You are on page 1of 15

IEEE TRANSACTIONS ON HUMAN FACTORS IN ELECTRONICS, VOL. HFE-8, NO.

3, SEPTEMBER 1967 187


REFERENCES monitor," in An Investigation of the Visual Sampling Behavior of
Human Observers, NASA Rept. CR-434, Appendix I, April 1966.
[1] J. I. Elkind and C. D. Forgie, "Characteristics of the human [5] W. B. Davenport and W. L. Root, Random Signals and
operator in simple manual control systems," IRE Trans. Auto- Noise. New York: McGraw-Hill, 1958.
matic Control, vol. AC-4, pp. 44-55, May 1959. [6] J. W. Senders, J. I. Elkind, M. C. Grignetti, and R. D. Small-
[2] D. T. McRuer and E. S. Krendel, "The man-machine con- wood, "An investigation of the visual sampling behavior of human
cept," Proc. IRE, vol. 50, pp. 1117-1123, May 1962. observers," NASA Rept. CR-434, pp. 22-26, April 1966.
[3] J. W. Senders, "The human operator as a monitor and con- [7] Ibid., pp. 27-33.
troller of multidegree of freedom systems," IEEE Trans. Human [8] R. J. Massa, "The role of short-term memory in visual in-
Factors in Electronics, vol. HFE-5, pp. 2-5, September 1964. formation processing," Proc. Symp. Models for Perception of
[4] R. D. Smallwood, "Some models for the human instrument Speech and Visual Form, Boston, Mass.

An Experimental Study of Human Operator Models


and Closed-Loop Analysis Methods for
High-Speed Automobile Driving
WALTER W. WIERWILLE, MEMBER, IEEE, GILBERT A. GAGNE, MEMBER, IEEE,
AND JAMES R. KNIGHT, MEMBER, IEEE

Abstract-The investigation described in this paper was aimed an example with which most people have had first-hand
at obtaining information about the way in which the human being experience. Yet, when asked to be precise, man-machine
controls an automobile at high speed on a winding road. A dynamic researchers are able to say surprisingly little about just
systems approach was taken, and three different classes of models
were postulated. The optimal models in these classes were then
how the human being does control an automobile. Auto-
obtained for data taken from three subjects in a driving simulation mobile driving, because of the multiplicity of visual cues
which projected a large moving roadway image. that are available to the driver and the subtleties in-
The three subjects were chosen in such a way as to make possible volved in the utilization of these cues by the driver, does
1) a comparison of driving behavior and compensatory tracking not fall within the realm of the usual man-servo analy-
behavior of subjects with similar backgrounds, and 2) a comparison
of driving behavior of a male subject with technical training and sis techniques so often used for aircraft manual control
driving behavior of a female subject without technical training. systems. Consequently, the usual control system methods
This behavior was quantitatively defined by synthesizing optimal do not appear to be applicable.
models from the experimental data, using the digital computer. The Previous work involving human control of the auto-
models were then used to perform close-loop man-machine system
servo-analyses of the automobile driving tasks and the correspond- mobile is sparse; however, work on related subjects such
ing tracking tasks. as steering response, suspension design, and anthropome-
It is clearly demonstrated that, under idealized conditions, try is being carried out at many institut-ions. In the field
standard control system techniques may be extended to make of human control, Senders and his associatesE11 develop
possible the quantitative study of the control of an automobile and experimentally verify models of the visual informa-
at high speed by a human operator.
tion content required for driving. Gordon 2' describes re-
sults of experiments which determine eye fixation posi-
I. INTRODUCTION tions on the road ahead. Biggs[3' describes a theoretical
OR MANY YEARS, man-machine researchers study of the visual presentation of the driver, but he
and engineers have been interested in the automo- presents no experimental results. Wohl'4] uses differen-
bile and the way in which a human being controls tial equation models of the vehicle and then postulates a
it. Automobile driving is repeatedly cited as a good ex- hypothetical model of the human's loop closure based
ample of a man-machine system, and the automobile is on a formula for alignment of the vehicle with the road;
again, no experimental results are reported. Finally,
Sheridan[51 postulates a group of hypothetical models for
Manuscript received October 15, 1966; revised May 8, 1967. This various "preview-type" man-machine control systems.
research was supported by the Cornell Aeronautical Laboratory. He then goes on to describe computer solutions that
The authors are with the Cornell Aeronautical Laboratory,
Buffalo, N. Y. closely imitate the traajectories of man-controlled ve-
188 IEEE TRANSACTIONS ON HUMAN FACTORS IN ELECTRONICS, SEPTEMBER 1967
hicles. In the driving case, his work appears most ap- equations of motion of an automobile. His work could
plicable to obstacle avoidance. be used in future studies involving refinement of human
The study described in this paper was aimed at deter- operator models in driving.
mination of models of human operators that are known Three different models were postulated for representa-
to be experimentally accurate in driving tasks. The study tion of the driver in the driving task. Since experiments
also was concerned with the use of these models in an were limited to forced-pace driving, time and distance
analysis of the closed-loop man-vehicle system, since so were linearly related. Thus, all models could be written
little is known about this type of loop closure. This study with time as the independent variable without loss of
differs from previous studies in that optimal linear con- generality. The first model consisted of the usual man-
stant-coefficient driver models were computed from ex- servo tracking model. It operated only on the present
perimental data, and in that the experimental models and past values of the error signal. Effectively, this
were used to examine loop closure.' model neglects the fact that the driver views the road
ahead, and it simply attempts to synthesize the driver's
II. EXPERIMENTAL APPARATUS AND PROCEDURE steering wheel commands from the vehicle's error signal.
The experimental data were gathered using simula- Thus,
tion rather than road test. Although an instrumented
automobile was available for research of the type de- rt
scribed herein, it was believed that environmental fac- 01(t) = e(X)g1p(t - X) dX (2)
tors could be controlled more effectively in the labora-
tory. Moreover, simulation possesses economic advantages or equivalently,
when control of traffic on a normal highway or prepara-
tion of a test track is taken into account. The driving 401(t) = 91,(X)e(t - X) dX (3)
simulation used in gathering the data is described in de-
tail elsewhere in this issue.17] The simulation incorpor- where +l (t) is the optimal approximation of the driver's
ates a projection system capable of displaying large steering wheel deflection in degrees for model 1, e(t) is
roadway images for viewing by the driver or subject. the lateral vehicle displacement error in feet, and gl,(t)
Because of the multiplicity of tasks and variations in is the impulse response of the optimal model. The func-
parameters encountered in normal driving, the experi- tion glp (t) is postulated to operate only on the past er-
ments performed had to be suitably limited and idealized. ror signal and is therefore physically realizable.
This paper describes the results obtained for high-speed, The second model (which has been suggested previ-
winding-road driving. A forced-pace speed of 100 ft/s ously[51) accounts for the driver's preview by including a
was chosen. This speed is equivalent to 68.2 mi/h, which second term which operates on the future roadway wave-
on most automobile speedometers would register at 70 form or signal. In this case, the model optimally proc-
mi/h, because of conservative calibration. At each given esses the error signal into the past and the roadway
instant of time, the roadway was displayed up to 450 ft waveform into the future in attempting to match its out-
ahead of the simulated vehicle. The automobile dy- put to the steering wheel deflections produced by the
namics were simulated using Wohl's small angle re- driver:
sults: [4]
Xo(s) = 0.185 (1)
1 02(t)= e(X)92p(t -X) dX + f Xi(X)92f(t -X) dX (4)
1(] s) - -cot

