In moving-base driving simulators, the sensation of the inertial car motion provided by the motion system is controlled by the motion cueing algorithm (MCA). Due to the difficulty of reproducing the inertial motion in urban simulations, accurate prediction tools for subjective evaluation of the simulator's inertial motion are required. In this article, an open-loop driving experiment in an urban scenario is discussed, in which 60 participants evaluated the motion cueing through an overall rating and a continuous rating method. Three MCAs were tested that represent different levels of motion cueing quality. It is investigated under which conditions the continuous rating method provides reliable data in urban scenarios through the estimation of Cronbach's alpha and McDonald's omega. Results show that the better the motion cueing is rated, the lower the reliability of that rating data is, and the less the continuous rating and overall rating correlate. This suggests that subjective ratings for motion quality are dominated by (moments of) incongruent motion, while congruent motion is less important. Furthermore, through a forward regression approach, it is shown that participants' rating behavior can be described by a first-order low-pass filtered response to the lateral specific force mismatch (66.0%), as well as a similar response to the longitudinal specific force mismatch (34.0%). By this better understanding of the acquired ratings in urban driving simulations, including their reliability and predictability, incongruences can be more accurately targeted and reduced.

@en

This article discusses a long short-term memory (LSTM) recurrent neural network that uses raw time-domain data obtained in compensatory tracking tasks as input features for classifying (the adaptation of) human manual control with single- and double-integrator controlled element dynamics. Data from two different experiments were used to train and validate the LSTM classifier, including investigating effects of several key data preprocessing settings. The model correctly classifies human control behavior (cross-experiment validation accuracy 96%) using short 1.6-s data windows. To achieve this accuracy, it is found crucial to scale/standardize the input feature data and use a combination of input signals that includes the tracking error and human control output. A possible online application of the classifier was tested on data from a third experiment with time-varying and slightly different controlled element dynamics. The results show that the LSTM classification is still successful, which makes it a promising online technique to rapidly detect adaptations in human control behavior.

@en

Users of automated vehicles will engage in other activities and take their eyes off the road, making them prone to motion sickness. To resolve this, the current paper validates models predicting sickness in response to motion and visual conditions. We validate published models of vestibular and visual sensory integration that have been used for predicting motion sickness through sensory conflict. We use naturalistic driving data and laboratory motion (and vection) paradigms, such as sinusoidal translation and rotation at different frequencies, Earth-Vertical Axis Rotation, Off-Vertical Axis Rotation, Centrifugation, Somatogravic Illusion, and Pseudo-Coriolis, to evaluate different models for both motion perception and motion sickness. We investigate the effects of visual motion perception in terms of rotational velocity (visual flow) and verticality. According to our findings, the SVC_I model, a 6DOF model based on the Subjective Vertical Conflict (SVC) theory, with visual rotational velocity input is effective at estimating motion sickness. However, it does not correctly replicate motion perception in paradigms such as roll-tilt perception during centrifuge, pitch perception during somatogravic illusion, and pitch perception during pseudo-Coriolis motions. On the other hand, the Multi-Sensory Observer Model (MSOM) accurately models motion perception in all considered paradigms, but does not effectively capture the frequency sensitivity of motion sickness, and the effects of vision on sickness. For both models (SVC_I and MSOM), the visual perception of rotational velocity strongly affects sickness and perception. Visual verticality perception does not (yet) contribute to sickness prediction, and contributes to perception prediction only for the somatogravic illusion. In conclusion, the SVC_I model with visual rotation velocity feedback is the current preferred option to design vehicle control algorithms for motion sickness reduction, while the MSOM best predicts perception. A unified model that jointly captures perception and motion sickness remains to be developed.

@en

This paper analyzes the effects of the helicopter dynamics on pilots' learning process and transfer of learned skills during autorotation training. A quasi-transfer-of-training experiment was performed with 10 experienced helicopter pilots in the SIMONA moving-base flight simulator at Delft University of Technology. Pilots had to control an in-house flight dynamics model setup to simulate two types of helicopter dynamics: (1) a "hard"dynamics characterized by a low autorotative flare index requiring high pilot control compensation and (2) a "easy"dynamics characterized by a high autorotative flare index with low pilot control compensation required. Two groups of pilots tested these types of dynamics in a different training sequence: hard-easy-hard (HEH group) and easy-hard-easy (EHE group). The main conclusion of this study proved that simulator training for autorotation can best start with pilots training in the most resource demanding condition. A more challenging helicopter's dynamics will require higher pilot agility and more rapid responses to his/her perceptual changes. This will result in pilots developing more robust and adaptable flying skills. Indeed, a clear positive transfer of training effect was observed in the experiment presented in this paper in terms of acquired pilot skills in the HEH group, but not the EHE group. Positive transfer was especially observed in terms of reduced rate of descent at touchdown. The two groups differed in the control strategy applied, with the HEH group having developed a control technique mimicking more closely the one adopted in a real helicopter.

