Sie sind auf Seite 1von 6

2011 14th International IEEE Conference on Intelligent Transportation Systems Washington, DC, USA.

October 5-7, 2011

An MPC approach to the design of motion cueing algorithms for driving simulators
Mauro Baseggio, Alessandro Beghi, Mattia Bruschetta, Fabio Maran and Diego Minen
Abstract Driving simulators play an important role in the development of new vehicles and advanced driver assistance devices. In fact, on the one hand, having a human driver on a driving simulator allows automotive OEMs to bridge the gap between virtual prototyping and on-road testing during the vehicle development phase. On the other hand, novel driver assistance systems (such as advanced accident avoidance systems) can be safely tested by having the driver operating the vehicle in a virtual, highly realistic environment, while being exposed to hazardous situations. In both applications, it is crucial to faithfully reproduce in the simulator the drivers perception of forces acting on the vehicle and its acceleration. The strategy used to operate the simulator platform within its limited working space to provide the driver with the most realistic perception goes under the name of motion cueing. In this paper we describe a novel approach to motion cueing design that is based on Model Predictive Control techniques. Two features characterize the algorithm, namely, the use of a detailed model of the human vestibular system and a predictive strategy based on the availability of a virtual driver. Differently from classical schemes based on washout lters, such features allows a better implementation of tilt coordination and to handle more efciently the platform limits.

I. INTRODUCTION In the automotive eld, there is an increasing interest on the development of dynamic driving simulator systems. Applications of such systems are in fact becoming more and more numerous and diverse, from drivers training and virtual vehicle set-up in the racing context, to the development of advanced driver assistance and accident avoidance systems. Automotive OEMs exploit driving simulators to cut down the costs for prototyping, by anticipating the on road vehicle behavior. Furthermore, such systems allow to ease the development process of the various vehicle components, by testing different hardware and software solutions, by resorting to sophisticated Hardware-In-the-Loop (HIL) tools, in a safe and realistic virtual environment. Given the wide range of applications, there is an ever growing need of developing small size, low cost dynamic simulators. In a different perspective, realistic dynamic simulators are crucial to develop detailed driver behavior models to devise accident avoidance strategies, for example by putting a standard driver in simulated dangerous condition and measuring his/her reactions. Also, assessment of driver performance under stress conditions (i.e., adverse weather conditions, endurance driving, etc.) can
This work was supported by VI-Grade Italy Mauro Baseggio, Alessandro Beghi, Mattia Bruschetta, Fabio Maran are with University of Padova, Department of Information Engineering Via Gradenigo 6/b, 35131, Padova, , Italy name.lastname@unipd.it Diego Minen is with VI-Grade Italy, Via lAquila 1c 33010, Tavagnacco, Udine, Italy diego.minen@vi-grade.com

be more effectively performed by using dynamic simulators with a high degree of immersion into virtual environments. To this regard, ever more effective Hazard Perception Test can be devised and used as requirements for achieving a driver license. In this scenario, the role of the motion cueing (MC) strategy, i.e. the algorithm for transforming vehicle accelerations into admissible motion commands to the platform, becomes more and more crucial to guarantee a realistic perception of the driving conditions. Motion cueing is a very complicated part of a dynamic simulator due to the complex nature of the human perception systems. In fact, it is not clear yet, from a physiological point of view, the role and priorities of stimula of different nature to the overall perception of accelerations and force. It is however well established that a coordinated visual-motion action is fundamental for achieving satisfactory performance of a MC algorithm. Given the above motivations the necessity of a perceptive model within the MC algorithm becomes evident. The classical approach to MC is based on high pass ltering the reference acceleration signal to keep platform excursions within the operational limits. Such approach is simple and has been implemented in many different ways over the years. However, it has some shortcomings:

it is a conservative approach, and as a consequence, the platform operational space is not fully exploited; it cannot explicitly handle hard constraints on the platform movements and accelerations; being a non model-based approach, set-up vehicle variations are hardly distinguishable from modication of the tuning of MC algorithm.

