1 Introduction↩︎

Since the inception of driving simulator technology, a key challenge has been to bring the driving simulator experience closer to the real vehicle experience. However, the realistic (unscaled) visual motion is still presented with scaled or even without any physical motion. This causes a mismatch between expected and perceived motion, eliciting motion sickness [1]. Motion sickness is a syndrome that arises as a consequence of a wide range of self-motion and orientation cues. It is characterised by symptoms of sweating, headache, dizziness, stomach awareness, where these symptoms usually grow in severity until nausea, retching and ultimately vomiting occurs [2]. The mechanisms behind the development and evolution of motion sickness have been studied extensively, relying heavily on models that predict sensory conflicts based on well-known mathematical models of vestibular and visual sensory integration [3]–[7].

In driving simulators, motion sickness typically occurs due to the intricate interplay of synthetic parallel visual, vestibular, and proprioceptive cues. A recent meta-analysis on simulator sickness [8] reviewed 41 studies and reported modest sickness levels across different simulator configurations. It found that improved visual fidelity significantly reduced sickness in motion-base simulators, but not in fixed-base systems. Mechanical motion presence had a minor, non-significant effect on sickness. Active driving reduced sickness compared to passive driving, although these comparisons were across different studies. Notably, active driving rarely induces sickness in real vehicles, highlighting a gap between simulation and reality. In a comparative study [9], three motion cueing strategies were assessed against a fixed-base setup. While the motion-cueing strategies improved perceived realism, they were also associated with increased sickness incidence, highlighting a trade-off between immersion and physiological comfort.

In this paper, we aim to develop a motion cueing algorithm (MCA) that can recreate specific forces from a real vehicle drive in a driving simulator, while explicitly minimizing the level of motion sickness. To the best of the authors’ knowledge, no prior work has established an MCA explicitly aimed at reproducing or reducing motion sickness in a simulator environment. For this, we need an MCA that can take into account both the simulator dynamics and the time evolution of motion sickness. New approaches to develop MCAs such as [10]–[12] based on model predictive control (MPC) are well suited as they solve a constrained optimal control problem and can take advantage of models of simulator dynamics and motion sickness development in their optimization. Traditionally, simple metrics of motion sickness severity are optimised, such as the motion sickness dose value (MSDV) in [13]–[16] for vehicle trajectory planning. While effective in optimising the magnitude of horizontal motion in real vehicles, such methods will not capture the effects of the frequency of motion, platform rotation, and conflicting visual and mechanical motion. To capture these effects, we selected the 6-DoF subjective vertical conflict (SVC) model from [4]. Comparing a range of models, this model best matched motion sickness data collected in several laboratory experiments and real-world driving studies [7]. In [17], we extended this SVC model with a personalised motion sickness accumulation model. We calibrated the accumulation parameters such that the model predicts sickness rated with the well-established MISC scale. Moreover, a population model was added to predict the sickness variation within a population. This paper presents a novel MPC-based MCA that mitigates motion sickness, using these motion sickness models to find a balance between motion sickness reduction and specific force recreation. We validate the MCAs in human-in-the-loop experiments, also including a reference adaptive washout MCA and a no motion condition.

The contributions of the work are outlined below:

• Incorporated a six-degree-of-freedom subjective vertical conflict motion sickness model, for the first time, directly into a motion cueing algorithm as part of its cost function.

• Formulated a multi-objective optimisation framework that jointly optimizes for motion sickness reduction and reproduction of reference vehicle motion in the driving simulator.

• Demonstrated that inclusion of the motion sickness model allows for an adjustable trade-off between motion fidelity and predicted motion sickness, enabling the algorithm to prioritise either motion fidelity or participant comfort based on application-specific requirements.

• Conducted a human-in-the-loop experiment confirming that the proposed algorithm effectively mitigates motion sickness, while having minimal impact on perceived motion fidelity.

2 Motion Cueing Algorithms↩︎

2.1 MPC-based MCA↩︎

The evaluation framework of the proposed motion cueing algorithm is shown in 1. The inputs are the reference vehicle specific forces, the reference sensory conflict representing motion sickness, and the yaw acceleration. In this paper, the reference sensory conflict is set to zero, resulting in an MPC that aims to minimise motion sickness.

The upper part of 1 represents the proposed algorithm combining hexapod dynamics and the sensory conflict generation model in the MPC framework. The considered hexapod dynamics consist of translational (fore-aft and sway) motion, and correspondingly, pitch and roll motion. The equations for the hexapod dynamics are defined in detail in 9 of the Appendix. Assuming yaw motion is a separate channel, it is controlled via a parallel washout channel to reduce computational complexity. More details regarding the control of yaw motion can be found in 11 of the Appendix. The vertical motion is neglected in this paper, assuming a perfectly flat road. The reference vehicle specific forces and yaw accelerations are preconditioned by scaling and limitations based on hexapod capability.

To calculate the sensory conflict, \(c_v\), the subjective vertical conflict model with visual rotational velocity (SVC-VR) is used [18]. It is based on the sensory conflict generation model proposed in [4] with further validation for motion sickness predictions in vehicles [7], [17]. The conflict in this model is a 3-DoF vector; therefore, the vector norm results in a scalar. In this case, the instantaneous conflict is used to predict sensory mismatch, without accumulation dynamics. Hence, the conflict term is effectively scalar and does not directly include the slow integration effects present in accumulation dynamics. Minimising this scalar conflict over the prediction horizon can reduce motion sickness stimuli, but does not replicate the full MISC integration (accumulation dynamics based on [17] shown on bottom right, 1), which is still computed offline for weight selection purposes. More details on the integration of the SVC model in the MPC can be found in 8 of the Appendix.

The proposed algorithm explicitly optimises the trade-off between motion sickness mitigation and motion perception fidelity. This is achieved by the introduction of the following components in the cost function: \(J_{MP}\) is related to the tracking performance, \(J_{MS}\) corresponds to the motion sickness through sensory conflict, and state costs \(J_{pen}\). The resulting cost function is defined as:

\[\label{eq:mpc95cost95split} J_c = J_{MP} + J_{MS} + J_{pen},\tag{1}\] with \[\begin{align} J_{MP} &= (f_{spec}-f_{ref})^\top W_{spec} (f_{spec}-f_{ref}), \\ J_{MS} &= w_c (c_v - c_{vref})^2, \\ J_{pen} &= (X-X_{ref})^\top W_x (X-X_{ref}), \end{align}\] where \(f_{spec} \in \mathbb{R}^2\), \(c_v \in \mathbb{R}\), and \(X \in \mathbb{R}^{n_x}\) contains all considered states related to the hexapod and the sensory conflict generation model.