@en

In driving simulation, the choice of a simulator, motion cueing algorithm, and associated set of tuning parameters for an experiment is typically made with an exclusive focus on the quality of the motion. In practice, many other metrics could affect this choice as well, such as tuning complexity, algorithm stability, or the financial costs of the simulation. Arguably, the complete motion cueing algorithm quality is thus more than the quality of the motion alone. This paper presents results of a survey which attempted to identify the most important metrics from the perspective of the main experiment stakeholders. Four stakeholder groups in typical driving simulator experi- ments are defined: The experimenters, motion cueing engineers, operators, and participants. All groups received the same survey, asking them to indicate how important various metrics are for them. Results show that, next to the quality of the motion, experimenters and participants are generally interested in reducing simulator sickness. The motion cueing engineers rank tuning effort and tuning complexity as most important metrics. Operators prefer an easy to use and overall stable motion cueing. A typical BMW experiment is discussed as example, which shows that the choice for a simulator and motion cueing algorithm can indeed differ when including these metrics in a trade-off, compared to when only motion quality is considered. The presented methods allow for a better, multi- faceted selection of the simulator, motion cueing algorithm, and associated tuning parameters, improving future driving simulation experiments. @en

Users of automated vehicles will move away from being drivers to passengers, preferably engaged in other activities such as reading or using laptops and smartphones, which will strongly increase susceptibility to motion sickness. Similarly, in driving simulators, the presented visual motion with scaled or even without any physical motion causes an illusion of passive motion, creating a conflict between perceived and expected motion, and eliciting motion sickness. Given the very large differences in sickness susceptibility between individuals, we need to consider sickness at an individual level. This paper combines a group-averaged sensory conflict model (as in Wada, et al., 2020) with an individualized accumulation model (as in Irmak, et al., 2022; Irmak, Pool, and Happee, 2020; Oman, 1990) to capture individual differences in motion sickness susceptibility across various vision conditions. This consideration of the effect of vision is crucial in driving simulators where there is a strong contribution of visual cues. The model framework can be used to develop personalized models for users of automated vehicles and improve the design of new motion cueing algorithms for simulators. The feasibility and accuracy of this model framework are verified using two existing datasets with sickening conditions in 1) an experimental vehicle with and without outside vision (Irmak, Pool, and Happee, 2020), and 2) comparing vehicle experiments with corresponding driving simulator experiments (Talsma, et al., 2023). Both datasets involve passive motion, representative of being driven by an automated vehicle. The model is able to fit an individual’s motion sickness responses using only 2 parameters (gain K1 and time constant T1), as opposed to the 5 parameters in the original model. This ensures unique parameters for each individual. Better fits, on average by a factor of 1.7 (for Accum 2 model), of an individual’s motion sickness levels, are achieved as compared to using only the group-averaged model (Accum 0 model). Furthermore, this model framework demonstrates robustness by accurately modeling various datasets with distinct motion and vision conditions. Thus, we find that models predicting group-averaged sickness incidence cannot be used to predict sickness at an individual level. On the other hand, the proposed combined model approach predicts individual motion sickness levels and thus can be used to control sickness @en

The human motion perception system has long been linked to motion sickness through state estimation conflict terms. However, to date, the extent to which available perception models are able to predict motion sickness, or which of the employed perceptual mechanisms are of most relevance to sickness prediction, has not been studied. In this study, the subjective vertical model, the multi-sensory observer model and the probabilistic particle filter model were all validated for their ability to predict motion perception and sickness, across a large set of motion paradigms of varying complexity from literature. It was found that even though the models provided a good match for the perception paradigms studied, they could not be made to capture the full range of motion sickness observations. The resolution of the gravito-inertial ambiguity has been identified to require further attention, as key model parameters selected to match perception data did not optimally match motion sickness data. Two additional mechanisms that may enable better future predictive models of sickness have, however, been identified. Firstly, active estimation of the magnitude of gravity appears to be instrumental for predicting motion sickness induced by vertical accelerations. Secondly, the model analysis showed that the influence of the semicircular canals on the somatogravic effect may explain the differences in the dynamics observed for motion sickness induced by vertical and horizontal plane accelerations.

