**Electromyographic Evaluation of Specific Elastic Band Exercises Targeting Neck and Shoulder Muscle Activation**

#### **Ying Gao 1,2, Lars A. Kristensen 3, Thomas S. Grøndberg 3, Mike Murray 3, Gisela Sjøgaard <sup>3</sup> and Karen Søgaard 3,4,\***


Received: 12 December 2019; Accepted: 16 January 2020; Published: 21 January 2020

**Featured Application: This electromyographic evaluation of elastic band exercises documents that the type and intensity of exercise played a role in the activity in three muscle groups of the neck and shoulder. However, importantly, shrugs and reverse flyes are simple to perform and also highly e**ff**ective training exercises as they induced high muscle activity of both shoulder and neck muscles at 12RM. The practical implications of these findings are that these simple exercises can be preferred for training programs targeting neck and shoulder muscle either for increased general strength or as prevention for work related pain and disorders.**

**Abstract:** Background: Specific strength training at a high intensity is effective in reducing work related neck/shoulder pain. However, it remains to be documented as to which exercises most specifically target neck and shoulder muscles at high activation level while using simple equipment as e.g., elastic bands. We hypothezised that selected exercises would specifically target the respective muscles, as follows: (1) shrugs and reverse flyes: the upper trapezius muscle, (2) cervical extension and lateral flexion: the upper neck extensor muscle, and (3) cervical flexion and rotation: the sternocleidomastoideus muscle. Methods: Eleven healthy males (25.9 ± 1.4 years, BMI 24.3 ± 1.4) with no neck/shoulder pain (VAS = 0) performed the six exercises with elastic bands at 12RM (repetition maximum) and 20RM in a randomized order. Electromyography was bilaterally recorded from the three muscles and it was normalized to maximal voluntary activation (%MVE). Exercises that evoke more than 60%MVE were considered as high intensity activation. Results: High muscle activation level was attained during 12RM in the upper trapezius muscle during shrugs (100.3 ± 29.8%MVE) and reverse flyes (91.6 ± 32.8%MVE) and in the upper neck extensor muscle during cervical extension (67.6 ± 29.8%MVE) and shrugs (61.9 ± 16.8%MVE). In the sternocleidomastoideus muscle, the highest activity was recorded during cervical flexion (51.7 ± 16.4%MVE) but it did not exceed 60%MVE. The overall activity was ~10% higher during 12RM when compared to 20RM. Conclusion: The simple exercises shrugs and reverse flyes resulted in high intensity activation of both the upper trapezius and neck extensors, while no exercises activated sternocleidomastoideus at high intensity.

**Keywords:** physical training intensity; therapeutic exercises; work related neck pain; musculoskeletal pain

#### **1. Introduction**

Neck and shoulder pain are frequent in the general population and within several occupational groups that are characterized by monotonous work tasks that require the stability of the head and neck in constrained postures [1–6]. In often male dominated jobs, helmets add to the weight of the head increasing the demands on the neck muscles and a high prevalence of neck pain is reported among e.g., construction workers, helicopter, and fighter pilots [7–9]. Physical exercise—in particular specific high intensity strength training at the job—has been shown to effectively reduce musculoskeletal pain in the neck and shoulders [10–13].

In clinical practice, at work sites, and in home-based rehabilitation the use of conventional resistance training equipment, such as mascines or free weight, may not be a possibility, as it requires high financial cost and advanced facilities. Thus, it is important to identify more feasible alternatives that are easy to perform at the workplace and still effectively target muscles in the neck and shoulder regions. One alternative is the use of elastic resistance bands that has recently gained interest, because they are simple to handle, low in cost, and easy to transport. However, with elastic bands, it might specifically for the neck extensor muscles be difficult to add enough external load as resistance to achieve the requested high activity level. To attain this, a special headband has been invented for more specific, but also more complex, elastic band neck exercises. These complex exercises have been applied in training interventions of helicopter pilots [8]. However, the attendance rate was low, which was possibly due to the more complex head band exercises; and, in that study, it was not investigated whether the more complex exercises actually resulted in a higher activation for the specific muscles when compared to the more simple neck exercises, like shrugs and reverse flyes [8,14]. The aim of the present study was to assess whether simple exercises as compared to the more complex exercises may suffice for sufficient activation of the neck and shoulder muscles.

Regarding the level of sufficient exercise intensity in strength training programs, this is often defined as the percentage of maximal voluntary force exerted [15]. In practice, one repetition maximum (RM) weight is often used as reference for the maximal weight that can be lifted in an exercise for a given number of repetitions. Activation levels of >60% of maximal voluntary contraction force (MVC) are generally recommended to attain a physiological effect of resistance training [15]. A higher force per contraction can be performed the lower the number of repetitions per set, e.g., at 1RM the performed force is close to 100% MVC, decreasing to around 60% MVC when performing 12RM—depending highly on training status. In well trained persons 60% MVC might be repeated more times, e.g., corresponding to 20RM.

To access the activation level, electromyography (EMG) is commonly used to measure the activity of the specific muscles that are involved in an exercise [16,17]; the bilateral upper trapezius muscle (mUTR), upper neck extensor muscles (mUNE), and sternocleidomastoideus muscle (mSCM) were the muscle groups included here. In the prevention and treatment of neck and shoulder pain such EMG measurements can ensure that a given mode and intensity of training will activate the target muscles to the sufficient level [18]. In addition, EMG measures allow for exploratory analysis, such as the development of muscle fatigue during multiple repetitions that might be an indicator of the long-term effect on hypertrophy, as well as the reduced perception of pain. The myoelectrical manifestations used as an indication of muscle fatigue is an increase in the amplitude and a shift in the power spectrum towards lower frequencies [19–21]. Previous studies with advanced exercises have presented important knowledge for selecting the most adequate exercises to target specific health effects [16,17]. However, studies using simple exercises have similarly shown an effect on both pain relieve and increase in strength [2,22]. The present study will expand on this by using simple (shrugs, SH, and reverse flyes, RF), as well as more advanced (cervical extension, CE and cervical flexion, CF), and quite difficult (cervical lateral flexion, CL and cervical rotation, CR) neck/shoulder exercises and compare the activation in the muscles specifically targeted with each exercise at 20RM and 12RM.

