steering force feedback for human–machine-interface automotive experiments

32 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 60, NO. 1, JANUARY 2011

Steering Force Feedback forHuman–Machine-Interface Automotive Experiments

Diomidis I. Katzourakis, Student Member, IEEE, David A. Abbink, Riender Happee, and Edward Holweg

Abstract—Driving-simulator fidelity is usually defined by thequality of its visual and motion cueing system. However, thequality of its haptic cues is also very important and is determinedby both hardware and control properties. Most experiments withhaptic steering systems employ commercially available systemsand do not address the system’s fidelity. The goal of this paper isto offer guidelines for the development of hardware, performanceevaluation, and system control in order to engineer realistic hapticcues on the steering wheel. A relatively low-cost solution for hard-ware is deployed, consisting of a velocity-controlled three-phasebrushless servomotor, of which its high-bandwidth control allowsfor a realistic representation of forces. A method is presented toovercome electromagnetic interference produced by the industrialservomotor and the controller through careful amplification andfiltering. To test the system, different inertia–spring–damper sys-tems were simulated and evaluated in time and frequency domain.In conclusion, the designed system allowed reproduction of alarge range of steering-wheel dynamics and forces. As a result,the developed system constitutes an efficient haptic device forhuman–machine-interface automotive experiments.

Index Terms—Automotive driving simulator, electromagneticinterference (EMI) filtering, haptic feedback, human–machineinterface (HMI), motor control, real-time (RT) dSPACE system,steering-system model, strain-gauge amplification.

I. INTRODUCTION

HUMAN-IN-THE-LOOP (HIL) driving simulators arewidely utilized by automotive manufacturers [1] and

researchers to reduce prototyping time and cost. Successfulapplications range from driver training [2] to human–machine-interface (HMI) system design for automotive-control applica-

Manuscript received December 15, 2009; revised April 26, 2010; acceptedApril 28, 2010. Date of publication September 7, 2010; date of current versionDecember 8, 2010. This work was supported in part by the Biomechanical En-gineering and Intelligent Automotive Systems Research Groups, by the Facultyof Mechanical, Maritime, and Materials Engineering, Delft University of Tech-nology, and by the Automotive Development Center, Svenska Kullagerfabriken(SKF). The Associate Editor coordinating the review process for this paper wasDr. Shervin Shirmohammadi.

D. I. Katzourakis and R. Happee are with the Biomechanical Engineer-ing and the Intelligent Automotive Systems Research Groups, Faculty ofMechanical, Maritime, and Materials Engineering, Delft University of Tech-nology, 2628 CD Delft, The Netherlands (e-mail: [email protected];[email protected]; [email protected]).

D. A. Abbink is with the Biomechanical Engineering Research Group, Fac-ulty of Mechanical, Maritime, and Materials Engineering, Delft University ofTechnology, 2628 CD Delft, The Netherlands (e-mail: [email protected]).

E. Holweg is with the Delft Center for Systems and Control and theIntelligent Automotive Systems Research Group, Faculty of Mechanical,Maritime, and Materials Engineering, Delft University of Technology, 2628Delft, The Netherlands, and also with the Automotive Development Center,Svenska Kullagerfabriken, 3430 DT Nieuwegein, The Netherlands (e-mail:[email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TIM.2010.2065550

Fig. 1. TU Delft moving-base simulator.

tions [1], [3]. Compared with real vehicle experiments, drivingsimulation provides enhanced repeatability, safety, unlimitedparameterization for the vehicle and the environment, and rel-atively lower cost. Nevertheless, designing a high-fidelity HILsimulator that is able to provide realistic cues to the driver ischallenging. It requires sophisticated vehicle-dynamics mod-eling, high-fidelity visualization, and sensor/actuation mecha-nisms [4].

Motion during driving is sensed through vestibular, audiovi-sual, and kinesthetic–haptic stimulants [5]. Poor quality of anyof the simulated cues can make it difficult or even impossibleto perform easy driving tasks, such as lane following [4]. Afundamental haptic cue is the feedback force at the steeringwheel. It renders the vehicle–road interaction and is consideredvery important for driving a vehicle [6], [7]. For example, thereduction of the self-aligning moment at the steering wheel,when the front tires approach their lateral force peak, is avaluable feedback to the driver [8].

This paper presents a framework to engineer a high-qualityforce-feedback (FF) steering system. Inspiration was drawnfrom a similar haptic-device design in automotive implemen-tations [4] and biomechanical applications [9]. The goal ofthis paper was to establish the apparatus for providing realisticsteering FF for HIL automotive experiments. The FF steeringwheel (FFSW) has been integrated into the moving-base simu-lator (see Fig. 1) of the Intelligent Automotive Systems ResearchGroup.

0018-9456/$26.00 © 2010 IEEE

KATZOURAKIS et al.: STEERING FORCE FEEDBACK FOR HUMAN–MACHINE-INTERFACE AUTOMOTIVE EXPERIMENTS 33

The rest of this paper is organized as follows: Section IIsummarizes the related work, whereas Section III describesbriefly the architecture of the system. Section IV focuses on theFF hardware and electronics, whereas Section V is dedicatedto the FF motor control. Section VI is occupied with the fi-delity analysis of the system. Section VII addresses automotivesteering-system modeling and HIL testing. Finally, Section VIIIconcludes this paper.

II. RELATED WORK IN FFSWs

FFSWs are popular within driving game simulators. Usually,they are cost-effective solutions for the wide market (e.g., Log-itech Momo [10]). Although some low-cost FFSWs can alreadyincrease simulation realism, they are not suited for high-fidelityhaptic research. The weak motors used offer little FF power,whereas reduction gearing imposes a noticeable high inertia.The steering angle is commonly determined through plain po-tentiometers achieving moderate angular resolution. Still, someresearchers have used low-fidelity FFSWs when studying theeffect of the FF [11] in cases where fidelity is of secondaryimportance. High-end steering wheels also exist but at a consid-erably higher price (Frex GP, Extreme Competitions Control,Inc., etc. [10]). More sophisticated commercial FFSWs are alsoavailable for HIL automotive research. Their enhanced perfor-mance is accompanied with a cost surpassing the price of anaverage automobile (a Moog FCS ECoL-8000 S actuator [12]).

