2015 IEEE/SICE International Symposium on System Integration (SII) December 11-13, 2015. Meijo University, Nagoya, Japan Haptic Shared Control for Backward Parking and Its Effect on Skill Increase in Novice Drivers Shintaro Tada and Takahiro Wada Abstract The present research aims to develop advanced driver assistance systems (ADASs) that achieve increase of driver s skill as well as increase of driver performance during using the ADAS. A haptic shared control (HSC) is a promising tool for such ADASs with skill learning function because it enables the human operator to interact and communicate continuously with the assist system through a haptic interface. We focused on the assist system by haptic shared control (HSC) for backward parking maneuvers, which are considered to be a demanding task for novice drivers. Some studies showed that backward parking assist by HSC achieves performance increase during assist control as well as increase of parking skill. However, no research shows the effect of the driver s skill and the method to use the HSC though it is thought to affect the effectiveness of the ADAS. The purpose of the present study was to investigate effect of the gain setting of our proposed parking assist system with HSC on the novice drivers performance increase during use of the system as well as their skill increases where the drivers were instructed to drive in accordance with the haptic guidance. Driving simulator experiments of different feedback gains demonstrated that skill im-provements were achieved when a higher gain is employed. I. INTRODUCTION Advanced driver assistance systems (ADASs), such as lane keeping assist system (LKAS), adaptive cruise control, and parking assist systems, have succeeded in reducing the driver s workload and reduce the possibility of the crashes. However, it has been pointed out that driver assistance systems might change a driver s behavior in some cases, thereby indicating behavioral adaptation. In fact, many research studies revealed that behavioral changes could be occurred with the ACC [1], with automatic braking systems [2], and in the automatic braking systems with individual adaptation functions [3] through some of them were driving simulator experiments. Such behavioral changes might decrease the chances of developing the driver s skills, because of a decrease in the driver s situational awareness and/or in the driver s activity in the control loop [7]. Therefore, the present research aims to develop a methodology to achieve compatibility between the decrease in the driver s performance during use of ADAS for reducing workload and an increase in the driver s skill after use of it. A haptic shared control (HSC) has drawn much attention for ADASs. The HSC encourages a driver to remain within This work was not supported by any organization S. Tada is with Graduate School of Information Science and Engineering, Ritsumeikan University, Kusatsu, Shiga, 525-8577, Japan is0122fr@ed.ritsumei.ac.jp T. Wada is with College of Information Science and Engineering, Ritsumeikan University, Kusatsu, Shiga, 525-8577, Japan twada@fc.ritsumei.ac.jp the control loop because it collaborates with the human driver in the given control task through a mutual connection via a haptic interface such as a steering wheel. It is thought that HSC is a promising tool for increasing skills or training given such physical interactions between human and ADAS. In fact, Hirokawa et al. [4] have proposed a parking assist system based on an HSC method and showed that the driver s parking performance was improved after using the proposed system. The control gain was fixed in a single value throughout the experiments. On the skill training or motor learning, adjusting level of assist is thought to be important in training system, for example, as shown in rehabilitation training context [5]. The level of controller gain of HSC can be adjusted in terms of the haptic feedback or level of haptic authority (LoHA, [6]). Based on this idea, Tada et al. [7] proposed a backward parking system with HSC and demonstrated that the performance of the backward parking was improved during assist as well as drivers skill was improved after use of it with some gains. However, subjects skill and a method how to use system were not fixed and the effect of these factors were unclear [7]. We hypothesize that effectiveness of such ADAS depends on individual skills and the how to use the system. Therefore, the purpose of the present research is to investigate the effect of gain setting of our proposed parking assist system with HSC on the novice drivers performance increase during use of the