where Xo (s) is the Fourier transform of the lateral dis- or equivalently,


placement of the vehicle in feet, 1D (s) is the transform of co o
the steering wheel deflection in degrees, and s is the
Fourier transform independent variable. This equation 2(t) g92,(X)e(t - X) dX + f 92f(X)xi(t -X) dX (5)
neglects factors which are normally associated with ve- where 2 (t) is the optimal approximation of the driver's
hicle dynamics, but the equation does account for the steering wheel deflection in degrees for model 2, x+(t) is
basic steering properties of the automobile.2 Segel[8] has the input roadway lateral displacement in feet, and g2p (t)
developed and experimentally verified rather detailed and g2f(t) are the impulse responses of the two parts of
1 In this paper, an optimal model is that model, selected from the optimal model. The function g2p(t) is physically
a general fixed-form class, which produces minimum integral realizable and operates only on the present and past
squared error in human operator characterization.j6l values of the error signal; g2f(t) is physically unrealiz-
2 Wohl's analysis actually specifies a higher gain than that
used able and operates only on the present and future values
in (1). However, when this higher gain was used in the simula-
tion, unrealistic results were encountered; the simulated vehicle of the input roadway waveform.
was difficult to control. The gain was then reduced, and a realistic
degree of controllability was obtained. Perhaps the discrepancy The third model is similar to the second, because it
lies in the neglect of such influences as tire bounce and flexure, as attempts to account for the driver's preview of the future
well as steering play, which might cause a reduction in steering
effectiveness or gain. roadway waveform. However, model 3 attempts to ac-
WIERWILLE et al.: CLOSED-LOOP ANALYSIS 189
count in addition for the vantage point of the driver by tion of the experiment. The first two subjects (A and B)
subtracting the present vehicle displacement xo(t) from were males with technical training (engineering), each
all future values of the roadway signal. In other words, of whom had been driving for approximately 10 years.
the unrealizable portion of the optimal model operates The third subject (C) was a female without technical
on the future roadway waveform with zero reference of training who also had been driving for approximately 10
that roadway taken as the vehicle's present lateral po- years. However, subject C's total driving time was much
sition. Thus, less than that of subjects A or B. Although for the pur-
t
pose of determining levels of significance a larger group
=fe(X)93,(t
+3(t)-J - X) dX of subjects would have been desirable, the amount of
computation required would then have been beyond the
practical economic limitations.
+ c [Xi(X) - XO(t)]93f(t -X) dX (6) Two different experiments were performed with the
or equivalently, three subjects. The general purpose of the first experi-
ment was to- determine the modifications that must be
made in transferring human tracking technology to the
03(t) = f q3p(X)e(t - X) dX analysis of the driving task. The general purpose of the
second experiment was to test the ability of the analysis
+ 93f(X)[Xi(t - X) - xo(t)] dX (7) methods to discern important differences in driving be-
havior. The first experiment, using only the two male
where +3(t) is again the optimal approximation of the subjects, was aimed specifically at determining the dif-
driver's steering wheel deflection in degrees for model 3, ferences between high-speed driving on a winding road
xo(t) is the vehicle lateral displacement in feet, g3p(t) (without obstacle avoidance) and compensatory tracking
is the realizable, and g3f(t) the unrealizable part of the in exactly the same system without preview. Subject A
two-part filter. performed the following sequence of tasks.
All models were computed on a high-speed digital com-
puter using the constant coefficient portion of the char- 1) The subject was instructed to drive in the simula-
acterization theory described.[6] In computing models 2 tion with the full preview of the roadway displayed. The
and 3, the optimization process for the realizable and instructions made it clear that the subject was not being
unrealizable portion was carried out jointly. Thus, the graded on accuracy. Instead, the subject was to try to
values of g2p1(t) and g2f(t), for example are chosen as an transfer his usual driving skills to the simulation by driv-
optimal combination and are dependent on each other ing down the right-hand lane of the simulated two-lane
for making the errors small. It must not be assumed highway.
that for the same driving data the realizable parts of the The subject performed the task during a 10-minute
three models are equivalent. practice session. After a 2-minute rest period, a 2-minute,
Subject data were taken in analog form using an ac- 24-second data-taking session followed. However, actual
curate chart recorder. These data were then converted model computation from the data began at the end of the
to digital form by means of a manually-operated tele- first minute of the data-taking session in order that a