Custom-made FFSWs have also been presented in the litera-ture regarding automotive simulation [4], [13] or drive-by-wireapplications [14]–[16]. Griffiths and Gillespie [13] explore thebenefits of augmented FF to share control between the driverand automated steering to support lane keeping. They use low-complexity vehicle dynamics and a current-controlled hapticwheel. Mohellebi et al. [4] propose a method that calculates thefeedback force for driving simulators and haptic augmentationto overcome the simulation latencies. The dc motor feedbacktorque can simulate the stiffness and the damping of the steeringcolumn.

An exceptional steer-by-wire conversion on a 1997Chevrolet Corvette has been presented by Yih and Gerdes [14](Stanford Dynamic Design Lab). This study aimed to improvethe handling characteristics of the Corvette through steer bywire. Using closed-loop system identification, they estimatedthe properties of the steering system (rack mass and damping)and derived a controller with minimal steering lag. Anotherpublication for the same vehicle from Switkes et al. [15]describes in detail the FFSW system and the control method.They present a realistic steering-system model for determiningthe feedback torque. The brushless dc motor they used hasa reduction ratio of 5:1, and the system’s properties wereexperimentally identified. For creating an additional sense ofdamping and inertia, they use the first and second derivativesof the steering angle. Furthermore, they evaluate the designof the drive-by-wire system for stability during lane keeping.Another haptic interface for steer-by-wire technology has beenpresented by Baviskar et al. [16]. They propose an adaptivesteer-by-wire controller with torque observers eliminating theneeds for torque measurement. Their experimental setup of the

FFSW involves a steering wheel in direct drive from a torque-controlled switched reluctance motor. The characteristics ofthe FF device (FFD) were identified by torque tests. Notethat, in the majority of this related work, the representation ofvirtual inertia and damping is accomplished by differentiatingthe steering angle, which is a method that introduces noise.Although the mentioned studies address the FFSW, they mostlyfocus on the effects of the FF on the driver and not on the deviceitself or its fidelity. This paper presents an implementation thatsurpasses the aforementioned limitations of realizing accuratelythe inertia and the damping by employing a torque-sensing ele-ment, avoiding the steering-angle differentiation, and applyingspeed control for the feedback motor. It provides a frameworkfor realizing a high-fidelity FFSW based on innovative conceptssummarized as follows:

1) architecture of an automotive simulator with the FFSW;2) feedback motor velocity control;3) FFSW experimental fidelity analysis;4) automotive steering-system modeling;5) noise-tolerant FFSW control method;6) simulation test analysis on the proposed methods.

III. ARCHITECTURE OF THE AUTOMOTIVE SIMULATOR

Here, we portray briefly the simulator’s architecture. Adetailed description can be found in [17]. The simulator isbased on a dSPACE real-time (RT) computer. It executes acommercial RT vehicle-dynamics model (VDM), developed onan open MATLAB/Simulink block, from the dSPACE Auto-motive Simulation Model (ASM) package. An overview of thesimulator is illustrated in Fig. 2.

A. Computing Components

The simulator consists of a midsize dSPACE computer withtwo DS1005 boards in master–slave topology. The dSPACEinterfaces with the environment via A/D (DS2002) and D/A(DS2102) boards. One desktop PC constitutes the user sta-tion providing the interface to the simulator. It also hosts theMATLAB/Simulink suite, where the VDM and control algo-rithms are being developed. The dSPACE simulator transmitsthe animation data over an Ethernet connection to three desktopPCs handling the graphics. Finally, three thin-film-transistormonitors compose a viewing angle of 135◦ (see Fig. 1).

B. Software for the Simulator

Software from MATLAB/Simulink and dSPACE is used forthe operation of the simulator. The dSPACE Test and Experi-ments software suite and the ASM VDM are used for RT pa-rameter handling, RT 3-D animation, and RT parameterizationof the virtual vehicle components (drivetrain or chassis). Thevehicle is an open MATLAB/Simulink model with 24 degreesof freedom (DOF). It incorporates semi-empirical tire models,suspension dynamics, steering-system models, etc. The VDMis being executed at 1 kHz at the “slave” DS1005 board.Communication with the environment through the D/A andA/D boards is performed through the “master” board at 5 kHz.


Fig. 2. Moving-base simulator components.

C. Motion Hardware

The simulator can render inertial and haptic cues through themotion of the platform and the FFSW, respectively.

a) The cockpit on the top of the platform (driver’s seat,FFSW, etc.) is mounted on a hydraulically controlledStewart platform, constructed at the Faculty of Mechani-cal, Maritime, and Materials Engineering, Delft Univer-sity of Technology (TU Delft). The signals controllingthe movement pass through an electronic array where in-appropriate signals are limited by saturation. The cockpitwas originally a commercial driving simulator, which wasmodified for the HMI-HIL experimental needs.

b) The FF steering motor (see Fig. 3) is an Ultract II,type 708303, high-response ac brushless servomotor. Themotor shaft is connected in a direct-drive fashion tothe steering wheel through an adjustable clamp mech-anism. The steering torque at the shaft is measured byfour symmetrically glued active strain gauges forminga Wheatstone bridge circuit. When the shaft is strained,the resistive changes of the bonded gauges unbalance thebridge, resulting in voltage deviation of a few millivolts.The Wheatstone bridge offers temperature compensationand linearity between strain and voltage measurements[18]. The torque signal requires further conditioning (fil-tering amplification) to become practical for usage, asexplained in Section IV. The FF motor is controlled bya digital-signal-processor-based AX-V brushless motorservo controller from Phase Motion Control. The default

Fig. 3. FF motor with torque sensing on the drive shaft.