The main hypothesis was that specific training exercises were requested to maximize the activation of the individual neck and shoulder muscles: with mUTR activation being the highest during SH and RF, mUNE during CE and CL, and mSCM during CF and CR. Further, it was hypothesized that the intensity of the exercises significantly affected muscle activity: 12RM, resulting in higher activity than 20RM during the specific training exercises. Finally, we hypothesized that the muscle activity in the concentric phase was higher than in the eccentric phase and fatigue development was present in each set of an exercise being shown as an increase in EMG amplitude and a decrease in EMG frequency in the targeted muscle groups.

#### **2. Materials and Methods**

#### *2.1. Study Design and Subjects*

This study was conducted over three separate days that were interspaced with at least three days with no exercises. On the first day participants were familiarized with the elastic band training exercises and practiced the proper technique under supervision. On the second day, individual loadings were determined for each training exercise during low intensity, high repetition sets (20RM), and during high intensity, moderate repetition sets (12RM). All of the exercises were performed while using concentric and eeccentric muscle contractions in a controlled manner, with slow to moderate lifting velocity corresponding to the execution in earlier intervention studies [16]. Subsequently, the actual duration of concentric and eccentric phases were measured and shown to have an equal duaration each of minimum three seconds. The length and color code of elastic bands were adjusted to reduce or increase the resistance to match 12RM and 20RM for each participant. The task was repeated until the participant was exhausted and/or not able to continue the exercise according to the basics of procedures in the literature to define the individual 12RM and 20RM loads for each exercise [15]. The participant was deemed not able to continue the exercise if making compensating adjustment in body posture or not obeying contraction time and range of movement. The exercise was repeated until the loading corresponding to 20RM and 12RM was determined. On the third day, the determined loadings for each participant were applied and all EMG measurements were performed, while the participant performed all exercises at the intensity of 20RM sets, followed by 12RM sets. The order of exercises was randomized and each exercise was separated by 3-minutes rest intervals to avoid a gradient of fatigue through the test protocol. At the end, isometric maximal voluntary contractions (MVC) were performed for the normalization of the EMG data. For each targeted muscle, three MVCs with a duration of 5 s and 1 minute rest between consecutive trials were performed to induce the maximal muscle activity in a seated position and according to standardized procedures described in details in a recent study [23]. In short, maximal activity in mUTR was obtained during resisted bilateral shoulder elevation, with resistance being placed above the acromions. Maximal activity for mUNE was obtained during a resisted cervical extension and for mSCM during a resisted cervical flexion with resistance being placed on the external occipital protuberance and above the eyebrows, respectively.

The subjects participating were a group of 11 healthy males (age 25.9 ± 1.4 years, height 183.6 ± 5.0 cm, body mass 82.1 ± 6.0 kg, BMI 24.3 ± 1.4), with no pain in neck or shoulders (VAS = 0). The inclusion of only males was justified by the aviation personell—who used these exercises in a previous RCT study—all being males [8]. All of the subjects were informed regarding the benefits and risks of the investigation prior to signing the institutionally approved informed consent document to participate in the study. The local Ethics Committee of Southern Denmark (S-20120121) approved the procedures in the study.

#### *2.2. Procedures*

#### 2.2.1. Electromyography (EMG)

The activity of the neck and shoulder muscles was measured through a bipolar surface EMG configuration (Neuroline 720-01-K, Medicotest A/S, Ølstykke, Denmark). The electrodes (Ag/AgC1, Ambu Blue Sensor, N-00-S/25, Denmark) were positioned bilaterally, in accordance with the SENIAM

recommendations and earlier studies [24–27]) with an inter-electrode distance of 20 mm. The electrodes were placed (1) above the upper trapezius muscle (mUTR), 20% medially from the centerpoint between the lateral edge of the acromion and seventh cervical vertebra, (2) above the upper neck extensor (mUNE) muscle, on the most prominent part of the muscle at the level of the fourth cervical vertebra, and (3) above the sternocleidomastoideus muscle (mSCM), over the main muscle belly at the 1/3 distance from the sternal notch to the mastoid process (Figure 1). Before affixing the electrodes, skin was shaved and cleaned with scrubbing gel (Acqua gel, Meditec, Parma, Italy) and 70% alcohol, to effectively lower the impedance to less than 10 kΩ.

**Figure 1.** Bilateral electrode placement above the upper trapezius muscle (mUTR) and upper neck extensors muscles (mUNE) (**A**) and sternocleidomastoideus muscles (mSCM) (**B**).

The EMG signal was pre-amplified with a gain of 400 and then transferred through an EMG probe (Direct Transmission System (DTS), Noraxon Inc, Scottsdale, AZ, USA) to a belt-receiver (TELEMyo ™ 2400T, DTS, Noraxon Inc, Scottsdale, AZ, USA) with a bandwidth of 10–500 Hz and a common mode rejection ratio better than 100 dB. The signal was sampled at 1000 Hz while using a 16-bit A/D converter (DAQCard Al 16XE 50, National Instruments, Austin, TX, USA) and stored on a computer via the laboratory interface (CED 1401, Spike2 software, Cambridge Electronic Devices, Cambridge, UK).

#### 2.2.2. Elastic Band Training Exercises

Six specifically tailored elastic band training exercises were performed targeting the neck and shoulder muscles. The choice of exercises was evidence-based and designed by an interdisciplinary team, including sports exercise training specialists, physical therapists, medical doctors, and chiropractors. The exercises were performed with a head harness (Neck Flex ®, Leesburg, VA, USA) while using different color-coded elastic resistance bands (Thera-Band®, The Hygenic Corporation, Akron, OH, USA). The individual exercises are briefly presented below, as a detailed description can be found in an earlier publication [14] and video material is available online: Neck exercises.