Recently, a novel approach to motion cueing has been proposed in [1], [2], based on a strategy already consolidated in the eld of industrial process control, namely, Model Predictive Control (MPC). MPC is a model-based control methodology that allows to handle limits on the working space and to exploit information on future reference signal. In this approach, a model of the human perception systems can be included in the motion cueing strategy and predictions of the future trajectory can be used to fully exploit the platform working area and generate accurate cues. In this paper we propose a MC algorithm for a small size dynamic simulator that is based on a specic implementation of the MPC. In particular, a detailed model of the human vestibular system is employed, as well as and availability of predictions of the future vehicle behavior as provided by a virtual driver developed in a detailed virtual prototyping environment. Performance of the algorithm is evaluated by comparison
692

978-1-4577-2197-7/11/$26.00 2011 IEEE

TABLE I P LATFORM PERFORMANCE . Range x y z Roll Pitch Yaw Position 1m 1m 0.3m 30deg 24deg 50deg Velocity 1.3m/s 1.3m/s 0.9m/s 112deg/s 61deg/s 61deg/s Acceleration 3.3m/s2 3.6m/s2 4.9m/s2 600deg/s2 600deg/s2 240deg/s2

with classical MC strategies, showing the effectiveness of the proposed approach. II. PROBLEM STATEMENT In Fig. 1 the platform considered in this study is represented. Its peculiarity is in the mechanical structure. By using linear actuators instead of the classic hexapodal structure, it is possible to achieve satisfactory results in physical simulation with a relatively small size hardware, that can t standard laboratories environments, whereas traditional, large dimensional, hexapodal platform require dedicated hangars.
Fig. 2. Scheme of motion cueing strategy.

Fig. 3.

Representation of MPC principle.

signal for the MPC algorithm; 3) compute via MPC the displacement signal d passed to the platform control system. III. MODEL PREDICTIVE CONTROL
Fig. 1. Platform sketch.

The architecture is based on three completely decoupled degrees of freedom (DOFs, longitudinal and lateral axis, and yaw), and three partially coupled DOFs. The simulator kernel, i.e. the vehicle dynamics physical engine, has been developed and extensively tested on the eld and provides a highly reliable representation of the real vehicle behavior. The screen covers more than 180 deg and moves in agreement with the platform to guarantee full immersion of the driver in the virtual environment. Finally, force feedback on the steering wheel and the braking system enhances the drivers feeling of the vehicle behaviour. The platform dynamic performance reported in Tab. I highlights the limitations of the operational space, with maximal linear excursions of 1 m. This fact makes the role of the MC algorithm crucial. The MC strategy has to provide the displacement references to the control system of the platform, which is assumed to be able to perfectly track the reference signals, with a xed time delay. The conceptual scheme of the MC procedure is shown in Fig. 2, and is composed by the following steps: 1) obtain the actual vehicle accelerations a from the simulation software; 2) obtain the perceived acceleration r by ltering a via the vestibular system model, thus generating the reference
693

Model Predictive Control (MPC) is an advanced control technique widely used in industrial applications [3], [4] since the 1980s. In recent years, robust and efcient implementations have been developed, as well as software tools in standard computational environments that ease the design of MPC algorithms. The main advantages of MPC can be summarized as follows: its underlying idea is simple and intuitive to understand; its the only generic control technique that efciently deals with constraints; it can handle Multi-Input Multi-Output (MIMO) systems without formally increasing the complexity of the problem; it can handle non linearities in both the model and the constraints. A. MPC basics Assume (discrete-time problem) that at time k a reference trajectory r(t|k), t k and a current measure of the output y(k), are available. Note that the current input is not yet computed. Now, suppose to have a model of the process to be controlled and that the state of the system (or its estimate) is available. We can therefore predict the future output y(k + i|k), i = 1, . . . , Np corresponding to the input sequence u(k + i|k), i = 0, . . . , Np 1 in a time-window of length Np , where Np is the prediction horizon length (Fig. 3).

The idea is to compute the input sequence u(k + i|k) that minimizes a cost function, e.g. a function of the tracking error (k + i|k) = r(k + i|k) y(k + i|k), while respecting a set of constraints. The input to be applied at time k is chosen as u(k) = u(k|k); at time k + 1 a new output y(k + 1) is measured and the algorithm is iterated applying only the rst element of the computed optimal input sequence. B. Process model In the literature, different implementations of the MPC principle have been proposed, with different model structures. In the application we are considering, the real-time constraints and the MIMO structure of the model are well described by a linear discrete state space model (sampled version of the continuous process) of the form xm (k + 1) = Am xm (k) + Bm u(k) y(k) = Cm xm (k). We assume that the input does not have a direct effect on the output (strictly proper system). State-space models are particularly well suited to design state estimators by using well-established tools of statistical ltering theory. In our case, as proposed by Wang [3], we consider as the element to be optimized the input difference u(k) = u(k) u(k 1): considering also the state difference xm (k), we can write a new state equation xm (k + 1) = Am xm (k) + Bm u(k). Considering the output difference y(k + 1) y(k) = Cm Am xm (k) + Cm Bm u(k) and dening a new, augmented state [xm (k)T y(k)T ]T , we obtain a new model x(k + 1) = Am C m Am x(k) :=