The weighting matrices are positive semidefinite and dimensionally compatible with the corresponding cost components:

• \(W_{spec} \in \mathbb{R}^{2 \times 2}\), weighting the two components of the specific force in \(J_{MP}\).

• \(w_c \in \mathbb{R}^{1 \times 1}\), weighting the scalar sensory conflict in \(J_{MS}\).

• \(W_x \in \mathbb{R}^{n_x \times n_x}\), a diagonal matrix weighting all states of the hexapod and the sensory conflict generation model in \(J_{pen}\), where \(n_x = 34\) (same as the size of \(X\)).

The state vector is defined in condensed form as

\[\label{eq:X95condensed} X = \begin{array}{ccccccccc} [\;\theta_{hex} & \omega_{hex} & s_{hex} & v_{hex} & f_{tilt} & \delta & a_{hex} & \alpha_{hex} \\ w_s & \hat{w}_s & v_s & \hat{v}_s & \tilde{v}_s & \hat{f}_{spec} &]^T\end{array}\tag{2}\]

where \(\theta_{hex}\) is the angular position of the hexapod (\(rad\)), \(\omega_{hex}\) is the angular velocity of the hexapod (\(rad/s\)), \(s_{hex}\) is the position of the hexapod (\(m\)), \(v_{hex}\) is the velocity of the hexapod (\(m/s\)), \(f_{tilt}\) is the force due to the tilt component of gravity (\(m/s^2\)), \(\delta\) is a dimensionless slack variable, \(a_{hex}\) is the linear acceleration of the hexapod (\(m/s^2\)), \(\alpha_{hex}\) is the angular acceleration of the hexapod (\(rad/s^2\)), \(w_s\) and \(\hat{w}_s\) are sensed and expected angular velocities in the motion sickness model (\(rad/s\)), and \(v_s\), \(\hat{v}_s\), \(\tilde{v}_s\), and \(\hat{f}_{spec}\) are verticality and specific force states in the motion sickness model (\(m/s^2\)). The hexapod states \((\theta_{hex}, \omega_{hex}, s_{hex}, v_{hex}, f_{tilt}, \delta, a_{hex}, \alpha_{hex})\) are two-element vectors in \(\mathbb{R}^2\), whereas the motion sickness model states \((w_s, \hat{w}_s, v_s, \hat{v}_s, \tilde{v}_s, \hat{f}_{spec})\) are three-element vectors in \(\mathbb{R}^3\).

To ensure physically feasible and safe motions, constraints are applied, including workspace limits on positions and orientations, as well as velocity and acceleration limits. Slack variables \(\delta\) allow for flexible handling of constraint violations. A description of constraints and slack variable implementation can be found in 12 of the Appendix. Additional workspace management strategies, such as washout and dynamic constraints, are described in 10 of the Appendix. The weight parameters and the motivation for selecting weight parameters can be found in 6 of the Appendix.

For this replay scenario, the algorithm is performed offline using the full horizon at each step to compute optimal inputs for the simulator. The prediction horizon has been selected as 3 seconds (60 steps and sampling time of 0.05 s). Shorter horizons reduced performance, while the selected amount of prediction steps and sampling time result in a trade-off between prediction accuracy and computational feasibility, with a real-time factor of approximately 2.1.

2.2 Adaptive Washout MCA↩︎

In the results below, we compare the MPC-based MCA to a state-of-the-art adaptive washout (AW) MCA introduced in [19]–[21], with MPC-based direct workspace management as used in [12].

This adaptive washout consists of the following components

• Translation channels: 1st order high-pass filter

• Rotation channels: 3rd order high-pass filter

• Tilt coordination: 2nd order low-pass filter

The used parameter settings for the adaptive washout are listed in 1. This configuration is consistent with settings used in our previous work [12].

2.3 Objective Evaluation of MCAs↩︎

This subsection evaluates the performance of the algorithms objectively. Based on the objective evaluation, suitable MCA conditions are selected for the experimental evaluation. Here we already use the driving scenario defined for the experiment, see 3.2.

The MPC-based algorithm under consideration addresses two primary objectives: accurate specific force tracking and minimisation of sensory conflict. To comprehensively explore the trade-off between these competing goals, we vary the ratio between the specific force tracking weight and the sensory conflict weight across the entire range, from pure conflict minimisation to pure force tracking. This is done by varying the sensory conflict weight, \(w_{con}\), between 0 to 1 while maintaining the relation \(w_{con} + w_{spec} = 1\).

2 illustrates the resulting performance metrics for different values of \(w_{con}\). The upper two graphs illustrate predicted motion sickness in terms of sensory conflict and MISC, whereas the other four graphs illustrate the corresponding workspace utilisation. In all graphs, the horizontal axis is defined as the specific force tracking error. Hence, the upper graphs represent the typical trade-off of sickness versus fidelity as a Pareto front.

The green triangle represents the no motion case, which practically coincides with the conflict weight of unity in our algorithm. Thus, the algorithm indicates the reduction of motion sickness to be the most when the simulator platform does not move at all. The blue dot represents the adaptive washout based MCA as described in 2.2.

For the experimental evaluation, it is important to ensure a clear distinction between the predicted MISC levels of the different algorithm configurations. Therefore, a minimum difference of at least one MISC level is desired. With this minimum MISC difference and other metrics in 2 in consideration, we have chosen two MPC-based MCAs for the experiment along with the adaptive washout and no motion condition. Hence, the four MCA conditions are:

• No Motion (NM): This is a baseline case, where there is no motion of the simulator, but there still are moving visuals.

• \(MPCw_{con0.9}\): This is a MPC-based MCA with a \(\omega_{con}=0.9\). This MCA is expected to be a good compromise, balancing sickness reduction and specific force reproduction.

• \(MPCw_{con0}\): This is a MPC-based MCA with a \(\omega_{con}=0\). This MCA is expected to have very good specific force reproduction while causing more motion sickness than \(MPCw_{con0.9}\).

• Adaptive Washout (AW): This is a state-of-the-art MCA widely used in industrial applications.

These MCAs and their expected performance are summarised in 2. In the experiment, speed-dependent road-induced vibrations were added as vertical motion in all conditions. For objective comparison of the selected algorithms, specific force tracking, shape similarity, and the median predicted MISC at the end of the experiment are considered.

