5

Tuning of ADAS Functions Using Design

Space Exploration

Abhishek Ravi1 , Hans Michael Koegeler1 and Andrea Saroldi2

1AVL List Gmbh, Austria
2 C.R.F. S.C.p.A , Italy

5.1 Introduction
AnADAS function developed within the DESERVE platform and the tuning of
this function for a particular application is discussed in this chapter. Based on
separating the software and tuning data, according to the standards described
in detail in Chapter 2, such a function can also be used for an alternate vehicle
or application use case. The opportunities as well as the potential challenges
are described, using a real world example, developed within the DESERVE
Project.

5.1.1 Parameter Tuning: An Overview

Tuning or calibration of vehicle components is essentially determining the
optimum attributes, which fulfill the legislative standards as well as refine the
car’s character to meet all the expectations of the driver for drivability and
comfort. Besides the comfort and legislative issues the vehicle tuning also
helps in brand differentiation and helps to determine the vehicle character.
In the tuning task for a specific component (e.g.: engine), the software and
the tuning data in the application layer of an Electronic Control Unit (ECU)
is separated which is illustrated in Figure 5.1. The resulting code is a hex file,
which can be flashed to the defined controller hardware which gives a big
flexibility in powertrain development. As an example, one engine hardware
can be put into more than 200 vehicle variants fitting for different countries,
different vehicles and/or different transmission systems – just by flashing a
different appropriate controller software.

77
78 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.1 Separation of software and tuning parameters in a control unit.

5.1.2 Industrial Tuning Applications: Challenges

and Opportunities
The engine – ECU has been the first mechatronic application in the automotive
world. It makes sense to have a short view on the historical development of
the tuning task in this field as illustrated in Figure 5.2.
In the past decades, the improving technology in the automotive sector
can be seen with cars having better engine performance, less consumption,
better handling and reduced emissions. But the improvement in technology
has come with increased complexity, especially in the tuning task.

Figure 5.2 History of powertrain tuning (calibration).

 5.1 Introduction 79

As can be seen in Figure 5.2, initially there used to be around 500

parameters which needed to be tuned, which was carried out by a single
engineer using the unit to be tested, which was then tested on a single test
vehicle. Initially, the powertrain was quite simple and the Engine – ECU was
the only one being considered.
With increasing legislative and user demands; the complexity of the tech-
nology, the number of involved interacting components (engine, gearbox and
electric engine) and also the number of functions controlling the interactions
between all the variable components increased dramatically. Further the tuning
allowed the derivation of many more vehicle variants with the same hardware
components but differing in the ECU-SW, wherein the functions in the SW
stay the same, just the tuning data are specifically developed.
This effect is also seen in the number of tuning parameters to be defined
in an engine calibration project, where around 50 k parameters have to be
defined – clearly assigned to many functions. So it is no longer possible to
have one person, who understands all the functions implemented and teams
of specialized persons are necessary, partly working in different areas of the
world. Thus the industry was confronted with several challenges and found
some responses.
For example, the management of tuning data becomes an issue. It must
be possible to track all the changes made to the tuning data by the different
engineers involved and bring all the tuning results into a single final tuning
result. The company should be able to ensure at Start of Production (SoP)
that:
1. All the tuning data are calibrated.
2. All the tuning data are calibrated with the correct settings to optimally
fulfill the desired, derivative use case.
These two requirements are very challenging, which explains the need of
“Tuning Data Management”. This topic itself is not further elaborated in this
chapter, but is supported by valuable literature [1, 2].
Another challenge lies in the tuning for single use cases: For example,
the emission tuning of an engine in a certain vehicle configuration for the
legislation of a specific country. There are about 5 to 10 strongly interacting
tuning parameters. E.g. an engine map to define the start of the combustion as
function of speed and load is counted as one of these parameters, and exhaust
gas recirculation rate, rail pressure, boost pressure, split patterns of the injected
fuel quantity are others, all either reducing the different kinds of emissions or
changing fuel consumption or noise.
80 Tuning of ADAS Functions Using Design Space Exploration

