Human Driving Skill Modeling Using Neural Networks for Haptic Assistance in Realistic Virtual Environments

09/12/2018 ∙ by Hojin Lee, et al. ∙ 0

This work addresses our research on driving skill modeling using artificial neural networks for haptic assistance. In this paper, we present a haptic driving training simulator with performance-based, error-corrective haptic feedback. One key component of our simulator is the ability to learn an optimized driving skill model from the driving data of expert drivers. To this end, we obtain a model utilizing artificial neural networks to extract a desired movement of a steering wheel and an accelerator pedal based on the experts' prediction. Then, we can deliver haptic assistance based on a driver's performance error which is a difference between a current and the desired movement. We validate the performance of our framework in two respective user experiments recruiting expert/novice drivers to show the feasibility and applicability of facilitating neural networks for performance-based haptic driving skill transfer.



There are no comments yet.


page 2

page 11

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

Haptic assistance (HA) provides an amount of assistive feedback in the form of tactile or kinesthetic stimuli. Recently, HA has been popularly facilitated in automobile technologies in addition to visual and auditory assistance [1]. Driving skills require coordinated dynamic controls of limbs via manual interfaces (steering wheel, accelerator/brake pedals, and so on). In particular, the kinesthetic feedback can deliver mechanical momentum and move the limbs of interest, supplying more direct, detailed, and continuous information. Hence, a number of studies have examined the effectiveness of HA (especially, kinesthetic) in virtual driving environments owing to the merits of allowing safe practice in simulated scenarios, especially those including risky driving situations.

Generally, driving with HA can be regarded to a human-machine cooperative task in a cybernetic human-machine shared control (HSC) [2, 3]. In HSC, information flows in both directions between two agents (a human and a machine) via mechanical contact under shared autonomy. First, a user can motorize an action via interfaces and the system reflects the corresponding action. Second, haptic interfaces intelligently delivers assistive information to a user, and then the user perceives the corresponding haptic information and reacts. Thus, most researchers have envisioned that in the simultaneous exchange of useful information, HA can offer two possibilities: task performance enhancement and skill training [4].

Fig. 1: The framework of driving skill modeling for HA.

For enhancing the task performance, the most representative, effective HA method is haptic guidance (often called haptic HSC) where external haptic stimuli are provided to the user in order to communicate control information on the desired movement. Several studies demonstrated haptic guidance can enhance the task performance of steering [5, 6, 7] and pedaling skills [8, 9, 10]. Here a HA system plays a role of a collaborator which encourages humans, mostly by demonstrating appropriate maneuvers and correcting their driving performance. To this end, the human driver and the system share a common goal: a successful driving that both agents perform an effective, safe and robust driving control. Therefore, HSC is considered as a bridge to an autonomous driving with improved performance and reduced effort [11], various car companies keep investigating this semi-autonomous strategy as advanced driver-assistance systems (ADAS) [12], such as a lane-keeping assistance system (LKAS), an intelligent parking assistance system (IPAS), and an adaptive cruise control (ACC).

HA is more widely utilized in motor learning and training, and rehabilitation applications. HA through a number of strategies including haptic guidance and other algorithms were investigated for training efficiency on various tasks [13, 14]. Here a HA system should play a role of a skill trainer; While a driving skill learner tries to drive properly, the system checks the current learner’s performance and provides haptic augmented feedback transferring knowledge of performance (KP [15]). Thus, the feasibility of HA, which the system aware of human performance errors continually to generate KP [16], lies special benefits to driving skill training. The effectiveness of HA under continuous observation of learner’s driving control was shown for several driving tasks, especially for curve-tracing and lane-keeping tasks [17, 18] and a reverse parking task [19].

Under the HSC manner, both the user and the HA system concurrently intercommunicate in real-time to endeavor an optimized control of the shared goal. The system continuously watches and analyzes a current user’s performance, and then exerts personalized haptic feedback, regarding a certain baseline of task performance, i.e. the ground truth. To this end, to design HA systems, the process of quantifying proper skill performance as a desired reference to a user’s current performance, usually called a modeling process, should be always preveniently required.

However, traditionally, the modeling was so far manually abstracted into deterministic forms without consideration of human agents, and also examined mostly for steering tasks as a sub-skill of driving, irrespective of pedaling tasks. Therefore, the modeling process in previous studies inevitably bears considerable differences from real and complicated driving, which requires simultaneous manipulation of the steering wheel and pedals. We face an immediate demand for reasonable modeling methods of complicated driving skills, which is a prerequisite of competent, intelligent haptic training systems.