Specific force tracking is the RMSE between the reference and the obtained total horizontal specific force (resultant of \(f_x\) and \(f_y\)) from the MCA. It portrays how closely the magnitudes of the reference and generated profiles match/align. For the AW method, the RMSE is 0.097 \(m/s^2\). The \(MPCw_{con0}\) configuration achieves lower RMSE values of 0.027 \(m/s^2\), indicating a closer match to the reference. In contrast, the \(MPCw_{con0.9}\) configuration results in higher RMSE values of 0.216 \(m/s^2\), reflecting the reduced motion magnitudes.

Shape similarity is the Pearson correlation between the reference and generated profile. This ensures the effectiveness of the algorithm. A manoeuvre with a similar shape to the obtained specific force results in a similar experience. Despite magnitude scaling, a higher shape similarity provides a better recreation of the manoeuvre. The AW method achieved shape similarity scores of 0.89. The \(MPCw_{con0}\) configuration yielded higher shape similarity scores of 0.99, indicating near-perfect alignment. The \(MPCw_{con0.9}\) configuration also performed well, with scores of 0.97.

The median predicted MISC at the end of the experiment quantifies the induced motion sickness severity of each MCA. The \(MPCw_{con0}\) and AW had similar MISC levels of 3.0 and 2.8, respectively. In contrast, NM produced the least MISC of 0.9. The \(MPCw_{con0.9}\) had an intermediate MISC of 2.0.

The selected MCA configurations were further analysed based on the rendered specific force profiles. 3 presents the specific force profiles produced by each configuration. More detailed graphs with rotation and displacement data are shown in 13 of the Appendix. The results show that the AW algorithm and the \(MPCw_{con0}\) configuration yield very similar profiles, whereas the \(MPCw_{con0.9}\) configuration results in noticeably lower specific force magnitudes.

3 Human-in-the-loop Evaluation↩︎

This section describes the human-in-the-loop driving simulator experiment and its subjective evaluation. Perceived driving simulator fidelity and motion sickness were evaluated from the perspective of passive users, representative of users of automated vehicles. This work utilises the Delft Advanced Vehicle Simulator (DAVSi), a 6-DoF moving-base driving simulator [22], capable of generating accelerations of up to 1 g in all directions at a frequency of 10 Hz.

3.1 Experimental Procedure↩︎

All participants provided informed consent before participating in the study. The human research ethics committee of TU Delft, The Netherlands, approved the experimental protocol (application number \(4819\)).

In total, 20 participants from the pool of students and employees of TU Delft participated in the study (mean age: 27.70 years, std: 3.42 years, 6 females, 14 males). The individual motion sickness susceptibility was evaluated using the motion sickness susceptibility questionnaire - short (MSSQ-Short) by [23], yielding a mean MSSQ score of 49.6 with a standard deviation of 24.9, which is well above the population mean of 12.9, indicating a high level of motion sickness susceptibility among the participants.

Each participant experienced all four selected motion cueing algorithm configurations, each in a separate simulator session, in a randomised order. To minimise carry-over effects and to mitigate motion sickness influence from previous exposures, the sessions were spaced at least 48 hours apart for each participant.

Before the experiment, the participants underwent a concise briefing session to familiarise themselves with the questionnaire and to understand the objective of the experiment. During the experiment, two-way communication was established between the experimenter and the participant via Bluetooth headphones and microphones.

Measures were taken to ensure that the driving scene remained their sole visual focus, including blocking side window views and part of the windshield to eliminate external cues indicating platform tilt.

During each session, participants reported their motion sickness levels every 30 seconds using the misery scale (MISC) [24], prompted by an auditory beep in the headphones. After completing a session, participants filled out an absolute grading questionnaire (7 in the Appendix) assessing different aspects of the driving experience.

To ensure a fair comparison, the sequence of algorithm exposures varied across participants. A latin square design was employed to balance the experimental testing order. Participants were permitted to withdraw from the experiment at any point; however, no individuals chose to exercise this option.

3.2 Scenario↩︎

The aim of this study is to investigate simulator sickness in relation to motion cueing in driving simulators. To this end, a virtual driving scenario was designed to simulate a naturalistic urban drive that would not typically induce motion sickness in real-world driving, but may do so in a simulator due to sensory mismatch and strong visual motion cues.

Since the ultimate goal is to develop strategies to mitigate simulator sickness, the experiment must first ensure a sufficiently high likelihood of inducing sickness. Given the naturalistic and non-aggressive nature of the driving scenario, participants were exposed to it for an extended period to allow the gradual onset of sickness symptoms. Based on this consideration, the total session duration was set to approximately 30 minutes.

The driving scenario consisted of 240-second laps, followed by a 10-second pause, repeated six times per session. This structure provided consistency across participants while maintaining a manageable session length and data segmentation for analysis. The urban driving scenario included vehicle speeds ranging from 0 to 70 km/h and consisted of a diverse set of manoeuvres: stop-and-go sequences at traffic lights and pedestrian crossings, moderate cornering, a tunnel section, and a double lane-change manoeuvre due to a road diversion caused by an accident. These dynamic events were chosen to replicate realistic driving conditions while introducing sufficient variations in acceleration to challenge the simulator’s motion rendering capabilities.

The scenario was created using IPG CarMaker, which provides a realistic virtual environment and high-fidelity vehicle dynamics simulation. CarMaker’s vehicle models are experimentally validated, ensuring that the generated accelerations are representative of actual driving conditions. These acceleration signals serve as the reference specific forces for the motion cueing algorithm that drives the simulator platform.

The resulting data shows lateral accelerations within the range of approximately \(\pm 4 m/s^2\), longitudinal accelerations up to \(1.25~\mathrm{m/s^2}\), and decelerations reaching \(-5 m/s^2\). These dynamic variations are used to evaluate the simulator’s motion cueing effectiveness and its relationship with simulator sickness development. The reference provided to all MCAs (MPC based and AW) was scaled down by a factor of 0.3 and 0.4 for the longitudinal and lateral vehicle accelerations, respectively. The choice of the scaling factors is coherent with our paper using MPC in conjunction with a frequency splitting MCA [12], where it was based on participant feedback obtained during pilot studies to enhance realism, and on other studies using similar scaling factors [25], [26].