Speed-V software is uploaded in the AX-V, transformingthe steering motor to a speed-controlled servo drive. Thecurrent loop for the motor is updated at 16 kHz and theposition loop at 4 kHz. The AX-V supplies the absoluteangular position of the shaft through a continuous analogsignal with a range of 0–10 V (representing 0◦–360◦).The analog position signal after being conditioned issupplied to the dSPACE A/D board. The AX-V platformwas developed specifically as a controller for the UltractII motor for optimal performance. Nonetheless, despitethe good specifications, the noisy nature of the motorcontroller imposes difficulties for connecting it to thesimulator. The challenge for achieving high accuracy inthe measurements and stability for control is addressed inthe following sections.

IV. FF HARDWARE EQUIPMENT

A realistic representation of haptic steering forces requiresa mechanism for adjusting the feedback torque. While directmotor-torque control is commonly utilized [4], [13], it is notaccurate enough for frequencies below 0.5 Hz [19]. A popularsolution for high-fidelity haptic applications is to applyvelocity-controlled motors [9]. The Ultract II motor employedfor the FF is powered by a three-phase source of alternatingvoltage. The speed can be adjusted by powering the motorthrough inverters. Inverters convert dc voltage to ac voltageat a specific amplitude and frequency. The three-phase voltagewaves are produced with pulsewidth modulation (PWM) [20]at 16 kHz. The PWM drive evokes noise lines at frequencieslinked to the switching frequency of insulated gate bipolartransistors [21]. This excites parasitic capacitive couplings atthe components of the system, spreading electromagnetic in-terference (EMI) currents to other components of the drivingsimulator [22]. Overall, when the system is in operation, itcreates noise to all circulation paths, including the ground andthe power supply.

The electric noise sampled from the A/D board with the FFmotor enabled can be easily identified in Fig. 4. It illustrates the


Fig. 4. Noisy sampled data from the A/D boards prior to the system’s redesign, recorded with a variable resistor used as the throttle pedal position sensor.

unfiltered voltage from the throttle sensor, which is a simplevariable resistor. The noise when the motor is enabled at timeequal to 2.55 s is illustrated at the top subplot. The same sensor,with the motor disabled, exhibited noise only at 50 Hz and itsharmonics, as displayed at the bottom subplot (the frequency ofthe main power line). The 16 kHz PWM (fpwm) signal is re-sponsible for the peaks around 1 kHz (and their harmonics), asdisplayed in the middle subplot. The 1 kHz peak is frequency-folded because of aliasing. The sampling frequency fs is at5 kHz; thus, the aliased frequency fa is expected to be presentat fa = fpwm − fs · C ⇒ fa = 16 kHz − 5 · 3 kHz = 1 kHz(C integer multiples of fs) [23].

The initial approach to reduce the noise was to apply second-order low-pass Butterworth active filtering (see Fig. 17, right)for all the sensors and to isolate the connections from andtoward the motor controller using two ISO-122 isolation am-plifiers. Although this approach improved the noise over thedSPACE lines, it did not sufficiently reduce the noise in thetorque measurements (the motor’s shaft; see Fig. 3). Spec-tral analysis revealed that the motor EMI was affecting thestrain gauges at frequencies above the bandwidth of the op-amp (1.5 MHz maximum) of the active filter. This resultedin the nonlinear operation of the op-amp, ruining the filter’sproperties.

Best results were finally achieved by simple analog RC fil-ters between the connections of the A/D board and the sensors.Concerning the torque sensor, individual output of the bridgewas RC-filtered and fed through a voltage follower (G = 1)to the difference amplifier (G = 470; see Fig. 5). A designguide from Texas Instruments [24] was used for designing thedifference amplifier.

The system performed best with the torque conditioningcircuit as close as possible to the motor shaft. Also, all theshieldings from the cables that carried analog signals wereconnected to the ground of the A/D board. The filters’ compo-nents are shown in Fig. 5, along with their resulting bandwidths[Fb (Hz); bold letters, inside dashed boxes]. The conditionedtorque is a fairly linear function of the measured voltage, asdisplayed in Fig. 6. The estimated torque from the voltage fol-lows closely the first-degree polynomial fitted with the polyfitfunction from MATLAB.

Two RC filters in series constitute a second-order low-passfilter for the steering angle (see Fig. 5). This filter has a lowbandwidth of 12.77 Hz. Nonetheless, a driver’s steering inputsare expected to be lower than the cutoff frequency of the filter.Groll et al. [25] denote that torque steering inputs from a driverare expected to be less than 4 Hz. Therefore, position inputsare expected in even lower frequencies, since they derive fromthe second integration of the driver’s torque. Yet, the steeringFF above 4 Hz will provide useful haptic information to thedriver about the vehicle’s state; thus, a high-bandwidth actuatoris always desired.

V. FEEDBACK MOTOR CONTROL

Here, we explain the velocity-control concept for the feed-back motor. The FFD was for simplicity initially tested inde-pendently of the VDM. It was programmed to simulate a virtualinertia–spring–damper (ISD) system (dotted sketch in Fig. 7)coupled to the motor and externally stimulated by the driver’storque Td. The virtual model consists of the torsion stiffnessKs, the moment of inertia Js, and the rotational damping bs.


Fig. 5. Optimal topology for the dSPACE sensor motor.

Fig. 6. Steering torque versus output voltage of the torque sensor.