In this work, we present a novel data-driven framework of (1) modeling a proper driving skill (steering and pedaling) from expert drivers and (2) utilizing the model for HA, in a freeway driving task (Fig. 1

). Our approach is to record how experts execute driving successfully without HA, and then train an adequate continuous model (representing a virtual expert) of associated variables from the collected professional data obtained from virtual environments . Specifically, we propose a useful methodological solution that extracts a behavioral model of expert drivers using artificial neural networks (NNs), and validate the strategy via user experiments. A NN is generally used to find a nonlinear function that explains an input/output relationship and identify a structure beneath a complex dynamic system. Zhang et al. showed that characterization of driving skill levels (expert, typical, and low-skill) are possibile, by analyzing individual driving manuvers using well-trained NNs and other pattern recognition algorithms 


Previously, Nechyba and Xu used NNs to model a human driving strategy from driving data collected from a simplified driving simulation using a mouse interface [21, 22]. Their NNs could produce a continuous predictive trajectory based on an individual motor behavior using experimental states and environmental variables as input. However, their studies did not involve realistic driving hardware, so the usability of their behavioral model for practical driving training has not been validated yet. In this study, we complete their NN-based modeling approach in a virtual driving simulator, and validate its practicality for HA.

Ii Simulator

Fig. 2: A visual scene of driving simulation.

This section describes a haptic driving simulator (Fig. 2 and 3) that we have developed for two purposes: (1) Driving Skill Modeling: The simulator records the expert’s driving data. It provides realistic driving experiences to acquire reliable data, including realistic torque feedback to the steering wheel and pedals; and (2) Haptic Assistance: The simulator generates torque feedback in order to assist a current driver’s driving skills with high fidelity. For realistic simulation, the simulator renders virtual driving environments including visual and auditory stimuli based on Vehicle Physics Pro (VPP [23]), a commercial vehicle physics engine running in the Unity 5 game engine (Fig. 2) with an update rate of 50 Hz. For the realistic simulation of car dynamics, a particular vehicle (Genesis, Hyundai Motors) was chosen to determine physical parameters of VPP, such as mass, dimension, steering and gear ratios, and engine power curves.

Fig. 3: Haptic driving training simulator.

Ii-a Hardware

The simulator consists of a large visual display, a steering wheel, an accelerator pedal, and a brake pedal (Fig. 3). All devices are fastened to an aluminum frame to imitate the real driving seat. We use a 55-inch LCD display (55LW6500, LG Electronics), and the distance from the display to the seat is about 1.2 m for a comfortable field of view of 60 . The simulator also uses a commercial steering wheel (SENSO-Wheel SD-LC, SensoDrive) to provide high-fidelity torque feedback. The maximum instantaneous torque and the maximum continuous torque are 16.58 Nm and 7.5 Nm, respectively.

We have custom-designed and built the accelerator and brake pedals with appropriate torque feedback capability. Two sets of AC servo motor (SGMGV-20A, Yaskawa Electrics) and servo pack (SGDV-18011A, Yaskawa Electrics) are used for independent torque feedback. The communication between the device and PC is done by MechatroLink-II network control board (PCI-R1604-MLII, Ajinextek). The maximum instantaneous torque and the maximum continuous torque of each motor are 27.8 Nm and 10 Nm, respectively.

For compact housing, both motors should be mounted in the same side, maintaining the alignment of the two rotation axes of the pedals. For this reason, while the accelerator pedal is directly connected to the motor with a coupler, the brake pedal is connected to another motor through a four bar mechanism. The loop formed by the four bar mechanism is designed to be a parallelogram for a simple kinematic relationship between the pedal and the motor. The steering wheel and the pedals are controlled with a sampling rate of 800 Hz.

Ii-B Realistic Torque Feedback Control

In our system, the steering angle is between and , and the steering ratio is 12.0:1. The driver cannot steer outside this range. The simulated steering torque is implemented to be similar to the real torque transmitted from the driving shaft to provide rich information about the road and vehicle status, as follows:


where is the self-alignment torque, and and

are the viscous and Coulomb frictions from the car dynamics. In four-wheel drive, the steering reactive torque can be estimated 

[24] as follows:


where and are the lateral forces applied to the left and right front wheels obtained from VPP. is the imaginary gain of torque transmission from the shaft. , and is constant, both in the opposite direction of steering wheel rotation. From (2), a user can perceive driving-like sensations on a road with respect to the direction and velocity of the virtual vehicle.

Our haptic pedals are controlled using a spring-damper impedance control scheme. Let the accelerator angle be . If is between and , it is normalized and sent to the throttle value of the virtual engine in VPP. The simulated torque to the accelerator is computed as follows:


where is a virtual damping torque and is a gravity compensation term. The spring-like torque is determined by


where is a virtual spring coefficient, and is the initial position of the accelerator pedal. pushes the driver’s right foot upward to deliver information about how much s/he is pressing the pedal from . is a unilateral feedback term to provide information as to the maximum angle such that


enables the driver to perceive the virtual endpoint at . We use .

The simulated torque to the brake pedal, , is computed similarly for the brake angle . The only difference was that the maximum brake angle .

We carefully tuned all the other parameters for realistic experiences. Their values are specified in Table I.

Steering Wheel Accelerator/Brake Pedals
(m) 0.75 , (Nm/degree) 0.2
(Nms/degree) 0.002 , (Nms/degree) 0.001
(Nm) 0.1
TABLE I: Constant Values for Driving Torque Feedback

Iii Modeling Using Neural Networks

To provide KP in trajectory learning tasks, e.g., an optimal (desired) trajectory should be given for the computation of task performance errors [14].

