecture is not explicitly reflected in the controller design methodology.
Given the consistent theme of performance improvement by integrated (or simply
multivariable) control, it is worth focussing further attention on algorithms employed,
and the extent to which these can support a modular and distributed architecture.
A number of authors have approached the design of integrated vehicle controllers
via a full-vehicle reference model. The dynamic influence of various actuators are
incorporated via input matrices in whole-vehicle MIMO model, which are typically of
low order. These are applied in an off-line control law design, resulting in improved
vehicle performance, for example in handling stability, when compared with the use of
independent controllers. Examples include model matched-control using a model-
inversion [15,16], robust H2 and H∞ design methods [17], and nonlinear predictive
control [18]. Whilst the these papers are purely based on simulation investigations,
there is a consistent (and perhaps obvious) conclusion that a properly designed
integrating controller improves handling performance over a wide range of
manoeuvres, strongly endorsing the conceptual g-g diagram of Tanaka et al.
Other authors have employed standard techniques such as direct output feedback
methods (single-loop analogue or discrete-time compensators) [3,7,19,20], sliding
mode control [5,21-23], model reference [16,24], fuzzy logic [25-27] and Artificial
Neural Networks [28,29]. Two particularly common formal control methods, both
essentially based on linear systems are those of robust H2 and H∞ control [7,15,17,30]
and optimal control [10,15,31]. These more ‘popular’ approaches are attractive from an
academic standpoint in that they can be easily related to commonly available linear and
nonlinear handling dynamic models. The robust algorithms use linear reference
models, with uncertainty bounds related to expected errors due to nonlinearity. The
optimal control papers typically use a linearised model to design a feedback law, and
test this in simulation of a nonlinear model, either directly, or using feedback signals
derived by reference to a nonlinear model with certain ‘desired’ handling
characteristics. Both of the more common methods are also attractive for further
development, because they offer a clearly ‘visible’ design technique well-suited to
calibration and development via explicit performance indices or frequency weighting
functions.
When focussing attention on control algorithms, it is easy to make the mistake of
identifying ‘integrated control’ with ‘multivariable’ control, hence overlooking the
many issues highlighted already in Section 1, especially those relating to control system
flexibility. In fact, most of the authors cited in the above fall into this category, the
exceptions being references [10,28]. In [10] the emphasis is on modular/distributed
control system design, with parallel development of subsystem model-based optimal
controls. Synthesis into a coordinated vehicle controller is then centralised, so there are
essentially two levels to the control hierarchy. In [28] a model-based control is
explicitly introduced to control the six rigid-body degrees of freedom, potentially

7
 T.J. GORDON ET AL.

including driver inputs for longitudinal and lateral control, or – as particularly described
in the paper – to integrate the vehicle motion within an automated vehicle highway
system. Lower level ‘neuro-compensator’ controls operate to help ensure the overall
vehicle behaves like the reference design model, and hence improve robustness. This
two-level hierarchical structure could clearly be developed with further levels of
cascaded control. These two approaches, based of specific algorithms, neatly highlight
the broad links between algorithmic structure and control architecture for integrated
controls; in these particular cases, the first defines a bottom-up distributed design
approach, while the second follows a top-down hierarchical structure.
Links between algorithms and supporting architectures are also to be found in
related application areas – as in unmanned air vehicles (e.g. [11,32]) and autonomous
underwater vehicles (e.g. [30,33]). In references [11,32] a structured hierarchical
structure is employed, and in both cases the key algorithmic approach to achieving
flexible control is via mode recognition and switching, with baseline control algorithms
designed off-line, the details of which are largely irrelevant, but might typically be
based on single-loop classical control, or multivariable ‘modern’ control. In [30] this
same approach to mode-switching is proposed, though the baseline algorithms are
specifically prescribed as robust multivariable H∞ control. A limitation with these
mode-switching algorithms is that the design stage must be used to pre-empt and
predict all likely failure events or changes in operating mode, something that greatly
increases the off-line design task and is particularly undesirable in the automotive
systems environment. In [33] a quite different approach is taken, with a number of
bottom-up behaviour-based approaches being considered. Key features of these
approaches will be taken up in the following sections, once the essential elements of
vehicle control architectures have been reviewed and refined.

3. VEHICLE CONTROL ARCHITECTURES

The term ‘architecture’ is widely used [2,6,8,11,12,34] but often loosely so, and with
different specific meanings by different authors. Here the intended meaning is as a
comprehensive representation of the global IVCS structure, both in its operation and its
design. The term therefore encompasses a range of relationships, or ‘topologies’,
connecting various sub-systems or modules. In increasing order of abstraction, there
appear to be five potentially distinct levels of topology that comprise the overall control
architecture:-
T1. the physical layout of microcontrollers, routers, communication links, sensors,
and actuators
T2. the causal connection and relative authority of control actions – incorporating
the connectivity of demands, reference signals and any coordination and
protection mechanisms
T3. the connection and flow of information – from sensors, state estimators, control
output, condition monitoring and diagnostics

8
 INTEGRATED CONTROL OF ROAD VEHICLES

T4. the structure of the control algorithms and methodologies that underpin the
above, for example that help guarantee stable and fault-tolerant operation of the
overall IVCS
T5. the underlying structure of the control functional design - as for example in
hierarchical or layered approaches where a modular design is used to cascade
or subsume control objectives

Again, the term ‘topology’ will be assumed to have a wider scope than is common,
not only to include the ‘geometric’ topology of network nodes and links, but also
relevant aspect of causality (A triggers B), ranking (A overrides B), and dependency (A
requires B in order to function).
Some authors refer to the above T2-topology as a ‘functional decomposition’ [6,35]
of the integrated control system, emphasising the difference between T1 and T2.
However this can easily be confused with T5, where functionality is also decomposed
into modules in the design process. Regarding T3, there is relatively little attention
given to this in the literature, except in the aspects of fault diagnosis and condition
monitoring [36-38].
The topology at one level does not uniquely constrain the topology at other levels.
For example, consider the traditional parallel/distributed approach seen in non-
integrated vehicle control systems. A parallel structure exists at all levels, except in T5
where a vehicle manufacturer typically specifies a hierarchy of functional objectives
and quantified performance targets. As a move towards ‘control system integration’ the
T1-topology may be changed from using separate wires for communication, to a
networked ‘architecture’ (topology) – as in Fig. 3. Here the shaded bubbles represent
microcontrollers associated with specific control functions, and the lines represent
physical communication channels. The second T1-topology remains compatible with
the traditional parallel design, but offers new opportunities at other levels. This simple
example underlines the fact that there really are a multiplicity of levels at which control
system topologies operate, and there is not a unique association of compatible
topologies; it also serves to demonstrate the simplistic nature of using a figure like this
as an overall conceptual model for integrated vehicle control!