In short the exercises are the following:


#### *2.3. Outcomes*

Analyses of the EMG signals were the outcome data. All data analysis was performed while using a custom-made script in Matlab (MathWorks Inc. USA) and Hedera 2.0 (University of Southern Denmark, Denmark). The peak EMG amplitude was calculated by root mean square (RMS) in a 500 ms moving window. For each repetition during exercises, the peak RMS was normalized to the peak RMS that was obtained during the MVC of each muscle as percentage of maximal voluntary EMG signal (%MVE) in standardized postures [23]. The 3rd, 4th, and 5th repetition of each exercise were averaged and used as the muscle activation for the exercises. The EMG activity in the concentric and eccentric phases of the muscle contraction was analyzed for each muscle during the exercise, resulting in the highest activation. The concentric and eccentric phases of each exercise were marked as an event in a separate channel in the Spike2 software, during each recording, As each repetition was performed with a 3 s concentric and 3 s eccentric phase, the midpoint was used to divide these phases. Additionally, the mean power frequency (MPF) was calculated in the last 1000 ms of the concentric phase of each contraction for the muscle groups during the exercise that represents the highest activation during the contraction.

#### *2.4. Statistical Analysis*

All of the data are presented as mean ± SD. The Q-Q plot and the Shapiro-Wilk test with skewness were used to test for normal distribution of data. Differences between left and right muscles during the symmetrical exercises SH, RF, CE, and CF, or between ipsilateral and contralateral muscles during CL and CR performed towards left and right side were examined while using paired-t test. In the case of no difference, the mean value of the two sides was used. If significant differences were found, the side specific values are reported and the highest used to represent the muscle activation level for that exercise. Regarding the main hypotheses of the muscle activity (%MVE) for each muscle group, one way repeated measures ANOVA was used to compare the different exercises (SH, RF, CE, CF, CL, and CR). If the assumption of Sphericity was violated, the Greenhouse-Geisser correction was used. Planned contrasts were used to localize the differences between the exercise, resulting in the highest activation of specific muscle group and the other exercises. Paired t-test was used for testing possible differences between the two exercise intensities. Furthermore, paired t-test was used to compare muscle activity in the concentric phase and eccentric phase of each muscle group. The fatigue development was tested by paired t-test to compare the third repetition and the last one repetition in normalized EMG amplitude (%MVE) and EMG frequency (MPF). A substantial addition in muscle activation level of at least 20% was requested in order to recommend an advanced, instead of a simple, exercise. This level is well above the SD of approx. 10% in mean level of trapezius activation both found within groups and variability of repeated EMG recordings during high intensity contractions, see, for instance, mean activation levels for trapezius in [17]. A power analysis showed that eight subjects would be sufficient to reveal a difference in muscle activation level of 20%. Within the recruiting period, we managed to recruit 11, who all conducted the full protocol. All of the results were analyzed using SPSS 16.0 statistical software (SPSS Inc, Chicago, IL, USA). A probability level of *p* < 0.05 (two-tailed) was considered to be statistically significant.

#### **3. Results**

Table 1 presents the muscle activation levels for the six exercises at two intensities. There were no differences between left and right side of muscles during the symmetrical exercises and no difference between ipsilateral and contralateral when the exercises CL and CR were performed to right versus left side. Therefore, the data were averaged over left and right muscles or ipsilateral and contralateral muscles, except during CL for mUNE and mSCM. This is due to the higher muscle activity during CL in ipsilateral than contralateral for mUNE (30.2% ± 13.6% vs. 12.8% ± 6.8%, *p* < 0.001, for 20RM, and 45.5% ± 14.6% vs. 17.2% ± 7.1%, *p* < 0.001, for 12RM) and for mSCM (25.6% ± 7.7% vs. 11.6% ± 6.3%, *p* = 0.001, for 20RM, and 41.5% ± 14.6% vs. 15.3% ± 8.4%, *p* = 0.001, for 12RM). Therefore, for CL the higher value of %MVE in ipsilateral of mUNE and mSCM were used as the results (Table 1).


**Table 1.** Mean muscle activity (%MVE) for representative muscle groups shown for each exercise at 12RM and 20RM intensities. Values are presented as mean ± SD. Values > 60% MVE are marked in bold. Italic values are unilateral data, see text.

The repeated measures ANOVAs revealed significantly different muscle activation between the types of exercise for: mUTR (F(1.099, 10.987) = 97.622, *p* < 0.001), mUNE (F(2.359, 23.59) = 29.038, *p* < 0.001), and mSCM (F(5, 50) = 18.045, *p* < 0.001), respectively. Figure 2 presents the contrast of muscle activity during the exercise, which induced the highest activation with other exercises. For mUTR, SH induced highest activity, and this was significantly higher than all of the remaining exercises, except RF. For mUNE, CE induced the highest activity and this was higher than during CF, CL and CR, but not significantly different from SH, and RF. For mSCM, CF induced the highest activity and this was higher than in all the remaining exercises. Furthermore, when compared with 20RM, the mean muscle activity of all exercises was significantly higher (relative increase 34%) during 12RM. For the

specific muscle groups the relative increase were 26% in mUTR (*p* < 0.001), 41% in mUNE (*p* < 0.001), and 35% in mSCM (*p* < 0.001). (Figure 3).

**Figure 2.** The muscle activity of the upper trapezius muscle (mUTR), upper neck extensors muscles (mUNE) and sternocleidomastoideus muscles (mSCM) were compared in six different types of exercises. The highest muscle activation (%MVE) was found during SH for mUTR, during CE for mUNE and during CF for mSCM. \* marks exercises with significantly lower activation than the exercise inducing the highest muscle activation.

**Figure 3.** Mean muscle activity (%MVE) for all exercises in the upper trapezius muscle (mUTR), upper neck extensors muscles (mUNE), and sternocleidomastoideus muscles mSCM during intensities of 12RM and 20RM. \* indicates significant difference between training intensities.

Analysis while considering activity separately in the concentric and eccentric contraction modes showed that the muscle activity of mUTR, mUNE, and mSCM in the concentric phase when compared to the eccentric phase was significantly increased by relative 75% for SH (*p* < 0.001), 32% for CE (*p* = 0.009), and 22% for CF (*p* = 0.002) (Figure 4).