We denote the current angle vector by

and the desired angle vector by . The error vector is . The current desired action is generally contingent upon the current driving situation and the past values of representing the driving history. For performance-based haptic transfer, we need a model that gives . We build such a model using NN train it to account for the driving data of experts.

Iii-a Approach

In [25], the dynamic nature of human control strategy is abstracted into a static mapping between input and output using NN. In fact, a dynamic system can be approximated using difference equations [26], such that


where represents a nonlinear map using NN, is the control vector, is the system state vector, and is a vector describing external environmental features at the time step . Then (6) can be rewritten to


where for an arbitrary vector . Using (7), we can predict a future value of at -step later from the current and previous system states and the exogenous environmental variables.

Iii-B Neural Network Design

From the expert’s driving data, we observed that the lane-keeping task does not require the manipulation of brake pedal and exclude from the control vector. Also, we do not consider the interdependence between the controls of steering wheel and accelerator pedal and train separate NNs for each. This allows us to use more compact networks still with accurate modeling results. Hence, in the model for the steering wheel, , and in the model for the accelerator pedal, . For the vehicle state, we use , where is the longitudinal velocity (m/s), is the angular velocity (degree/s), and is the engine’s revolution per minute (RPM) of the virtual car.

To define the environmental features, we rely on ; Fig. 4), which is the 2D Euclidean distance from the driver’s position to the road boundary of the road in the direction of -30, -15, 0, 15, and 30 to the driver’s frontal direction. The angular values of were determined considering the driver’s field of view () within the simulated vehicle. The maximum value of is set to 60 m. Then the environmental feature vector , where


represents the future hazard of collision in the -th direction.

Then the two NNs, and , for the steering wheel and the accelerator pedal, can be written as


and are trained using and , respectively, in the expert’s driving data as the output. Therefore, the output variables and should be similar to the expert’s respective and used in the training, representing the expert’s predictive driving behavior under the vehicle state and the environment state .

Considering that the human motion bandwidth is less than 5 Hz [27], we use the same constants for all the input vectors: and . Then all NNs predict 0.2 s future values of execution from the five current and previous variables in 50-Hz simulations. Before training, all input vectors, , , and are normalized.

Fig. 4: Five distances from the driver’s perspective.

Iii-C Data Acquisition and Training Results

We designed 25 two-lane paths to collect driving trajectories and other important variables for driving skill modeling. Each path consists of three segments with a total length of 600 m. The first and third are a 200-m straight segment. The second segment is curved with the curvature , where is the radius, is the arc length, and is the angle in radian, as shown in Fig. (a)a. of the second segment is 200 m, but each path has varying from -180  to 180  with 15  step (Fig. (b)b). So results in a 600-m long straight path.

Fig. 5: Driving paths. (a) Variable definitions. (b) Examples.

Five driving experts (; all males; age 25–51 years, M 37.6, SD 10.8; driving experience 5-30 years, M 15.2, SD 10.3) participated in the data acquisition. In each trial, the expert was instructed to complete driving following a given path, while staying only in the first lane of the path and maintaining 60 km/h velocity on the speedometer. Each trial took about 36–40 s, and every expert completed 6 trials for each path (150 trials per expert).

We trained all networks using MATLAB (R2017a, MathWorks). The training used a network training function of gradient decent backpropagation with an adaptive learning rate and a transfer function of hyperbolic tangent sigmoid. The initial learning rate was 0.5. Every NN consisted of 4 hidden layers with 32, 16, 8, and 4 nodes. We pooled the input-output data of all the expert drivers for NN training. The data were partitioned into training, validation, and test sets in the proportion of 70%, 15%, and 15%, respectively. The training was terminated if the root mean squared error (RMSE) of predicting a test set decreased and was saturated into 1% and 4.5% for

and , respectively. These values were determined by trials and errors.

Iv Experiment I: Modeling Validity

We could model the driving behaviors of the expert drivers using NNs. Experiment I was to validate whether our model successfully captured the representative driving skills, with the following research questions: : Can our model work for other general driving environments?; : Can our model represent particular driving behaviors different from other drivers’s style?; and : Can we prove that expert drivers have better driving performance than of novice drivers in an objective manner?

Iv-a Data Acquisition

Our NN models were trained with the expert drivers’ data collected along the 25 simple paths (Section III-C). The first goal of Experiment I was to validate whether the NN models can be applied to general, longer, more complex driving environments. To this end, we designed a long path as a sequence of randomly generated straight and curved segments similarly to [22].

Each straight segment had one parameter, length . Each curved segment had the radius of curvature and its sweep angle as parameters. The parameters were randomly chosen from 100–150 m ( and ) and 45135

(negative for right curves) for each segment. A straight segment was followed by a left/right curve with equal probability 0.5. A left (right) curve was followed by a straight segment with probability 0.4 and a right (left) curve with probability 0.6. The total length of the path was 4 km. From many randomly-generated paths, we selected two representative paths respectively consisting of 23 and 22 segments for our experiments (Fig. 