the classic formulation of a quadratic problem (QP) [3], that is, 1 J = U T HU + U T F. 2 Note that the length of the prediction horizon Np can differ from the control horizon Nc : for the sake of simplicity, we will consider Np = Nc . Np is a fundamental parameter both in terms of computational complexity of the problem and stabilizing properties of the resulting control law. The weights Q and R, on the tracking error and on the input respectively, allow to tune the control law to obtain the desired behavior. To deal with constraints, limitations on the system inputs U and outputs Y can be written in terms of constraints on the input variations U [3] A U b . As a consequence, the QP becomes a constrained QP, for which a variety of solving algorithms are available in literature. This is a key step to ensure that the control problem, and consequently the MC algorithm, can be solved in real time. State estimation is also an important element of the MPC scheme since it represents the key to actually close the control loop. In the considered setup, since the platform has its own position controller, state estimation is not necessary. Nevertheless, it can be used to close a second control loop in case the platform control system is linear. IV. THE VESTIBULAR SYSTEM The vestibular system is located in the inner ear and is composed by the semicircular canals and the otolith organs. The former sense the angular rotation and the latter linear motion. Accurate mathematical models of the two systems have been derived starting from the 70s for application to MC of ight simulators. Zacharias [5], in a survey written in the 1979, reported most of the results nowadays available. Telban and Cardullo in 2005 [6], [7] published a simplied transfer function model with estimates of the corresponding parameter values. For the semicircular canal, the transfer function that can best relate the sensed angular velocity to the acceleration stimulus in a MC control problem is the following: WSCC (s) = 80s2 (s) = 5.73 (s) (1 + 80s)(1 + 5.73s) (1)

0 Bm x(k) + u(k) I C m Bm

y(k) = 0

I x(k)

where the control input is u(k). C. Cost function and state estimation The optimal input sequence (k+i|k), i = 0, . . . , Np 1 u is computed by minimizing a cost function of the form [3] J(U ) = (Rs Y )T Q(Rs Y ) + U T SU + U T RU where the tracking error, the input variations and the control inputs are weighted by matrices Q, R and S respectively. Rs , Y , U , and U are matrices of appropriate sizes obtained from the vectorization of the reference signal rs (k + i), the output prediction y(k + i|k), the input u(k + i) and its difference u(k + i), all considered on a horizon of Np samples. The cost function can be rewritten in order to depend only on the input difference matrix U , obtaining
694

The otoliths are described in terms of the following transfer function that relates the sensed response to the specic force stimulus: f (s) 1 + 10s WOT H (s) = = 0.4 (2) f (s) (1 + 5s)(1 + 0.016s) A. Tilt coordination An important component of perception in a dynamic simulator is given by tilt coordination. Otoliths are not capable to discriminate between gravitational and longitudinal forces. Hence, by using a non-zero pitch angle and

without any other visual reference, it is possible to provide the driver in the simulator with a fake longitudinal acceleration sensation. The same holds true for roll and lateral acceleration. Such approach goes under the name of tilt coordination. Taking into account this effect is crucial to reproduce low frequency behavior with a reduced range working area. Tilt coordination is particularly relevant when predictions of the future vehicle trajectories are available, since, differently from classical cueing strategies, it can yield to less conservative strategies and thus to better exploitation of the platform working space. In the perception model, because of linearization, tilt coordination is nothing but a further contribution in the otoliths model WOT H (s) due to the pitch angle in the longitudinal direction and to the roll angle in the lateral direction. B. The complete model In order to use the perception models in the MPC approach, state space realization of WOT H (s) and WSCC (s) are obtained and coupled with the tilt coordination contribution for all the 6 DOFs. The resulting system can be written as x =AV EST x + BV EST u y =CV EST x + DV EST u where the input u is composed by the three applied longitudinal accelerations and the three angular velocities, i.e u = [ax , ay , az , x , y , z ]T . The overall state vector x is x = xSCC xOT H vx px vy py vz pz
T