system as well as their skill increases where the drivers were instructed to drive in accordance with the haptic guidance. II. BACKWARD PARKING ASSIST METHOD BY HSC A. Overview of proposed assist method We propose a driver assistance system that uses HSC to assist a driver to manually park a car by driving backwards, from the time the car is stopped and the gear is changed to reverse driving until the final parking position is reached (Fig. 1). The proposed method is basically same as [7] except for haptic feedback controller (6). Please refer to [7] in detail. B. Desired vehicle trajectory generation method The desired trajectory for backward parking that connects the starting and final parking positions of the backward parking operation is generated by using a third-order Bezier curve (1) by adopting [4] as shown in (Fig. 1). p(u) := (1 u) 3 p S + 3u(1 u) 2 p S + 3u 2 (1 u)p G + u 3 p G, (1) 978-1-4673-7241-1/15/$31.00 2015 IEEE 461
predicted vehicle position t p [s] ahead using the second order prediction. Θ d (s) = K c E(s), (7) 1 + T s where s denotes the Laplace operator, K c is a gain, T is a time constant of the first order lag system. Θ d is Laplace transformation of θ d (t). E(s) is Laplace transformation of ϵ t p (t), which is defined as the distance between the vehicle positon predicted by the preview model at t p [s] ahead on time t, and the desired trajectory. III. METHOD Fig. 1. Desired trajectory from P S to P G where u [0, 1] is a control parameter and p S, p G denote the position vectors whose coordinate expressions using the road fixed frame o xy are given by (2), (3). [ ] [ ] p x S = S xs l = S cos ϕ S (2) p G = y S [ ] x G y = G y S l S sin ϕ S [ ] xg + l G cos ϕ G, (3) y G + l G sin ϕ G where ϕ S and ϕ G denote the yaw angle of the vehicle at positions P S and P G, respectively. Scalars l S and l G represent the distance from P S and P G to points p S and p G, respectively. The desired vehicle trajectory is uniquely determined by determining l S and l G. The desired trajectory or parameters l S, l G, are determined by solving the following optimization. [l S, l G ] = arg min L(l S, l G ), (4) s.t. 1 κ max r min (5) where L(l S, l G ) is path length from P S to P G. r min is the minimum turning radius of the given vehicle and κ max denotes the maximum curvature of the trajectory. We confirmed that the generated vehicle trajectory was close to those driven manually by individuals who were accustomed to the driving simulator. A. Design We investigated the effect of the gain settings of the proposed assist system: Equation (6) on the changes in the driver s performance during and after use of the system. The independent variable of the experiments was the combination of the C s gain, and there were three levels of gain conditions A, B, and C as listed in Table I. C s = 1 was chosen as the TABLE I GAIN CONDITIONS Condition A B C C s gain 0 0.5 1 largest value within the stable region. Condition A means no assist system. The gain condition was a between-subject factor; thus, each subject was assigned to only one gain condition. B. Experimental apparatus Experiments were performed using a stationary driving simulator (DS), which has four LCD displays at the front and sides, and one at the rear as shown in Fig. 2. The drivers could see the computer graphics displayed on the rear monitor through the room mirror. A 250 W brushless DC motor (Maxon Corporation) was attached to the steering shaft to generate torque around the axis. A torque sensor C. Haptic shared control The torque around the steering wheel in the assist control is determined by (6). τ das (t) = C s e(t) (θ(t) θ d (t)), (6) where θ(t) is steering wheel angle at time t and θ d (t) is the desired steering angle, which is determined by a preview driver model to track the given vehicle desired trajectory. Scalar e(t) ( 0) is the distance between the vehicle position at the current time t and the nearest point on the desired trajectory. Scalars C s denote a gain. The preview model determines the desired steering angle θ d as (7), which uses the error between the desired trajectory and a Fig. 2. Driving Simulator (Kyowa) was also installed on the steering wheel to measure torque exerted on the steering shaft. Computer graphics were generated by Unity 3D (Unity Technologies, Inc.). Vehicle motion was calculated by Carsim (Mechanical Simulation Corp.). 462