Assuming a physical implementation of the system inside thedotted box in Fig. 7 and applying Newton’s second law, onecan derive the following dynamical equation:

Jsθs = Td − Ksθs − bsθs. (1)

The actual simulated system, however, is described insidethe dashed rectangle in Fig. 7. The resultant torque is thesubtraction of the driver’s torque Td and the model’s torque.Td is measured from the torque sensor on the shaft (see Fig. 3).Dividing the resultant torque by the virtual inertia results in thedesired angular acceleration d2θdes/dt2. Its integral, i.e., thedesired angular velocity dθdes/dt, is used as the velocity com-mand for the feedback motor. The damping force is calculated

Fig. 7. ISD virtual model for FF testing. (Dotted square) Physical model.(Dashed rectangle) Actual simulated system.

as the product of bs and the desired angular velocity, not thederivative of the steering-wheel angle θs that other researchersuse [15]. The dynamical equation of the simulated system isdepicted as

Jsθdes = Td − Ksθs − bsθdes. (2)

The rationale behind this is that the steering angle is ananalog signal with a noise ripple of 0.1◦ (after being analog


Fig. 8. System identification scheme.

and digitally filtered). The differentiation of the noisy steering-angle signal results in large errors, which, when multiplied witha high damping value, can lead to system instability. At thesame time, the integration of the desired angular acceleration(see Fig. 7) for obtaining the desired angular velocity filters outthe torque noise ripple, which has a magnitude of 0.02 N · m.The usage of the motor velocity command as the actual steeringvelocity also coheres with the haptic controller design guide ofSchouten et al. [9].

If the virtual inertia Js (see Fig. 7) is set below 0.001 kg · m2

and, at the same time, the steering wheel is grasped extremelystrongly by the driver (thus rigidly coupling the arm inertia tothe virtual dynamics), then an effect named “contact instability”[9] occurs, provoking the system to become unstable. Nonethe-less, the system is still safe for the subject since it will becomestable as soon as the grasp becomes less stiff.

Both the steering-angle θs (V ) and torque Td (V ) voltages(see Fig. 5) are software-conditioned before being utilized inthe virtual model. This includes second-order filtering with abandwidth of 195.5 Hz for the torque and 19.5 Hz for the steer-ing angle and further conversion from voltages to actual values.

VI. FFD FIDELITY

The performance of the designed hardware and controller ofthe FFD was evaluated by simulating several virtual models(i.e., changing the parameters for inertia, damping, and stiff-ness). The response of the steering-wheel system to torqueperturbations was measured and validated in two ways. First,the measured signals were compared in the time domain withsimulated signals that describe the response of the virtualsystem to the torque perturbation. Second, the virtual modelswere estimated back using closed-loop system identificationtechniques [26]–[28].

Fig. 8 illustrates the system identification scheme. The virtualdynamics of several ISD systems without load (i.e., no sub-ject holding the steering wheel) were perturbed by multisinetorque disturbances. The virtual stiffness was realized usingthe steering-angle signal θrealized from the motor controller.Although the experiments were conducted with no human load,the measured torque was also included in the simulated model.That is because, in practice, an external torque is caused by thecombined inertia of the steering wheel and the motor’s shaft;

Fig. 9. (Upper subplot) Torque disturbance and (lower subplot) correspondingspectral power. It consists of 30 logarithmically spaced clusters of frequencypoints starting from 0.65 Hz.

Fig. 10. Time response of the actual FFD (“Real”) response and a simu-lated ISD system (“Simulated”) with the same parameter settings and initialconditions.

damping is also introduced by the supporting ball bearing (seeFig. 8, Load block).

To enable accurate system identification, the torque dis-turbance signal was designed in the frequency domain as amultisine signal with an optimized crest factor, so as not tointroduce bias or variance in the estimated spectral densities[26]. Furthermore, for an improved SNR, the power of thedisturbance signal was distributed over a limited number offrequencies within the bandwidth. The disturbance signal con-sisted of 30 logarithmically spaced clusters of frequency points(a cluster contained two adjacent frequencies for averaging) in abandwidth of 0.65–20 Hz. Fig. 9 shows the time and frequency-domain representations of the designed torque perturbationsignal.

A. Time-Response Analysis

The most simple fidelity analysis can be done by comparingthe time response of the steering angle θrealized (see Fig. 8) ofthe real system and the simulated system when both are excitedwith the same torque disturbance. An example of time-domainanalysis is illustrated in Fig. 10.


Fig. 11. Normalized RMS angle error between the actual response fromthe FFD and that of a simulated ISD system. Plotted for multiple parametersettings. The radius of the sphere is the square of the RMS error (for bettervisualization).

Multiple parameter settings (inertia, damping, and stiffness)were tested, and the RMS between the “real” and “simulated”angles of the ISD system was obtained. The normalized RMSis shown in Fig. 11. The radius of the sphere representing theerror’s magnitude is the square of the RMS error (for bettervisualization). The minimum and maximum RMS errors were0.03 and 0.2 rad, respectively. Overall, it can be concluded thatthe required dynamics were realized accurately over a largerange of parameters. Notice that the error increases for smalldamping values. Definitely, the time-domain analysis can givean insight for the dynamic performance of the FFD. Yet, high-frequency responses cannot be evaluated solely from a time-domain evaluation. Thus, frequency-domain analysis was thesecond stage in the fidelity assessment process.

B. Frequency-Domain Analysis

Another way to visualize the developed system’s ability toaccurately emulate the desired stiffness, damping, and inertiais to estimate the realized dynamics from the measured signalsand to compare them with the desired dynamics. This visualiza-tion has an advantage over the time domain in that it also showsthe performance to simulate dynamics at higher frequencies(where damping and inertia dominate).

To achieve the aforementioned comparison, an identificationprocess described by Schouten et al. [9] and de Vlugt et al. [28]was adjusted to the particularities of the current system. Duringindividual experiments, the measured torque Tsensor, the torquedisturbance Tdisturbance, the realized angle θrealized (steeringangle realized after the signal conditioning), and the perturba-tion torque signals Tstim (Tstim = Tdisturbance + Tsensor) weresampled at 500 Hz for 60 s (the names of the signals correspondto Fig. 8). Two frequency response functions (FRFs) wereestimated: the admittance of the programmed ISD [ISD(f)],resulting from Tstim to the θactual signal, and admittanceLoad(f)−1, resulting from Tsensor to the θactual signal. Notethat Load(f)−1 is the inverse FRF of the load’s impedance sothat the causal admittance is shown. In the estimation process,

the actual steering angle θactual was not available; thus, it wastaken to be equal to the realized steering angle θrealized.