So one can imagine, that it is just not possible to measure the emissions
and the fuel consumption of all the feasible combinations of say 8 of such
parameters on an engine. (A similar issue faced with ADAS functionality)
Such tasks are typically performed on engine test beds and chassis dynos
and have to be finally validated on the road again. With the latest legislation
(Real Driving Emissions, RDE) even the certification will be done on the road
giving additional challenge [3–5].
Figure 5.3 illustrates the generalized development environment, which
allows the engineer to reproduce maneuvers and then double check the results
of tuning work. In the manual tuning method, the engineer operates the UUT
with a certain setting of control parameters in certain maneuvers. The engineer
observes the behavior of the UUT and performs a judgment according to
his experience. Then the next setting is defined with the intention to better
approach the desired behavior. This process becomes complex when there are
many relevant tuning parameters [6].
In this trial and error method, the quality of tuning and the optimization
results depend on whether the engineer considers all the parameters that are
relevant for the desired behavior and the relevant start point. There is a strong
dependence on the experience of the engineer. There are also limitation on the
number of tests that can be conducted, due to the testing time, complexity and
cost factors. The final results are highly subjective, as the decision making

Figure 5.3 Illustration of a generalized development environment and manual tuning

process.
 5.1 Introduction 81

process lacks traceability and a reuse is not possible for future projects, e.g.
tuning an ADAS setup for a different drive mode. As a result, a methodology
to increase the efficiency and the quality of the tuning work at the same time,
the so called “Design of Experiment” method (DoE) was adapted accordingly.
Within the DESERVE context this methodology was applied as “Design
Space Exploration” for Simulation environments, which are excellent devel-
opment environments for tuning of ADAS Functions.
The model-based approach was used with two objectives:
• Firstly, to find an optimum tuning result.
• Secondly, to validate an existing tuning result under a big variety of use
cases, which will happen during the lifetime of a vehicle.

5.1.3 Model-based Tuning

Model-based tuning is a statistical, model-based approach which reduces
the amount of actual experiments/test runs needed to accurately describe the
behavior of the UUT within the design space. This method helps to choose the
position of the test data points in order to generate behavior models with
an efficient low number of measurements. Such models are then utilized
to develop an accurate and robust tuning according to specific optimization
target(s). In Figure 5.4 the entire method is illustrated again for the generalized
development environment.

Figure 5.4 Model-based tuning task illustrated.

82 Tuning of ADAS Functions Using Design Space Exploration

In a model-based tuning task the below steps are followed:

• The user begins with a task planning for the measurement series, where
the targets for the tuning task are determined. Based on the targets, the
relevant input parameters which are considered to influence the observed
UUT response are selected. AVL CAMEO is used for the test plan
generation. This is based on a one time set up process, in which CAMEO
is connected to the development environment. Thus CAMEO gets access
to set tuning parameters in the UUT, observe responses of the UUT and
to start/stop maneuvers and to take measurements after maneuver. The
development environment hosting the UUT could be in the form of a test
bed, a hardware-in-the-loop (HiL) or even a vehicle simulation software
like IPG Carmaker in combination with an ADAS-function prototype
programmed in MATLAB.
• Once the targets have been defined the next important step is to make
the test matrix. In order to get a full picture of the area to be investi-
gated, the Design of Experiments (DoE) is used [7]. It is a systematic
technique which allows varying all the parameters simultaneously while
answering the two important questions of every tuning activity: Firstly,
how many tests are needed to cover the entire design space? And
secondly, at which locations in the design space test points are needed
to effectively get modelling equations valid throughout the entire design
space. There are many DoE designs available to us in AVL CAMEO,
but COR DoE methodology [8] was used in the current example exer-
cise. Besides setting up the test design, it is also important to set the
limits for the test and appropriate actions when the limit is violated.
These topics are addressed further on in the example discussed in
Subsection 6.2.1.
• With the test plan and limits decided the tests are run, where the necessary
parameter settings are uploaded to the UUT by CAMEO, and after the
test, the required measurement results were stored in CAMEO. The raw
measured data check is then carried out in order to check the plausibility
and feasibility of measurement. It is a necessary check to get a rough
idea of how the measurements compare against expected values, and
also observe possible errors which could have occurred during the test
execution.
• The measurements are modeled empirically to obtain behavior models of
the UUT. In this content, modeling means more or less to fit a function –
like a polynomial equation for example – into the measured responses in
order to estimate the response function of any point in the design space.
 5.1 Introduction 83