C. Experiment scenario In the experiments, the subjects were asked to perform a backward parking operation from a certain point, which is referred to as start point of backward driving. Please note that there were two conditions in the starting point. For one condition, the participants drove forward and he/she had to determine the start point of backward driving, which is called self-selected condition. For another condition, the participants started to driver from a given start point of backward driving, which is called fixed condition. In the fixed condition, the participants did not drive forward direction. There were several cars around the host vehicle, including cars on both sides of the final parking position and the parking spaces in front of the final parked area as shown in Fig. 3. Fig. 3. Location of host vehicle and other vehicles composed of the trial 11 through 20, the subjects performed the parking operation with the assist control under each given gain condition. In the during-assist phase, the participants started backward driving from a fixed point and he or she did not drive forward direction. c) After-assist with fixed point phase: At this phase, which is composed of the trial 21 through 23, they again performed parking without any assistance to evaluate the extent of the increase in the driver s skills with the same start point of backward driving with the during-assist phase. The driving trials started backward from a fixed starting point without forward driving as phase b). d) After-assist with self-selected point phase: At this phase, which is composed of 24 through 26, they again performed parking without any assistance to determine the extent of the increase in the driver s skills when participants determined the start point of backward driving. Because condition A was essentially performed without assistance, none of the trials performed under this condition were performed with assistance through the experiments. The subjects in condition B and C were asked to execute backward driving by following the steering torque generated by the assist system and not oppose the assistance in the during-assist and after-assist with fixed point phases. In the during-assist and after-assist with fixed point, the starting point of backward and final parking positions were fixed as P S = [13, 1.5] T and P G = [8.2, 4] T through the experiments. For before-assist phase, data was collected only by the last 5 trials from 6 th to 10 th trial. Each subject participated all experiments in a same day. Subjects took 5 minute rest between the phases. D. Subjects Eighteen subjects (fifteen males and three females), aged 19 to 23 years old who were in possession of a driver s license, participated in the experiments. Each subject was assigned to one of gain condition so that driving frequencies are as same as possible among conditions. All subjects gave written informed consent in advance. E. Experimental procedure First of all, each subject was asked to perform several parking trials without any assistance from the system to become accustomed to the driving simulator, followed by twenty six parking trials as shown in Table II. There are four driving phase as follows. a) Before-assist phase: At the before-assist phase, which is composed of the trial 1 through 10, subjects performed backward parking without assistance to determine the driver s initial skill to perform backward parking. In the beforeassist phase, the participants drove forward from a given point and he or she had to determine where to change the direction of the drive should change from the forward into backward direction, which is called self-selected start point of backward driving. b) During-assist phase: At the during-assist phase, which is TABLE II PRESENCE OF ASSIST AND START POINT OF BACKWARD DRIVING IN THE NUMBER OF TRIALS Gain Condition Phase Trials A B, C start point of backward driving before-assist 1-10 no assistance no assistance self-selected during-assist 11-20 no assistance assistance fixed after-assist with fixed point 21-23 no assistance no assistance fixed after-assist with self-selected point 24-26 no assistance no assistance self-selected F. Evaluation method Vehicle variables including the vehicle position, yaw angles, and velocity, were recorded as well as the steering wheel angle and the torque around the steering wheel angle τ. In the present paper, the root mean square (RMS) of the error between the desired trajectory, the resultant vehicle trajectory, yaw rate, and driver s torque were used to evaluate the driver s parking performance. In addition, start point of backward driving in before-assist phase and after-assist with self-selected point phase were analyzed. 463