@en

Current capabilities for predicting skill retention, i.e., the extent to which human operators retain learned skills over time, at an individual level are limited due to a requirement for large data sets and methods that can extract relevant patterns in highly dimensional data. This paper investigates the application of Extreme Gradient Boosting (XGBoost) decision tree models for predicting a high-resolution individual skill retention curve. For this, a large skill-based tracking experiment dataset is used to extract different feature classes and train an XGBoost predictive model. To identify the robust predictors, the effects of the different features on the model's output are analyzed using SHapley Additive exPlanations (SHAP). Furthermore, the proposed XGBoost model is trained using both the experiment dataset and a matched synthetic dataset, with both approaches evaluated on the experiment data. Overall, the available experiment dataset was found to include too few retention measurements, and too significant between-group differences, to extract a reliable prediction model. On the synthetic dataset, the XGBoost model was found to accurately capture individuals' skill retention curves, where the features that contributed most (21%) to the prediction model's accuracy were found to be the considered learning curve parameters. Overall, this paper shows that experiment data of skill-based tracking tasks can be used to predict skill decay curves using XGBoost, but that more research and data are needed to achieve sufficient accuracy and reliability at an individual level for practical applications.@en

An improved understanding of pilot’s control behavior adaptations in response to sudden changes in the vehicle dynamics is essential for realizing adaptive support systems that remain effective when task characteristics suddenly change. In this paper, we replicate, extend, and validate the ‘adaptive pilot model’ proposed by Hess to verify its effectiveness for predicting human adaptive behavior in pursuit tracking tasks. The model relies on a Triggering function, that compares the current tracking performance to a stored nominal (pre-transition) state, and an Adaptation mechanism which determines new adapted human operator gain settings proportional to the magnitude of the off-nominal error occurrences. For model validation data from a previous experiment were used, where ten participants performed a pursuit tracking task with transitions in controlled element dynamics from a single to a double integrator, and vice versa. Overall, with an added human operator delay and participant-specific inner- and outer-loop gain adjustments, the model was found to accurately describe the measured steady-state tracking behavior for the participants in our data set. The results for the time-varying single integrator to double integrator transitions showed that the model can capture the transient control behavior of participants. However, the adaptive logic could only be tuned to activate for participants that had a pre-transition crossover frequency above 0.9 rad/s. Furthermore, the model was not able to capture the change in control behavior for transitions from a double to a single integrator. Here, as no distinct degradation in tracking performance occurs for such a transition to a more easily controlled system, the model's proposed Triggering logic will not activate. Further investigation and more experiment data are required for improving the applicability of the model's adaptive logic and to enable more accurate prediction of adaptive human control behavior.@en

In the design of human-like steering support systems, driver models are essential for matching the supporting automation's behavior to that of the human driver. However, current driver models are very limited in capturing the driver's adaptation to key task variables such as road width and visibility (i.e., 'preview' of the road ahead). This paper uses a recently proposed, novel control-theoretical model for centerline tracking to investigate driver steering in lane-keeping tasks with restricted and unrestricted preview, in an attempt to substantially extend this model's validity. Using data from a tailored driving simulator experiment, three driver control loops (feedforward, heading and position feedback) are separately quantified using system identification techniques. The results show that when preview is restricted, drivers use all of the remaining preview to anticipate the curves of the road ahead, and are no longer able to 'smooth' tight curves in the road trajectory (i.e., corner cutting). When sufficient preview and lane width are available, the time to line crossing increases, and steering behavior is less aggressive and more intermittent, or more 'satisficing'. The novel driver steering model captures these adaptations very well (over 95% of the steering actions) and can thereby be instrumental in realizing human-like steering automation and support systems.

@en

Due to the non-deterministic nature of longitudinal human driver behaviour, motion cueing algorithms currently cannot fully utilize the workspace of driving simulators. This paper explores the possibility of using various predictor variables to predict longitudinal driving behaviour. Through the development of a logistic regression model, it is shown that a combination of the current vehicle velocity, the speed limit eight seconds ahead and the accelerator pedal deflection yields the most accurate estimate of the probabilities that drivers will accelerate or decelerate. Based on these probabilities, a driving simulator was linearly pre-positioned in combination with a classical washout algorithm. The perceived motion incongruence was subjectively evaluated by the drivers (N = 34), testing: (i) no pre-positioning, (ii) pre-positioning, and (iii) pre-positioning with an increased longitudinal classical washout gain enabled by the pre-positioning. Results show that the pre-positioning improves the margins with respect to the longitudinal workspace limits (better workspace management), without affecting the motion incongruence ratings. When using the increased margins to increase the longitudinal gain, however, no significant reduction in motion incongruence ratings was observed. This is likely due to the small motion space of the hexapod motion system used in the current study. However, this paper shows that longitudinal driving behaviour can be accurately predicted and can enable improved workspace utilization for driving simulators. @en