6). Compared to the short, simple, predetermined paths used in the NN training (Section III-C), the two paths in Fig. 6 are long, arbitrary and complex, randomized (, , and ). Thus, we deem the two paths appropriate for our experiments.

Fig. 6: Driving paths used in Experiment I (left) and II (right).

The same five experts (EX: ) and 18 new novice drivers (NO: ; all male; 18–28 years old, M 22.8, SD 3.0) participated in collecting new driving data. The latter participants either did not have driving licenses or had licenses but with very little driving experience, e.g., young individuals who had not own and driven a car/motorcycle in the past two years. We controlled the novice drivers’ gender and age since they are important factors for motor learning (the same participants also participated in Experiment II).

As practice, participants drove in three 600-m short paths (). Then they completed driving in the 4-km long path while staying only in the first lane of the path at 60 km/h. The driving data of each participant was applied to the NN models ( and ) to obtain the trajectories of the predicted device angles ( and ).

Iv-B Performance Measures

Iv-B1 Modeling Performance

For each participant, the NN output and represent the control action that the experts would do after steps given the control vectors and , the vehicle state , and the environmental state of that participant; see (9) and (10). Hence, the following two errors indicate how different the participant’s driving action is from the predicted output of the experts’ action:


Let be an operator for computing the root mean square of all available samples of in the sequence . Then, the normalized RMSE, and , for each individual driving data are defined by


where , , , and are the maximum and minimum device angles from the experts’ training data used for NN modeling (also used for the training data normalization (Section III-B). and quantify the similarity of the participant’s driving skill to that of the five experts captured in the NN models.

Iv-B2 Objective Skill Performance

The driving skill of each participant is broken down to steering and pedaling performance. The steering performance is evaluated by a distance error and an angle error of the virtual vehicle as defined in Fig. 7. The distance error is the distance between the current car position and the closest point on the (invisible) midline of the first lane. The angular error is the angle between the car heading direction and the road frontal direction at the closest point on the midline of the first lane. Then we use and as measures for the steering performance.

For the pedaling performance, we first define a vehicle velocity error by where  km/h. In our simulator, the target speed 60 km/h imposed on the participants corresponds to the actual vehicle speed of when the needle of the speedometer reaches 60 km/h from the driver’s perspective. is used for a measure of the pedaling performance. Since the initial vehicle velocity is 0 km/h, is computed using only the velocity samples obtained after the vehicle speed first reaches . Additionally, as a measure of pedaling efficiency, we compute , where , focusing on the pedaling speed. increases if the participant operates the pedal more abruptly.

Fig. 7: Driving errors (, , and ) used in Experiment I and II. In our simulation, and  s is the look-ahead time.
(a) Expert E4, Right: magnified at 0–40 (s).
(b) Novice N11, Right: magnified at 0–40 (s).
Fig. 8: Examples of the recorded trajectory (black, dotted) and the desired trajectory (red, solid) in Experiment I.

Iv-C Results and Discussion

Fig. 9: Mean (left) and

(right). Error bars represent standard errors. Asterisks indicate statistically significant differences.

(a) Distance error
(b) Angular error
(c) Vehicle velocity error
(d) Pedaling speed
Fig. 10: Mean objective skill measure for the steering wheel (a and b) and the accelerator pedal (c and d). Error bars represent standard errors. Asterisks mean significant differences.

Fig. 8 shows examples of an expert () and a novice () who achieved a median performance of and among the respective groups. The expert’s driving trajectories seem to be in better agreement with the desired trajectories.

The means of the six performance measures are shown in Fig. 9 and 10. We applied Welch’s

-test (unequal sample sizes and unequal variances) to assess the effect of participant group (

EX and NO) on each measure. Results are: : EX (1.55 %) NO (2.55 %), , ; : EX (2.18 %) NO (5.78 %), , ; : EX (0.34 m) NO (0.46 m, , ; : EX (1.01 ) NO (1.45 ), , ; : EX (1.93 km/h) NO (2.18 km/h), , ; and : EX (1.58 degree/s) NO (4.27 degree/s), , .

These results provide answers to our research questions.


Our NN models ended training on the experts’ driving data along the simple paths when RMSE was less than 1 % and 4.5 % for and (Section III-C). In this experiment, the predictive errors of our NN models for the more complicated and general path were also very low with mean and mean for the driving data of the same experts. Hence, our models constructed using many simple paths can effectively predict experts’ desired driving behaviors for complicated paths composed of arbitrary curves and straight lanes.


For both model predictive errors and , the experts showed better performance than the novice participants with statistical significance. This result, since our NNs, and , were built using the same experts’ data, suggests that our model can represent the specific, common driving skills of expert drivers, which are especially different from those of other novice drivers.


The skill performance measures, and for steering and and for accelerator pedaling, all showed better performance with the experts than with the novice participants. The differences in and were statistically significant. Thus, the better driving skills of the experts were adequately captured with the four measures.