Such a model helps understand the reaction of the UUT to the parameter
tuning, and the interaction of the different tuning input parameters and
the output measurements. The confidence and prediction intervals of the
empirical models are observed to evaluate the model quality. Models in
CAMEO also allow extrapolation in defined ranges beyond the design
space covered by measurements to observe the UUT behavior at points
where tests could not be run based on equipment limitations or time/cost
constraints.
• Based on the optimization target, optimization algorithms can be imple-
mented for a single objective or multiple objectives. The engineer can
decide if the results meet the targets and constraints and in case of multiple
objectives decide on a suitable tradeoff between the different desired
targets (Pareto front).
• Before, the results from the analysis are accepted a final verification test is
carried out. Tests are run at least on the point of the decided optimum, but
can also be extended on parameters settings of ten or more points spread
across the Pareto front. If these verification measurements match the
modeled results then the empirical models are accepted and the engineer
can use the optimization results as the desired tuning setting.

5.1.4 Model-based Validation

A model-based validation is a task carried out to test and evaluate the
robustness of the results from the tuning task. The UUT is run at the parameters
settings obtained from the tuning task, but tested for an alternate use case
and the response is evaluated. For example; if say a diesel engine was tuned
to operate at an economy mode and a sport mode with strong limits set on
NOx emissions. Economy mode encourages the engine to conserve fuel while
sacrificing power, while the Sport mode encourages the engine to provide
greater power while making compromises on fuel economy, with the engine
running more at the higher RPMs. The engine is initially tuned at driving
conditions imitating an urban environment and lower altitudes, and from the
tuning tasks the input parameters settings like the rail pressure, injection
pressure, injection timing etc. are selected to operate the engine at the two
targeted modes while sticking to the NOx limits. In the validation test run the
engine is first run at the economic mode and then sport mode, but now the use
case is in hilly road conditions and higher altitude. The engine performance is
evaluated with respect to power and emissions, while the road and altitude of
operation is varied. The target is to see if tuning settings could be extrapolated
84 Tuning of ADAS Functions Using Design Space Exploration

or extended to alternate use cases. It also gives further information on how the
engine tuned for urban conditions would perform on rugged hilly conditions.

5.2 Demonstrative Example

A map-based ACC-Function (developed by the DESERVE Partner CRF)
running in a commercially available MiL Environment (IPG-Carmaker +
MATLAB Simulink) has been used as an example. The calibration tool of
AVL CAMEO was connected to this environment in order to tune the function
for a Fiat 500L.

5.2.1 Function: An Overview

A map-adaptive autonomous cruise control (ACC) was developed to:
• Control the vehicle velocity in order to enter and exit curves in a
comfortable and safe manner.
• Complete the drive maneuver in the least amount of time.
The controller function controls the vehicle speed by sending jerk request
(see Figure 5.7). Jerk is the rate of change of acceleration. Hence the jerk
request signals from the controller function are converted into the vehicle
acceleration and speed. For the reference maneuver a digitized road was used
and a reference speed curve was determined, which is the maximum speed
at which this road can be safely maneuvered. The function tries to ensure
that, the vehicle follows this reference speed profile as closely as possible
without exceeding it. The target speed was set at 130 km/h for the ACC.
A demonstrative speed profile is shown in Figure 5.5 for a sample settings in

Figure 5.5 Velocity profiles for a sample test run using the control function.
 5.2 Demonstrative Example 85

Figure 5.6 Function developed using IPG carmaker and MATLAB simulink.

the ACC function. It can be seen that the vehicle velocity tries to follow the
reference velocity while never exceeding it. The vehicle velocity is not able
to exactly replicate the reference velocity due the road conditions, the vehicle
limitations and the control function settings.
The function was developed using IPG Carmaker for Simulink and has
been illustrated in Figure 5.6. IPG Carmaker for Simulink is integrated into
MATLAB/Simulink and necessary modification were done by adding the
custom Simulink blocks developed for the current use case.