Aerodynamic model identification remains essential for simulator operations and control system design and operations. In this paper, state-of-the-art methodologies for aerodynamic model identification and validation are presented, together with a number of novel applications of the identified models that were recently investigated and developed at the Faculty of Aerospace Engineering of Delft University of Technology. In particular, this paper focuses on methodologies for identifying models of aerodynamic stall from flight data, as well as multivariate spline-based aerodynamic model identification methods, together with their applications in flight simulation and advanced flight control.@en

High levels of vehicle automation are expected to increase the risk of motion sickness, which is a major detriment to driving comfort. The exact relation between motion sickness and discomfort is a matter of debate, with recent studies suggesting a relief of discomfort at the onset of nausea. In this study, we investigate whether discomfort increases monotonously with motion sickness and how the relation can best be characterized in a semantic experiment (Experiment 1) and a motion sickness experiment (Experiment 2). In Experiment 1, 15 participants performed pairwise comparisons on the subjective discomfort associated with each item on the popular MIsery SCale (MISC) of motion sickness. In Experiment 2, 17 participants rated motion sickness using the MISC during exposures to four sustained motion stimuli, and provided (1) numerical magnitude estimates of the discomfort experienced for each level of the MISC, and (2) verbal magnitude estimates with seven qualifiers, ranging between feeling ‘excellent’ and ‘terrible’. The data of Experiment 1 show that the items of the MISC are ranked in order of appearance, with the exception of 5 (‘severe dizziness, warmth, headache, stomach awareness, and sweating’) and 6 (‘slight nausea’), which are ranked in opposite order. However, in Experiment 2, we find that discomfort associated with each level of the MISC, as it was used to express motion sickness during exposure to a sickening stimulus, increases monotonously; following a power law with an exponent of 1.206. While the results of Experiment 1 replicate the non-linearity found in recent studies, the results of Experiment 2 suggest that the non-linearity is due to the semantic nature of Experiment 1, and that there is a positive monotonous relation between MISC and discomfort in practice. These results support the suitability of MISC to assess motion sickness.

@en

Improved understanding of human adaptation can be used to design better (semi-)automated systems that can support the human controller when task characteristics suddenly change. This paper evaluates the effectiveness of a model-based adaptive control technique, Model Reference Adaptive Control (MRAC), for describing the adaptive control policy used by human operators while controlling a time-varying system in a pursuit-tracking task. Ten participants took part in an experiment in which they controlled a time-varying system whose dynamics changed twice between approximate single and double integrator dynamics, and vice versa. Our proposed MRAC controller is composed of a feedforward and a feedback controller and an internal reference model that is used to drive an adaptive control policy. MRAC's adaptive control gains, the internal model parameters, and the learning rates were estimated from the experiment data using non-linear optimization aimed at maximizing the quality-of-fit of participants' control outputs. Participants' control behavior rapidly changed when the dynamics of the controlled system changed, in particular for transitions from single to double integrator dynamics. The MRAC model was indeed able to accurately capture the transient dynamics exhibited by the participants when the system changed from an approximate single to a double integrator, however, for the opposite transition the MRAC gains were always adapted too slowly. Therefore, in its current form, our MRAC model can be used to approximate human adaptation in pursuit tracking tasks when a change in the dynamics of the controlled system requires significant (rate) feedback controller adaptation to maintain satisfactory closed-loop control performance. @en

BMW’s new driving simulation center operates multiple motion-base simulators – each with a different kinematic configuration – to serve various experiment use-cases and requirements of simulator users. The selection of a simulator for each experiment should ideally be based on their relative strengths and weaknesses. To support this decision-making process, subjective and objective predictions of motion cueing quality can be used. This paper provides an example comparison of four motion-base driving simulators. The kinematic configurations of the simulators considered differed in the additional presence of a yaw-drive and/or a linear xy-drive. The comparison is made by calculating offline, optimization-based motion cueing with perfect prediction capabilities (the ‘Oracle’) for nine urban drives. A prediction of subjective motion incongruence ratings is made for each simulator. In addition, an error type identification method is used (identifying scaling, missing cue, false cue and false direction cue errors) and evaluated per simulator. As Oracle can fully utilize the available workspace, the employed evaluation methods provide an insight in the fundamental capabilities of each simulator. Both the modelled ratings and the error type analysis show the benefits of adding a xy-drive in urban use-cases: predicted ratings reduce by 19% (i.e., better), while scaling and missing cue errors in the yaw rate are reduced when adding a yaw-drive. The presence of both of these additional motion systems allow for practically one-to-one and therefore error-free motion cueing. The proposed methods provide a straight-forward, yet insightful basis for simulator selection. The presented methods can be extended towards the analysis of multiple motion cueing algorithms and/or other usecases for systematically selecting the best-suited motion cueing method. @en