V Experiment Ii: Haptic Transfer

Our NNs can provide the desired device angle vector . We can design any performance-based HA, exerting haptic feedback to correct current movement, by utilizing the performance error vector . Among various HA, the most representative approach is haptic guidance, where external haptic stimuli are provided to learners concurrently during training in order to communicate information on the desired movement [28, 29]. To this end, we conducted an experiment comparing two performance-based haptic guidance for a driving skill; the one obtains from the NN and another conventionally formulates from a fixed environmental situation.

The followings are research questions: : Can haptic guidance be implemented with NN?; : Can haptic guidance implemented with our NN transfer experts’ driving behavior?; and : Can haptic guidance implemented with our NN provide the competitive performance to conventional haptic guidance?

V-a Methods

We report three different methods tested in the experiment.

V-A1 : No Haptic Guidance

A driver completes driving, receiving only realistic driving feedback (Section II-B).

V-A2 : Haptic Guidance with Neural Networks

A driver completes driving, receiving assitive haptic feedabck using NN. First, and (50 Hz) have been smoothend to and by moving average filters to command semi-continuous feedback (800 Hz). Let the desired device angles be predicted experts’ behavior, i.e. . Then, PID-based steering feedback to deliver can be computed as follows:


where is the recent time when becomes zero. Whole steering feedback is replaced by,


where provides stable feedback with increased viscosity and without the Coulomb friction.  Nm/degree,  Nms/degree, and  Nms/degree, and . By trials and errors, the gains have been appropriately tuned for two purposes; (1) to exert the steering wheel feedback strongly so that the virtual vehicle can complete driving only with pedal manipulations (similarly to autonomous steering), but also (2) to enable a driver to overcome the feedback to adjust device angles.

Since the driver’s foot and the acclerator pedal are not in full contact in any time, for assistive pedaling feedback, an unidirectional torque rather than PID-based feedback is utilized. to deliver is as follows:


which replaces in to . From (18), the accelerator pushes the driver’s foot upwoards when a driver pushes it more than . Then, whole pedaling feedback is altered to:


V-A3 : Conventional Haptic Guidance

A driver completes driving, receiving conventionally-designed assitive hpatic feedabck. In comparison to , determines by external environments. The same torque control equations (15) and (18) are adopted. For steering feedback, predictive haptic guidance ([17, 18]) was adopted. The predictive haptic guidance is based on the observation that a driver determines his/her driving based on a prediction. This method considers two error terms, a look-ahead direction error and the distance error (Figure 7), and determines a desired angle , as follows:


Using and  degree/m and the same torque gains of , this method can also support the vehicle to complete driving only with pedal manipulations.

For the accelerator, there exist several applicable algorithms [9, 10], but none of them guarantees effectiveness of training. Hence, we implemented a simpler, deterministic feedback which only provides overspeed cues. Let  km/h be a criterion of overspeed. We computes as follows:


From (22), the drivers perceives a impulse-like feedback from the right foot when the vehicle velocity exceeds .

V-B Experimental Protocol

Every participant (the same in Experiment I) completed three different driving trials in a complicated path different from the path in Experiment I (Fig. 6), by receiving corresponding assistive feedbacks. Since there are total possible permutations from three conditions, novices of three each was assigned to the same presentation order.

After each trial, the participant was asked to answer the following questions for both steering and pedaling feedbacks, respectively, on a 7-point Likert scale: (1) Was the training effective for driving? (Effectiveness); (2) Was the training comfortable/uncomfortable? (Comfort); (3) Was the training fun? (Fun); (4) Do you think a longer training under the corresponding feedback can help to improve your skill (Helpfulness). Thus, there were total 24 questions (4 questions 2 devices 3 conditions) for each participant. Every participant was paid KRW 15,000 ( USD 13) after the experiment.

V-C Results and Discussion

This section reports the quantitative (the same metrics in Experiment I) and qualitative results of Experiment II. For a statistical analysis, we applied a repeated measures ANOVA with methods as a within-subject factor. Tukey’s multiple testing was conducted as a post-hoc test for significant effects.

V-C1 Behavioral Similarity

We computed the predictive errors ( and ) for each resulted trajectory (Fig. 11). If a driving behavior is similar to experts’ behavior, then the errors decrease. The ranking of is . Since the assumption of sphericity had been violated from the Mauchly’s test (, ), the Greenhouse-Geisser estimate of sphericity () was used for recomputation of statistics. In results, there exists a significant difference (, ), and in the post-hoc test, (, ) and (, ). The ranking of is , and the assumption of sphericity was not violated (, ). There exists a significant difference (, ), and in the post-hoc test, (, ) and (, ). Two haptic guidance methods showed significant differences from .

Fig. 11: Mean (left) and (right) for each method. Error bars represent standard errors. Asterisks indicate statistically significant differences.
(a) Distance error
(b) Angular error
(c) Vehicle velocity error
(d) Pedaling speed
Fig. 12: Mean objective skill measure for the steering wheel and (a and b) and the accelerator pedal (c and d). Error bars represent standard errors. Asterisks mean significant differences.