**Figure 4.** The muscle activity (%MVE) of the upper trapezius muscle (mUTR), upper neck extensors muscles (mUNE), and sternocleidomastoideus muscles (mSCM) in the concentric phase (CON) and eccentric phase (ECC) during shugs (SH), cervical extension (CE), and cervical flexion (CF), respectively. \* Indicates significant difference between CON and ECC within exercise.

Analysis of fatigue development showed that, in each set of an exercise, a significant increase occurred in %MVE and a significant decrease occurred in EMG frequency (MPF) during SH, CE, and CF for each muscle groups of mUTR, mUNE, and mSCM, respectively (Table 2).

**Table 2.** The fatigue development of mean muscle activity (%MVE) and mean power frequency (MPF) for the upper trapezius muscle (mUTR), upper neck extensors muscles (mUNE), and sternocleidomastoideus muscles mSCM at intensities of 12 and 20RM. #Values are presented as mean ± SD for the exercises shrugs (SH), cervical extension (CE), and cervical extension (CF) that resulted in the highest activation of each muscle.


#### **4. Discussion**

#### *4.1. Main Finding*

Specific activation patterns were identified in the three neck/shoulder muscles that were studied, in accordance with the hypothesis exercise type. Interestingly, the simple exercises, SH and RF, elicited muscle activities of >60% MVE in mUNE as well as in mUTR. The more advanced exercise, CE, showed

numerically slightly higher—but not significantly—activation of mUNE, and CF produced the highest activity in mSCM, but this did not attain >60% MVE neither during 20RM nor 12RM and can therefore, not be recommend in case of mSCM insufficiencies. The two muscle groups most often attracting work related pain are mUTR and mUNE [1,3–5]. It is therefore interesting that both muscles can be intensively activated by two simple exercises (SH and RF), as this might potentially counteract the work related pain. The CL and CR, which are the most difficult exercises to perform correctly, did not elicit significantly higher percentages of muscle activation than any of the other exercises and can therefore be disregarded, unless for extra training variation. Thereby, our main hypothesis was rejected.

Earlier studies have reported significantly higher activation levels of mUTR during SH than RF [17,28], but this was not the case in the present study. The similar activation levels of mUTR during SH and RF may have great clinical importance due to the difference in force resistance that is required for performing the exercises. When performing the SH exercise for all subjects, a combination of three high resistance elastic bands were needed, whereas only one low resistance elastic band was used during RF. Thus, the SH exercise requires significant grip strength and core stability, which might be challenging for some populations. The mUTRs role as a scapula up- and outward rotator in combination with the significant external moment arm during the RF exercise may explain the high muscle activation in spite of low external resistance. Therefore, the RF exercise might, for some populations, be a more appropriate choice of exercise as compared to the SH.

The activity of mUNE was not significantly higher in CE compared to SH and RF. To our knowledge this is the first study to demonstrate high activity of neck muscles during SH and RF. This observation implies that the use of sophisticated exercises might not be needed as simple exercises effectively can activate both neck and shoulder muscles. This finding is supported by an earlier finding on high (>60% MVE) muscle activity of m. splenius capitis during lateral raise, which is similar to the RF that was used in the present study [29]. As splenius capitis is a part of the upper neck extensor muscle group, it will contribute to the gross activity measured as mUNE in the present study. Based on functional anatomy, the cervical spine will be forced in a protracted position during SH and RF, when performed with an external loading [30–32]. The neck muscles must be activated in order to stabilize and keep a neutral positio. This essential role of neck stability during SH, RF, and lateral raise may thus be a possible explanation for the high muscle activity in neck extensor muscle group. In this regard, it must be noted that the stabilizing effect arrives from static muscle contractions and the subsequent physiological effect might not be the same as from dynamic muscle contractions during the CE exercise [33–35]. Studies using wire electrodes to separate activity in the deep and superficial neck extensor muscles have pointed out the importance of activity specifically in the deep extensor muscles for patients with chronic neck pain [36]. Since surface EMG records from all of the underlying muscles, we cannot rule out that the same activity in SH and CE represent different contribution from the neck extensor muscles that may be of importance for specific patient groups. However, this distinction might not be equally important for the purpose of general strengthening of the neck extensors to relieve work related pain.

#### *4.2. Contraction Intensities*

In general, it is assumed that muscle activation increases with increased external load. Our hypothesis regarding a higher intensity during 12RM as compared 20RM was confirmed. We found that the EMG activity increased by 9% MVE from an overall mean of 28% MVE at 20RM to 37% MVE at 12RM. This corresponds to the estimated increase from 60% of 1RM at 20RM to the 70% of 1RM at 12RM. However, this general increase consisted of largely different increases for each of the targeted muscles. The muscles have a rather different function in each exercise and a high degree of co-contraction occurring in the neck region. Only for the mUTR would the 20RM evoke intensities >60% MVE that is normally considered to be most efficient for a training effect. From a pain reduction perspective, the high intensive exercises have been shown to be most effective and, therefore, the 12RM might be the best gain for the work time invested into exercises at the workplace [11,18].

#### *4.3. Muscle Activation Patterns*

The exercises included a concentric and an eccentric phase and EMG recordings clearly demonstrated that the highest activation occurred in the concentric phase, which is in concert with general physiological findings of the recruitment pattern explained by the force-velocity relationship [37,38]. Elastic band training exercises has been shown to compare well to activation patterns seen during traditional dumbbell exercises [29] and this underpins that elastic band exercise training might well substitute training with more advanced equipment. EMG recording also demonstrated that, during each set of contractions, significant muscle fatigue developed in all recorded muscle groups. In spite of the difference in activation level between the 12RM and 20RM in the first contraction of the sets then at the last contraction of the 12RM and 20RM both attained levels >60%MVE, which may potentially evoke an increase in muscle capacity [15]. Hereby, the present hypotheses regarding the contraction mode and fatigue EMG patterns were confirmed.