3.3 Results↩︎

3.3.1 Subjective Realism Ratings↩︎

Based on participant realism scores collected during the experiment, a subjective evaluation of the motion cueing algorithms was conducted. 4 presents a summarised boxplot of the ratings for the different motion cueing configurations. The results indicate that the three conditions with platform motion (\(MPCw_{con0.9}\), \(MPCw_{con0}\), and AW) received closely clustered ratings across all realism criteria. In contrast, the \(NM\) (no motion) configuration consistently received the lowest scores, indicating that participants perceived it as the least realistic. These findings imply that the absence of motion cues in the \(NM\) configuration significantly reduced the sense of immersion, making it feel less like a real vehicle drive. Conversely, the other configurations were perceived to deliver comparable levels of realism and immersion, reinforcing the value of motion-based cues in enhancing perceived driving fidelity.

4 highlights statistically significant differences between the configurations, calculated using analysis of variance (ANOVA), using asterisks (‘*’). These markers indicate that each motion-based configuration exhibits a statistically significant difference in perceived realism when compared to the \(NM\) configuration. However, no significant difference is observed among the motion-based configurations themselves, further supporting the conclusion that they provide a comparable level of immersive experience.

3.3.2 Motion Sickness Ratings↩︎

The experimental motion sickness (MISC) as a function of time is presented in 5, along with the predictions from the model framework. Coloured lines represent experimental (solid) and predicted (dashed) results for the four configurations used in the experiment. The black dotted line shows the predicted motion sickness if the scenario would be driven in a real vehicle with external vision. Dropout rates and variance are predicted using the SVC model in conjunction with our population-based motion sickness model [17], [18].

The experimental data aligns well with the MISC levels predicted by the model framework. Although experimental values do not exactly match the model predictions, they follow a similar trend across configurations (see 5). The SVC model generally overestimates MISC levels relative to the observed data, with the exception of the \(MPCw_{con0}\) configuration, where the predictions are lower than the measured values at the end of the experiment.

As predicted by the model, in the experiment, the \(NM\) configuration results in the lowest levels of motion sickness. Slightly higher levels are seen with the configuration, followed by \(AW\), and the highest levels are recorded for . This trend confirms that increasing the weight on sensory conflict leads to a reduction in motion sickness, as expected from the offline analysis in 2. Additionally, only one participant dropped out in the \(MPCw_{con0.9}\) condition, compared to six in \(MPCw_{con0}\) and four in the \(AW\) condition. This confirms that the \(MPCw_{con0.9}\) is capable of extending the exposure time in the simulator.

The experimental MISC data were checked for condition effects and order effects using the mean MISC over the last lap. For this analysis, participants who dropped out had their MISC level assumed to be 6 from the point of dropout until the scheduled end of the experiment. The statistical significance, calculated using analysis of variance (ANOVA), is tabulated in 3, and boxplots are shown in 6. There was a significant overall effect of conditions on MISC, and there was a significant difference between the pairs (a) \(NM\) and \(MPCw_{con0}\) (b) \(NM\) and \(AW\) (c) \(MPCw_{con0.9}\) and \(AW\) (d) \(MPCw_{con0.9}\) and \(MPCw_{con0}\), whereas, no significant differences were found between (a) \(NM\) and \(MPCw_{con0.9}\) and (b) \(AW\) and \(MPCw_{con0}\). However, there were no significant differences because of order effects, i.e., the order in which the participant was subjected to the algorithms did not have a significant effect on the output of the experiment.

In 7, we compare the histogram of individual MISC levels at the end of the experiment, as predicted by the model framework and as observed in the experimental data. While there are some deviations in the exact occurrence across MISC levels, the overall trend is reasonably well captured by the motion sickness model. The most notable discrepancy occurs at MISC level zero, where the model underestimates the proportion of participants reporting no motion sickness, failing to predict a MISC of zero for a substantial segment of the population.

4 Discussion↩︎

This work addresses a pivotal gap in the domain of driving simulator motion cueing by directly integrating motion sickness mitigation as an explicit control objective. Historically, the development of MCAs has emphasised fidelity in motion perception, often overlooking the physiological consequences of simulator use, especially motion sickness. While prior works have explored high-fidelity motion rendering through MPC, to the best of the authors’ knowledge, no work has incorporated advanced motion sickness models, such as the 6-DoF subjective vertical conflict model (SVC) [4], to anticipate and suppress motion sickness.

The algorithm introduced here not only enhances the perceptual accuracy of motion (via specific force tracking), but also integrates a dedicated cost term that helps to minimize human discomfort through sensory conflict. This dual-objective formulation represents a significant methodological advancement.

A notable strength of this work is the experimental validation of this predicted improvement. Through simulation, objective metrics (like RMSE and shape similarity), and human-in-the-loop evaluation, the proposed solution was shown to be robust across both algorithmic performance and perceptual quality. The experimental results confirm that sensory conflict minimisation, even at the expense of reductions in specific force magnitude, can lead to a statistically significant improvement in user comfort without degrading immersion and realism. The successful prediction of MISC levels using a population-based SVC and sickness accumulation model further emphasises the algorithm’s predictive validity and practical utility.

4.1 MISC Prediction↩︎

In a novel contribution to the field, this work presents the first-ever quantitative prediction of absolute motion sickness levels using the SVC model in combination with a sickness accumulation model in driving simulator motion cueing. Here we build on our recent papers where we combine the SVC model with an accumulation model and estimated individual parameters to capture experimental MISC values in vehicle and driving simulator experiments [17], [18]. Prior research has predominantly focused on trends or qualitative sickness development; in contrast, our approach enables the prediction of absolute MISC scores. As sickness varies between individuals, we generalise our model using population-representative parameters. As a result, we provide predicted average MISC responses, accompanied by the standard deviation to capture inter-individual variability. This methodology offers a new standard for objective, model-based evaluation of motion sickness in simulator contexts.

As shown in 5, the MISC predictions (represented by dashed lines) follow a trend similar to the experimental data (represented by solid lines). Additionally, the predicted MISC levels remain close to the actual values recorded at the end of the experiment.