Driver Driver
Sensor Sensor Sensor Sensor

Vehicle network

Actuator Actuator
Actuator
Actuator

Fig. 3. Parallel versus network implementation of a parallel design topology

9
 T.J. GORDON ET AL.

Basic issues of the physical topology are also highlighted in the paper by Schmidt et al
[36]; by drawing sub-system boundaries in places where the communication traffic is
low, the effect of an unavailable link can be minimised. Once again, there is clearly no
essential link from this to the supporting control algorithms (T4), or functional
decomposition in design (T5).
Following the self-imposed limitations described in the Introduction, focus will
primarily be on the topology levels T2 and T5, though these will be related to the other
levels where relevant. Indeed, in the future it may be that T4 will provide the ‘defining’
topology: without a proper mathematical formulation at the algorithmic level, it is hard
to see how key properties of stability, reliability etc. can be validated for an integrated
control system design.
Both T2 and T5 may be classified to some extent by the degree of centralisation.
The T5-topology of a centralised controller may look something like that on the left of
Fig. 4. This represents a total vehicle functional control hierarchy, with arrows
representing the cascade of control objectives from the vehicle level to supporting
control functions. A less centralised approach is represented on the right of the figure,
where some more restricted objectives give rise to a number of parallel hierarchies. In
all of these cases however the downwards arrows indicate that the lower level control
design cannot be carried out until the higher levels have been suitably defined.

Braking Objectives (Limit) Handling Objectives

W hole Vehicle Objectives

Wheel Speed Sensors Yaw Rate Sensor

Sensing and Information

Integrated Control ABS Controller 4WS Controller

Actuation : ABS, Rear Steer, Active Suspension, etc ABS Actuation 4WS Actuation

Fig. 4. Single versus parallel functional hierarchies

We now consider three common T2-topologies in decreasing order of centralisation,

and return to the major role of T5 in Section 4.

3.1 Centralised Control

Here a central ‘total vehicle controller’ is responsible for making all control decisions.
Therefore, as depicted in Fig. 5, control, sensor and diagnostic information is routed
vertically between the two layers of controller and vehicle hardware (actuators and
basic mechanical components). The horizontal arrows represent mechanical or other
physical interactions directly within the vehicle, the major aspects of which should be
modelled and accounted for in the controller design. While the flexibility limitations of

10
 INTEGRATED CONTROL OF ROAD VEHICLES

such centralised control for vehicles is already clear, there is the potential advantage
that stability and performance predictions can be made in a unified manner at the design
stage.

Central Multivariable Controller

Subsystem 1 Subsystem 2 Subsystem 3 Subsystem N

Vehicle Hardware

Fig. 5. Centralised control structure

There are many examples in the literature where integrated vehicle control is identified
with this architecture, typically linked to the application of a global multivariable
control formalism – [7,8,15-18,31,39-41]. When the T2-topology is implemented
directly into hardware (T1-topology) any desired fail-safe redundancy of
microcontrollers, power converters etc. increases the number and cost of control
components [42]. The option preferred by Kelling and Heck [42], in the context of a
brake-by-wire system, is to distribute the brake ECU’s to the wheel locations. In the T4
sense, the control topology remains centralised, but each brake ECU calculates global
controls in parallel, and these are communicated between the ECU’s and a
validation/voting system provides supervision and failure mode capability. In the T2
(and T1) sense the control topology is actually in the form of supervisory control.

3.2 Supervisory Control

Supervisory control represents a range of topologies intermediate between fully
centralised and fully decentralised, and is established by adding an intermediate layer of
local control to the former, or adding a level of supervision to the latter – Fig. 6.
Subsystem controllers may then be designed and validated relatively independently.
One particular factor is whether the intermediate layer is fully dependent on the
supervisory layer – and hence whether effective control is possible in the event of a
communication failure between them. For example in the braking system described by
Kelling and Heck [42] not only is the latter possible, the (voting-based) supervisory
control function is physically distributed – there is no physical central control ‘module’
at all! On the other hand, in the system described in [6], while some lower level
functionality (sensor and sub-system diagnostics) is devolved to the intermediate layer,
the essential control task remains centralised, and the loss of communication envisaged
could not be tolerated by the IVCS.

11
 T.J. GORDON ET AL.

Supervisor or Master Controller

Controller 1 Controller 2 Controller 3 Controller N

Subsystem 1 Subsystem 2 Subsystem 3 Subsystem N

Vehicle Hardware

Fig. 6. Supervisory control

The role of the Master Controller may be to supply set-points within a familiar cascade
structure [2], or provide mode switching [30] or both [11]. In the ‘Open Control
Platform’ of Wills et al [11], the lowest layer incorporates actuators and low-level
control (e.g. set-point tracking) and the supervisory layer works at the ‘mission’ level,
including all sensor input, system condition monitoring, situational awareness etc. The
intermediate layer consists of ‘traditional’ multivariable controllers, e.g. for the
(helicopter) flight dynamics control. For this particular topology, the routing of all
sensor functions through the highest layer means that again, communications failure to
the supervisor (or a major processing failure within it) would cause overall system
dysfunction. As mentioned in Section 2, the same general approach is taken by Katebi
and Grimble [30] for underwater vehicle control. While it is clear that this type of T2-
topology supports a basic form of reconfigurable control, doing so in a more general
case, in response to non preset scenarios, is a challenging task requiring a degree of
‘intelligence’ not addressed in these references.
Returning to the control of road vehicles, in the work of Freuchte et al [2], relating
to General Motors’ “Project Trilby” on integrated vehicle control, the same general
three layer T2-topology is proposed, though in this case the supervisory layer is further
resolved into those aspects that adapt to changing situations within the vehicle, and
those that adapt to the driver; a third sub-layer relating to the external highway
environment would also seem a natural extension to these ideas. This work also
provides a reminder of the essential role of human factors issues in integrated vehicle
control, though in [2] much of the description is aspirational in nature.
From the above, it is clear that the three-layer supervisory (T2) control topology has
a number of attractive features:
• it can potentially operate even if the supervisor malfunctions or high-level
communications fail
• it can be extended to a multi-layered hierarchical structure (see Section 4)

12
 INTEGRATED CONTROL OF ROAD VEHICLES

• it generates a requirement for consistent interfaces and consistent design

function concepts that is reassuring to a conservative industry
• it is compatible with a modular approach to design and enjoys a close
relationship to a familiar supporting T5 topology – the hierarchical Systems
Engineering framework common in the automotive industry.

However at the T1 level, the topology tends to be expensive in terms of sensors,

communications and control hardware. More importantly though, there are alternative
topologies which enable a much greater degree of flexibility and fault tolerance.

3.3 Decentralised and Heterarchical Control

Controller 1 Controller 2 Controller 3 Controller N

Subsystem 1 Subsystem 2 Subsystem 3 Subsystem N

Vehicle Hardware

Fig. 7. Decentralised and heterarchical control

Decentralised control (Fig. 7) has an obvious parallel structure and is essentially the
basic non-integrated case of vehicle controllers (designed and) implemented for each
subsystem independently, and with no inter-system communication. Instead of adding
supervisory action, ‘integration’ may be achieved by the addition of communications
between the controllers, allowing each to interact with some or all of the others; this
gives rise to the so-called heterarchical architecture sometimes considered in the
flexible control of manufacturing systems - e.g.[43]. The obvious lack of global
structure would seem to make this approach wholly unsuitable for safety critical control
in road vehicles, and it does not seem to have been considered in this context at all. Of
course it might be that a suitably constrained algorithmic structure would make this T2
topology feasible in the future, but for all of its obvious flexibility, the lack of an
operational control structure makes it unlikely to be treated seriously.

4. MULTI-LAYER ARCHITECTURES

Here we consider a general layered architecture which is naturally abstracted from the
hierarchical two and three-layer control (T2) topologies of Section 3. This leads to a
single class of functional (T5) topologies, but which admits alternative structures in
terms of the flow of control actions (T2).