#### *4.4. Study Strenghts and Limitations*

We consider this study as a first step in the investigation of recommendable training and rehabilitation exercises for the neck/shoulder muscle based on elastic bands. The strengths of the study is the direct measures using EMG to assess the activation of representative neck muscle groups and instead of subjective assessment. Additionally, the design of the study using randomization of the exercises is a strength, since this ensures that the fatigue development within each type of exercise is not biased by its timewise performance.

The limitations of present study relate to the population only including a small group of males similar to majority of employees in the job groups specifically challenged with high exposures of the neck muscles due to helmet wearing or high accelerations (fighter pilots, helicopter pilots, construction workers, etc.). Additionally, the population consist of healthy subjects without serious neck/shoulder disorders. Therefore, extrapolation of our results to patients with neck/shoulder injury should be undertaken with caution. We cannot conclude that patients suffering from neck/shoulder pain will show similar muscle activation patterns when performing the exercises that are studied here.

Furthermore, EMG activity during the dynamic exercises was normalized to measurements of EMG activity during static MVCs to quantify the level of muscle activation. Given the inherent methodological limitations that are associated with surface EMG, only a rough estimate of the absolute level of muscle activation can be inferred while using this method. Fine wire electrodes for some of the deeper layers of neck extensor muscles could have further detailed the contribution of activity from each of the neck extensors. Finally, more muscles and contraction intensities could be studied to give an even more complete activation pattern for optimal recommendations of training exercises.

#### *4.5. Future Perspectives*

While considering the limitations above, future studies should include participants with work-related neck pain in a randomized controlled trial preferably designed with three groups: (A) training only using the simple exercises, (B) training with a variety of exercises, including the advanced and difficult exercises, and (C) a control group. Importantly the training volume, progression, undulation, and duration should be similar for groups A and B, as we have exemplified in previous studies [9–11,13]. The primary outcome should be neck pain testing the hypothesis that groups A and B would have equally positive effects. Additionally, secondary outcomes should be included, like muscle strength and adherence to the exercises; the latter in particular to test the hypothesis implied in the present study that adherence for group A would be larger that for group B.

#### **5. Conclusions**

The simple exercises (SH and RF) were highly effective, as they induced high muscle activation level of both shoulder and neck muscles and induced fatigue within the set of exercises. The more advanced neck exercises (CE and CF) and quite difficult exercises (CL and CR) did not, in general, add significantly higher muscle activation levels. The practical implications of these findings are that these simple exercises can be preferred for self-administered training programs—e.g., at the worksite—targeting neck and shoulder muscle either for increased general strength or as prevention for work related pain and disorders.

**Author Contributions:** Responsible for the conceptualization of the study were, M.M., G.S. and K.S.; the methodology, software, and validation was taken care of by M.M., L.A.K. and T.S.G.; formal analysis, M.M., L.A.K. and T.S.G.; investigation and experimental work, M.M., L.A.K. and T.S.G.; resources, G.S., K.S. and Y.G.; data curation and statistics, Y.G.; writing—original draft preparation, Y.G.; writing—review and editing, all authors.; visualization, Y.G., L.A.K. and T.S.G.; supervision, G.S. and K.S.; project administration, K.S. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Article* **Configurable 3D Rowing Model Renders Realistic Forces on a Simulator for Indoor Training**

#### **Ekin Basalp 1,\*, Patrick Bachmann 1, Nicolas Gerig 2, Georg Rauter 1,2 and Peter Wolf <sup>1</sup>**


Received: 31 December 2019; Accepted: 17 January 2020; Published: 21 January 2020

**Abstract:** In rowing, rowers need outdoor and indoor training to develop a proficient technique. Although numerous indoor rowing machines have been proposed, none of the devices can realistically render the haptic, visual, and auditory characteristics of an actual rowing scenario. In our laboratory, we developed a simulator to support rowing training indoors. However, rendered forces with the initial rowing model, which was based on a simplified fluid dynamic model that approximated the drag/lift forces, were not perceived realistic enough for indoor training by expert rowers. Therefore, we implemented a new model for the blade–water interaction forces, which incorporates the three-dimensional rotation of the oar and continuously adjusts drag/lift coefficients. Ten expert rowers were asked to evaluate both models for various rowing aspects. In addition, the effect of individualization of model parameters on the perceived realism of rowing forces was elaborated. Based on the answers of the experts, we concluded that the new model rendered realistically resistive forces and ensured a smooth transition of forces within a rowing cycle. Additionally, we found that individualization of parameters significantly improved the perceived realism of the simulator. Equipped with a configurable rowing model, our simulator provides a realistic indoor training platform for rowers.

**Keywords:** 3D force modeling; rowing biomechanics; robot-assisted training; individualized training; virtual reality simulator; training of experts; sports engineering; rowing simulator; tendon based parallel robot; transversal vibration control

#### **1. Introduction**

In rowing races, success is determined by the minimum time to complete a pre-defined course track of commonly 2000 m [1]. Thus, mathematically, the main performance metric for competitive rowers is the average boat speed to finish the course [2]. Attaining a faster boat speed requires not only a physical effort but also an appropriate boat setup and a skillful technique to transfer the propulsion forces into speed [3–5]. Therefore, fitness and technique trainings are crucial for competitive rowers [5].

Outdoor rowing training may not always be feasible since performance of rowers can be affected by unpleasant weather conditions [6,7]. Indoor training conducted with rowing machines offers a valuable alternative to outdoor practice [8]. Over the past decades, various rowing machines have been developed to yield a realistic rowing training [9]. Overall, the developed machines can be grouped in four categories depending on the technical aspects: indoor rowing tanks, ergometers, ergometer simulators, and virtual reality simulators [10].

For indoor-tank rowing, two large unit water-tanks are used next to a fixed rowing boat [11]. Since the oars directly interact with the real water, the resistance at the oar blade is naturally present. However, water tanks require specially constructed large spaces, which might be challenging to maintain.