Interestingly, the adaptive washout and the configuration \(MPCw_{con0}\) yield very similar MISC predictions. However, in the experiment, adaptive washout was more sickening. Likewise, the experimental MISC fluctuates more across time, whereas the predicted MISC increases more gradually, providing scope to enhance the accumulation model and its parameters. MSSQ was reported to be higher in the participants of this experiment, indicating an above-average susceptibility, but the experimental MISC was not higher than the model-predicted MISC. However, MSSQ is an imperfect predictor of susceptibility, see [27] for a recent discussion. Certain instances of false cues may have a higher/bigger impact on motion sickness than others, which the model may not be able to discern/quantify (such as prepositioning or coupled translational and rotational motion). However, even with these discrepancies, the trend of prediction of motion sickness severity by the SVC model still matched the measurements. This is particularly evident comparing \(MPCw_{con0}\), \(MPCw_{con0.9}\) and \(NM\), where the experimental results follow the trend predicted by the SVC model.

A generalised rather than individual-specific motion sickness model was used in the MCA. Nevertheless, the use of generalised parameters still proves valuable for reducing motion sickness. This is because it is more important to lower the sickening stimuli than to precisely predict their impact on MISC levels. In the developed algorithm, sensory conflict is minimised over all time points in the MPC horizon and the slow accumulation process is ignored. One of the reasons for doing this is to make the problem computationally feasible. The offline predictions and the experiment demonstrate that this effectively reduces sickness over the entire experiment duration.

4.2 Realism↩︎

This work evaluates realism through both subjective and objective measures. The objective metrics include specific force tracking and the shape similarity factor. 4 indicates that the perceived realism is not significantly impacted by the different motion cueing algorithms. Only no motion had reduced realism perception. Realism remained consistent across the model predictive control (MPC) algorithms and the adaptive washout algorithm, even when the overall motion was substantially reduced, as evidenced by the reduction in specific force RMSE reported in 2. These findings suggest that shape similarity, as calculated by the Pearson correlation coefficient, is a more reliable indicator of perceived realism than the specific force RMSE. This implies that the temporal shape of the motion signal plays a more critical role in realism perception than the magnitude of the experienced forces. Additional data is needed to determine whether realism remains consistent across the entire weight spectrum, particularly in the range \(w_{con}=0\) to \(0.9\).

It would have been interesting to query the perceived realism continuously, as done in [28], to get a better understanding of where exactly in the maneuver the realism breaks down instead of getting a single value over the entire drive. However, in the current study, this could not be done as the participants were already reporting motion sickness levels, and we did not want to increase the mental load of the participants, at the risk of disrupting immersion.

4.3 Motion Sickness Manipulation↩︎

A key finding is that the algorithm is capable of significantly reducing motion sickness while preserving the perception of motion. Subjective results revealed that a conflict weight of \(w_{con} = 0.9\) resulted in an average end MISC level of 2 across participants. In contrast, both the adaptive washout and the MPC configuration with only specific force tracking yielded an average MISC level of 3. This reduction was achieved without any significant reduction of perceived realism. Another important finding is that the MPC-based algorithm, with no restriction on motion sickness (\(MPCw_{con0}\)), outperforms the adaptive washout algorithm in terms of specific force reproduction.

Although this work specifically targets the reduction of motion sickness, the algorithm can also be configured to recreate motion sickness in real vehicles. For example, the model predictions in 5 indicate that, in this scenario, replicating real-vehicle motion sickness within a driving simulator would already be nearly achieved with either \(AW\) or \(MPCw_{con0}\). To minimise motion sickness, the sensory conflict is given a reference of zero. However, by feeding the actual sensory conflict, calculated from real vehicle data, into the MPC as reference, motion sickness can be intentionally reproduced. This could prove valuable for assessing how sick an occupant may feel when exposed to specific automated driving styles, and for conducting human-acceptance or behavioural studies within a simulated environment, especially in cases where real vehicle and simulator responses are directly compared.

The experimental validation confirmed that the no-motion case resulted in the least amount of sickness, as predicted by the SVC model. However, no motion was rated as the least realistic. This aligns with [9], where the perceived fidelity improved with the addition of motion to the simulator platform, accompanied by an increase in motion sickness. However, this does not align with the meta-analysis of [8], where enhanced visual fidelity reduced motion sickness for moving base, but not for fixed base, indicating that increasing the visual fidelity may be responsible for inducing more sickness. In our experiment, the frontal view was the main focus, thus, the peripheral vision was blocked. This blocking of peripheral vision may have limited the visual influence on participants’ sickness levels. In the SVC model used in this study, the visual input is restricted solely to the visual rotational velocity, which is assumed to be perfectly congruent with the actual rotational velocity of the visual stimuli presented on the screen. This simplification may be a limitation, underscoring the need for more comprehensive modelling approaches that more accurately reflect the complexity of visual systems.

4.4 Yaw↩︎

In this study, yaw motion was controlled via a traditional washout algorithm, which does not consider motion sickness in its design. This decision was made to simplify the MPC computation and limit computational load. A recent study found missing and false yaw cues to affect perceived fidelity and sickness with a similar magnitude as missing and false longitudinal and lateral cues, although the significance of these effects was not reported [29]. Future work may explore integrating yaw dynamics into the MPC framework to potentially improve sickness mitigation and deliver more precise rotational cueing.

5 Conclusion↩︎

The developed motion cueing algorithm demonstrates a strong capability to reduce motion sickness while maintaining a realistic perception of vehicle motion. Experimental results show that increasing the weight for sensory conflict within the model predictive control framework leads to a noticeable reduction in motion sickness experienced by participants. This confirms the effectiveness of penalising sensory conflict as an explicit control objective in an MPC-based motion cueing algorithm.

Subjective evaluations further reinforce the findings, with participants rating motion-based configurations significantly higher in realism compared to the no motion baseline. Even low-amplitude motion significantly enhances perceived motion fidelity compared to no-motion conditions.

In summary, this work marks a significant advancement in motion cueing design by successfully integrating a perceptually-informed motion sickness model directly into the control framework. The algorithm not only proves effective in reducing motion sickness through both simulation and human-subject validation, but also demonstrates that motion fidelity can be preserved while optimising for comfort. This achievement paves the way for more immersive and tolerable long-duration driving simulation experiences, setting a new benchmark for future motion cueing algorithms.

6 MPC Algorithm Weight Settings↩︎

For the simulations presented in this chapter, model predictive control (MPC) is implemented using
ForcesPro [30], using the Primal dual interior point (PDIP) algorithm. The maximum iterations are chosen to be 200, to ensure convergence and avoid sub-optimal solutions. The optimisation has been performed on Intel(R) Xeon(R) W-2223 CPU @3.60GHz with 32GB RAM.
In this work, we consider two major contribution terms that define the primary objective of the algorithm, along with several minor terms that help guide the MPC toward the desired performance. The two primary objective terms should have the highest contribution in the cost function.