13
 T.J. GORDON ET AL.

Fig. 8 shows a number of levels or layers of vehicle control functionality or

‘behaviours’. Higher layers involve more integrated functionality, in other words a
more coordinated use of actuators. The bottom layer, Layer 0, represents the minimum
level of functionality for the vehicle to operate at all, whether or not any electronic
control systems (e.g. drive-by-wire) are included. Because, for simplicity, the sensor
and information structures are not being dealt with in any detail, the interface to such
systems is simply represented on the left side of the figure in terms of a multiplexed
‘bus’ or networked server from which information is interchanged. There are many
possibilities for how information is structured and exchanged, for example with basic
sensor signals being supplemented by state and disturbance estimates from dedicated
information processing systems, or via the control blocks shown, or both.

Layer 5 IVCS
6 DOF: roll, pitch, yaw, bounce, lateral and longitudinal motion
Overall Control
Vehicle Manufacture Settings
Vehicle Sensing ;Driver Status/ Intentions; Road, Environment and Other Vehicles

Layer 4 Collision Lane Detection

Behaviour Avoidance and Tracking
Sensor and Information Bus

Layer 3 Straight Safety

Neutral Parking
Line Distance
Behaviour Steer Aid
Stability from Obstacles

Layer 2
Anti - Dive Anti - Squash Yaw Stability Roll Stability Ride Comfort Load Transfer Forward Speed
Basic Behaviour

Integrated
Layers

Layer 1 Controllable Non Integrated

ABS TCS RWS
Basic Software
Suspension Layers (mostly
single loop)

Layer 0 Brakes Power Power Source Transmission Suspension Steering

‰ EMB Regeneration ‰ ‰ Automatic ‰ Active ‰ By - Wire
‰ ‰ ‰ ‰ ‰
Combustion Engine
‰
Hardware
EHB Semi-Automatic Semi-Active Power-Assisted
‰
HEV
CVT/IVT

Fig. 8. Multi-layered vehicle control design architecture

The blocks shown are not proposed as a definitive list of vehicle control functions, nor
as a IVCS standard; rather the figure illustrates the kind of increasing levels of
‘integration’ in the higher layers that might be implemented in practice, and it would be
for each vehicle manufacturer (or systems integrator) to define and develop such
functionality in any particular case. It is of course quite possible to extend the layering
upwards to include multi-vehicle system behaviour and interaction with other roadway
systems; e.g. in [44] such extensions are considered within a layered architecture, albeit
in the aerospace context.
Each block represents an individual ‘behaviour’ [45,46] that is to be integrated into
the overall control system operation. While in principle a ‘behaviour’ could be any

14
 INTEGRATED CONTROL OF ROAD VEHICLES

type of operational activity that has no obvious overall function (e.g. increase driver
braking demand signal), in any realistic application in IVCS, designed behaviours
would all have some desirable function. For example in Layer 1 ‘ABS’ could represent
the limited function of avoiding excessive longitudinal slip within braking a single
wheel by releasing or holding brake pressure at the wheel cylinder, whilst ‘Yaw
Stability’ in Layer 2 might represent the use of brakes to control excessive oversteer.
And it becomes a matter of design philosophy whether the control of understeer via
braking is treated as a separate behaviour, and similarly the use of drive-torque
distribution for yaw control might be treated as a separate behaviour.
This figure does not include any vertical arrows to define the flow of control
commands, and there are very many ways to achieve this. So Fig. 8 represents a
multiplicity of control topologies that potentially share a common layered structure,
rather than a single prescribed architecture. The essential feature is the system
modularity, and hence design flexibility – each block represents a ‘sub-system
controller’ that may be designed separately to achieve desired functional requirements.
Three associated control topologies are now considered. It is important to recognise
that these cases have implications not only for the connectivity and protocols that
determine the overall control authority (and perhaps the available classes of control
algorithm) but also the design methodologies that underpin them; these have significant
implications for broader aspects such as flexibility discussed above.

4.1 Hierarchical Control

In this top-down approach, the required vehicle level behaviour is initially defined,
broken down into sub-tasks, sub-sub tasks etc., in a traditional ‘Systems Engineering’
cascade. Higher layers are responsible for the overall goals and objectives of the
integrated vehicle system, while lower layers are responsible for solving the resulting
sub-problems, e.g. by tracking reference input signals. A serial structure is typical [6,
34,35,47,48] with communication only possible between adjacent layers, as in the
simplified Fig. 9.

Vehicle motion controller

driver
command
& ABS TCS ……………
vehicle
sensors

Brake controller Transmission …………..

Actuators

Fig. 9. Hierarchical Integrated Vehicle Control System

15
 T.J. GORDON ET AL.

The figure illustrates a feasible T2-topology, with the T5 structure still ‘visible’ in the
layering. The flow of control actions is top-down, with higher layers directly acting as
inputs – typically as reference signals – to lower layers. Strictly, all modules must be
operational for the IVCS to function, so if the physical (T1) topology is similar or
identical to the control topology, hardware redundancy becomes be necessary for all of
the modules shown. As has already been discussed, in a pure form the architecture is
inherently inflexible from the design perspective – for example the addition of a new
actuator potentially leads to the redesign of the entire IVCS.

4.2 Subsumption Architecture

The subsumption architecture [45,46] is derived via a bottom-up design methodology.
As originally proposed, behaviours (control functional objectives) operate in parallel
within any layer, and without higher level supervision. Higher layers intermittently take
over or ‘subsume’ control from lower layers, and hence there would be no single
coordinating vehicle level controller – different control functions, across all layers
operate asynchronously and in parallel. The layering simultaneously determines
authority over control actions (higher layers dominate), design sequencing (lower layers
first) and fault response (regress to lower level competency). The fundamental property
of this architecture is that higher layers subsume the actions of lower ones by applying
a bias to, or completely suppressing, lower level control commands. Lower levels
continues to function as higher levels are added, ‘unaware’ of any interference with
their activity. A feasible flow of commands is represented in Fig. 10, where the circular
nodes represent a switching function: when the higher layer activates a control signal it
predominates, replacing that of any lower layer(s).

Vehicle motion controller

driver
command
& ABS TCS
vehicle
sensors

Brake controller Transmission

Actuators

Fig. 10. Subsumption based control topology

The introduction of additional switching nodes, plus relaxing the strict hierarchy of
layer-to-layer communication greatly increases the number of ways that a series of
control modules or behaviours can be assembled, and justifiably raises questions of
increased complexity and unpredictable outcomes. On the other hand, the idea that just

16
 INTEGRATED CONTROL OF ROAD VEHICLES

the basic Layer 0 functionality is sufficient for the vehicle to operate, and that both
design and implementation of higher layers is incremental, offers distinct advantages in
terms of flexibility and fault tolerance; for example, only the lowest layers absolutely
require hardware redundancy for fault tolerant operation; when higher level functions
are compromised by faults or failures, the idea would be to shut down the behaviours
immediately affected.
Although formally introduced in the context of layered behaviour-based control, the
above approach is already familiar in the context of standard ABS control systems – see
for example [49,50]. In practice an ABS system may intermittently override a driver
demand for increasing braking torque whenever there is danger of locking the wheels,
by either releasing or holding pressure at the wheel cylinders. The natural scope for
fault detection and management of failure modes is also implemented in practice – once
a fault is detected, the ABS is simply disengaged and a warning communicated to the
driver [49].

4.3 Hybrid Architectures

The architectures shown above are in ‘pure form’ and can be easily generalised and
hybridised. For example, the connectivity in Fig. 11 combines both hierarchical and
subsumption features. In the vehicle context it seems natural to retain a subsumption
type of control at lower levels, whilst introducing a more hierarchical approach within
the highest layers.