5.2.2 Design Variables

In order to tune the function for the reference maneuver, four input parameters
or design variables were selected (see Figure 5.7). As per the terminology used
in CAMEO these tunable input parameters will be referred to as the variation
parameters. The variation parameters selected for the tuning task are:
• Acceleration Maximum (A MAX) limits the maximum positive accele-
ration the vehicle can have while safely completing the maneuver. The
negative acceleration is not limited in order for the vehicle to generate
the necessary breaking force in case of obstacles.
• Jerk Maximum (J MAX) limits the maximum positive jerk request from
the controller function in order to meet the reference velocity curve. But
only the positive jerk given by the engine and responsible for positive
acceleration is limited, while there is no lower limit for the negative jerks
for reasons mentioned previously.
86 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.7 Function overview.

Figure 5.8 Illustration of the kinematic variables A MAX and J MAX.

 5.2 Demonstrative Example 87

Figure 5.8 illustrates the kinematic parameters, with acceleration being

the derivative of velocity and jerk the derivative of acceleration.
• Forward Time (FORWARD TIME) is a gain factor to transform the
jerk request from the controller function to an acceleration request. Even
though the controller function is based on jerk and sends the desired
jerk requests for the vehicle, the interface to control vehicle motion is
based on acceleration. Hence to control the vehicle the desired value
of acceleration is required. In order to obtain the desired accelera-
tion from the request jerk, one has to look forward for a given time
which is called Forward Time. Mathematically it can be defined by the
formula.
A req = A 0 + J req*FORWARD TIME
A req is the Acceleration request
A 0 is the current vehicle acceleration
J req is the Jerk request generated by the controller function
• Jerk Horizon (J HOR) is a parameter used to determine when the
controller function sends the necessary jerk requests and the required
jerk magnitude in response to an approaching curve. To define what
is “near” and “far” (with respect to the distance from the approaching
curve) for the controller function, the parameter J HOR is used, where
HOR stands for the horizon points (of the electronic horizon) to be
considered. J HOR is always a negative value, and values closer to zero
make the controller respond to the approaching curve when it is further
away with a smaller deceleration demand. Higher negative value tells the
controller to respond when the approaching curve is closer in proximity
but with a larger deceleration. A pictorial representation is given in
Figure 5.9.
The black line represents the target velocity set for the controller and
the reference velocity curve is given in red. As explained previously the
controller tries to control the vehicle speed (in blue) as close as possible
to the reference speed.
The mathematical expression “A MAX + J HOR*time” determines
the funnel of the vehicle velocity curve shape (shown in blue). More
negative J HOR give the velocity curve a sharper shape, while values
closer to zero give the velocity curve a flatter shape.
The range of the variation parameters examined in the tuning task have been
shown in Table 5.1.
88 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.9 Illustration of the design variable (variation) J HOR.

Table 5.1 Range of variation parameters used in the tuning task

Design Variable From To
A MAX (m/sˆ2) 1 5
FORWARD TIME (s) 0.1 2
J HOR (m/sˆ3) –5 –0.2
J MAX (m/sˆ3) 1 3

5.2.3 Key Performance Indicators (KPI)

The output variables to demonstrate the effectiveness of our tuning task to
meet the targets are described below and illustrated in Figure 5.10:
• Mean Speed: The mean of the vehicle speed in each test run is indicative
of the sportiness of the driving experience. A higher mean speed helps
finish the test maneuver in less amount of time, and makes the driving
experience sportier.
• Speed below reference: The reference speed curve is the maximum speed
with which the vehicle (Fiat 500L) can maneuver the digital test track
without leaving the road for the reference use case. Hence to ensure
vehicle safety it was ensured that the vehicle speed during the tuning
task was always below the reference velocity.
 5.2 Demonstrative Example 89

Figure 5.10 Key performance indicators.

• Jerk RMS: Vehicle jerk which is the rate of change of vehicle acceler-
ation, is indicative of the driving comfort. Lower rate of change of jerk
gives a comfortable ride, so the root mean square of the jerk in a test run
is a good indication of the driving comfort.