Since the cost terms have different units, their relative contributions cannot be directly determined by the weights alone. To address this, we normalise the cost terms. This normalisation facilitates the assignment of weights, enabling each term to be tuned based on its priority. Specifically, a preliminary simulation is used to estimate the maximum expected value of each cost term. Each term in the cost function is then divided by its corresponding maximum, ensuring that the normalised errors lie within comparable ranges.

This approach allowed us to explore the full trade-off spectrum, or Pareto front, between competing objectives such as motion perception (via specific force tracking) and motion sickness mitigation (via RMS sensory conflict). While the normalisation does not provide a universal or mathematically exact scaling, the estimated maxima—derived from representative scenarios—offer a robust basis for comparative analysis.

Should future scenarios diverge significantly from those used during this normalisation phase, the scaling bounds can be re-evaluated and updated accordingly.

The final selected weights are tabulated in 4.

The \(w_{s}\) and \(w_{\theta}\) correspond to the parameters k1 and k3, respectively, in Section 10, that affect the shape of the non-linear weight function. \(w_{\omega},\;w_{j}\), \(w_{ang,j}\) and \(w_{\delta}\) are the penalisation weights for angular velocity, translational jerk, angular jerk and the slack variable. Note that all weights are represented as \(2\;\times \;2\) diagonal matrices to apply separate penalties in the longitudinal and lateral directions, except for the slack variable weight \(w_{\delta}\), which remains scalar since it does not relate to spatial dimensions.

For the initial analysis, a penalisation weight of unity is set for specific force tracking. Additionally, the penalisation weights on the angular orientation and the platform displacement is a dynamical non-linear weight which changes based on the platform state (see 10 under workspace management) the penalisation weight chosen for this quantity just scales the overall non-linear shape of the weight based on the platform state.

6.1 Penalisation Weight for Angular Orientation↩︎

The weight on angular orientation was varied between 1e-4 and 1e-1. The weight of 1e-1 provides deteriorated specific force tracking performance, while the weights 1e-2, 1e-3 and 1e-4 provide almost identical responses. Hence, the weight of 1e-4 is chosen as it provides desirable performance, while keeping a lower contribution in the cost term.

6.2 Penalisation Weight for Angular Velocity↩︎

The weight on angular velocity was also varied between 1e-4 and 1e-1. While all the weights provided a very similar response in the specific force generation, the weights had very different responses for the angular velocities. The weight of 1e-1 provides almost no excessive motion, whereas all other weights show excess oscillations in angular velocity. Thus the authors choose the weight of 1e-1 for the penalisation on angular velocity.

6.3 Penalisation Weight for Displacement↩︎

The weight on translational displacement was varied between 1e-4 and 1e-1 as well. While a weight of 1e-4 tracks the specific force desirably for the majority of the simulation, at various instances, it performs a jerky motion when it reaches the limits of the workspace. With the weight 1e-3 we obtain desirable specific force tracking performance, however oscillations are observed in the angular velocity with this setting. The weight 1e-1 and 1e-2 provide a very similar response with no excess oscillation in the angular velocity. Thus due to its lower contribution towards the overall objective function, 1e-2 is chosen as the preferred weight for the simulations.

6.4 Penalisation Weight for Translational Jerk↩︎

The weight on translational jerk was varied between 1e-4 and 1e-1 as well. All the explored weights provided near identical responses. Thus the weight of 1e-4 is used to have the lowest possible contribution to the objective function, while providing a desirable performance.

6.5 Penalisation Weight for Angular Jerk↩︎

The weight on angular jerk was varied between 1e-5 and 1e-2 as well. The weight of 1e-2 renders a profile that does not follow the reference specific force properly. While the weight of 1e-3 traces the specific force, it exhibits oscillations in the tilt rate. The weights 1e-5 and 1e-4 provide a desirable specific force tracking, with the weight of 1e-4 also attaining slightly lower tilt rate values. Thus the weight of 1e-4 is selected for the simulations in this work.

7 Questionnaire↩︎

Participants rated realism on a 5-point Likert scale with the following questions:

• How closely did the ride’s motion correspond to the video?

  [0 = Not at all, 5 = Completely coherent]

• How close did the cornering feel compared to a real car?

  [0 = Not at all, 5 = Exactly like a real car]

• How realistic did the acceleration and deceleration feel compared to a real car?

  [0 = Not at all, 5 = Exactly like a real car]

• Were there any unnatural motions that did not match real driving?

  [0 = Not at all, 5 = A lot of them]

8 Motion Sickness Model↩︎

The MPC-based algorithm needs a motion sickness prediction model to calculate a motion sickness metric (\(J_{MS}\)) for the cost function (\(J_{c}\)). For this, we use the Subjective Vertical Conflict with Visual Rotational velocity (SVC-VR) as described by [18]. The model was first introduced by [4] and later validated and found to be favourable for motion sickness predictions in vehicles for a population by [7]. This model ([4]) predicts the motion sickness incidence (MSI), which is a group-averaged metric representing the percentage of people who will develop motion sickness. This is not suitable as we need a scale based on the severity of motion sickness in an individual. This way we will have better control on how much the severity of sickness varies with different algorithms. For this, we used the model by [18], which adapted the model by [4] to predict motion sickness for an individual in MIsery SCale (MISC). This model accepts specific forces, angular rotations and visual flow (visual angular rotation) as inputs and gives motion sickness in MISC as output.

It is important to mention that the platform’s motion is different from the visual cues. In this study we simulate conditions with out of the window view. Thus, the vehicle angular velocities are given for the vision angular velocity input to the SVC model, while the vestibular system is provided with the platform motion.

The motion sickness score output from the accumulation model has a large time constant due to the slow dynamics of the accumulation model. The high time constant of the accumulation model implies that a long prediction horizon would be required for it to function effectively. However, as demonstrated in [15], optimising for motion sickness over a short horizon can still lead to effective reduction of accumulated sickness over longer durations—provided that an appropriate short-term proxy, such as instantaneous sensory conflict, is used. As we optimise the motion at each time instant, we want to calculate the instantaneous response to the motion stimuli which drives motion sickness. This is the sensory conflict that can be obtained by the first half part of the model, often termed the ‘sensory conflict generation’ model shown in 1.