reader machine. The data were read into the digital com- steady-state model might be obtained.
puter and then automatically plotted back out to insure 2) All preview was removed from the display by block-
the correctness and accuracy of the conversion. ing the light from the projector. Only the present value
Computation of all models was based on the use of 84 of the centerline appeared on the screen. (The sidelines
seconds of simulation data, sampled at a rate of 20 per and roadside posts were turned off electronically.) The
second. The first and last 12 seconds of these data were subject was then instructed to track the simulated left-
consumed in avoiding initial and final condition tran- side hood ornament to the spot on the screen. In this task,
sients in the computed results. Because models 2 and 3 the subject was told that he was to minimize the error in
operate on future roadway waveforms and past error aligning the hood ornament with the spot on the screen.
waveforms, final conditions at the end of the data run The same type of practice and data-taking session
must be taken into account in the same fashion as initial procedure was followed as that used in task 1.
conditions at the beginning.
The optimal models computed by the digital computer Subject B performed the tasks in reverse order so that
were plotted automatically by means of subroutines gradual learning and fatigue effects might be minimized
which called for the use of a special digital plotter. The to the extent possible with only two subjects.
use of this equipment makes possible the direct plotting The second experiment was aimed at determining
of the various waveforms used for computation of a given differences in driving behavior between one of the male
model, the step response and impulse responses of the subjects with technical training and the female subject
model, and the transfer function of the model. without technical training. In this experiment subject C
Three subjects were chosen for the data gathering por- performed task 1 above. Then the task 1 results of sub-
190 IEEE TRANSACTIONS ON HUMAN FACTORS IN ELECTRONICS, SEPTEMBER 1967
jects A and C, which were taken under identical circum- able portions of the optimal models, while Fig. 10 is a
stances, were compared.3 plot for the unrealizable portion of an optimal driving
It is important to note that, in general, drivers vary model.
in their choice of a lateral position on a highway. Some A number of important patterns become evident upon
will drive with the car near the centerline while others careful'examination of the numerical results of Table I
drive with the car centered in the lane or nearer to the and the graphical results of the various plots. These pat-
right. Also, in the tracking task (task 2), the position of terns and possible explanations for them are the subject
the subject's head will determine the apparent desired of the remainder of this section.
zero-error position of the system. To avoid mainfesta- The runs computed for subj ects A and B show that
tions of these effects in the model computation, it was these subjects would tolerate larger vehicle lateral de-
decided to remove the mean values in the roadway wave- viations while driving than they would while tracking
form, the vehicle lateral position waveform, and the (column 4). Apparently, when looking down the high-
vehicle lateral error waveform. The removal of these way, these subjects were less concerned with their pre-
mean values makes possible the comparison of dynamics cise road position than they were while tracking. Subject
between runs in such a way that a subject's choice of C, who was not technically trained, produced smaller
zero position is unimportant, provided that it is consistent vehicle errors while driving than did subject A. Indeed,
during the data-taking period. subject C's driving performance, if based on error alone,
was superior to both the driving and tracking perform-
III. RESULTS OF THE EXPERIMENTS
ance of subjects A and B.
A great amount of computer output data were ob- A possible explanation for subject C's smaller error is
tained for the five simulation data runs. In addition, 51 that, since this subject was introducing more lead equali-
plots of important aspects of the models were drawn by zation into the system than subjects A or B while driv-
the computer. It clearly becomes necessary, therefore, ing, her errors would have a tendency to be smaller. T'his
to reduce these data to concise form for the presentation. lead would be the result of her slight oversteering of the
Table I contains numerical information regarding the vehicle. (The greater lead is evident in both the complex
raw data, the models, and the closed-loop man-machine frequency response of the models representing subject C
systems for all of the runs in this experiment. The col- and in the peak values of the step responses of the
umns have been numbered for ease in referencing the models. See Fig. 9 and column 7 of Table I.) Thus, it is
following discussions. Columns 4 through 8 were com- not difficult to explain the difference in error for the
puted by the digital computer, or were obtained by direct driving task between subjects A and C. Explanation