V-C2 Objective Skill Performance

We computed the objective skill measures (, , , and ) for each resulted trajectory (Fig. 12). The ranking of is mmm, and the assumption of sphericity was not violated (). There exists a significant difference (, ), and in the post-hoc test, (, ). The ranking of is , Since the assumption of sphericity had been violated (), the Greenhouse-Geisser estimate of sphericity () was used for recomputation. In results, there exists a significant difference (, ), and in the post-hoc test, (, ), (, ), and (, ).

The ranking of is km/hkm/hkm/h, and the assumption of sphericity was not violated (). There exists a significant difference (, ), and in the post-hoc test, (, ). The ranking of is degree/sdegree/sdegree/s, and the assumption of sphericity was not violated (). There exists a significant difference (, ), and in the post-hoc test, (, ). and (, ). In summary, showed better performance of but worse performance of than . showed better performances of and , but worse performance of and than . In comparison between two haptic guidance methods, achieved better performance than in . However, they have no difference in other measures.

(a) Steering Wheel
(b) Accelerator Pedal
Fig. 13: Subjective scores responded from questionnaires (7-point Likert scale). Error bars represent stanard errors. Asterisks indicate significant differences.

V-C3 Qualitative Results

We computed the mean scores for each subjective question (Fig. 13

). For a statistical analysis, we applied Kruskal-Wallis test. Dunn’s post-hoc nonparametric test was conducted as a post-hoc test for significant effects. For the steering wheel, the rankings of the effectiveness, comfort, fun, helpfulness scores are:

(), (), (), and (), respectively. For the accelerator pedal, the ranks of the subjective scores are: (), (), (), and (), respectively. The significant differences are observed in the effectiveness/comfort/helpfulness scores of steering feedback. In the post-hoc test, the subjects reported that two haptic guidance methods are felt more effective than . However, they reported that only is felt more comfortable, and is felt less helpful than .

These results provide answers to our research questions.


We successfully implemented haptic guidance which involves performance error vector , utilizing which is a predicted outcome from NNs.


Receiving the steering feedback based on the NNs, the novices could steer the vehicle with decreased predictive errors (), which indicates that the novices had similar steering behavior to experts. However, receiving the pedaling feedback, the novices moved the accelerator with increased predictive errors (), which indicates that the novices had awkward pedaling behavior distinct from experts. Therefore, the steering feedback could, but the pedaling feedback could not effectively transfer experts’ behavior.


Both haptic guidance methods helped the novices to achieve better steering performance of than driving without guidance, which implies that the guidance methods can be adequately applied to the skill transfer. The predictive haptic guidance improved performance of but vitiated performance of , compared to haptic guidance with NN. Hence, it is inconclusive to assert which method is the better. The qualitative results also support that two guidance methods have competitive effectiveness. The effectiveness of the two methods can be vary depending on implementation details, e.g., tuning the parameters.

In contrast, our implementation of pedaling guidance using NN was inappropriate for the skill transfer. Only the pedaling feedback providing overspeed cues achieved better pedaling performance of , whereas haptic guidance with NN failed to show an improvement. Both haptic guidance methods increased , which indicates that the novices abruptly moved the accelerator pedal when the assistive feedback was given.

Vi General Discussion

Vi-1 Driving Skill Chracteristics

Point-to-point human movements consist of a gross, less accurate transfer motion with slower responses and fine, more accurate corrective movements with faster responses [30, 31], in a speed-accuracy tradeoff. We used four measures for analysis of driving performance, and each measure is closely related to an ability to corresponding subskills: for motion-initiating, for fine-tuning (both for steering), and for motion-initiating, for a fine-tuning (both for pedaling). A driving skill is a vague mixture of the subskills.

captured experts’ subskill of gross steering more effectively (Experiment I). Usually, haptic guidance is effective in transferring gross skills by providing kinetic references with specific timing and force [32, 18]. Hence, haptic guidance could successfully transfer experts’ steering behavior of gross motions (Experiment II). In contrast, captured experts’ subskill of fine pedaling more effectively (Experiment I). Haptic guidance may not be effective in transferring experts’ pedaling behavior of fine motions, which leads a failure of pedaling skill transfer (Experimet II).

We used the same configuration of  s for both NNs, which provides desired information in 5 Hz updating frequency. However, the sensing accuracy and dexterity of lower limbs are often regarded as lower than those of hands [33]. Moreover, the simultaneous nature of driving which requires manipulation of both steering and pedaling which impose learners a selection of learning either one of them. Therefore, compared to , may have too frequent assistive feedback, which resulted in ineffective skill transfer. Thus, we suggest that a NN with leading less frequent feedback may mediate better facilitation of haptic guidance for pedaling.