Vehicle motion controller

driver
command
& ABS TCS
vehicle
sensors

Brake controller Transmission

Actuators

Fig. 11. Hybrid topology

An ambiguous feature can be seen in Figs. 10 and 11 – where multiple arrows enter a
single node, and the subsuming commands arise from the same layer, it is uncertain
which control predominates. And even without this problem, is it always the case that
higher layers ‘know best’ in subsumption? While it might be argued that higher level
controls will only occasionally interrupt, and the lower level functionality is not
severely disturbed, this ‘interventionist’ approach is clearly not absolutely necessary.
To sidestep further discussion, it is obvious that the ‘subsumption nodes’ shown above
are special cases of a situation where multiple systems are potentially competing for the

17
 T.J. GORDON ET AL.

use of a smaller number of actuators, and some form of coordination mechanism is

needed. Subsumption (highest rank wins) is only one of many possibilities.
Before considering further coordination mechanisms, two essential features should
be noted in the above: (i) all actuator signals are uniquely sourced from a coordination
node or a control block – there is no ambiguity in the actuator input, (ii) there is no
expectation that coordinators have ‘greater knowledge’ of the overall vehicle control
task than the subsystem controllers feeding actuation demands. Indeed to require this
would make the coordinator blocks more like ‘total system controllers’, and hence
regress from a flexible modular control design methodology to a complex and highly
redundant variant of the centralised/hierarchical topology.

5. COORDINATED CONTROL

This section considers how state-of-the-art coordinated control can be applied to

vehicle systems. The focus now shifts away from architectures (some form of multi-
layer design architecture will now be assumed) and control algorithms (any number
might be used for sub-system controller design); emphasis is instead on the
coordination mechanisms that ‘glue together’ the overall IVCS.
Fig. 12 depicts four basic coordinator types, which are representative of the methods
described in the literature of behaviour-based control (e.g. [33,46]). In Fig. 12(a) the
coordination block is one of ‘pure subsumption’ as introduced in the previous section;
the highest ranked (highest level) non-zero command is transmitted to the output. This
is an example of competitive ‘conflict resolution’ where a single behaviour dominates
and takes control of an actuator. Fig. 12(b) represents another competitive approach
where the largest (modulus) activation signal is chosen – this approach is essentially
where the coordinator ‘listens to the loudest voice’.
Both (a) and (b) clearly involve step changes in control, as the coordinator switches
between modes, and there is a clear danger that this will create undesirable transients in
the vehicle dynamic response. Of course it is possible to low-pass filter the output, or
include a ramped transition within the coordinator rather than an instantaneous switch.
Cooperative coordination occurs when some kind of superposition or averaging is
carried out. This could be in the form of simple averaging, or via a non-linear
interpolation function provided by an Artificial Neural Network (Fig. 12c). In either
case, the linear or non-linear function weights need to be chosen to ‘optimise’ the
coordinated control performance in some sense. The potential advantage of using a
nonlinear interpolation function is that actuator saturation may be avoided, and that
smooth transitions can be made between different modes of coordination as input signal
amplitudes change. It is also quite feasible to include additional state or sensor
information as input into the neural network; this would allow for example a greater
emphasis on safe and stable performance when the vehicle is moving at high speed.
However, if taken to extremes, a neural network coordinator could itself become a

18
 INTEGRATED CONTROL OF ROAD VEHICLES

complex non-linear multivariable controller, overly-dependent on the detailed control

activity of the supplying sub-system controllers.

Behaviour 4 Behaviour 4

R
R

TO
TO
Voting
Behaviour 3 Behaviour 3 based

NA
NA (Neural)

DI
DI

Respond
ACTION
Behaviour 2 Behaviour 2

OR
OR

of highest
active

CO
CO

behaviour
Behaviour 1 Behaviour 1
ACTION

(a) (b)

Response
Behaviour 4 of Behaviour 4
R

R
behaviour
TO

TO
with the
highest Fuzzy
NA

Behaviour 3 Behaviour 3

NA
activation Rule
signal based
MAX (B1, B2,B3,B4)
DI

DI
ACTION ACTION
Behaviour 2 Behaviour 2
OR

OR
CO

CO
Behaviour 1 Behaviour 1

(c) (d)

Fig. 12. Examples of control action coordination

Fig. 12(d) represents a fuzzy rule-based coordination function, which though also
inherently nonlinear, has the advantage that the underlying rule-base may be easily
understood in everyday terms. For example a ‘fuzzy subsumption’ coordinator may use
membership functions for ‘small’ and ‘large’ input signals in place of the ‘crisp’ ‘off’
and ‘on’ states for normal subsumption. The same logical decisions are used (highest
level predominates), but the fuzzy system again provides smooth transitions.
Fuzzy coordination is also very natural in the case where input sub-system
controllers are themselves fuzzy systems; in this case the inputs may remain fuzzy
variables, in the form of aggregated fuzzy sets; the coordinator can then performs
additional aggregation, and defuzzification is only applied at the final step before
actuation [51].
The use of fuzzy control and (crisp) subsumption is now demonstrated in a simple
example of integrated vehicle control. Three layers are evident in Fig. 13, which are
now described in turn. Note that the coordinator marked ‘F’ uses fuzzy aggregation of
fuzzy inputs, while those marked ‘S’ use standard subsumption with the filled arrow
( ) predominating. Also note that in this case the ‘lower’ ABS/TCS behaviour
subsumes control from the ‘higher behaviour’ of Yaw Stability; this is strictly against
the philosophy of the subsumption approach to behaviour-based control, where higher

19
 T.J. GORDON ET AL.

layers normally predominate. (Here there is clearly no point locking wheels to attempt
to improve yaw stability, when this would overrides the ABS controller and actually
tend to reduce lateral stability!)

Yaw stability control

Sensors
ABS TCS
and state
estimation
F

S Brakes

Driver

S Powertrain

Fig. 13. Fuzzy coordinated control of vehicle dynamics

Layer 0 – Basic hardware

This covers the basic driver-to-actuator control of steering, driving torque and brakes,
which for simulation comprises a vehicle simulation model and a driver model. For
simplicity, the vehicle model has three major degrees of freedom – forward, lateral and
yaw - and a ‘stiff suspension’ that eliminates body bounce, roll and pitch, but allows
lateral and longitudinal load transfer, and thereby the essential tyre nonlinearities. There
are also four dynamic wheel-spin degrees of freedom and nonlinear tyre characteristics
modelled using a combined slip Pacejka formula. As well as the dynamic degrees of
freedom, the model includes simple controls and transients for wheel braking, driveline
torque and front-wheel steer. The vehicle model is described further in reference [52],
as is the combined longitudinal/lateral driver model. The model is based on a simple
combination of a ‘reference vector field’ feedforward policy, plus proportional-integral
(PI) feedback control. The driver model operates throughout the whole linear and
nonlinear range of the vehicle dynamics, but is deliberately ‘unsophisticated’ – for
example the PI tracking of reference yaw rate works on the assumption that ‘more
steering implies more yaw velocity’ which may well not be the case, for example with
limit understeer behaviour. Full details of the driver model are given in [52].
While it is possible to criticise the integrated modelling of driver and vehicle to
assess closed-loop dynamic behaviour, on the grounds that the results depend on the
specific details of the driver model, its use is becoming increasingly widespread (e.g.