observation of computer plotted results. Columns 9 of the lower error in driving for subject C as compared
through 15 were computed using standard servo methods with tracking of both subjects A and B is more difficult.
on the computer results. Bode plot analysis produced Here, it is possible that subject C was using the previewed
the results in columns 9 through 11, while Nichols chart signal to overcome her own reaction time lag and that she
analysis produced the results in columns 12, 14, and 15. tried to minimize vehicle error even though instructed
Use of these methods made it unnecessary to fit low- otherwise. Since no preview was available for task 2
order human transfer functions to the experimentally ob- (tracking), subjects A and B were unable to overcome
tained amplitude and phase plots. their own reaction time lag as effectively. Consequently,
In addition to the numerical data of Table I, repre- A and B had larger errors in tracking than C had in driv-
sentative computer plots have been included in this ing. Other factors tend to support this explanation. First,
paper. Figs. 1 to 3 show the signals as plotted by the column 8 shows that for tracking, subjects A and B had
digital computer for given data runs and model solutions. delayed or nonminimum phase responses, indicating a
Figs. 4 to 6 exhibit the step responses and impulse re- reaction time lag. Subject C did not have any apparent
sponses of both the realizable and unrealizable portions effects of reaction time in her model responses. Secondly,
of the optimal models.4 Finally, Figs. 7 to 9 are computer a previous investigationt91 has shown that in some cases
plots of the gain and phase characteristics for the realiz- a prediction display may be used to reduce tracking er-
rors. In conclusion, then, the possible explanation is
3 For both tasks and for all data runs, precisely the same road-
that, although all three subjects overcame their reaction
way signal was used. All practice sessions were performed with time lags in the driving task, only subject C weighted
a second roadway signal with identical statistical characteristics. error heavily. She was, therefore, able; to take advantage
4 In Figs. 4 to 6, the reader is cautioned not to confuse these
model impulse and step responses with human responses to an of the preview and to reduce her error below the non-
impulse or step. These models have been obtained for humans preview tracking error level.
driving or tracking a winding road; therefore, they do not cor-
respond to human responses to steps or impulses. Also, in Figs. Column 6 provides information about the goodness
4 to 6, "past" response refers to response of the realizable portion of fit of the three models. The numbers in this column
of the optimal model, and "future" response refers to response
of the unrealizable portion. Finally, in Figs. 4 and 6, the zero represent a ratio: the integral squared value of the error
point in time is at 2.5 seconds for the future responses. In other between the model output and the steering wheel signal
words, the correct value of running time for future responses is
obtained by subtracting 2.5 seconds from each value of the abscissa. is the numerator, and the integral squared output of the
WIERWILLE et al.: CLOSED-LOOP ANALYSIS 191
steering wheel signal is the denominator. The ratio is driving models such as model 2 are more accurate than
multiplied by 100 for presentation as a percentage. It is the usual servo models for human tracking.
clear from this information that model 2 produces the Columns 7 and 8 show that considerable difference
least error for all three sets of driving data. As might exists between realizable portions of the step responses
be expected, model 2, which accounts for preview, is of driving models and the step responses of tracking
superior to model 1. However, contrary to intuition, models. First, as mentioned previously, the tracking
model 3, which additionally accounts for the operator's model responses are initiated with a hesitation which may
vantage point, is inferior to model 2. Since model 2 is be either an approximate delay or a nonminimum phase
both simpler in composition and more accurate in char- (negative going) initial response. In contrast, none of the
acterization than model 3, closed-loop analyses were not driving models exhibits this characteristic; there is an
performed for model 3. immediate positive-going response of every driving model
The errors which model 2 produces in characterizing to a step input. Secondly, the peak values of the step re-
the human operator in this driving simulation are defi- sponses are much smaller for driving than for tracking.
nitely within tolerable limits for closed-loop analysis. Step response peak values are indicative of the amount of
Model 2, when used for characterizing the human oper- lead compensation introduced by the operator if his
ator in the driving task, actually has smaller errors than reaction time lag is considered as part of the. system fol-
model 1 when used for characterizing the operator in the lowup dynamics. Thus, a much greater amount of lead
tracking task with the same system dynamics. Thus, is used in tracking, but the effect of this lead on the