We converted a Simulink implementation of the SVC-VR model into ordinary differential equations, which our MPC solver can use. The ordinary differential equations for the SVC-VR model are given below:

\[\begin{align} \dot{v}_{s} &= &\frac{f_{spec}-v_s}{\tau} - \omega_{s} \times v_{s}\\ \dot{\omega}_{s} &= & \dot{\omega} - \frac{\omega_{s}}{\tau_{d}}\\ \dot{\hat{v}}_{s} &= &\frac{\hat{f}_s-\hat{v}_s}{\tau} - \hat{\omega}_{s} \times \hat{v}_{s}\\ \dot{\hat{\omega}}_{s} &= &\frac{(K_{\omega,c} + K_{\omega}) \dot{\omega} + K_{\omega,vis} \dot{\omega}_{vis}- \frac{K_{\omega,c}}{\tau_{d}}(\omega_{s} - \hat{\omega}_{s})}{1 + K_{\omega,vis} + K_{\omega,c}} \notag \\ && -\frac{\hat{\omega}_{s}}{\tau_d}\\ \dot{\tilde{v}} &= &K_{vc} (v_{s} - \hat{v}_{s}) +K_{g,vis} (v_{vis} - \tilde{v})\\ \dot{\hat{f}}_{s} &= &K_{vc} (v_{s} - \hat{v}_{s}) +K_{g,vis} (v_{vis} - \tilde{v}) \notag \\ & & + K_{ac} (f_s -v_s - \hat{f}_{s} +\hat{v}_{s}) + K_{a}\dot{a}\\ \nonumber \end{align}\]

where \(\hat{f_s}\) represents the estimated specific force vector, \(v_{s}\) and \(\hat{v}_s\) are the sensed subjective vertical and estimated subjective vertical, respectively. \(\omega_{s}\) and \(\hat{\omega}_{s}\) are the sensed angular velocity and estimated angular velocity respectively. \(\tau = 5\;s\), \(\tau_{d} = 7\;s\), \(K_{\omega,c} = 10\), \(K_{a,c} = 1\), \(K_{v,c} = 5\), \(K_a = 0\), \(K_\omega = 0\), \(K_{\omega,vis} = 10\), \(K_{g,vis} = 0\) are the parameters used for the SVC model taken from [17]. Here, \(\tau\) and \(\tau_{d}\) are the time constants corresponding to the otoliths and the semicircular canals respectively. \(K_{\omega,c}\), \(K_{a,c}\), and \(K_{v,c}\) are the vestibular feedback gains, \(K_a\) and \(K_{\omega}\) are the anticipatory gains, \(K_{\omega,\text{vis}}\) and \(K_{g,\text{vis}}\) are the visual feedback gains.

The sensory conflict is derived by calculating the Euclidean norm or 2-norm of the difference between the sensed subjective vertical and estimated subjective vertical:

\[\begin{align} c_v &= &\left\lVert v_{s}-\hat{v}_s \right\rVert_2 = \left( \sum_{i=1}^3 (v_{s_i}-\hat{v}_{s_i})^2 \right)^{1/2} \label{sen95con95ch5} \\ J_{MS} &= & w_{con} \quad {c_v}^2 \end{align}\tag{3}\] where, \(w_{con}\) is the weight on the conflict, \(c_v\), to create the sensory conflict term quantifying motion sickness (\(J_{MS}\)) used in the objective function.

The sensory conflict is one-dimensional. The accumulation of this sensory conflict drives the overall motion sickness scores. Minimising this sensory conflict over the MPC time horizon will result in a reduction of motion sickness.

To demonstrate the effectiveness of the motion cueing algorithm, we also predict MISC over the entire experiment duration. Here we use the entire model framework by [18], including the ‘conflict accumulation’ model as shown in the bottom right in 1. To predict the variance in MISC across a population, we simulate the MISC for 1000 individuals with varying motion susceptibilities. These parameters are sampled from the parameter distribution described in [17]. In that work, a 3-component probability distribution of parameter sets was generated for the model by [18], which can be sampled according to the desired motion sickness susceptibility. These parameter sets have been shown to generalise well for MISC predictions in new driving scenarios [17]. This makes it ideal to use for testing and selecting algorithms.

9 Hexapod/Driving Simulator Dynamics↩︎

The motion of the hexapod platform is defined in a state-space form to facilitate implementation in the MPC. The base states include hexapod position (\(s_{hex}\)) and angular orientation (\(\theta_{hex}\)). These base states are added to the state-space model with the relation

\[\dot{x}_{hex} = A_{hex} x_{hex} + B_{hex} u_{hex}\]

where the state vector, \(x_{hex}\), comprises of the position, \(s_{hex}\), translational velocity, \(v_{hex}\), angular orientation, \(\theta_{hex}\), and angular velocity, \(\omega_{hex}\), of the hexapod and the input vector, \(u_{hex}\), comprises of translational acceleration, \(a_{hex}\) and angular acceleration of each euler angle, \(\alpha_{hex}\). The matrices \(A_{hex}\) and \(B_{hex}\) represent the double integrator system of the inputs, adapted from [31].

As the algorithm is designed for both longitudinal and lateral degrees of freedom, each state comprises of components in x and y directions (roll and pitch for orientation). In this study, positive values correspond to forward, left, and upward orientations along the x, y, and z axes, with counterclockwise rotations indicated as positive.

To achieve realistic motion perception (one of the primary objectives of the algorithm), vehicular accelerations are tracked using specific forces generated by the driving simulator. The specific force consists of two components that arise through translational accelerations and gravity. This specific force encapsulates the combined effects of accelerations and gravity as perceived by the human via the otoliths (part of the vestibular system). Therefore, the specific force is calculated at the estimated head coordinate system, thereby incorporating the effects of platform rotation.

The translational component is the acceleration of the platform. The gravitational force vector at the estimated head location, \(G_{loc}\), is defined by the relation:

\[\label{gravitation95vector95ch5} G_{loc} = R^{T} [0\;0\;g]'\tag{4}\]

Here \(R\) is the transformation matrix that resolves gravitational force to the vectors corresponding to longitudinal, lateral and vertical body reference frame directions and \(g\) is the acceleration due to gravity, acting in the inertial vertical direction.

The total specific force is defined as: \[f_{spec} = a_{hex} - G_{loc}\]

where \(a\) is the translational acceleration of the platform. The tilt component, \(G_{loc}\), provides an additional pseudo acceleration to the occupant of the simulator. The specific force is the quantity to be tracked to achieve realistic motion perception.