The sampled signals Tsensor(t), Tdisturbance(t), Tstim(t), andθrealized(t) were transformed with the fast Fourier transformto the frequency-domain signals of TSensor(f), TDisturbance(f),TStim(f), and ΘRealized(f), respectively. The frequency signalswere used to estimate the cross-spectral densities, i.e.,

ΦTDisturbanceΘRealized = T ∗Disturbance(f) · ΘRealized(f) (3)

ΦTDisturbanceTSensor = T ∗Disturbance(f) · TSensor(f) (4)

ΦTDisturbanceTStim = T ∗Disturbance(f) · TStim(f). (5)

The hat (∧) denotes estimate, and the asterisk (∗) denotescomplex conjugate. The FRFs of ISD(f) and Load−1(f) canbe estimated by dividing the appropriate cross-spectral densities[9], [28], i.e.,

ISD(f) =ΘRealized(f)

TStim(f)=

ΦTDisturbanceΘRealized

ΦTDisturbanceTStim

(6)

Load−1(f) =ΘRealized(f)TSensor(f)

=ΦTDisturbanceΘRealized

ΦTDisturbanceTSensor

. (7)

The FRFs of (6) and (7) without any further processing willbe referred as “measured data” in the rest of this section. Onlyfrequencies with an expected high SNR from the FRFs wereused for the model estimation. Those were the frequenciescontaining more than half of the maximum power of the powerspectral density of the disturbance signal [see (8)]. In thefollowing, f denotes the frequency vector, and k is the indexof that vector, i.e.,

ΦTT =T ∗Disturbance · TDisturbance

= |TDisturbance|2fk : ΦTT(fk) ≥ 0.5 · Max(ΦTT). (8)

The outliers from the resulting FRF were also removed forbetter performance. An outlier was considered to be more thanthree standard deviations away from the mean [29]. Finally,the resulting FRF was fitted to a second-order (ISD) systemwith the invfreqs function from MATLAB’s Signal ProcessingToolbox. This function is suitable for identifying continuous-time filter parameters from frequency responses.

Fig. 12 shows an example of such a frequency-domain plotbased on closed-loop system identification for one of the testeddynamics: an underdamped ISD with inertia = 0.01 kg · m2,stiffness = 0.2 (N · m)/rad, and damping = 0.05 (N · m)/rad/s.The desired programmed dynamics (cyan line) are comparedwith the measured data (magenta asterisks). To further substan-tiate the findings, an ISD model was fitted to the measured datain the frequency domain according to the process describedabove, which yielded nice fits (blue circles). The figure showsthat the estimated model resembles closely the programmedone. This was true for all tested conditions.

The estimated parameters were compared with the simu-lated parameters (same parameters as in Fig. 11). In the leftboxplot in Fig. 13, one can see the relative error between theprogrammed and fitted model parameters for the ISD FRF. Theprocess exhibited satisfying results, given the sensitive nature of


Fig. 12. FRFs of the admittance (in radians per newton meter) from an un-derdamped system setting [inertia = 0.01 kg · m2, stiffness = 0.2 (N · m)/rad,and damping = 0.05 (N · m)/rad/s]. (Upper subplot) Gain and (lower subplot)phase are shown for the FRFs for (cyan line) desired dynamics, (magentaasterisks) measured data, and (blue circles) estimated ISD model fitted to themeasured data.

Fig. 13. Left boxplot; relative error between the programmed and fitted modelparameters for the ISD FRF. Right boxplot; estimated parameters from theLoad(f)−1 FRF.

the described method. The stiffness demonstrates the smallestrelative error compared with damping and inertia. Its effectis dominant in low frequencies and, thus, can be easily bothsimulated and estimated.

As mentioned earlier, the experiments were performed withno external load on the system. However, the steering-wheelassembly is expected to introduce dynamics estimated by theLoad−1(f) FRF. This is illustrated in Fig. 14 and resem-bles a second-order system with inertia and damping witha gain drop of 40 dB/dec. The estimation process yieldedrealistic values with medians of μ(inertia) = 0.017 kg · m2,μ(damping) = 0.0937 (N · m)/rad/s, and μ(stiffness) = 0 (N ·m)/rad presented at the right boxplot in Fig. 13.

Both time- and frequency-domain analyses showed that thedeveloped system can accurately simulate a large range of linear

Fig. 14. FRFs of Load−1 denoting the load’s admittance (in newton meterper radian). It shows the gain from (magenta asterisks) measured data and (blueline) estimated second-order system fitted to the measured data.

Fig. 15. Overview of relevant elements in the steering-system model.

dynamics. Thus, the performance of the FFD system can becharacterized as suitable for high-fidelity FF HMI experiments.

VII. AUTOMOTIVE STEERING-SYSTEM MODELING

After assuring that the designed system could accuratelysimulate a wide variety of linear ISD models, the final challengeto enable realistic FF on the steering wheel was addressed tocouple the FFD to the ASM VDM. However, realizing thedynamical connection of real measured data (steering angleand driving torque) with a virtual model is a difficult process.For example, the ASM VDM assumes noise-free values of thestate variables, which is improbable when real measurementsare being involved.

A. Classical Steering-System Dynamic Model

To model the steering system, a power-assist steering systemis chosen, as illustrated in Fig. 15. It is composed of a steeringcolumn, a rack with gearing, and a power-assist motor [30]. Itsdynamic behavior is described by the following (the parametersare described in Table I):

Jsθs =Td − Ktb(θs − θsc) − bsθs (9)

Jscθsc =Ktb(θsc − x/Rs) + Tassist (10)

mrx =(Ktb(θsc − x/Rs) + Tassist)

Rs− brx − Fr (11)

Tassist =Ga(θs − θsc) = −Tm. (12)


TABLE ISTEERING-SYSTEM PARAMETERS

This particular model involves the majority of the importantparts that can be found on most real steering systems. Equation(12) describes the assist torque Tassist as a function of the strain(θs − θsc) on the torsion bar multiplied by an assist gain Ga.The resulting torque from the torsion bar, as well as the assisttorque, acts directly on the steering rack through a gear withradius Rs. Two equal-sized frictionless gears are assumed todeliver the motor’s torque Tm directly to the steering column.The ASM steering-system model that was employed assumesthat force Fr acting to the rack through the tie rod is realizedthrough the vehicle’s suspension. Fr is derived from 6-DOFtorques and forces from the tires.