TABLE I
NUMERICAL RESULTS OBTAINED FROM THE DRIVING AND
TRACKING EXPERIMENTS PERFORMED

1 2 3 It 5 6 7 8 9 10 11 12 13 14 15
192 IEEE TRANSACTIONS ON HUMAN FACTORS IN ELECTRONICS, SEPTEMBER 1967

MODELING ERROR tNHEEL POSITION - MODEL OUTPUT)

cn
Gi
.,, A (.s 'tl- /\ 'Jr-It- -. J"\ ., /'
go0 M,ODOUPU 2.V A \1 2. 'k 4
V-41 V -
iv li
1

sa r*
5

MODEL OUTPUT

STEERING WHEEL RNGLE

RORDWRY

SECONDS

VEHICLE ERROR
in

-- --,
- .1.. 5i au a 7 32 __
4. 40 5A
SECOWNDS

Fig. l. Computer plot of waveforms associated with the computation of optimal model
no. 2 for subject A in the driving task (task 1).
al.: CLOSED-LOOP ANALYSIS
193
WIERWILLE et

MODELING ERROR (WHEEL POSITION - MODEL OUTPUT)

A t (\PV4LAAKAA I ~WiA AA A MAr.[NA,A A

MODEL JUTFUT

A Af IA iMA

STEERING WHEEL RNGLE

flf2gf\I

VEHICLE ERROR

Fig. 2. Computer plot of waveforms associated with the computations of optimal model
no. 1 for subject A in the tracking task (task 2). (The roadway waveform was identical
to that shown in Fig. 1.)
194 IEEE TRANSACTIONS ON HUMAN FACTORS IN ELECTRONICS, SEPTEMBER 1967

MODELING ERROR (WHEEL POSITION - MODEL OUTPUT)

MODEL OUTPUr

STEERING WHEEL RNGLE

SECONDS

RORDWRY

SECONDS

VEHICLE ERROR
in

84
SECONDS

Fig. 3. Computer plot of waveforms associated with the computation of optimal model
no. 2 for subject C in the driving task (task 1).
U
FUTURE IMPULSE RESPONSE

0.4 0. . 0.3 1.0 0.1 1.2 1.3 1.21- 1. 1.7 0.3 2 ..


0.2
0.1
:Q 03.
SE Sz

FUTURE STEP RESPONSE


u-

I' 14 1s 1'7
17 13I 13' 2.
2.0 .
'1 2.. 2. -22.4 2.3
0 0t 0'. 1'.
1.1 S 0.2 LLJ~~~~~~~~~~~~~~~~~~~~~~~~~~.
0.3 0.4 0.' 0,. 0.7 0.3 I1.0...2.3
'S
1.0 1.1 ~~~~~~~~~~~~~~~~~~~~~~SECONDS

PAST IMPULSE RESPONSE

2.2 2.4 2.!


0.1 0 .2 0.3 04 0.1 03 -0 0. .'. . '. '. 1 13 20. 2.1
O0. S2.3

PAST STEP RESPONSE

2.2.3 2.4
0.3 1.0 1.1 1.2 t.4 1.5 1.3 1.7 1.3 1.3 2.02
0.2
0.10.0 0.3 0.4 0.S 0.3 0.7 0.9 t.3
SECONDS

Fig. 4. Impulse and step responses of optimal model no. 2 for subject A in the drivingto task (task 1). (Note that the future
responses begin at 2.5 on the abscissa and move into the future the left.)

cc
to
PAST IMPUL5E RESPONSE

V
m
H

;r
Q
n
0

PAST RESPONSE
S
SCALE- z0
m
0

z
Q
Ii.;;
zF3
CD
Q
0
H
8 -L.1- 0
Q
aj .1 c a. o). o.4 0.5 0.5 0.7 0.5 0.9 1.X 1.1 1.2 13 1. .5 X6 .7 ..0 19 2.0 2. t 2.Z 2.3 2.4 2.5
SECONDS C12
0
H4

Fig. 5. Impulse and step responses of optimal model no. 1 for subject A in the tracking task (task 2).
FUTURE IfWULSE RESPONSE

I°n.. -21 h 0 1.5


,1Ou.S 1.2
H I
a1.1 *0 1.2 1.4 1.0 1.5 1.? 1.3 1.3 2.0 2.1 is LI 2.2
SE~ ~ ~ ~ ~ ~ ~3CONDS
0
0
FUTURE STEP RESPONSE
0
t'd
0

.
o eI.I _
0 io~~~~~~~~~~~~-
01 01.2 0.. 0.4 0.5 0.3 0.1 0.3 0. 1. 0 1.1I . 1. 1.4 1e. 1. 1.7 1.3 1. i .1 L L2 2.2 2.4 .s
SCCNDS z
02
t:f

PRST IMPULSE RESPONSE

SECND

SCFLE1 PRST STEP RESPONSE

G..