Cyberneticists develop mathematical human control models which are used to tune manual control systems and understand human performance limits. Neuroscientists explore the physiology and circuitry of the central nervous system to understand how the brain works. Both research human visuomotor control tasks, such as the pursuit tracking task. In this paper we discuss some commonalities and differences in both approaches to better understand the adapting human controller. Special attention is given to Adaptive Model Theory, which studied adaptive human control using several linear and nonlinear control engineering techniques. The insights gained yield schemes and concepts which pave the way for key future work on this topic.

@en

This paper proposes a new method that estimates the three-dimensional stochastic wind velocity for an aircraft equipped with a Pitot-static tube and airflow vanes. Since the performance of most state estimators, e.g., the extended Rauch-Tung-Striebel smoother, relies on the process and measurement noise covariance settings, the proposed method employs the expectation-maximization approach to estimate the noise covariance matrices to improve the estimation accuracy. Numerical simulations demonstrated that the proposed method can successfully estimate the noise covariance matrices, especially for the noise covariance of the wind velocity, using the measurement data and reconstruct the wind velocity offline. Additionally, the smoothed true airspeed, angle of attack, and angle of sideslip data are more accurate compared to the direct measurements. This feature is also beneficial for other applications such as the aerodynamic model identifications of aircraft.@en

Nonlinear dynamic inversion (NDI) is a nonlinear feedback linearization technique that has been widely applied to flight control systems [1,2]. Using state feedback and the inverted nonlinear system dynamics, NDI can significantly reduce controller development costs by avoiding gain scheduling and Jacobian linearization at a multitude of operating points. However, the control performance of NDI is directly dependent on required detailed knowledge of the model. As a simplified and enhanced NDI method [3], incremental nonlinear dynamic inversion (INDI) [4,5] has been proposed to reduce the model dependency and improve the robustness against model uncertainties. Instead of using a global nonlinear model, in INDI the dynamic inversion is implemented on a locally linearized system model that is updated at every sampling period, for which the control input is calculated in an incremental manner. Unlike NDI, for which full knowledge of the complete system dynamics is needed, INDI only requires explicit knowledge of the system’s control effectiveness matrix.@en

Mathematical human controller (HC) models are widely used in tuning manual control systems and for understanding human performance. Typically, quasi-linear HC models are used, which can accurately capture the linear portion of HCs' behavior, averaged over a long measurement window. This paper presents a deep learning HC skill-level evaluation method that works on short windows of raw HC time signals, and accounts for both the linear and non-linear portions of HC behavior. This deep learning approach is applied to data from a previous skill training experiment performed in the SIMONA Research Simulator at TU Delft. Additional human control data is generated using cybernetic HC model simulations. The results indicate that the deep learning evaluation method is successful in predicting HC skill level with 85-90% validation accuracy, but that training the classifier solely on simulated HC data reduces this accuracy by 15-25%. Inspection of the results especially shows a strong sensitivity of the classifier to the presence of remnant in the simulated training data. In conclusion, these results reveal that current quasi-linear HC model simulations, and in particular the remnant portion, do not adequately capture real time-domain HC behavior to allow effective training-data augmentation.

@en

Better understanding of manual control requires more research on human anticipatory feedforward behaviour. Recent advances include a human control model for preview tracking, and a subsystem identification (SSID) technique that uses a candidate pool approach to identify the human feedforward and feedback responses. This paper discusses the performance of the SSID method when estimating the preview control model parameters. Through simulations of a preview task with two controlled element dynamics, the SSID performance with different remnant noise levels and candidate pool densities is quantified. We demonstrate its successful application to the preview model and show that its performance deteriorates for higher noise levels. While the feedforward parameters are estimated accurately, the high-frequency compensatory feedback dynamics cannot be reliably determined. Future work focuses on alternative formulations for using SSID to estimate preview model parameters. Since in manual control the closed-loop magnitude decreases at higher frequencies, effects of manipulating the weightings of the closed-loop fitting cost values at these frequencies must be further analyzed.

@en