Ergometers provide a useful strength training opportunity indoors. On an ergometer, a rower interacts with a handle to simulate the resistance. Rendered resistance is mainly dependent on the pulling speed of the handle. Ergometers generate the desired resistance forces through friction-based interaction of a flywheel with air or water [12]. Since ergometers built with a stationary footrest were found to increase inertial forces and joint injuries compared to the on-water rowing [13], moving footrest systems were also proposed for ergometers such as Rowperfect® Indoor Sculler [14] and Concept2® Dynamic [15]. Thus, the boat's acceleration and deceleration at each stroke can be simulated on such ergometers.

Ergometers do not render the essential kinematic and dynamic characteristics of an oar movement. Thus, advanced ergometer simulators that incorporate the actual handling of a rowing oar with three degrees of freedom were developed [16–18]. In ergometer simulators, rowers manipulate one (sweep) [16] or two shortened oars (sculling) [17] as in the case of actual rowing boats. The rolling motion of an actual boat can also be simulated on ergometer simulators. However, rowing simulators do not incorporate actual water dynamics to render the resistive forces. Resistance is rendered with an air fan and disc brake system, and the resultant force perceived by the rower is mainly dependent on the speed of the oar. Additionally, none of the ergometer simulators provide any visual or auditory displays to bring the realism of the training closer to the on-water rowing experience. Thus, virtual reality simulators were developed to augment the realism of indoor rowing training [19–22].

The virtual rowing training platform SPRINT was developed at Scuola Superiore S.Anna, Pisa, Italy [19,20]. The SPRINT platform included two shortened oars with three degrees of freedom (DoF) each, a fixed foot stretcher, and a sliding seat. Visual rendering of the rowing scenario was coupled to the output of a mathematical model that combined the movement of boat, oars, and a rower. Rendering of hydrostatic vertical forces and oar weight was realized by a mass–spring system. The vertical force model only depended on the vertical rotation of the oar, i.e., the effect of longitudinal rotation of the oar was neglected. Rendering of the horizontal forces in the virtual water was based on a modified form of a dissipation mechanism used in Concept2® ergometers. Due to this mechanism, haptic rendering of the horizontal forces did not capture the fluid-dynamics characteristics of the oar blade–water interaction. Therefore, the haptic rendering of the overall forces in the SPRINT platform was not based on the actual kinematic and dynamic features of a rowing movement.

In our laboratory, we also developed a rowing simulator to support realistic training for sculling and sweep rowing indoors [23]. Thanks to the real-time control and dedicated hardware setup, our rowing simulator has been used for training of rowers [22] and motor learning related research such as augmented feedback designs [24,25] and robotic training strategies [26,27].

The haptic rendering of the rowing forces incorporated the effect of the 3-DoF rotation of an actual rowing oar and of the boat dynamics [28]. The propulsion forces were calculated according to an empirical model [29], which approximated the drag and lift forces acting on the blade [21]. Although the model was previously shown to be suitable for technique training [22], expert rowers suggested that the realism of the rendered resistive forces on the simulator could be improved. Therefore, in this study, we proposed a new three-dimensional force model that incorporates 3-DoF rotation of the oar and accounts for a smooth blade transition between air and water.

The new force model was implemented on our rowing simulator and assessed by ten expert rowers in a training experiment. In the first part of the experiment, rowers were asked to qualitatively assess both the existing oar-blade model [21] and the new model. We hypothesized that the new blade force model would be perceived as more realistic than the existing model. In the second part of the experiment, the benefit of individualizing the force model on expert rowers' acceptance of the rowing

simulator as a training platform was investigated. We hypothesized that the individualization of model parameters would result in higher realism of perceived rowing forces of the simulator.

#### **2. Materials and Methods**

#### *2.1. Participants*

Ten expert rowers (five females, five males; age = 18–27 years, mean age = 21.8 years; experience of rowing = 5–15 years, mean experience = 7–8 years) were recruited for the experiment. Participants were either students at ETH Zurich and/or affiliated with a rowing club in Zurich or neighboring cities. We set the inclusion criteria for the 'expertness' in rowing as (i) intensive training with port-side sweep rowing for at least 3 years; and (ii) participation in at least one national and one international regatta as a port-side rower.

The current study was approved by the ETH Zurich Ethics Commission (EK 2017-N-75). Each participant was invited for a 2-h session conducted on a single day. Expert participants were verbally instructed about the goal, experimental protocol, assessment by means of a questionnaire, and the potential risks that might occur. In addition, participants were informed that they could withdraw from the study at any time without stating a reason and without facing any consequences. All participants signed a consent for participation and were reimbursed for their commitment with a gift (maximum value of CHF 20).

#### *2.2. Rowing Setup*

In the study, our custom-made rowing simulator was used (Figure 1) [21]. The rowing simulator consisted of an instrumented, shortened, single scull boat (Stämpfli Racing Boats AG, Zürich, Switzerland). The boat was placed on a platform in the middle of three 4.4 m × 3.3 m sized screens, i.e., one screen in front of the stern and one screen to each the left and right side of the boat [30].

g

**Figure 1.** Rowing simulator setup with a shortened single scull boat and three screens surrounding the boat. The person rowing in the boat is the first author of the paper.

A visual scenario representing a plain lake environment along with a line of buoys defining the course was displayed with three projectors (Projection Design F3+) on the large screens. The visual scenario was programmed in Unity (Unity Technologies ApS, CA, USA) and synchronized to the kinematic input from the participant with a minimum update rate of 30 fps.

The auditory rendering of the rowing scenario was developed in C++. Two speakers (DELL A525 Zylux Multimedia Computer Speaker System), which were placed behind the bow of boat, were used to deliver the task inherent, virtual water–oar interaction sounds with an update rate of 30 Hz.

The haptic rendering of the rowing simulator was realized by a tendon-based parallel robot [31]. The robot consisted of five ropes attached to the end-effector, i.e., shortened oar, in parallel, and five actuation units that were placed on a fixed frame [21]. The ropes were guided from the respective motors through fixed deflection units to control the motion of the oar for rowing application. At the other end of the oar, the handle of the oar was held by the user with both hands.