20
 INTEGRATED CONTROL OF ROAD VEHICLES

[20,22,53,54]) and provided the driver model is sufficiently general and robust, the
resulting trends (as described below) are most likely to be meaningful.

Layer 1 – Non-integrated ABS/TCS control

The ABS controller uses a 5×5 matrix of control rules, based on a computer error
between wheel slip λ and a desired (reference) value. The control output is defined,
based on the error and the (discrete-time) change in error. As is common in fuzzy
inference systems, triangular membership functions for input and output are associated
with fuzzy sets such as ‘positive big’ (PB) etc.; for full detail see the Appendix, Table
A1. The ABS controller is applied to each wheel independently, as would be expected
for the low level non-integrated control.
The TCS controller is similarly based on fuzzy inference (Table A2) except that
both brakes and driveline are controlled, and a logic table is required for each output.
Also, since both ABS and TCS potentially ‘compete’ for the use of the brakes, a fuzzy
aggregation coordinator is required, as in Fig. 12 (even though ABS and TCS generally
operate under very different driving conditions).

Layer 2 – Integrated yaw stability control

This is a fuzzy/sliding mode controller, the error signal being determined by deviations
from a sliding surface, and a rule-base in constructed to reduce this error and hence
ensure stable control – see Appendix. It is designed only to correct oversteering
vehicle behaviour, and in line with the strategy being demonstrated here, if understeer
corrections were required, an additional controller could be defined, and fuzzy
aggregation applied to coordinate the two sets of demands; for simplicity this is not
included in the current example. The yaw controller operates on the brakes, and only
when vehicle yaw rate exceeds a certain threshold. A combination of brake actuators is
selected in order to generate a moment that acts to reduce the instantaneous vehicle yaw
velocity.

The following results demonstrate some expected performance benefits of including the
two-layer integrated control, compared to relying on driver skill and basic vehicle
hardware. In both cases the driver attempts to negotiate an ‘S-bend’ with the vehicle
entry speed ‘too high’ for stable vehicle motion along the desired path, the centre-line
of the track. Simulation starts at the entry to the first bend, with a vehicle speed of 30
ms-1; the driver recognises the problem (very late) and attempts to brake towards a
reference speed of 20 ms-1, reducing to 17 ms-1 through the curves (sufficiently slow for
stable centre-line path following) then increasing to a new target speed of 25 ms-1 on
the second straight section. Fig. 14 shows the vehicle paths, and Fig. 15 presents a
number of dynamic response variables (dotted line is without integrated control in both
figures).

21
 T.J. GORDON ET AL.

R = 50 m

1st Leg 2nd Leg

R = 50 m

integrated
none

Fig. 14. Vehicle Path – severe braking through an S-bend

The vehicle paths are not too dissimilar, though it is clear that the vehicle without
integrated control veers off the track on the first bend. In both cases the vehicle
manages to return close to the centreline on the second bend, but the basic vehicle is
much harder to control, as can be seen in Fig. 15 where a number of dynamic responses
are displayed. In the steering plot, it is clear to see that the driver's work load is
reduced when integrated vehicle control is used, especially in the second half of the
simulation.
Vehicle lateral stability is most clearly represented in the plots of vehicle side-slip
angle and yaw velocity; compared to the basic vehicle, these variables are very well
controlled, even under aggressive driving. The basic (rear-wheel drive) vehicle is hard
to control by the driver when maximum acceleration is demanded in the second half of
the simulation; indeed an oversteer condition leads to a yaw oscillation under closed-
loop control, which grows until the driver completely loses stable control of the vehicle
later in the simulation.
There is almost no oscillation on the forward speed as the longitudinal slip is
controlled to be the optimal one for the specific surface. Considering the acceleration,
the vehicle with integrated control allows slightly less and constant forward
acceleration than the vehicle without integrated control, keeping its stability.

22
 INTEGRATED CONTROL OF ROAD VEHICLES

(a) Steering input (b) Pedal demand

200
1
integrated
150 none 0.8

0.6
100

normalised pedal position

0.4
50
0.2
steer [deg]

0 0

−0.2
−50
−0.4
−100
−0.6
integrated
−150 −0.8 none
1st Leg 2nd Leg 1st Leg 2nd Leg
−1
−200
1200 1250 1300 1350 1400 1450 1500 1200 1250 1300 1350 1400 1450 1500
track position [m] track position [m]

(c) Vehicle forward speed (d) Vehicle side−slip angle

30 20

integrated
target speed 15 none
25

10
20
5
slip angle [deg]
speed [m/s]

15
0

10 −5

−10
5
integrated
1st Leg 2nd Leg none −15
0 1st Leg 2nd Leg
1200 1250 1300 1350 1400 1450 1500
track position [m] −20
1200 1250 1300 1350 1400 1450 1500
track position [m]

(e) Yaw rate (f) g−g diagram

2 0.4

1.5 0.2

0
longitudinal acceleration [g]

−0.2
yaw rate [rad/s]

0.5
−0.4
0
−0.6
integrated
−0.5 none
−0.8

−1 −1
integrated
1st Leg 2nd Leg none
−1.5 −1.2
1200 1250 1300 1350 1400 1450 1500 −1 −0.5 0 0.5 1
track position [m] lateral acceleration [g]

Fig. 15. Dynamic response effect of the integrated vehicle control system

23
 T.J. GORDON ET AL.

The plot in Fig. 15(f) is a g-g diagram, which shows a clear and expected advantage of
the integrated control (compare Fig. 2). In particular it is possible to simultaneously
brake and steer in a controlled fashion near the limits of friction. It is worth noting that
the apparently greater longitudinal accelerations possible for the basic vehicle are
merely due to poor vehicle side-slip angle control; in this case, when the vehicle rotates
towards the cornering direction, lateral accelerations attain an additional component
that appears to be longitudinal in the vehicle-fixed axis system. In any case, these
excursions of acceleration in the forward direction are clearly very badly controlled,
and coincide with loss of stability on the second straight section of track.
The IVCS presented here is quite rudimentary; for example there are no
circumstances under which wheel lockup is allowed; subsumption provides a
permanent bias towards the basic ABS/TCS functionality. However, in the presence of
large slip angles (and in the absence of an active steer degree of freedom) the direction
of the force vector at any wheel can only be properly controlled by inducing such
lockup or wheel-spin.
However the example serves to demonstrate that useful integrated vehicle control
can be implemented by adhering to a multi-layer ‘bottom up’ design architecture,
implemented by a relatively simple coordination policy applied to independently
designed sub-system controllers.