5.2.4 Test Maneuver

The test maneuver consisted of 5000 m test run on a digitized road imitating
the road between Ceva and Savona in Italy run on IPG Carmaker for Simulink
(CM4SL). IPG Carmaker environment is illustrated in Figure 5.11. The top
left is the Carmaker for Simulink main GUI, showing details about the vehicle,
simulation speed, time and distance of maneuver etc. The bottom left imitates
the car instrumentation. The top right is time based plot of car speed and the
vehicle jerk. The bottom right is the IPG Movie which illustrates the overall
test run in a movie.

5.2.5 Test Run Overview

The test run overview is illustrated in Figure 5.12. The test parametrization
was done in AVL CAMEO, where a space filling DoE design with the four
variations was used. The variations were then uploaded to CM4SL through
90 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.11 IPG Carmaker test environment.

Figure 5.12 Test run overview illustrating the work flow.

the CAMEO-Carmaker Interface, where the test maneuver was run for each
variations setting. AVL CAMEO then stores the measurement parameters
observed as the KPIs for further evaluation.
 5.2 Demonstrative Example 91

During parametrization there were limits set on the minimum (–2 m/sˆ3)
and maximum (2 m/sˆ3) acceptable vehicle jerk values. Whenever the vehicle
jerk value violated the limits the test run at that test point was halted
and no measurements were recorded. This affected the overall DoE design
effectiveness with a reduced design space and as a result reduced measurement
points. To overcome this challenge a COR DoE (Customized Output Range)
method was utilized, which is an iterative method where first alternate test
points were added by CAMEO to maintain the DoE design. Then based on
these preliminary measurements the design space was further modified and
additional test points were added in the relevant variation space to improve
the final information from the measurements. Design space modification. The
AVL CAMEO interface is illustrated in Figure 5.13, where the image to the
left illustrates the overall test parametrization while the image to the right
shows the test run window.

5.2.6 Raw Data Plausibility Check

Before the mathematical modeling of the selected output measured variables,
the raw measurements were checked for plausibility. Firstly, the measured
variables were checked for any outliers as shown in Figure 5.14 for mean

Figure 5.13 Left image illustrates the test preparation window while the right image illustrates
the test run window.
92 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.14 Checking for outliers in the measured variables.

speed. The measured values were within the acceptable range. The figure also
shows that the repetition points (a select number of test conditions, usually the
start condition which are repeated to check the reproducibility of test results)
shown in green were perfectly reproduced.
The effect of design space modification, due to limit violations and the
design correction by COR DoE method can be seen in Figure 5.15. In a
certain range of variations for A MAX, J HOR and FORWARD TIME there
are no test points. Limit violations encountered when tests were carried out at
these range of points are the reason why they were skipped by AVL CAMEO.
Conversely a greater density of test points in certain ranges of variations show
where the COR DoE added alternate or additional test points.

5.2.7 Meta Modelling

The raw data plausibility check was followed by empirical modeling of the
output variables. The automatic modeling in CAMEO gave reasonable results
with a neural networks model with local model order 2, as can be seen in
Figure 5.16 which is the Measured (Predicted) plot which shows the fit of the
model to the measurement points. If there is a perfect match all points will lie
along the black line, but in our case the measurement points are reasonably
close to the black line.
 5.2 Demonstrative Example 93

Figure 5.15 Check of DoE design and the boundaries of variation parameters.

Figure 5.16 Figure depicting the quality of empirical modeling.

After checking the quality of modeling, the intersection plots were used
which represent a cut through the multidimensional model, showing the
influence of each variation depending on the values of the other variations.
In Figure 5.17 the influence of the variation parameters on Speed Mean and
94 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.17 Intersection plot highlighting the influence of each variation on the output
variables and their interaction.