Using the same principle that was described in Section V,the FF can be realized by integrating the right-hand side of (9)divided by the steering-wheel moment of inertia. The resultingdesired angular velocity dθdes/dt in the following:

θdes(t) =

t∫0

Td − Ktb(θs − θsc) − bsθdes

Jsdτ (13)

(in correspondence with Fig. 7) will constitute the feedbackmotor command. Again, for the damping estimation, dθdes/dtshould be used.

Utilizing the above principle with a realistic stiffness valuefor the torsion bar, which would be in a kilonewton-by-meter-per-radian scale, creates numerical challenges. A 0.1◦ ripple er-ror in the steering-angle measurement multiplied by a stiffnessof 1 kN · m/rad results in a 0.1(π/180) · 1000 = 1.75 N · mfluctuation in the resultant torque (see Fig. 7). The initialattempt to overcome this issue was to increase the damping andto use smaller values for the torsion-bar stiffness (values from50 to 200 N · m/rad). This approach, besides deteriorating thesteering feel, failed to achieve satisfactory results. Hence, a newsteering-system model with reduced dynamics was developed,where the feedback control was less vulnerable to the noise, andat the same time, it was safer for the driver holding the steeringwheel.

B. New Simulator Steering-System Model

A new steering-system model with reduced dynamics wasimplemented into the system, which is appropriate for being

adopted at the simulator and to operate with the velocity-controlled motor. The system, although simple, achieves real-istic FF. The dynamical equations are the following:

Jsθdes =Td − RsFr + Tassist − Ks_d(VX) · θs − bs_d(VX)

· θdes − (1/ (|Td| + 0.2)) · bhigh_damp · θdes (14)

x = θsRs (15)

x = θdesRs(θdes : motor velocity command) (16)

Tassist =Ga_d · Td. (17)

Equation (14) is the backbone of the system and consists ofmultiple terms, including rack forces Rs · Fr, the power-assistforce Tassist, stiffness and damping terms of both functionsof the longitudinal velocity, and an optional “safety” term.The innovation of this model lies on the fact that the highstiffness at the torsion bar has been bypassed; the feedbackstiffness component derives now from the tires’ forces throughthe Rs · Fr term. The former promotes stability because thesteering-rack displacement is calculated as the product of thesteering angle θs times the pinion gear radius Rs [see (15)].This allows for the fluctuation error at the steering angle tobe damped through the tire dynamics without creating arti-facts in the feedback forces. Additionally, the Js in (14) ishigher compared with that in (9) since it renders the aug-mented inertia of the steering wheel, the steering column,and the rack mass. The higher inertia reduces the magnitudeof the high-frequency FF noise caused by the steering-angleripple.

Equation (14) also includes the stiffness term Ks_d(VX),which increases with the longitudinal velocity VX . It serves asa firm aligning component to the steering wheel to improve theon-center feel on high velocities. The damping term bs_d(VX)also increases with VX so as to compensate for the increasingstiffness. The power-assist torque Tassist is determined as themeasured driver’s torque Td times the gain Ga_d [see (17)].Thus, the actual strain on the shaft of the FFSW is used andnot a simulated strain originating from the noisy signal of thesteering angle as in (12). Finally, the steering-rack velocitydx/dt, which is necessary for the vehicle’s suspension dynam-ics, derives from the feedback-motor desired velocity commanddθdes/dt times the Rs [see (16)].

Regarding the “safety” term (1/(|Td| + 0.2))bhigh_damp,suppose that the parameters of the steering system have beenadjusted for high-magnitude FF (e.g., small-assist gain Ga_d orlarge Rs). If the driver releases the steering wheel, |Td| dropsto zero. Thus, the safety term can inhibit any extreme motionof the steering wheel by increasing the feedback damping.However, when the driver is holding the steering wheel, thisterm has a limited effect.

C. Full System Evaluation With a Human Driver

We present here an application example of the proposedsystem. A double-lane change maneuver with a driver in theloop utilizing the proposed feedback model was performed at


Fig. 16. Driver-in-the-loop double-lane change maneuver.

68 km/h. The final FFSW system was perceived as to feel likea real car. Results are plotted in Fig. 16, where the trajectoryof the vehicle is illustrated at the top subplot (solid line). Thedashed line represents the center of the lane. Time t and yaw an-gle YA(◦) are also displayed. In the bottom subplot in Fig. 16,dynamical states of the steering system are being displayed incorrespondence with the top subplot. State Model_Torque isdefined as follows:

Model_Torque=−RsFr+ Tassist−Ks_d(VX)·θs− bs_d(VX)

· θdes − (1/ (|Td| + 0.2)) · bhigh_damp · θdes (18)

and is the right-hand side of (14) when Td = 0(Model_Torque’s numeric negative has been plotted). Thephase differences between the coupled state variables can bediscerned from the sequence of the dynamic events.

The driver initially applies torque Td. The resultant torque(Td − Model_Torque) results in a steering angular velocity anda steering angle, first and second integrals, respectively, ofthe resultant torque divided by inertia Js. The above plots ofthe feedback torque, the steering angle, and the model torque(which represents the human–tire–road interaction) resemblethe results obtained from real-vehicle double-lane change ex-periments conducted by Velenis et al. [31]. By altering theconstant parameter values of the steering system in (14)–(17),

Fig. 17. (Left) Difference amplifier and (right) second-order low-pass filter.

one can achieve a virtually infinite number of different steeringfeelings.