IV. RESULTS A. RMS of vehicle trajectory error A one way ANOVA of RMS vehicle trajectory error at before-assist phase showed that the gain condition had no significant effect (F (2, 15) = 0.755 p = 0.487), there is no evidence that the parking skills differ between the gain conditions. Fig. 4 shows the mean of the RMS vehicle trajectory error for each condition. Error bars show the point, the post hoc test showed that RMS error in condition B was smaller than condition A (p = 0.093). B. RMS of yaw rate A one way ANOVA of RMS yaw rate at before-assist phase showed that the gain condition had no significant effect (F (2, 15) = 0.588 p = 0.568), there is no evidence that the parking skills differ between the gain conditions. Fig. 5 shows the mean of the RMS yaw rate for each condition. Error bars show the standard deviation. Fig. 4. RMS of the vehicle trajectory error standard deviation. The two way ANOVA of the mean of RMS trajectory error by gain setting condition and driving phase condition as well as their interaction showed that the main effects were significant in driving phase (F (3, 20) = 21.352, p = 0.000) and interaction (F (6, 20) = 2.626, p = 0.029), while it was marginally significant for gain setting condition (F (2, 15) = 2.081, p = 0.081). The one way ANOVA of the RMS trajectory error by driving phase for each gain condition showed that the simple main effects were significant in condition B (F (3, 20) = 7.425, p = 0.002) and condition C (F (3, 20) = 8.424, p = 0.001), but not in condition A (p = 0.057). In condition B, the post hoc test by the Bonferroni method showed that RMS error was significantly smaller during-assist and after-assist with fixed point than before-assist (p = 0.003, p = 0.005). In condition C, the test showed that it was significantly smaller duringassist, after-assist with fixed point, and after-assist with selfselected point than before-assist (p = 0.001, p = 0.007, p = 0.048). For condition A, the post hoc test showed that RMS error after-assist with fixed point was marginally smaller than after-assist with self-selected point (p = 0.070). The one way ANOVA of RMS trajectory error by gain setting condition for each driving phase showed that the simple main effects driving phase were significant in duringassist (F (2, 15) = 4.693, p = 0.026) and after-assist with self-selected point (F (2, 15) = 8.480, p = 0.003) but not for other two phases. In during-assist phase, the post hoc test by the Bonferroni method showed that RMS error was significantly smaller with condition C than condition A (p = 0.031). In after-assist with self-selected point, the test showed that it was significantly smaller with condition C than condition A (p = 0.003). For after-assist with self-selected Fig. 5. RMS of yaw rate The two way ANOVA of the mean of RMS yaw rate by gain setting condition and driving phase condition as well as their interaction showed that the main effects were significant in driving phase (F (3, 20) = 5.766, p = 0.002) and gain setting (F (3, 20) = 5.185, p = 0.019) but not interaction (F (6, 20) = 1.706, p = 0.142). The one way ANOVA of the RMS yaw rate by driving phase for each gain condition showed that the simple main effects were significant in condition B (F (3, 20) = 3.203, p = 0.045) and condition C (F (3, 20) = 7.011, p = 0.002), but not in condition A. In condition B, the post hoc test by the Bonferroni method showed that RMS yaw rate was smaller during-assist than before-assist (p = 0.074). In condition C, the test showed that it was significantly smaller duringassist, after-assist with fixed point, and after-assist with selfselected point than before-assist (p = 0.005, p = 0.011, p = 0.010). The one way ANOVA of RMS yaw rate by gain setting condition for each driving phase showed that the simple main effects driving phase were significant in during-assist (F (2, 15) = 6.682, p = 0.008) and after-assist with self-selected point (F (2, 15) = 4.197, p = 0.036) but not for other two phases. In during-assist phase, the post hoc test by the Bonferroni method showed that RMS yaw rate was significantly smaller with condition C than condition A (p = 0.007). In after-assist with self-selected point, the test showed that it was significantly smaller with condition C than condition A (p = 0.041). C. RMS of driver s torque A one way ANOVA of RMS driver s torque at beforeassist phase showed that the gain condition had no significant effect (F (2, 15) = 0.340 p = 0.101), there is no evidence 464