Ref: Perceived long. acc. x axis Perceived longit. acc. 0.5

m / s2

0.5

1 0 5 10 time [s] 15 20 25

Fig. 4. Perceived acceleration tracking based on MPC approach without look ahead

A. Comparison with classical Motion Cueing Algorithm To better understand the potential of MPC for designing MC algorithms, in this Subsection we show that it is possible to reproduce the behavior of classical MC algorithms, based on the combination of lters with a washout strategy, by appropriately tuning the MPC parameters. To this aim, we set up a simulation based on experimental telemetry data. To better highlight the properties of the algorithms, we concentrate on the pure longitudinal dynamic with no pitch action, hence with no tilt coordination. This is a reasonable condition since in the classical approach prediction is not exploited and tilt coordination cannot be effectively performed. To mimic what happens in the classical approach, telemetry data are high pass ltered after linear scaling, and used to compute the reference signal for the MPC. The wash-out action is performed by the MPC strategy by weighting the position of the platform in the cost function. In agreement with the classical approach, no look ahead is used. Tuning of the weights is performed so as to reproduce a similar behavior of the classical and MPC algorithms. In Fig. 4 tracking of the perceived accelerations obtained by using the MPC approach is shown. Tracking is not fully satisfactory because of the wash-out action. In Fig. 5 and Fig. 6 the comparison of the real platform acceleration (classical MC) and the MPC based one is presented, both in terms of accelerations and positions. The reference trajectory for for the MPC algorithm is also reported. In this segment of the circuit the two strategies appear to be almost equivalent, in terms of both acceleration and position signals. It is interesting to note that the washout action can be clearly identied as a deviation from the reference trajectory in both strategies (after time t = 15 s in Fig. 5), and it can be interpreted as generating a low frequency tracking error. B. Using the look ahead In the look ahead approach, we make use of a virtual driver to predict the future telemetry and feed such signal as a reference to the MPC based MC algorithm, also considering the hard constraints provided by the platform operational space. In Fig. 7 a look ahead prediction over a time interval of 2 s is used. The comparison with the classical approach is reported in Figs. 8 and 9 in terms of accelerations and position signals, respectively. The MPC based algorithm with
695

(3)

where the actual angles, positions and velocities are obtained by integration from the inputs u and xOT H and xSCC are the state variable for the dynamical systems associated with the otoliths and semicircular canals. To impose a set of constraints in a simple manner we choose y = [, f , vx , px , vy , py , vz , pz , , , ]T , where v and f are v the vectors of perceived angular velocities and longitudinal acceleration along all the DOFs. V. RESULTS We report here some results obtained by using data form a non professional driver training session on the Silverstone track with a GP2-class car. Since the platform is almost decoupled for all the 6 DOFs, we consider here the longitudinal dynamics only, hence the longitudinal acceleration, that has to be reproduced as faithfully as possible (in terms of driver sensations) by operating the platform with longitudinal and pitch motions. With reference to the algorithm structure as described in Section III, Fig. 2, we introduce an intermediate step, that is, longitudinal accelerations from the simulator are appropriately scaled to obtain a feasible prole for the platform.

1 0.5 m / s2 0 0.5 1 0 5 10 15 time [s] Ref. longit. acc. x axis (scaled) Applied longit. acc. MPC Applied longit. acc. classical MC 20 25 30 m / s2

1 0.5 0 0.5 1 1.5 0 Ref. longit. acc. x axis (scaled) Applied longit. acc. MPC Applied longit. acc. classical MC 5 10 time [s] 15 20 25

Fig. 5. Comparison of applied acceleration tracking without look ahead: MPC based and classical motion cueing
0.3 0.2 0.1

Fig. 8. Comparison of applied acceleration tracking with look ahead: MPC based and classical motion cueing.
0.4 0.3 0.2 0.1 [m] 0 0.1

0 0.1 0.2 0.3 0 Longit. displ. x axis with MPC Longit. displ. x axis with classical MC 10 20 30 time [s] 40 50 60 70

0.2 0.3 0.4 0 5 10 15 20 25 time [s] 30 35 40 45 50

Fig. 6. Comparison of actual position without look ahead: MPC based and classical motion cueing

Fig. 9.