VIII. CONCLUSION

This paper has described the development of a high-fidelityFF steering device aimed to enable HIL automotive simula-tions. The developed system allows the FF to be delivered tothe driver through a speed-controlled three-phase brushless ser-vomotor with a torque sensor on the motor shaft. The developedsystem was able to simulate a wide range of virtual dynamics,allowing realistic FF from a large variety of steering systems.The system’s performance has been investigated by frequency-and time-domain techniques and has been found to be reliableand accurate.

The system has been connected to a dynamic model of asteering system developed to achieve smooth FF with accuracyof 0.02 N·m.

An extension of this paper will involve a more realisticsteering-system model based on real vehicular data, incorpo-rating coulomb friction in the steering column and rack, and arealistic torque map instead of a constant gain. A video with theFFD in operation is available online [32].

APPENDIX

The transfer functions for the difference amplifier and for thefilter (see Fig. 17) are respectively given in (A.1) and (A.2),shown at the bottom of the page.

ACKNOWLEDGMENT

The authors would like to thank Dr. F. van der Helm,C. Droogendijk, and M. Gerard.

Difference amplifier

Vout_amp(s) =(

R1 + R2R3 + R4

)R4R1

V 2(s) − R2R1

V 1(s) R1=R3−−−−−−→R2=R4

Vout_amp(s) =R2R1

(V 2(s) − V 1(s)) (A.1)

Filter

Vout_filt(s) = − R2/R1s2R2R3C1C2 + s (R3C1 + R2C1 + (R3R2C1/R1)) + 1

Vin(s) (A.2)


REFERENCES

[1] D. Toffin, G. Reymond, A. Kemeny, and J. Droulez, “Role of steer-ing wheel feedback on driver performance: Driving simulator andmodeling analysis,” Vehicle Syst. Dyn., vol. 45, no. 4, pp. 375–388,Apr. 2007.

[2] J. C. F. de Winter, P. A. Wieringa, J. Kuipers, J. A. Mulder, andM. Mulder, “Violations and errors during simulation-based driver train-ing,” Ergonomics, vol. 50, no. 1, pp. 138–158, Jan. 2007.

[3] M. Mulder, D. A. Abbink, and E. R. Boer, “The effect of haptic guidanceon curve negotiation behavior of young, experienced drivers,” in Proc.IEEE Int. Conf. SMC, Oct. 12–15, 2008, pp. 804–809.

[4] H. Mohellebi, A. Kheddar, and S. Espié, “Adaptive haptic feedback steer-ing wheel for driving simulators,” IEEE Trans. Veh. Technol., vol. 58,no. 4, pp. 1654–1666, May 2009.

[5] A. Kemeny, “Simulation and perception,” in Proc. Driving Simul. Conf.,Paris, France, Jul. 1999, pp. 21–28.

[6] A. Liu and S. Chang, “Force feedback in a stationary driving simulator,”in Proc. Human Factors Ergonom. Soc. Annu. Meeting, Vancouver, BC,Canada, Oct. 1995, pp. 1671–1675.

[7] D. Toffin, G. Reymond, A. Kemeny, and J. Droulez, “Influence of steeringwheel torque feedback in a dynamic driving simulator,” in Proc. DrivingSimul. Conf., Dearborn, MI, Oct. 2003.

[8] R. Rajamani, Vehicle Dynamics and Control. New York: Springer-Verlag, 2006, p. 408.

[9] A. C. Schouten, E. de Vlugt, J. J. B. van Hilten, and F. C. T. van der Helm,“Design of a torque-controlled manipulator to analyse the admittance ofthe wrist joint,” J. Neurosci. Methods, vol. 154, no. 1/2, pp. 134–141,Jun. 2006.

[10] Live for Speed, Race Simulator, “Wheels”, LFS Manual Wiki. [Online].Available: http://en.lfsmanual.net/wiki/Wheels

[11] T. Nagai, N. Cohen, Y. Moriguchi, and Y. Murakami, “Dual-drivernetworked fire truck simulator with multimodal display including forcefeedback steering and rotating motion platform,” in Proc. IEEE Int. Work-shops Enabling Technol.: Infrastructure Collaborative Enterprises, 2007,pp. 424–430.

[12] D. A. Abbink and M. Mulder, “Exploring the dimensions of haptic feed-back support in manual control,” J. Comput. Inf. Sci. Eng., vol. 9, no. 1,p. 011006, Mar. 2009.

[13] P. G. Griffiths and R. B. Gillespie, “Shared control between human andmachine: Haptic display of automation during manual control of vehicleheading,” in Proc. IEEE HAPTICS Symp., 2004, pp. 358–366.

[14] P. Yih and J. C. Gerdes, “Modification of vehicle handling characteristicsvia steer-by-wire,” IEEE Trans. Control Syst. Technol., vol. 13, no. 6,pp. 965–976, Nov. 2005.

[15] J. P. Switkes, E. J. Rossetter, I. A. Coe, and J. C. Gerdes, “Handwheelforce feedback for lanekeeping assistance: Combined dynamics and sta-bility,” Trans. ASME, J. Dyn. Syst. Meas. Control, vol. 128, no. 3, pp. 532–542, Sep. 2006.

[16] A. Baviskar, J. R. Wagner, D. M. Dawson, D. Braganza, and P. Setlur, “Anadjustable steer-by-wire haptic-interface tracking controller for groundvehicles,” IEEE Trans. Veh. Technol., vol. 58, no. 2, pp. 546–554,Feb. 2009.

[17] D. Katzourakis, M. Gerard, E. Holweg, and R. Happee, “Designissues for haptic steering force feedback on an automotive simulator,” inProc. IEEE Int. Workshop Haptic-Audio Visual Environ. Games, 2009,pp. 1–6.

[18] A. L. Window, Strain Gauge Technology. New York: Springer-Verlag,1992, pp. 151–154.

[19] ABB, Technical guide No. 1 Direct Torque Control—The World’s MostAdvanced AC Drive Technology.