In the setup, the shortened oar was longitudinally fixed inside the oarlock with a clamp and bearing, which restrained the end-effector to move on a spherical surface. Therefore, the user could manipulate the oar in the three remaining rotational degrees of freedom. These three rotational oar movements were the horizontal oar angle θ, the vertical oar angle δ, and the longitudinal oar angle ϕ. Although the robot rendered forces in all three directions at the end-effector, the longitudinal forces were not transferred to the user due to the mechanical constraints of the oarlock. Rendered forces at the end-effector were haptically displayed to the user as torques along the horizontal oar angle θ and the vertical oar angle δ. The longitudinal oar angle ϕ was only controlled by the user.

The hardware configuration and rowing model application on our simulator was set up for portside sweep rowing. In sweep rowing, an even number of rowers are seated in the boat. Each rower holds a single oar at the oar handle with both hands. The haptic rendering of the rowing forces for this study was calculated for the portside and mirrored onto the bow side in the Matlab/Simulink® model to simulate a pair on the rowing simulator.

#### *2.3. 3D Modeling of Rowing Forces*

#### 2.3.1. Sub-Phases of a Rowing Stroke

Rowing is a cyclic movement that is executed with the coordination of a complete set of body limbs to control an oar motion to propel the boat [32]. One rowing cycle, i.e., stroke, can be partitioned into four sub-phases, which features specific kinematic and kinetic characteristics. A cycle begins with the catch, followed by drive, release, and the recovery [1]. In the catch, the rower rapidly lifts the oar handle to insert the blade into the water and prepares for application of propulsive forces. During drive, the rower pulls the oar handle towards his/her body by a well-coordinated movement of legs, trunk, and arms, while the seat is moved towards the bow and the oar blade is moved through water with power. In the release, the oar handle is swiftly pushed down to remove the blade from the water and oar is feathered to reduce the air resistance [33]. In the recovery, when the blade is fully in the air, the rower flexes trunk and legs while extending the arms towards the stern of the boat to prepare for the next cycle [32].

#### 2.3.2. Simulator Coordinates for the Rowing Models

To define the motion of the oar and resulting rowing forces, four different coordinate systems were defined. A boat-fixed coordinate system, i.e., χ*Boat* = *Bx*, *By*, *Bz* , was defined on the boat's center of gravity such that *By* was directed towards the stern and *Bz* upwards (Figure 2). An oar-fixed coordinate, i.e., χ*Oar* = *Ox*, *Oy*, *Oz* , was defined on the tip of the oar blade and was rotated with the blade movement. A special configuration of χ*Oar* was defined as the initial oar coordinate system, i.e., χ*initial Oar* <sup>=</sup> *Oi <sup>x</sup>*, *O<sup>i</sup> <sup>y</sup>*, *O<sup>i</sup> z* , when the portside oar was in the initial configuration (θ = 0◦, δ = 0◦, ϕ = 0◦) and perpendicular to the boat. Vertical (δ), longitudinal (ϕ), and horizontal (θ) rotations of the oar were the mathematically positive angles, i.e., defined in a counterclockwise (CCW) direction around *Oi <sup>x</sup>*, *O<sup>i</sup> <sup>y</sup>*, *O<sup>i</sup> <sup>z</sup>* axes, respectively. To calculate the forces acting against the oar blade resulting from the relative blade velocity, a stream coordinate system, i.e., χ*Stream* = *Sx*, *Sy*, *Sz* , was defined on the blade such that *Sy* = −*By* and *Sz* = *Bz*.

**Figure 2.** Coordinate systems and forces acting on the blade in the BIFC (binary blade immersion, fixed drag/lift coefficients) model. Forces due to horizontal oar movement are shown in the left pane (top view). the vertical force on the blade is shown in the right pane (back view of the boat).

Calculated rowing forces at the portside (right) oar blade were mirrored onto the bow side (left) blade, simulating identical rowers for a pair boat. As a result, the rotation of the simulated boat around the *Bz*-axis was assumed to be 0◦, and the linear displacement of the boat was constrained to be along the *By*-axis for simplicity.

#### 2.3.3. Description and Implementation of Base (1D) Models of Rowing for Blade–Water Interaction

Both rowing models presented in this paper were originally based on the modeling of drag and lift forces acting on a blade immersed into the water, which was provided by [34] as follows:

$$F\_{Drag} = \frac{1}{2} \mathcal{C}\_{D} \rho A \left(\overrightarrow{V}\_{o/w}\right)^{2} \tag{1}$$

$$F\_{Lift} = \frac{1}{2} \mathcal{C}\_{L} \rho A \left(\overrightarrow{V}\_{o/w}\right)^{2} \tag{2}$$

where ρ = 1000 kg <sup>m</sup><sup>3</sup> is the density of the water; <sup>→</sup> *Vo*/*<sup>w</sup>* is the relative blade velocity between blade and water; *A* is the projected area of the blade (calculated perpendicular to the front of the blade); *CD* and *CL* are the angle of attack dependent dimensionless drag and lift force coefficients, respectively.

Both equations were based on an experiment in which the blade was inserted into the water tank and held fixed against a one dimensional (1D) incoming flow [34]. Thus, the relationship between the forces and the velocity of the blade was only considered on the horizontal plane.

An alternative version of the 1D rowing model described by Equations (1) and (2) was implemented for our rowing simulator by [30]. In the simplified version of the model, drag and lift forces acting on the oar blade in the horizontal plane were approximated as follows:

$$F\_{Drag}^{H} = \mathbb{C}\_{D}(a\_{H}) \left(\overrightarrow{V}\_{\text{Corr}\_{\theta}}\right)^{2} \tag{3}$$

$$\mathbf{F}\_{Lift}^H = \mathbb{C}\_L(\alpha\_H) \left(\overrightarrow{V}\_{\text{Our}\_{\theta}}\right)^2 \tag{4}$$

where

$$\mathbb{C}\_D(\alpha\_H) = 2\mathbb{C}\_L^{\text{max}} \sin^2 \alpha\_H \tag{5}$$

$$\mathbf{C}\_{L}(\alpha\_{H}) = \mathbf{C}\_{L}^{\max} \sin 2\alpha\_{H} \tag{6}$$

and the angle of attack <sup>α</sup>*<sup>H</sup>* was defined between the longitudinal axis of the oar and <sup>→</sup> *VOar*<sup>θ</sup> , i.e., the oar blade velocity that depends on the 1-DoF horizontal rotation of the oar (θ). *Cmax <sup>L</sup>* was calculated from the maximum value of lift coefficient *CL* whose values were presented based on α*<sup>H</sup>* in [34].