This shapes our cost function term for the MCA defining the motion perception term (\(J_{MP}\)), which is given by:

\[J_{MP} = (f_{spec}-f_{ref}) \;w_{spec} \; (f_{spec}-f_{ref})^T\]

To reduce motion sickness stimuli (the secondary objective of the algorithm), sensory conflict (3 ) needs to be minimised. Thus, the MPC includes the hexapod dynamics and the 6-DOF SVC model to predict the development of motion sickness over the prediction horizon.

Thus the complete analytical description of the dynamics of motion sickness development, through the platform motion, includes the following states: \[x = [{\theta}_{hex}, {\omega}_{hex}, {s}_{hex}, v_{hex}, {f}_{tilt}, f_{all}, \hat{f}_{all}, {\hat{v}}_{s}, {\omega}_{s}, {\hat{\omega}}_s, {v}_{s}]\]

Here, \(f_{all}\) denotes the vector of specific forces in all three directions (\(x\), \(y\), \(z\)); \(\hat{f}_{all}\) represents the estimated specific forces in those directions; and \(f_{tilt}\) corresponds to the tilt-generated specific force, which is equal to the local gravitational vector \(G_{loc}\).

10 Workspace Management↩︎

The MPC considers these constraints over the prediction horizon to optimally use the workspace and generate realistic motion. Two additional strategies are employed here for effective workspace management: washout and dynamic constraints.

Washout: The simulator platform has the maximum potential of recreating the specific forces at its neutral position. To ensure the platform remains near its neutral position, we penalise its states in the cost function. In this work, we use non-linear weights (based on the platform orientation and position) for the washout instead of constant weights. This allows a single non-linear setting for all scenarios rather than tuning the washout weights for each scenario.
The non-linear weights are defined as \[\begin{align} &w_{s} &= \frac{k_1}{k_2(|s_{hex}|-s_{lim})^2+\Delta} \label{adaptive95wt} \\ &w_{\theta} &= \frac{k_3}{k_2(|\theta_{hex}|-\theta_{lim})^2+\Delta} \end{align}\tag{5}\]

where \(k_{1}\), \(k_{2}\) and \(k_{3}\) define the shape of the weight function, \(s_{lim}\) and \(\theta_{lim}\) are the defined limits for the platform for displacement and tilt angle. \(\Delta\) (here \(0.01\)) is a small value added to the denominator to avoid singularity. The selected values are \(k_1 = 1\), \(k_2 = 50\) and \(k_3 = 0.1\), these values were manually tuned to ensure that the penalisation is low near the neutral position, while high, close to the platform limits.

Dynamic constraints: In this study, we incorporate dynamic bounds on the platform position and orientation via the constraints proposed in [32], as ‘braking constraints’.

The formulation of the constraints is \[\begin{align} &s_{hex, min} \leq s_{dyn} \leq s_{hex, max} \\ &\theta_{hex, min} \leq \theta_{dyn} \leq \theta_{hex, max} \end{align}\] \[\begin{align} &s_{dyn} = s_{hex} + c_v v_{hex} T_{dyn, s} +0.5 c_u a_{hex, tran} T_{dyn, s}^2\\ &\theta_{dyn} = \theta_{hex} + c_w \omega_{hex} T_{dyn, \theta} +0.5 c_u a_{hex, rot} T_{dyn, \theta}^2 \end{align}\]

where, \(c_v = 1, c_w = 1, c_u = 0.45, T_{dyn, \theta} = 0.5\), \(T_{dyn, p} = 2.5\) and \(s_p, \theta_p\) limits are \(0.3\;m\) and \(20\;deg\) respectively. The selected values were adopted from [33]. When the platform approaches its limits, the platform’s acceleration and velocity reduce, to stay within the workspace envelope.

11 Yaw Channel↩︎

The fifth DoF (yaw) is controlled separately using a parallel washout channel, ensuring reduced computational complexity. For the simulator used in this work (DAVSi), the control commands we require to provide are yaw position velocity and acceleration. This is done by passing the acceleration through a high-pass filter to obtain the desired platform yaw acceleration. To ensure that the yaw angle returns back to it’s neutral position at the end of the simulation, we use a second order high pass filter instead of a first order filter. The second-order high-pass filter used for this purpose is given as

\[HP(s) = \frac{s^2}{s^2+2\nu_{yaw}s +\nu_{yaw}^2}\]

Here, \(\nu_{\text{yaw}}\) denotes the cut-off frequency of the high-pass filter. A value of 0.0159 Hz is used, consistent with the configuration used for our previous studies. Additionally, for simplicity, the damping ratio is kept to be 1 (critically damped).

As yaw motion also affects motion sickness, the yaw prediction for the future should also be communicated to the MPC. As the yaw washout is highly computationally efficient, it can calculate the solution to the reference yaw for the prediction horizon almost instantly. In our implementation, we include yaw information as online data for communication with the MPC.

12 Constraints↩︎

Constraints in the MPC are added to ensure that the hexapod’s motions remain physically feasible and safe. These include workspace limits on position and orientation to prevent the platform from exceeding mechanical boundaries, as well as limits on velocities and accelerations, to avoid abrupt or excessive movements that could cause discomfort or destabilise the system. For defining such constraints, the workspace limits of Delft Advanced Vehicle Simulator (DAVSi) are used (5).

Additionally, to minimise the perception of platform motion and avoid introducing false cues, it is essential to limit platform rotation rates below the human perception threshold. This threshold is typically considered to lie between 2–4°/s [22], [25], [34]. In this work, a value of 3°/s was selected based on subjective feedback from participants during a pilot study.

Rather than enforcing this as a hard constraint, we model it as a soft constraint to allow occasional violations when necessary to improve specific force tracking performance. This approach enables flexibility in generating higher specific forces without introducing excessive perceived motion. The tilt-rate constraint is therefore defined as:

\[\begin{align} -\omega_{thd} \leq \omega_{hex} + \delta\;\omega_{hex} - \delta \leq \omega_{thd} \end{align}\]

Here, \(\omega_{thd}\) is the selected perception threshold (3°/s for both pitch and roll rate), and \(\delta\) is a positive slack variable included in the cost function. While violations of the threshold are permitted, they are penalized to encourage minimal deviation, balancing perceptual fidelity with motion cueing performance.

13 Comparison of MCAs↩︎

References↩︎