6. TOWARDS DYNAMIC COORDINATION IN IVCS

The advantages of the above general approach to integrated vehicle control is clear: it
provides a modular and incremental design method, providing adequate scope for
separating out local and integrated design responsibilities. Within vehicle development,
its use would rely heavily on simulation – so the modular approach would mean sub-
system suppliers needing to provide validated dynamic model of the closed-loop
controller/hardware, to allow the IVCS to be developed in simulation. This latter
development would include the detailed choice of T2/T3 architectures, the types of
coordinator and the tuning of these coordinators (e.g. via choice input weights).
One problem with the approach is the very large amount of freedom still available
for integration. If an inappropriate type of coordination is chosen for example, the
overall system may exhibit poor performance, or even generate dynamic instabilities in
certain circumstances. One particular issue is that the coordination function is
relatively decoupled from the performance objectives of the subsystem controllers,
providing what might be seen as an undesirable freedom to degrade the overall control.
Hence the purpose of this section is to explore further possibilities (ones that do not so
far appear to have been the subject of detailed research).
The essential features of a coordinator are that (i) it improves the ‘integrated
dynamic performance’ of the vehicle and (ii) it makes decisions based directly on
information obtained from the sub-system controllers (e.g. control signal amplitude).
This includes the possibility of requiring a ‘richer’ form of control information from

24
 INTEGRATED CONTROL OF ROAD VEHICLES

sub-system controllers than the desired actuator commands. In fact this is already the
case with coordination via fuzzy aggregation, where the fuzzy inputs contain
significantly more information than the (crisp) control demands.
Where sub-system controllers are based on formal optimal control, an enriched
control signal may comprise - in addition to the actuator demands - local sensitivity
information in the form of an ‘internal sub-system Hamiltonian function’ [10]. If the
Hamiltonian function from the optimal control is computed in the subsystem, as a
function of control output, then this immediately provides a dynamic sensitivity
function for the additional cost of any sub-optimal control. Summing the input
sensitivity functions φ1 and φ2 , and minimising the result provides a ‘dynamically
optimised’ control signal u (t ) ; the coordinator formally recognises and balances the
control objectives of the sub-system controllers. Fig. 16 illustrates this situation where
a scalar control is coordinated between two sub-systems. In reference [10] the sub-
systems (simulated active suspension corner actuators) were controlled via linear
optimal control (LQR), and the minimisation was carried out off-line: an explicit
closed-form integrated controller was obtained, and performance was similar to an
equivalent ‘global’ centralised controller derived from a full vehicle dynamic model.
However, the on-line minimisation proposed here has the major advantage that the
coordinator needs no specific information of the sub-system states and models. In the
case of linear optimal control theory, the sensitivity functions are quadratic, and even in
the multivariable case the solution is closed-form, requiring only one matrix inversion.

u1
sub-system 1
φ1 ( ∆u ) minimise : u
φ1 (u − u1 )
u2 + φ2 (u − u2 ) φ (∆u )
sub-system 2
φ2 ( ∆u )

Fig. 16. Hamiltonian (cost sensitivity) coordinator concept

In the figure, an output sensitivity function is also shown, as this would be necessary if
the coordinator output were to be fed into a second Hamiltonian coordinator. Clearly
weighting parameters can also be included to vary the relative importance of the sub-
system demands. Although it will not be discussed further here, one major advantage
of this approach is the potential to develop overall stability criteria for the IVCS. This
might be developed from the Lyapunov functions directly available from the sub-
system controllers. In any case, provided the sub-system control criteria are compatible
in some sense, this dynamically based control integration should always work to

25
 T.J. GORDON ET AL.

improve the overall system performance compared to non-integrated control, promoting

safety, stability and confidence in the resulting IVCS.
In the above, two types of dynamic coordination have been suggested – a weak form
based on fuzzy aggregation, and a more fundamentally based approach based on formal
(linear) optimal control theory. To be really feasible in practice, such coordination
needs to be applicable to a wider range of subsystem control methods. The key
requirement is that there exists a form of error criterion, where minimising this criterion
automatically means achieving the desired performance outcome of the sub-system
controller. One form of control for which this is applicable is sliding mode control,
where deviation from the sliding surface can be used as a measure of error. And more
generally, where nonlinear or adaptive control makes explicit use of a Lyapunov
function, the dynamic change in this function can also be developed as a measure of
error for coordinator operation. The fundamental development required for control
implementation is that such criteria need to be explicitly available in real-time, rather
than simply implicit within the controller design.
Perhaps the most common type of low-level dynamic control currently employed on
vehicles (and elsewhere in industrial control systems) is single-loop PID control. There
is unlikely to exist any universal formula for costing perturbations in such controllers;
the ‘dynamic cost sensitivity’ is likely to be highly dependent on the application.
However it is easy to conceive of a ‘retrofit’ sensitivity analysis, even if this is not
carried out within the original design. Controls are perturbed in system simulation and
performance objectives assessed over time in response to these perturbations.
Modelling of the performance dependency may then be undertaken via a variety of
techniques (system identification, neural networks, polynomial response surfaces, etc.).
It has already been pointed out that the layered design architectures offer
possibilities for robust and fault-tolerant vehicle control systems, based on the simple
idea that if faults develop in hardware or software that is specific to higher layers, these
can be shut down and the lower layers continue to function. So only the ‘basic
hardware’ needs to incorporate expensive component redundancy. But fault tolerant
behaviour may also be derived from the ‘intelligent’ use of coordination functions.
This has been suggested in [10] where a simple actuator malfunction is seen to be
compensated for by (on-line) optimisation that takes account of diagnostic information
on the failure. The major advantage is that this dynamic fault compensation requires no
a-priori knowledge of anticipated failure modes.

7. CONCLUSIONS

This paper has reviewed examples from the literature on Integrated Vehicle Control
Systems (particularly chassis control systems) and attempted to discern the main
structural features of such control, and establish where there is scope for new
developments, not just theoretically, but also taking account of the practical commercial

26
 INTEGRATED CONTROL OF ROAD VEHICLES

constraints in vehicle development. Related literature from robotics and aerospace

vehicles has also been taken into account.
One simple but important conclusion is that control system architecture has several
levels of meaning, and in this paper five fairly distinct levels have been discerned. This
means that, with the possible exception of fully decentralised or fully centralised
concepts, it is insufficient to formulate the control architecture in terms of a single
topology. A related conclusion is the integrated control is a very much richer concept
than multivariable control, with requirements that it should:-
1. be modular - so that design may be distributed between teams, and for example
there is no need to redesign existing systems when additional sub-systems, sensors
or actuators are added
2. respect the intellectual property of suppliers - sub-system controllers need to
collaborate, without the need to make public all algorithmic details
3. avoid excessive complexity (e.g. as measured by real-time processing,
communication and memory requirements) – hence to grow ‘reasonably’ with the
number of actuators and sensors on the vehicle
4. incorporate fault detection, diagnosis and tolerance
5. possess an open architecture that permits equal status ‘access’ to additional
systems, both internal and external.

The most promising candidate for the required structured approach is through layered
design, similar to that employed in ‘behaviour-based’ control. The term ‘sub-system’
controller has been used interchangeably with the ‘control behaviour’, both essentially
referring to the delivery of a specific form of dynamic performance capability, rather
than some specific use of an actuator or a subset of states. In this approach the design
of control coordinators becomes the major task for vehicle control integration, and
some specific types have been explored. One such method has been demonstrated,
where a combination of fuzzy sub-system controllers, with coordination via fuzzy
aggregation and (crisp) subsumption has been seen to develop integrated control
behaviour at the vehicle level.
The excessive design freedom of this approach has motivated the need for seeking
closer links between sub-system controllers and the associated control coordinators, and
one potential approach has been suggested.
There is one key aspect of IVCS not covered in this paper, and this concerns the role
of the driver and his or her connection to the control architecture. In the multi-layered
architectures, the driver may clearly interact simultaneously will all layers of the
integrated control system. This follows from the sequential design methodology of
adding competency with each new layer, and is quite different from typical hierarchical
approaches where the driver is to interact only at the highest level in the hierarchy (e.g.
[6,34,47]).
The final conclusion is that there still remains a significant research challenge to
properly understand and implement IVCS’s in future vehicles, even without

27
 T.J. GORDON ET AL.

considering the greater challenges of understanding how such advanced systems should
best interact with the driver, external vehicles and highways systems.