Vi-2 Application to Various Haptic Assistance

In this study, we selected haptic guidance among the variety of performance-bsed HA. However, haptic guidance has a demerit named the guidance hypothesis: excessive concurrent augmented feedback may make learners dependent on the feedback and reduce their focus during the training, rather interfering with retention of the learned skill [34]. In Experiment II, the novices reported that even they think the concurrent haptic feedback might not be helpful to a longer training, which is an attribution to the guidance hypothesis.

The human performance error vector formulated by NNs is adaptable to other HA. For example, error amplification providing the haptic stimuli that increase trajectorial errors [35], or haptic disturbance (an extension to error amplification) providing random, unpredictable force fields [36] can employ the same performance errors. Hence, they are possible candidates for our approach based on modeling of experts’ skill using NNs, which can induce more effectiveness of driving skill training.

Vii Conclusions

We developed a haptic driving training simulator providing realistic experiences, to accomplish modeling and transferring a driving skill. In our simulator, performance-based haptic feedback can be delivered to a learner to assist the training of simultaneous manipulation of both a steering wheel and an accelerator pedal. To design proper haptic feedback, an adequate optimized model of the skill is ncessary, and we used NNs to extract a driving expert’s motor behavior for modeling of the skill. To this end, we validated our model with predictive errors and proved objective performance of haptic guidance using the model via human experiments. In results, our appoach showed a potential to transferring experts’ skill. We are planning to conduct a user study to figure out the educational effectiveness of our approach in a longer training.

We note that still opportunities of other famous approaches, such as Learning from Demonstration (LfD) [37]

, remain for efficient reference modeling for the driving skill. Moreover, several machine learning techniques based on a human decision behavior, such as Hidden Markov Model (HMM) 

[21] would be a novel addition for our modeling approach for more difficult, decision-based driving tasks. The area of these studies is the direction our research should proceed in future.