Jerk RMS can be observed. The confidence interval of the model is displayed
in the green dotted line and colored section. The narrow confidence interval
shows a high quality fit. The green bar on the x axis for each variation shows
the total design space, and as the confidence interval of the model in the
extrapolated region is also narrow, it shows good extrapolation capability of
the model. Now looking at the intersection plots, it can be noticed that J HOR
and A MAX have a strong influence on the output parameters. The more
 5.2 Demonstrative Example 95

negative the J HOR, the later the vehicle reacts to an approaching curve.
Hence it is still travelling at a high speed before decelerating to approach
the curve safely. Hence a higher mean speed is observed, but the resulting
braking produces higher vehicle jerk reducing the driving comfort. Influence
of A MAX can be a bit counter intuitive but it can be seen that A MAX is
used to calculate J HOR. The higher A MAX, the less negative is J HOR.
Hence for higher A MAX values J HOR is closer to zero hence a smoother
and slower ride. It can also be observed that higher FORWARD TIME allows
for a smoother and slower ride, which is because the controller can take more
time to achieve the desired acceleration.

5.2.8 Optimization
From the intersection plot, it is possible to manually find values of the
variations which give a comfortable ride or sporty ride or an acceptable com-
promise. But it is quite easy to miss the optimum or an acceptable compromise
when working with multiple input variations, hence the optimization tool in
CAMEO was used. In the current tuning scenario, the target was to be able
to isolate two modes of operation, comfort mode and sporty mode. Hence
a multi objective optimization was chosen with limits set on the minimum
desired mean speed of 115 Km/h and maximum acceptable JERK RMS of
0.28 (Figure 5.18).

Figure 5.18 Optimization setting window in AVL CAMEO.

96 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.19 Trade-off plot between comfort and speed.

The result is plotted in a trade-off plot as shown in Figure 5.19, where

the steel blue is the pareto front, the blue points indicates the measurement
values and the other yellows points are random space filling points. The pareto
front shows the possible optimum trade-off solutions which can be considered
equally good as the only way to improve on objective would be to compromise
on the second objective. So by observing the pareto front it is possible to define
an optimum for comfort mode and an optimum for sporty mode of operation
Table 5.2.
In Figure 5.20: Sporty mode vs comfort mode: the vehicle performance
when operating at the two modes can be observed. The red velocity curve is
the reference velocity and blue velocity curve is the actual vehicle velocity.
It can be observed that the actual velocity is always below reference velocity
which was the safety requirement. Also the velocity changes in comfort mode

Table 5.2 Variations values for comfort and sporty mode

A MAX FORWARD TIME J HOR J MAX SPEED Mean JERK RMS
Comfort 4.99 1.94 –0.84 1.0 115 0.09
Sporty 3.88 1.37 –1.84 3.36 120 0.28
 5.2 Demonstrative Example 97

Figure 5.20 Sporty mode vs comfort mode.

is more gradual with no sharp peaks unlike in sporty mode where there are
rapid fluctuations in vehicle velocity. This behavior is also mirrored in the
acceleration values in both operation modes. The vehicle jerk curves (red plot
is the jerk request generated by the controller and blue the actual vehicle jerk
response) show much lower values in vehicle jerk for comfort mode while the
sporty mode show sharp and frequent peaks in jerk value.

5.2.9 Verification
The pareto front consists of points a majority of which are from the model
extrapolation. In order to verify the robustness of the model to accurately
extrapolate, ten random points were selected from the pareto front and for the
corresponding variation values the test runs were rerun. The results from these
test runs were evaluated as verification points in CAMEO. The Figure 5.21
shows the extrapolated model (in red) and its prediction interval (in blue), and
the measured verification points and its modeling (in green). The measured
verification points lie within the prediction interval of the model, showing the
extrapolation accuracy of the model.
98 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.21 Verification plot to see how well the measured results from the verification run
fit the model results.

5.3 Model-based Validation

Once the reference tuning task is completed, it has to be tested, if the tuning
results are still acceptable, when not running the reference use case but
for varying road characteristics. Will the comfort mode still allow for a
comfortable drive also for different road situations? It would be unfeasible
to run simulations on thousands of different roads, besides making it difficult
to realize the influence of a specific road. In the current method the two tuning
 5.3 Model-based Validation 99

modes are fixed and a system variation of a digitized road is performed using
the model based approach to validate our tuning results.
The digitized road is shown in Figure 5.22, where the lengths of the straight
sections (L1, L2, L3, and L4) and curvatures (R1, R2, R3) were varied while
keeping the total maneuver length to 5000 m. The controller settings were
fixed to run at first comfort mode and then sporty mode, and the resulting
measurement output variables are shown in Figure 5.23.