REFERENCES

1. Anon.: Daimler-Chrysler HITECH REPORT, Research and Technology Issue 2 (2002), p.67.
2. Freuchte, R.D., Karmel, A.M. and Rillings, J.H.: Integrated vehicle control. IEEE Vehicular
Technology Conference, 2 (1989), pp.868-877.
3. Tanaka, H., Inoue, H. and Iwata, H.: Development of a vehicle integrated control system. Proc. of the
XXIV FISITA Congress, June 1992, London. (1992), pp.63-72.
4. Wallentowitz, H.: Integration of chassis and traction control systems. what is possible - what makes
sense - what is under development. AVEC 92 (1992), pp.1-7.
5. Kiencke, U.: Integrated vehicle control systems. IFAC: Intelligent Components for Autonomous and
Semi-Autonomous Vehicle, Toulouse, France (1995), pp.1-5.
6. Coelingh, E., Chaumette, P. and Andersson, M.: Open-interface definitions for automotive systems.
SAE 2002 World Congress (2002) 2002-01-0267.
7. Hirano, Y., Harada, H., Ono, E. and Takanami, K.: Development of an integrated system of 4WS and
4WD by H-infinity control. SAE Journal 930267 (1993), pp.79-86.
8. Busshardt, J., Chaumette, P., Hac, A., Hoang-Gia, S., Hoying, J., Oakley, R. and O'Connell S.: An
integrated chassis control system to enhance vehicle stability. Progress in Technology. 77 (1999),
pp.373-382
9. Ghoneim, Y.A., Lin, W.C., Sidlosky, D.M., Chen, H.H., Chin, Y.K. and Tedrake, M.J.: Integrated
chassis control system to enhance vehicle stability. International Journal of Vehicle Design, 23
(2000), pp.124-144.
10. Gordon, T.J..: An integrated strategy for the control of a full vehicle active suspension system.
Vehicle System Dynamics, 25 (1996), pp.229-242.
11. Wills, L., Kannan, S., Sander, S., Murat Guler, M., Heck, B., Prasad, J., Schrage, D., and
Vachtsevanos, G.: An open platform for reconfigurable control. IEEE Control Systems Magazine,
21(3), (2001), pp.49-64.
12. Shladover, S.: Review of the state of development of advanced vehicle control systems (AVCS).
Vehicle System Dynamics vol.24 (1995), pp.551-595.
13. Abu-Lebdeh, G.: Integrated adaptive-signal dynamic-speed control of signalized arterials. Journal of
Transportation Engineering, 128 (2002), pp. 447-451.
14. Sato, S., Inoue, H., Tabata, M. and Inagaki, S.: Integrated chassis control system for improved
vehicle dynamics. AVEC 92 (1992), pp.413-418
15. Nagai, M., Yamanaka, S. and Hirano, Y.: Integrated control of active rear wheel steering and yaw
moment control using braking forces. JSME International Journal, 42 (2) Series C (1999), pp.301-
308.
16. Wang, Y.Q. and Nagai M.: Improvement of vehicle handling and stability by an integrated four-
wheel-steer and yaw moment control system. Progress in Technology, 77 (1999), pp.173-182.
17. You, S.S. and Chai, Y.H.: Multi-objective control synthesis: an application to 4WS passenger
vehicles. Mechatronics (1999), pp.363-390.
18. Horiuchi, S., Okada, K. and Nohtomi, S.: Effects of integrated control of active wheel steering and
individual wheel torque on vehicle handling and stability - a comparison of alternative control
strategies. Vehicle System Dynamics Supplement 33 (1999), pp.680-691.
19. Mastinu, G., Babbel E., Lugner P., Margolis D., Mittermayr, P. and Richter, B.: Integrated controls
of lateral vehicle dynamics. Vehicle System Dynamics (Supplement) 23 (1994), pp.358-377.
20. Rodic, A.D. and Vukobratovic, M.K.: Contribution to the integrated control synthesis of road
vehicles. IEEE Transactions on Control Systems Technology, 7 (1999), pp.64-78.

28
 INTEGRATED CONTROL OF ROAD VEHICLES

21. Tai, M. and Tomizuka, M.: Robust lateral control of heavy duty vehicles for automated highway
systems. Proc. 1999 14th IFAC Triennial World Congress, Beijing, (1999), pp.37-42.
22. Mokhiamar, O. and Abe, M.: Active wheel steering and yaw moment control combination to
maximize stability as well as vehicle responsiveness during quick lane change for active vehicle
handling safety. Proc. Inst. Mech. Eng. 216 Part D (2002), pp.115-124.
23. Lim, E.H.M. and Hedrick, J.K.: Lateral and longitudinal vehicle control coupling for automated
vehicle operation. American Control Conference (1999), pp.3676-3680.
24. Komatsu, A., Gordon, T.J. and Best, M.C.: 4WS control of handling dynamics using a linear optimal
reference model. Proceedings of the 5th International Symposium on Advanced Vehicle Control
(AVEC2000) , Ann Arbor, USA, (2000), pp.253-260.
25. Economou, J.T. and Colyer R.E.: Modelling of skid steering and fuzzy logic vehicle ground
interaction. American Control Conference (2000), pp.100-104.
26. Lee, C.: Fuzzy logic in control systems : fuzzy logic controller -Part I and Part II. IEEE Trans. Syst.
Man Cybern. 20 (1990), pp.404-434.
27. Zeyada, Y.F., ElBeheiry, E.M. and ElArabi, M.E.: Fault diagnosis and mu-estimation for integrated
chassis control of vehicles in emergency maneuvers. Journal of Engineering and Applied Science,
Cairo, 47 (2000), pp.665-682.
28. Rodic, A.D. and Vukobratovie, M.K.: Design of an integrated active control system for road vehicles
operating with automated highway systems. International Journal of Computer Application in
Technology, 13 (2000), pp.78-92.
29. Nwagboso, C.O., Ouyang, X. and Morgan, C.: Development of neural-network control of steer-by-
wire system for intelligent vehicles. Heavy Vehicle Systems, 9 (2002), pp.1-26.
30. Katebi, M.R. and Grimble, M.J.: Integrated control, guidance and diagnosis for reconfigurable
autonomous underwater vehicle control. International Journal of Systems Science, 30 (1999),
pp.1021-1032.
31. Harada, M. and Harada, H.: Analysis of lateral stability with integrated control of suspension and
steering systems. JSAE 20 (1999), pp.465-470.
32. Evans, J., Hodge, W., Liebman, J., Tomlin, C. and Parkinson, B.: Flight tests of an unmanned air
vehicle with integrated multi-antenna GPS receiver and IMU; towards a testbed for distributed
control and formation flight. Proceedings of the ION-GPS Conference, Nashville, (1999).
33. Carreras, M., Batlle, J., Ridao, P. and Roberts, G.N.: An overview on behaviour-based methods for
AUV control. 5th IFAC Conference on Manoeuvring and Control of Marine Craft (MCMC2000).
Aalborg, Denmark (2000), pp.141-146.
34. Palkovics, L.A., and Fries, A.: Intelligent electronic systems in commercial vehicles for enhanced
traffic safety. Vehicle System Dynamics, 35 (2001), pp.227-289.
35. Phillips, A.M.: Functional decomposition in a vehicle control system. American Control Conference
(2002), pp.3713-3718.
36. Schmidt, E.H., Wright, C.D. and Zwerne, J.S.: Required elements of integrated vehicle control
systems. Proc. International Congress on Transportation Electronics, SAE paper 901170 (1990),
pp.463-471.
37. Rago, C., Prasanth, R., Mehra, R.K. and Fortenbaugh, R.: Failure detection and identification and
fault tolerant control using the Imm-Kf with applications to the eagle-eye UAV. Decision & Control
Tampa, Florida, USA (1998), pp.4208-4213.
38. Bartolini, A., Sheppard, J.W. and Munns, T.E.: An application of diagnostic inference modelling to
vehicle health management. IEEE Systems Readiness Technology Conference Piscataway, NJ, USA
(2001), pp.706-717.
39. Nagai, M. Hirano, Y. and Yamanaka, S.: Integrated robust control of active rear wheel steering and
direct yaw moment control. Vehicle System Dynamics (Supplement) 28 (1998), pp.416-421.
40. Ono, E., Takanami, K., Iwama, N., Hayashi, Y., Hirano, Y. and Satoh, Y.: Vehicle integrated control
for steering and traction systems by mu-synthesis. Automatica 30 (1994), pp.1639-1647.