The authors would like to thank…


  • [1] Y. Gaffary and A. Lécuyer, “The use of haptic and tactile information in the car to improve driving safety: A review of current technologies,” Frontiers in ICT, vol. 5, p. 5, 2018.
  • [2] M. Mulder, M. R. v. Paassen, and E. R. Boer, “Exploring the roles of information in the manual control of vehicular locomotion: From kinematics and dynamics to cybernetics,” Presence-Teleop. Virt., vol. 13, no. 5, pp. 535–548, 2004.
  • [3] D. A. Abbink and M. Mulder, “Neuromuscular analysis as a guideline in designing shared control,” in Advances in haptics.   Vienna, Austria: InTech, 2010, pp. 499–516.
  • [4] M. K. O’Malley, A. Gupta, M. Gen, and Y. Li, “Shared control in haptic systems for performance enhancement and training,” J. Dyn. Syst. Meas. Contr., vol. 128, no. 1, pp. 75–85, 2006.
  • [5] M. Steele and R. B. Gillespie, “Shared control between human and machine: Using a haptic steering wheel to aid in land vehicle guidance,” Proc. Hum. Factors Ergon. Soc. Annu. Meet., vol. 45, no. 23, pp. 1671–1675, 2001.
  • [6] B. A. C. Forsyth and K. E. MacLean, “Predictive haptic guidance: Intelligent user assistance for the control of dynamic tasks,” IEEE Trans. Visual Comput. Graphics, vol. 12, no. 1, pp. 103–113, 2006.
  • [7] L. Saleh, P. Chevrel, F. Claveau, J.-F. Lafay, and F. Mars, “Shared steering control between a driver and an automation: Stability in the presence of driver behavior uncertainty,” IEEE Trans. Intell. Transp. Syst., vol. 14, no. 2, pp. 974–983, 2013.
  • [8] D. A. Abbink, M. Mulder, F. C. Van der Helm, and E. Boer, “Measuring neuromuscular control dynamics during car following with continuous haptic feedback,” IEEE Trans. Syst. Man Cybern. B Cybern., vol. 41, no. 5, pp. 1239–1249, 2011.
  • [9] M. Mulder, D. A. Abbink, M. M. van Paassen, and M. Mulder, “Design of a haptic gas pedal for active car-following support,” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 1, pp. 268–279, 2011.
  • [10] A. H. Jamson, D. L. Hibberd, and N. Merat, “The design of haptic gas pedal feedback to support eco-driving,” in Proceedings of the International Driving Symposium on Human Factors in Driver Assessment, Training, and Vehicle Design, 2013, pp. 264–270.
  • [11] M. Mulder, D. A. Abbink, and E. R. Boer, “Sharing control with haptics: Seamless driver support from manual to automatic control,” Hum. Factors, vol. 54, no. 5, pp. 786–798, 2012.
  • [12] S. M. Petermeijer, D. A. Abbink, M. Mulder, and J. C. de Winter, “The effect of haptic support systems on driver performance: A literature survey,” IEEE Trans. Haptics, vol. 8, no. 4, pp. 467–479, 2015.
  • [13] D. J. Reinkensmeyer. and J. L. Patton, “Can robots help the learning of skilled actions?” Exerc. Sport Sci. Rev., vol. 37, no. 1, pp. 43–51, 2009.
  • [14] L. Marchal-Crespo and D. J. Reinkensmeyer, “Review of control strategies for robotic movement training after neurologic injury,” Journal of neuroengineering and rehabilitation, vol. 6, no. 1, p. 20, 2009.
  • [15] R. A. Magill, Motor Learning: Concepts and Applications, 6th ed.   New York, NY, USA: McGraw-Hill, 2000.
  • [16] H. Heuer and J. Lüttgen, “Robot assistance of motor learning: A neuro-cognitive perspective,” Neurosci. Biobehav. Rev., vol. 56, pp. 222–240, 2015.
  • [17] L. Marchal-Crespo and D. J. Reinkensmeyer, “Haptic guidance can enhance motor learning of a steering task,” J. Mot. Behav., vol. 40, no. 6, pp. 545–556, 2008.
  • [18] H. Lee and S. Choi, “Combining haptic guidance and haptic disturbance: An initial study of hybrid haptic assistance for virtual steering task,” in Proceedings of the IEEE Haptics Symposium, 2014, pp. 159–165.
  • [19] M. Hirokawa, N. Uesugi, S. Furugori, T. Kitagawa, and K. Suzuki, “Effect of haptic assistance on learning vehicle reverse parking skills,” IEEE Trans. Haptics, vol. 7, no. 3, pp. 334–344, 2014.
  • [20] Y. Zhang, W. C. Lin, and Y.-K. S. Chin, “A pattern-recognition approach for driving skill characterization,” IEEE Trans. Intell. Transp. Syst., vol. 11, no. 4, pp. 905–916, 2010.
  • [21] M. C. Nechyba and Y. Xu, “Human control strategy: Abstraction, verification, and replication,” IEEE Control Systems Magazine, vol. 17, no. 5, pp. 48–61, 1997.
  • [22] M. C. Nechyba, “Learning and validation of human control strategies,” Ph.D. Thesis, The Robotics Institue, Carnegie Mellon University, 1998.
  • [23] Vehicle Physics Pro (VPP).
  • [24] T. Hiraoka, O. Nishihara, and H. Kumamoto, “Steering reactive torque presentation method for a steer-by-wire vehicle,” Rev. Automot. Eng., vol. 29, pp. 287–294, 2008.
  • [25] M. C. Nechyba and Y. Xu, “Human skill transfer: Neural networks as learners and teachers,” in Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, vol. 3, 1995, pp. 314–319.
  • [26] K. S. Narendra and K. Parthasarathy, “Identification and control of dynamical systems using neural networks,” IEEE Trans. Neural Networks, vol. 1, no. 1, pp. 4–27, 1990.
  • [27] T. L. Brooks, “Telerobotic response requirements,” in Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, 1990, pp. 113–120.
  • [28] R. B. Gillespie, M. S. O’Modhrain, P. Tang, D. Zaretzky, and C. Pham, “The virtual teacher,” in Proceedings of the ASME Dynamic Systems and Control Division, 1998, pp. 171–178.
  • [29] D. Feygin, M. Keehner, and F. Tendick, “Haptic guidance: Experimental evaluation of a haptic training method for a perceptual motor skill,” in Proceedings of the International Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems, 2002, pp. 40–47.
  • [30] E. Burdet and T. E. Milner, “Quantization of human motions and learning of accurate movements,” Biol. Cybern., vol. 78, no. 4, pp. 307–318, 1998.
  • [31] B. Corteville, E. Aertbelien, H. Bruyninckx, J. De Schutter, and H. Van Brussel, “Human-inspired robot assistant for fast point-to-point movements,” in Proceedings of the IEEE International Conference on Robotics and Automation, 2007, pp. 3639–3644.
  • [32] L. Marchal-Crespo, S. McHughen, S. C. Cramer, and D. J. Reinkensmeyer, “The effect of haptic guidance, aging, and initial skill level on motor learning of a steering task,” Exp. Brain Res., vol. 201, no. 2, pp. 209–220, 2009.
  • [33] E. Velloso, D. Schmidt, J. Alexander, H. Gellersen, and A. Bulling, “The feet in human–computer interaction: A survey of foot-based interaction,” ACM Comput. Surv., vol. 48, no. 2, pp. 21:1–21:35, 2015.
  • [34] R. A. Schmidt, “Frequent augmented feedback can degrade learning: Evidence and interpretations,” in Proceedings of the NATO Advanced Study Institute, J. Requin and G. E. Stelmach, Eds., 1990, pp. 59–75.
  • [35] J. L. Emken and D. J. Reinkensmeyer, “Robot-enhanced motor learning: Accelerating internal model formation during locomotion by transient dynamic amplification,” IEEE Trans. Neural Syst. Rehabil. Eng., vol. 13, no. 1, pp. 33–39, 2005.
  • [36] J. Lee and S. Choi, “Effects of haptic guidance and disturbance on motor learning: Potential advantage of haptic disturbance,” in Proceedings of the IEEE Haptics Symposium, 2010, pp. 335–342.
  • [37] B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,” Rob. Auton. Syst., vol. 57, no. 5, pp. 469–483, 2009.