that the parking skills differ between the gain conditions. Fig. 6 shows the mean of the RMS driver s torque for each condition. Error bars show the standard deviation. condition C was smaller than condition B (p = 0.058). D. Start point of the backward driving Fig. 7 shows an example of parking trajectories from the start point of the backward driving to the end of parking maneuvers for each condition in before-assist and afterassist with self-selected point where the start points of the backward driving were determined by participants. In the figure, circles represent the start point of backward driving. As seen from these figures, the start points of backward Fig. 6. RMS of driver s torque The two way ANOVA of the mean of RMS driver s torque by gain setting condition and driving phase condition as well as their interaction showed that the main effects were significant in driving phase (F (3, 20) = 56.700, p = 0.000), gain setting (F (3, 20) = 9.911, p = 0.002), and interaction (F (6, 20) = 10.772, p = 0.000). The one way ANOVA of the RMS driver s torque by driving phase for each gain condition showed that the simple main effects were significant in condition B (F (3, 20) = 35.390, p = 0.000) and condition C (F (3, 20) = 20.822, p = 0.000), but not in condition A. In condition B, the post hoc test by the Bonferroni method showed that RMS driver s torque was significantly smaller during-assist, after-assist with fixed point, and after-assist with self-selected point than beforeassist (p = 0.000, p = 0.000, p = 0.003) and after-assist with fixed point and after-assist with self-selected point than during-assist (p = 0.000, p = 0.000). In condition C, the test showed that it was significantly smaller during-assist, after-assist with fixed point, and after-assist with self-selected point than before-assist (p = 0.000, p = 0.003, p = 0.003) and after-assist with fixed point and after-assist with selfselected point than during-assist (p = 0.009, p = 0.009). The one way ANOVA of RMS driver s torque by gain setting condition for each driving phase showed that the simple main effects driving phase were significant in during-assist (F (2, 15) = 25.472, p = 0.000), after-assist with fixed point(f (2, 15) = 9.786, p = 0.002), and after-assist with self-selected point (F (2, 15) = 8.516, p = 0.003) but not for before-assist phases. In during-assist phase, the post hoc test by the Bonferroni method showed that RMS driver s torque was significantly smaller with condition B and condition C than condition A (p = 0.000, p = 0.000). In after-assist with fixed point, the test showed that it was significantly smaller with condition C than condition A (p = 0.002). In after-assist with self-selected point, the test showed that it was significantly smaller with condition C than condition A and condition B (p = 0.004, p = 0.028). For after-assist with fixed point, the post hoc test showed that RMS error in [a] Condition A [b] Condition B [c] Condition C Fig. 7. Example of parking trajectories from start point of backward driving driving were converged in higher gain setting condition. Variance of the start point of backward driving is defined by (8) d := 1 N x x i, (8) N i=1 where N is number of sampled data points. A one way ANOVA showed that the gain condition had no significant effect on the variance of start point of backward driving point (F (2, 15) = 0.814 p = 0.462), there is no evidence that the parking skills differ between the gain conditions at the before assist. Fig. 8 shows the variance of start points of backward driving before-assist condition and those afterassist with self-selected point. Error bars show the standard deviation. The effect of the driving phase on the variance of start points of backward driving was analyzed. The one way ANOVA for each gain condition showed that the simple main effects of the driving phase on the variance of start 465
Variance of directionchange point [m] Fig. 8. 1.5 1.2 0.9 0.6 0.3 0 before assist * after assist A B C Gain Condition Variance of start points of the backward driving point of backward driving were significant in condition C (F (3, 20) = 8.011, p = 0.018), variance after-assist was smaller, but not conditions A and B. The one way ANOVA for each phase showed that the simple main effects of the driving phase on the variance of start point of backward driving were significant after-assist (F (2, 15) = 5.179, p = 0.019), but not before assist. The post hoc test by the Bonferroni method showed that the variance was significantly smaller in condition C than that in condition A in after-assist with self-selected point condition (p = 0.020). V. DISCUSSION Decrease in RMS vehicle error during assist and after assist in both conditions B and C but not in condition A suggests that performance of backward parking was increased during assist and the effect remains after use of the system, which means driver s skill was increased. The fact that same tendency was found in drivers torque RMS