29
 T.J. GORDON ET AL.

41. Brennan, S. and Alleyne, A.: Integrated vehicle control via coordinated steering and wheel torque
inputs. American Control Conference (2001), pp.7-12.
42. Kelling, N.A., and Heck, W.: The BRAKE project – centralized versus distributed redundancy for
brake-by-wire systems. SAE 2002 World Congress (2002) 2002-01-0266.
43. Duffie, N.A., Chitturi, R. and Mou.,J.: Fault-tolerant heterarchical control of heterogeneous
manufacturing system entities. Journal of Manufacturing Systems, 7 (1988), pp.315-328.
44. Keller, K., Wiegand, D., Swearingen, K., Reisig, C., Black, S., Gillis, A. and Vandernoot, M.: An
architecture to implement integrated vehicle health management systems. IEEE Systems Readiness
Technology Conference, Piscataway, USA, (2001), pp.2-15.
45. Brooks, R. A.: A Robust Layered Control System for a Mobile Robot. IEEEJournal of Robotics and
Automation, 2(1), (1986), pp. 14-23
46. Brooks, R.A.: Integrated systems based on behaviors. SIGART Bulletin 2:4, August (1991), pp. 46-
50.
47. Girard, A.R., De Sousa, J.B., Misener, J.A. and Hedrick, J.K.: A control architecture for integrated
cooperative cruise control and collision warning systems. IEEE Conference on Decision and Control
(2001), pp.1491-1496.
48. Smakman, H.: Functional integration of active suspension with slip control for improved lateral
vehicle dynamics. AVEC 2000 Ann Arbor, Michigan, USA (2000), pp.397-404.
49. Austin, L. and Morreu, D.: Recent advances in antilock braking systems and traction control systems.
Proc. Inst. Mech. Eng., Part D Journal of Automotive Engineering, 214 (2000), pp.625-639.
50. Bosch.: Driving Safety-Systems, Robert Bosch GmbH, 2nd Ed., SAE, Warrendale, USA, 1999.
51. Hagras, H., Callaghan, V., Colley, M. and Carr-West, M.: A behaviour based hierarchical fuzzy
control architecture for agricultural autonomous mobile robots. The International Conference on
Computational Intelligence for Modelling, Control and Automation, Vienna-Austria (1999), pp.172-
177.
52. Gordon, T.J., Best, M.C. and Dixon, P.J.: An automated driver based on convergent vector fields.
Proc. Inst. of Mech. Eng., Part D Automobile Engineering , 216D4 (2002), pp.329-347.
53. Beji, L. and Bestaoui, Y.: An adaptive control method of automated vehicles with integrated
longitudinal and lateral dynamics in road following, robot motion and control. (RoMoCo'01)
Bukowy Dworek, Poland (2001), pp.201-206.
54. Becker, U., Rodic, A. and Schnieder, E.: Integrated modeling of driver-vehicle dynamics for use in
designing of driver-assisted control systems. Automatisierungs- und Assistenzsysteme fur
Transportmittel, Brunswick, Germany (2002), pp.108-129.
55. Buckholtz, K.R.: Use of fuzzy logic in wheel slip assignment - part I: yaw rate control. SAE Journal
(2002), 2002-01-1221.

30
 INTEGRATED CONTROL OF ROAD VEHICLES

APPENDIX

The simulation study contained in Section 5 is based on the vehicle and driver models
described in [52]; the (T2) control topology is described in Section 5, together with the
of form of the subsumption coordinators, so this Appendix is provided to summarise the
fuzzy sub-system controllers.
The ABS controller is a fuzzy inference system based on the 25 rules given below in
Table A1, together with standard triangular membership functions. The TCS is also a
fuzzy controller with 25 rules for braking actuation (Table A2) on the driven wheels
and another 25 rules for the control of driving torque (Table A3). The TCS controller
also makes use of standard triangular membership functions.

Table A1. ABS control rules

error
ABS PB PS ZR NS NB
PB NB NB NB NS PB
Change

error

PS NB NB NS NS PB
of

ZR NB NB ZR ZR PB
NS NB NS ZR PS PB
NB NB NS ZR PB PB

Table A2. TCS control rules (brake control)

error
TCS-Brake PB PS ZR NS NB
PB PB PB ZR ZR ZR
Change

error

PS PB PS ZR ZR ZR
of

ZR PB ZR ZR ZR ZR
NS PS ZR ZR ZR ZR
NB PS ZR ZR ZR ZR

Table A3. TCS control rules (drive torque control)

error
TCS-Drive PB PS ZR NS NB
PB NB NB ZR ZR ZR
Change

error

PS NB NB ZR ZR ZR
of

ZR NB NS ZR ZR ZR
NS NB NS ZR ZR ZR
NB NB NS ZR ZR ZR

31
 T.J. GORDON ET AL.

Table A4. Yaw stability control rules

YAW Outer-Front Outer-Rear Inner-Front Inner-Rear
N4 ZR P4 ZR P4
N3 ZR P4 ZR P3
N2 ZR P4 ZR P2
N1 ZR ZR ZR P1
ZR ZR ZR ZR ZR
P1 P1 ZR ZR ZR
P2 P2 ZR P4 ZR
P3 P3 ZR P4 ZR
P4 P4 ZR P4 ZR

For these two longitudinal slip controllers, the error is defined as

e = λ − λref
where λ is the actual or estimated longitudinal wheel slip and λref is the target
reference value.
The yaw stability controller is a combined sliding mode-fuzzy controller based on
that presented in reference [55]. The error and change of error signals are combined in a
sliding surface defined by the equation
s = e + γ e&
where e(t ) is a tracking error in yaw velocity r :
e = r − rref
and the reference yaw rate is given in terms of lateral acceleration and forward speed:
ay
rref = .
vx
The sliding mode signal s (t ) is presented to the fuzzy logic controller, which
commands distributed wheel braking, designed to reduce the error towards zero. The
output signals from the fuzzy controller are labelled Outer-Front, Outer-Rear, etc.,
where ‘Outer’ and ‘Inner’ are defined relative to the cornering direction of the the
roadway (or intended vehicle path).
The fuzzy control has 9 rules with 4 outputs for each rule – Table A4. The input
membership function are equally distributed across the fuzzy input ‘universe’ (-1,1). N4
is a Z function, P4 is an S function, all the others are standard Gaussian functions. The
output membership function are all Gaussian functions with universe is defined from 0
to 1, and 1 corresponds to maximum braking.

32