U-

0.0 0 L2 0.5 04 .5 s 0. 0.
070 0.3 1. A 1.0 1.I 1.2 1.3 1.4 1.5 1.6 1.7 1.3 1.3 2I 2I 2.2 2.5 2.4 2.l
SECONDS
Fig. 6. Impulse and step responses of optimal model no. 2 for subject C in the driving tasks (task 1). (Note that the future
responses begin at 2.5 on the abscissa and move into the future to the left.)
198

aAI

I~
IX

GAIN
PHASE

I
20 i PH~ ASEIO
IEEE TRANSACTIONS ON HUMAN FACTORS IN

20

30
- I

III T
ELECTRONICS,

la - - - - -- -- - - ----- - - -- -7 - ---- - --- ,-- - -- -- -S-D


--

II IIII
SEPTEMBER

IIII
1967

II 27

-lo - - ----- --- r -- - r - - Y-r ~-


Z

-20 - - -- - ----- r- - r - - --- - r - -- ----- -\- - to


GAIN!

--------- --------- --------- -------- . -[an

-4a
I
.
II
7---------r--------- -30 - - - - - -- - r- - ir - - - - -- --- r- -------- r t- ~--27
L \
1 . II III 1111
-Z -I a -3 -2 -1o 2
LOGRYITHM OF FREQUENCY IN RRD/SEC LOGARITHM OF FREQUENCY IN RRD/SEC
Fig. 7. Gain and phase plot of the realizable part of optimal Fig. 9. Gain and phase plot of the realizable part of optimal
model no. 2 for subject A in the driving task (task 1). model no. 2 for subject C in the driving task (task 1).

.0 - , , 111
-,-- - I - -

c F - - - - - - -r
- - - - X- - - - - - - -r
- - -\i 9Ss
- - - - -

,0 ~~~~n

g.L- LORRTH EF FfE.EC PHASE25E

es,f r o er-s
LOGARITHM OF FREQUENCY IN RAD/SEC

Fig. 8. Gain and phase plot of optimal model no. 1 for subject A Fig. 10. Gain and phase plot of the unrealizable part of optimal
in the tracking task (task 2). model no. 2 for subject A in the driving task (task 1). (Written
as Ger(-jw).)