suggest that such increases in performance and skill was achieved with smaller effort. For RMS yaw rate, similar results were observed but there was significant differences between before-assist and after assist in condition C but not in condition B, suggesting that higher gain condition has greater effect on skill improvement after use of the system. These tendencies agrees with the results of Tada and Wada [7] with high impedance gain and low another gain condition. The fact that variance in start point of backward driving was decreased after use of the system in only condition C implies the system affects drivers skill where drivers start backward driving. This tendency agree with Hirokawa et al. [4], in which parking skills were quantified by final vehicle position and orientation in the parking space, and time required for the parking and start point of backward driving to converge by the scale car simulation experiment [4], though that paper employed different haptic controller. In order to know the cause of skill increase on start point of backward driving, driver s behaviors including gaze points should be investigated. Note that subjects with variety of driving experience participated in the experiments as well as the participants were not instructed about their reaction to the assist torque in those research studies in [4] [7]. The contribution of the present paper is as follows: we firstly showed that the parking * assist method with HSC could increase driving performance during assist and the novice driver s skill could be increased after use of it if the gain setting is relatively high when they drive in accordance with the assist system. However, the effect of driver s experiences and driver s reaction against the HSC on the skill increase, these are limitation and future direction of the study. VI. CONCLUSIONS The backward parking assist system with HSC was proposed and the effect of the gain setting of the proposed method on novice drivers parking performance during assist and after use of it in the case that the participants were instructed to drive in accordance with the assist torque displaying the error from the desired vehicle trajectory. The experimental results using driving simulator demonstrated the performances were increased during assist and after use of it in terms of vehicle trajectory error, yaw rate, and driver s steering torque when gain setting is relatively high. Decrease of variance of start point of backward driving was also observed. From these results, the present paper demonstrated that the parking assist method with HSC could increase driving performance during assist and the novice driver s skill could be increased after use of it when they drive in accordance with the assist system. The limitation of the paper is that it is unclear whether the driving experience or driver s initial skill affect the effect of the proposed assist method as well as driver s reaction to the assist torque. The individuals preference for the desired trajectory affects the effectiveness of the proposed method. These are important to build effective assist system that takes driver s skill increase into account. VII. ACKNOWLEDGMENTS This work was partially supported by a JSPS KAKENHI Grant-in-Aid for Scientific Research (A) Grant Number 26242029. REFERENCES [1] M. Hoedemaeker and K. A. Brookhuis, Behavioural Adaptation to Driving with An Adaptive Cruise Control (ACC), Transportation Research Part F, Vol. 1, No. 2, pp. 95-106, 1998 [2] M. Itoh, Y. Fjiwara, and T. Inagaki, Driver Behavioral Change through Interactions with an Autonomous Brake System, Transactions of the Society of Instrument and Control Engineers, Vol. 47, No. 11, pp. 512-519, 2011 (in Japanese) [3] S. Hiraoka, T. Wada, S. Tsutsumi, and S. Doi, Automatic Braking Method for Collision Avoidance and Its Influence on Driver Behaviors, in Proc. FAST-ZERO, Tokyo, No. 20117375, 2011 [4] M. Hirokawa, N. Uesugi, S. Furugori, T. Kitagawa, and K. Suzuki, Effect of Haptic Assistance on Learning Vehicle Reverse Parking Skills, IEEE TRANSACTIONS ON HAPTICS, Vol. 7, No. 3, pp. 334-344, 2014 [5] T. Wada and T. Takeuchi, A Training System for EMG Prosthetic Hand in Virtual Environment, Proceedings of Human Factors and Ergonomics Society 52nd Annual Meeting, Sept., NYC, pp. 2112-2116, 2008 [6] D. A. Abbink, M. Mulder, and E. R. Boer, Haptic shared control: Smoothly shifting control authority?, Cogn. Technol. Work, Vol. 14, No. 1, pp. 19-28, 2012 [7] S. Tada, T. Wada, Effect of Skill Development of Haptic Shared Control in Backward Parking Assist, IEEE International Conference on Vehicular Electronics and Safety, 2015 466