Figure 5.22 Digitized road used for the validation run.

100 Tuning of ADAS Functions Using Design Space Exploration

Figure 5.23 Measurements comparison when run on comfort mode (in blue) and sporty
mode (in red).

It can be seen in Figure 5.23: Measurements comparison when run on

comfort mode (in blue) and sporty Mode (in red) that for the sporty mode the
resulting drive comfort is lower as indicated by the higher JERK RMS. The
length of the straight portions do not influence the JERK RMS for comfort
mode as strongly as in the sporty mode. The curvature of the turns seem to
influence the output in both the operation modes. A JERK RMS limit of at
least 0.35 is expected, and it can be seen that the limit is maintained in both
the modes of operation for majority of the design space. In the sporty mode
the controller is set to maintain a higher vehicle speed and responds to the
oncoming curve only when it is close, hence the longer the straight sections, the
larger the jerk experienced when it decelerates rapidly to approach the curve
followed by a strong acceleration on leaving the curve. For the comfort mode,
the controller is set to focus on keeping the vehicle jerk close to minimum.
The validation task showed that, if the function (our UUT) is kept constant and
the simulation environment is changed, the function still manages to meet the
expected vehicle jerk targets. The influence of ‘L4’ on the jerk behavior needs
to be further investigated as it strongly increases the vehicle jerk fluctuations at
higher values especially for the sporty mode. To further explore and investigate
 References 101

the influence of test track characteristics on the function response, it can be

tested on a variety of road types and test tracks. This assists in the further
improving the function performance.

5.4 Conclusions
Virtual tuning of an ADAS function developed on a MiL environment using
an optimization tool can be a powerful combination for the development of a
brands driver assistance system. The classical approach relies on a subjective
tuning of the ADAS function on a proving ground and public roads, which can
be supported and accelerated by using a virtual tuning environment. Using
DoE methods supported by AVL CAMEO, it was possible to increase the
number of tuning tests compared to a manual tuning, and also the number of
target parameters and tests needed to match them. The possibility to use the
developed function for alternate use cases by separating the software and the
tuning data is precondition for tuning works in general.
Independent of that also in the validation process a model-based approach
can be very helpful, as the test coverage for a certain use case can be extended
to a wide range of possibly occurring variants of that use case. The robustness
of the key performance indicators considered as relevant can be estimated.

Acknowledgement
We would like to thank Mr. Andreas Saroldi from CRF for providing theADAS
function.

References
[1] M. Paulweber and K. Lebert: Instrumentation and Test Systems: Power
Train development, Hybridization and Electrification. Chapter 5.4.2
Application Data management, Springer View, 2016.
[2] AVL Tuning Data Management Software: http://www.avl.com/creta
(called at 2016 01 20).
[3] H.-M. Koegeler; A. Fürhapter; M. Mayer; K. Gschweitl: DGI-Engine Cal-
ibration, Using New Methodology with CAMEO. In SAE NA, Capri –Italy,
23–27. September 2001.
[4] E. Castagna; M. Biondo; J. Cottrell; H. Altenstrasser; Ch. Beidl;
H.-M. Koegeler; N. Schuch: Multiple Tier 3 Engine Applications based
on global modelling. In MTZ 6/2007.
102 Tuning of ADAS Functions Using Design Space Exploration

[5] T. Fortuna; H.-M. Koegeler; M. Kordon; V. Gianluca: DoE and Beyond-

Evolution of the Model based Development Approach, in ATZ world-
wide, Springer, 2015.
[6] H. M. Koegeler; B. Schick; P. E. Pfeffer; A. Contini; M. Lugert;
T. Schöning: Model Based Steering ECU Calibration on a Steering in
the Loop Test Bench, in Chassis Tech 2015.
[7] D. C. Montgomery: Design and Analysis of Experiments, John Wiley
and Sons.
[8] A. Rainer; H. M Koegeler; D. Rogers: Iterative DoE – Improved
emission models and better optimization results within a shortened
measurement time, in PMC, 2014.