systedois reduced somewhat by compensation for reac- show convincing evidence that the bandwidth for a track-
tion time lag. ing system is considerably greater than the bandwidth
Columons 9 to 15 reveal a great deal of information for driving when all parameters except display preview
about the closed-loop man-machine system for driving as and subject instruction are held constant. For subjects
compared with thracking. First, the results of closed-loop A and B, tracking bandwidth (to -3 dB) is approxi-
analyses for models 1 and 2 do not differ significantly mately twice that for driving. Even subject C, who
from each other for any given driving data run. Thus, placed heavy emphasis on lateral error during the driving
the inclusion of the unremalizable filter in the overall task, did not attain the bandwidth associated with the
driving model aids in reducing the modeling error, but tracking tasks. However, note that C's bandwidth was
does not seriously change the results obtained for a greater than either A's or B's for t.he driving task; this
closed-loop -analysis. It should. be noted that, in the use greater bandwidth is probably another indication of C's
of model 2, only the realizable part of the model is inside stronger emphasis on reducing the errors.
the control loop. Thus, the unrealizable part of the mnodel No clear-cut pattern exists in the computation of
may be temporarily neglected when performing a closed- maximum frequency-response amplitude, Mp (column
loop stability analysis. 12), or in the approximate computations of damping
The maj or difference between the closed-loop man- ratio, (column 13). However, all of these values are in
machine driving system and the closed-loop man-machine an interval which implies reasonably good man-machine
tracking system is in bandwidth. Columns 11 and 15 closed-loop stability.
WIERWILLE et al.: CLOSED-LOOP ANALYSIS 199
Table I indicates that subjects A and B reacted in previous closed-loop man-machine analyses. There is,
much the same way to each task. Since the task order for nevertheless, a definite need to describe the human oper-
B was the opposite of that for A, it may be concluded ator's action in the high-speed driving task by polynom-
that gradual learning and fatigue did not cause signifi- ial-form transfer functions of reasonably low order. These
cant changes in the results of the experiment. transfer functions might serve as vehicle design aids and
Observation of the various representative plots in Figs. could also be helpful in the development of automatic
1 to 10 makes it possible to obtain additional experi- highway driving equipment. In the latter case, transfer
mental results. The results discussed in regard to the functions of human operators in the same driving situa-
plots included in this paper were also found consistently tion could be copied in the automatic equipment, thereby
in the computed plots which are not included in this paper. resulting in closed-loop operation that simulates the
Figs. 1 and 3 show that the unrealizable portions of human driver and produces riding qualities with which
the optimal models for driving weight future roadway the passengers are familiar.
signals up to approximately 3 seconds into the future. Polynomial transfer functions were fitted to the driv-
Thus, some input information is derived from the first ing and tracking models of subject A. Emphasis has been
300 feet of roadway in front of the, simulated vehicle. placed primarily on the accuracy at frequencies between
Since no sharp peaks exist on the impulse response plot, 0.5 and 5.0 radians per second. These models would also
it may be concluded that the subjects did not process in- fit subject B without appreciable error within this fre-
formation from any single point in front of the simulated quency band. For driving, the transfer functions are
vehicle. It is possible that the subjects did focus attention
on an area a certain distance down the road; however, 0.412(1 +
their processing of this information would then neces- G2p(s) =(10~[ .10 ±(602 (8)
sarily include a spreading or interpolating effect which f1 + ti1 +
rlae s +
would cause more uniform weighting of the future road-
way. It is also interesting to note that the realizable for the realizable portion and
portions of the optimal models contribute heavily to the
total model output. Thus, error signal remains important
in obtaining a good model. 0.541(1 + 440)(-s)(1 0 375)
The steering wheel signals for driving and for track- G2f(S) = .0 0.375 (9)
ing are quite different from one another. Motions of the - -2(0.80) 8 1521
steering wheel are fewer in number for driving than for
tracking. However, in the driving task, subject C's steer- for the unrealizable portion.5 Note that it is unnecessary
ing wheel motions have sharper leading edges and are to include a delay or nonminimum-phase factor in G2, (s)
more numerous than subject A's motions. since the human operator overcomes reaction time lag
Figs. 7 to 9 verify the conclusions reached earlier re- by using the previewed roadway signal. The unrealizable
garding the lead that is introduced by the subjects- in portion is written with all poles and zeros in the right-
driving and in tracking. In addition, these plots show half plane or at the origin, except for one zero. This
that certain small "antiresonances" exist in the frequency network must be nonminimum phase in the unrealizable
response plots as seen in Figs. 7 and 9. It is difficult to sense in order that the larger phase shifts at higher
ascertain whether these are simply a result of errors in frequencies may be taken into account. Equations (1)
characterization of the operator or whether the operator and (8) may be used to compute the realizable portion
actually exhibits a type of slight antiresonance. These of the closed-loop transfer function of the man-machine
antiresonances were found in most of the driving models driving system. The resulting expression after factoring
and in none of the tracking models. is

Xo(s) (1 + O--,o) (10)


Xi(s) - (1+ 0 +2(0.434) + s2 ][±2(0.67) +± 2 1
0.115/ 0807(0.807) 6.6 (6.6)2
IV. LOW-ORDER TRANSFER FUNCTION MODELS FOR HIGH- if the unrealizable portion of the system is neglected.
SPEED DRIVING However, neglecting the unrealizable portion does not
All of the previously described analyses have been per- change the pole locations. Fig. 11 shows a block dia-
formed in such a way that it has been necessary to obtain
transfer functions of the human operator ipoy inpolynomial
malthat he cpolnYi°omial
tha the mapproxi(mation
condition lim.,o has Ga2(S()
G2f(s) = 0 for chosens
has beenThis
been satisfied. condi-
form. Thus, the source of error which results from the tion, when satisfied by the unrealizable portion of the model, insures
approximation has not been introduced in the ~isthat
polynomialapproximation
polynomial the zero-frequency gain of the total closed-loop driving model
unity.
200 IEEE TRANSACTIONS ON HUMAN FACTORS IN ELECTRONICS, SEPTEMBER 1967
HUMAN OPERATOR MODEL

I- - - G2f (S)

r (14S0)l-2(o5os+.5 (1.5)2]
0.541 (I +_) (-S) (I -03s
.~~~~~~a 2_o .)
z'

xo(s)

Fig. 11. Block diagram of closed-loop man-machine system for


driving. (Data from subject A used.)

HUMAN OPERATOR MODEL

0.4L 12 (I - 2 (I +o.o-5)
(l+20) l+6-)[ 6.0 +(G60)2]

Fig. 12. Block diagram of closed-loop man-machine system for


tracking. (Data from subject A used.)

gram of the closed-loop man-machine driving system nonminimum phase factor into this transfer function to
that includes both the realizable and the unrealizable obtain the correct phase response.
portions of the model. When this transfer function is substituted into the
A polynomial transfer function was also fitted to the closed-loop expression and the result factored, the fol-
gain and phase plots for subject A while tracking, thus lowing closed-loop man-machine transfer function re-
making it possible to compare the cases for driving sults:

Xo( ()) 2
(
(12)
(l + 102U + ( 3)s + (2S3 1[ + 2(8,)s + S2
and tracking. In the tracking case, the transfer function The bandwidth of (12) is considerably greater than that
is of (10), as previously indicated by the analysis of Sec-
tion III. Fig. 12 is a block diagram of the closed-loop
man-machine tracking system.
0.412(12)1 +005 Transfer functions may also be fitted to subject C's
.05) gain and phase plots if desired. However, the antireson-
+ 2(0.707) s+
Gj.(s) + + S2
ances, if not accounted for by the polynomial transfer
20 06.0 (6 .0)2 functions, would make the approximation slightly less
(11) accurate than those for subject A above.
and, of course, there is no unrealizable portion of this V. CONCLUSIONS
model. Here, again, emphasis has been placed on the Perhaps the most important conclusion that can be
accuracy at frequencies between 0.5 and 5.0 radians per drawn from this investigation is that it is feasible to
second. Note that it is necessary to introduce a delay or study human control of an automobile with servo tech-
WIERWILLE et al.: CLOSED-LOOP ANALYSIS 201
niques that are only slightly more complicated than were found. The female subject produced smaller lateral
the usual man-servo methods. Of course, many aspects vehicle errors and larger closed-loop system bandwidth
of the driving situation have been idealized in the process than did the male subject. However, the female subject
of application of these methods. Even with these ideali- used more abruptly steering wheel motions than the
zations, the essence of the man-machine control system male subject.
is preserved. Techniques such as the ones used in this Finally, all tracking models exhibited the hesitancy
investigation should make it possible to design future property resulting from reaction time lag that has been
land vehicles in a way which provides for the dynamic characteristic of models determined in previous studies.
responses of the human. However, refinement and further In contrast, none of the driving models exhibited this
extension of the techniques will probably be necessary to hesitancy, possibly because the subj ects used the pre-
encompass a more complete range of automobile con- viewed information to overcome their reaction time lags.
trol problems.
The preliminary experiments of this investigation show
that extended servo models of the operator's responses ACKNOWLEDGMENT
to vehicle error and future roadway possess sufficient ac- The authors wish to thank J. H. Doolittle and Dr.
curacy for closed-loop man-machine analysis. Indeed, W. C. Schultz, of the Avionics Department of the Cornell
errors in driving models under idealized driving condi- Aeronautical Laboratory, for their untiring administra-
tions may generally be made smaller than errors in con- tive and technical support of this work.
ventional tracking models. Closed-loop servo analysis
using these models leads to results which give every ap-
pearance of accurately assessing the closed-loop dynam- REFERENCES
ics of the man-vehicle combination. Also, little difference [1] J. W. Senders, A. B. Kristofferson, W. Levison, C. W. Diet-
rich, and J. L. Ward, "An investigation of automobile driver in-
in closed-loop system parameters exist for driving models formation processing," Bolt Beranek and Newman Inc., Cam-
which account for operator preview and for driving bridge, Mass., Rept. 1335, April 26, 1966.
[2] D. A. Gordon, "Experimental isolation of drivers' visual in-
models which do not; but, characterization error in- puts," Public Roads, vol. 33, pp. 266-273, February 1966. (See
creases significantly for models that do not account also Human Factors, vol. 8, pp. 129-137, April 1966.)
[3] N. L. Biggs, "Directional guidance of motor vehicles--a
for operator preview. preliminary survey and analysis," Ergonomics, vol. 9, pp. 193-202,
In a comparison of driving with tracking, it was found May 1966.
[4] J. G. Wohl, "Man-machine steering dynamics," Human Fac-
that tracking results in a closed-loop man-machine band- tors, vol. 3, pp. 222-228, December 1961.
width that is approximately twice as wide as the driving [5]T. B. Sheridan, "Three models of preview control," IEEE
Trans. Human Factors in Electronics, vol. HFE-7, pp. 91-102,
closed-loop bandwidth. The decreased bandwidth in driv- June 1966.
ing is caused by a shift of emphasis from lateral error [6] W. W. Wierwille, "A theory of optimal deterministic char-
acterization of time-varying human operator dynamics," IEEE
minimizing to a driving control criterion allowing larger Trans. Human Factors in Electronics, vol. HFE-6, pp. 53-61, Sep.
lateral error. In servo terms, the decreased bandwidth is tember 1965.
[7] W. W. Wierwille, G. A. Gagne, and J. R. Knight, "A labora-
caused by the human operator's tapering off of lead tory display system suitable for man-machine research," this issue,
equalization at a lower frequency. These results are page 250.
[8] L. Segel, "On the lateral stability and control of the auto-
based on the simulation data obtained for two tech- mobile as influenced by the dynamics of the steering system," J.
nically trained male subjects with driving experience. Engrg. for Industry, Trans. ASME, vol. 33, pp. 283-294, August
In a comparison of driving models for a male subject 1966.
[91 W. W. Wierwille, "Improvement of the human operator's
who was technically trained and for a female subj ect tracking performance by means of optimum filtering and predic-
who was not technically trained, important differences tion," IEEE Trans. Human Factors in Electronics, vol. HFE-5,
pp. 20-24, September 1964.

You might also like