Actual position with look ahead.

look ahead allows to achieve better tracking performance without low frequency error. The platform working area is also fully exploited. Observe that the main difculties in the replication of perceived accelerations is associated with the necessity of reproducing constant accelerations over long periods of time, that is, to velocity peaks. C. Longitudinal-Pitch coupling By coupling the longitudinal and pitch DOFs we can test the proposed strategy in a more complex and interesting operating condition. We expect to observe the effect of tilt coordination that allows to reproduce the low frequency component. The MPC is now tuned by taking the physical limits of the platform as constraints and by choosing the weights in the cost function to exploit at best the platform working space. The reference signal is still supposed to be
1 0.5 0

Reference: long. acc. x axis Perceived longit. acc.

known in advance, to introduce tilt coordination, and no preltering of the telemetry data is performed. In Fig. 10 the reference signal, that is, the driver perceived longitudinal acceleration, is compared with that perceived in the platform, showing that an almost perfect tracking is achieved. In Fig. 11 and Fig. 12 the pitch angle and the longitudinal position of the platform are shown. Tilt coordination is clear when analyzing the platform behavior during the two braking events. The pitch angle gives the low frequency component of the longitudinal acceleration. The effect of tilt coordination is even more evident in Fig. 13 by comparing the scaled version of the original longitudinal acceleration signal with the only actual longitudinal acceleration of the platform. The platform acceleration produces the high frequency contribution only. In Fig. 14 the virtual vehicle jerk signal along the longitudinal axis is compared with the platform one, to underline the quality of the result. Finally in Fig. 15 the frequency contributions due to pitch and pure longitudinal dynamic are shown. As we expected the pitch, in a natural way, is used to cover the low frequencies of the perceived acceleration. VI. CONCLUSIONS

m / s2

0.5 1 1.5 0

10

time [s]

15

20

25

Fig. 7. Perceived acceleration tracking based on MPC approach with look ahead.

In this paper we describ the design of a MC algorithm for a small size dynamic driver simulator, that is based on MPC techniques. The most relevant features of the developed algorithm are the possibility of implementing look-ahead strategies by generating reference trajectories by means of an advanced virtual driver and the use of a detailed model of the human vestibular system to accurately reproduce the drivers
696

Fig. 10. Perceived acceleration tracking in combined longitudinal and pitch simulation.

Fig. 14.
1400 1200 1000 800 600 400 200 0

Longitudinal jerk in combined longitudinal and pitch simulation.

Overall perception Longitudinal Contribution Pitch Contribution

0.1

0.2

0.3

0.4

0.5 0.6 Frequency [Hz]

0.7

0.8

0.9

Fig. 15. Fig. 11. Longitudinal displacement in combined longitudinal and pitch simulation.

Frequency contribution of Longitudinal and Pitch DOF.

perception. Results show that the proposed algorithm exhibits better performance with respect to standard algorithms based on wash-out lters both in avoiding platforms movements that induce misleading cues to the driver and in exploiting the platform operational space. The implementation of the developed algorithm on a real device is undergoing. This will allow to assess the algorithm performance by analyzing also real drivers response. Results of such analysis will be reported in future publications. R EFERENCES
Fig. 12. Angular displacement in combined longitudinal and pitch simulation. [1] M. Dagdelen, G. Reymond, A. Kemeny, M. Bordier, and N. Mazi, Model-based predictive motion cueing strategy for vehicle driving simulators, Control Engineering Practice, vol. 17, no. 9, pp. 9951003, 2009. [2] B. Augusto and Loureiro, Motion cueing in the chalmers driving simulator: A model predictive control approach. [3] L. Wang, Model predictive control system design and implementation using MATLAB. Springer Verlag, 2009. [4] J. Maciejowski, Predictive control: with constraints. Pearson education, 2002. [5] G. Zacharias, Motion cue models for pilot-vehicle analysis, bolt beranek and newman inc Cambridge ma control systems dept, Tech. Rep., 1978. [6] J. A. Houck, R. J. Telban, and F. M. Cardullo, Motion cueing algorithm development: Human-centered linear and nonlinear approaches, NASACR, vol. 213747, no. May, 2005. [Online]. Available: http://hdl.handle.net/2060/20050180246 [7] R. Telban, W. Wu, F. Cardullo, and L. R. Center, Motion Cueing Algorithm Development: Initial Investigation and Redesign of the Algorithms. National Aeronautics and Space Administration, Langley Research Center, 2000.

Fig. 13. Actual longitudinal acceleration comparison in combined longitudinal and pitch simulation.

697

Das könnte Ihnen auch gefallen