[20] J. Lepka and P. Stekl, 3-phase AC induction motor vector control us-ing a 56F80x, 56F8100 or 56F8300 device design of motor controlapplication, Chandler, AZ, 2005, Application note. [Online]. Available:http://www.freescale.com/files/product/doc/AN1930.pdf

[21] R. Corton, B. Cassoret, and J. F. Brudny, “Characterization of three-phaseharmonic systems generated by PWM inverter switching application toinduction machine magnetic noise reduction,” in Proc. IEEE IEMDC,2001, pp. 442–447.

[22] A. Kempski and R. Smolenski, “Decomposition of EMI noise into com-mon and differential modes in PWM inverter drive system,” J. Elect.Power Quality Utilization, vol. 12, no. 1, pp. 53–58, 2006.

[23] J. G. Proakis and D. G. Manolakis, Digital Signal Processing, 4th ed.Englewood Cliffs, NJ: Prentice-Hall, Apr. 2006, pp. 28–30.

[24] B. Carter, A single-supply op-amp circuit collection, Texas Instru-ments, Dallas, TX, Application report, SLOA058. [Online]. Available:http://focus.ti.com/lit/an/sloa058/sloa058.pdf

[25] M. V. Groll, S. Mueller, T. Meister, and R. Tracht, “Disturbance com-pensation with a torque controllable steering system,” Vehicle Syst. Dyn.,vol. 44, no. 4, pp. 327–338, Apr. 2006.

[26] R. Pintelon and J. Schoukens, System Identification: A Frequency DomainApproach. New York: IEEE Press, 2001.

[27] L. Ljung, Perspectives on System Identification, Online Tutorial.[Online]. Available: http://www.control.isy.liu.se/~ljung/seoul2dvinew/plenary2.pdf

[28] E. de Vlugt, A. C. Schouten, F. C. T. van der Helm, P. C. Teerhuis, andG. G. Brouwn, “A force controlled planar haptic manipulator for posturecontrol analysis of the human arm,” J. Neurosci. Methods, vol. 129, no. 2,pp. 151–168, 2003.

[29] Mathworks, Data Analysis: Removing Outliers, Matlab Help Files.[30] A. Zaremba and R. I. Davis, “Dynamic analysis and stability of a

power assist steering system,” in Proc. ACC, Seattle, WA, Jun. 1995,pp. 4253–4257.

[31] E. Velenis, D. Katzourakis, E. Frazzoli, P. Tsiotras, and R. Happee, “Sta-bilization of steady-state drifting for a RWD vehicle,” presented at the10th Int. Symp. Advanced Vehicle Control, AVEC, Loughborough, U.K.,2010.

[32] TUDelft Video. [Online]. Available: http://www.youtube.com/watch?v=3lHV8jsXjIo

Diomidis I. Katzourakis (S’08) received theDiploma of Engineering in computer engineeringand informatics from the University of Patras, Patras,Greece, in 2006 and the M.Sc. degree in electron-ics from the Technical University of Crete, Chania,Greece, in 2008. He is currently working towardthe Ph.D. degree in the Biomechanical Engineer-ing, Faculty of Mechanical, Maritime, and MaterialsEngineering, Delft University of Technology (TUDelft), Delft, The Netherlands.

He is currently with the Intelligent Automo-tive Systems Research Group, TU Delft. His research interests focus onhuman–machine interface in extreme driving conditions, sensors and instru-mentation, embedded systems, and robotics.

David A. Abbink received the M.Sc. and Ph.D.degrees in mechanical engineering from the DelftUniversity of Technology (TU Delft), Delft, TheNetherlands, in 2002 and 2006, respectively. Histhesis “Neuromuscular Analysis of Haptic Feedbackduring Car Following” was awarded the best Ph.D.thesis in the area of movement sciences in TheNetherlands.

For three years, he worked on a research projectfunded by Nissan, during which he helped developand evaluate a force-feedback gas pedal to support a

driver with a car following, which has recently been released on American andJapanese markets. He is currently an Assistant Professor with the Delft HapticsLaboratory, Faculty of Mechanical, Maritime, and Materials Engineering, TUDelft.

Dr. Abbink was a recipient of the Dutch “Veni grant” to further stimulate hiswork on the design of human-centered haptic guidance.

Riender Happee received the M.Sc. and Ph.D.degrees in mechanical engineering from the DelftUniversity of Technology (TU Delft), Delft,The Netherlands, in 1986 and 1992, respectively.

He introduced biomechanical human models forimpact and comfort simulation with the commer-cial software MADYMO with TNO Automotive.Since 2007, he has been an Assistant Professorwith the TU Delft, leading automotive projectson human–machine interfacing for extreme driving,cooperative driving, driving-simulator fidelity, and

driver observation, and biomedical projects on neuromuscular stabilizationof the neck and the lumbar spine. He is also currently with the IntelligentAutomotive Systems Research Group, TU Delft.


Edward Holweg received the M.Sc. and Ph.D. de-grees in robotics from the Delft University of Tech-nology (TU Delft), Delft, The Netherlands, in 1990and 1996, respectively.

Prior to his current positions, he was responsiblefor the Automotive Development Centre and theMechatronics Research and Development Group forSvenska Kullagerfabriken (SKF), Nieuwegein, TheNetherlands. Some of the accomplishments duringthis time period include the SKF/Bertone drive-by-wire demonstration vehicle Filo, the General Motors

drive-by-wire concept vehicles AUTOnomy and Hy-wire, and the developmentof load-sensing wheel-end bearings. Since 2007, he has been the Director forProduct and Systems Development with the Automotive Development Center,SKF. He is responsible for all activities related to the development of theautomotive products within SKF. Since 2008, he has been a Part-Time Professorwith the Delft Center for Systems and Control and the Head of the IntelligentAutomotive Systems Research Group, Faculty of Mechanical, Maritime, andMaterials Engineering, TU Delft.

steering force feedback for human–machine-interface automotive experiments

Documents