In principle, the force model presented in [34] with Equations (1) and (2), and the interaction forces calculated with Equations (3)–(6) are based on the assumption that the magnitude of drag and lift forces depend quadratically on the blade velocity. To our knowledge, Equations (3) and (4) were initially mentioned for drag and lift oar blade forces in [29] based on the authors' interpretation of the results from the experiments and force model presented in [35]. In [35], the authors empirically found an approximation of the drag *CD* and lift *CL* coefficients that depend on the angle of attack (Equations (5) and (6)), which were further considered for the modeling of rowing in [29,36], and calculation of the efficiency of rowing oars [37].

In [30], the oar blade was assumed to be squared (vertical with respect to water surface, i.e., ϕ = 0◦), and the resultant force acting on the blade, i.e., *F*θ, was only dependent on the horizontal (θ) movement of the oar.

$$F\_{\theta} = \sqrt{\left(F\_{D\text{rag}}^{H}\right)^{2} + \left(F\_{Lift}^{H}\right)^{2}} = 2\mathcal{C}\_{L}^{\text{max}}\left(\overrightarrow{V}\_{\text{Our}\_{\theta}}\right)^{2}\sin\alpha\_{H}\tag{7}$$

However, for a real rowing stroke, the resultant forces on a blade do not only depend on the horizontal rotation of the oar θ but also on the vertical δ and longitudinal rotation ϕ. Thus, two rowing models that incorporate three rotations of the oar are presented in the following sub-sections.

#### 2.3.4. 3D rowing Model with Binary Blade Immersion and Fixed Drag/Lift Coefficients (BIFC)

In 2010, the 1D rowing model implemented by [30] was upgraded by [21] to a 3D rowing model, which incorporates the 3-DoF rotation of the oar (Figure 2). In [21], Equation (7) that was used for the θ angle-dependent forces was also applied to calculate the δ angle-dependent forces on the oar blade, i.e., *F*δ, due to the vertical rotation (δ):

$$F\_{\delta} = 2C\_{L}^{\max} \left( \overrightarrow{V}\_{\text{Corr}\_{\delta}} \right)^{2} \tag{8}$$

where *VOar*<sup>δ</sup> = *lOar* . δ, and *lOar* is the length of the oar.

The underlying assumption in Equation (8) was that the oar blade moved in the water with a feathered orientation (parallel with respect to the water surface, i.e., ϕ = 90◦). No additional angle of attack was employed since the influence of an incoming stream in this direction was omitted. *F*<sup>θ</sup> and *F*<sup>δ</sup> were combined with the longitudinal oar angle (ϕ) to calculate the horizontal (*F<sup>H</sup> Oar*) and vertical (*F<sup>V</sup> Oar*) rowing forces based on three rotational angles of the oar blade as follows:

$$F\_{\rm Our}^H = F\_0 \cos^2 \varphi + F\_\delta \frac{\sin 2\varphi}{2} \tag{9}$$

$$F\_{\rm Our}^V = F\_0 \frac{\sin 2\varphi}{2} + F\_\delta \sin^2 \varphi \tag{10}$$

from which the overall oar blade force (*FOar*) could be calculated as

$$F\_{\rm Corr} = \sqrt{\left(\overset{\rm H}{F}\_{\rm Corr}\right)^2 + \left(\overset{\rm F}{F}\_{\rm Cur}\right)^2} = F\_0 \cos qr + F\_\delta \sin qr. \tag{11}$$

Finally, to calculate the *F*<sup>θ</sup> and *F*δ, the value of the constant *Cmax <sup>L</sup>* was chosen as 42.25 <sup>N</sup> (m/s) <sup>2</sup> for a big blade area (*A*) of 0.13 m2 based on the calculations in [29,37].

#### 2.3.5. 3D Rowing Model with Linear Blade Immersion and Adjusted Drag/Lift Coefficients (LIAC)

In 2019, we developed a new 3D rowing model (linear blade immersion, adjusted drag/lift coefficients or LIAC), which was essentially based on Equations (1) and (2). In the model, *FDrag* was defined in the opposite direction of the oar blade velocity while *FLi f t* was perpendicular to drag. During a continuous movement of the oar blade in the water, the magnitude and the direction of *FDrag* and *FLi f t* change based on the values of drag (*CD*) and lift (*CL*) force coefficients, respectively. Furthermore, *CD* and *CL* were dependent on the instantaneous angle of attack. For the horizontal plane (P*xy*), we used a look-up table based on [34], which delivered the values of *CD* and *CL* depending on the calculated instantaneous angle of attack. Extending the model to the vertical plane (P*yz*), we employed a separate instantaneous angle of attack (Figure 3).

**Figure 3.** Coordinate systems and forces acting on the oar blade in the LIAC (linear blade immersion, adjusted drag/lift coefficients) model. Blade forces on the horizontal plane are shown in the left pane (top view). Blade forces on the vertical plane are shown in the right pane (side view of the boat). In the right pane, the enlarged circle shows the cross-section of the blade.

As a first step, the relative oar-blade velocity vector was defined in the boat-fixed coordinate system (χ*Boat*) as <sup>→</sup> *Vxyz* = *VOarx VOary VOarz T* . To calculate the angles of attack for the oar blade motion in the horizontal (P*xy*) and vertical (P*yz*) planes, the projection of <sup>→</sup> *Vxyz* vector onto these planes were described as <sup>→</sup> *Vxy* = *VOarx VOary <sup>T</sup>* for the horizontal and <sup>→</sup> *Vyz* = *VOary VOarz <sup>T</sup>* for the vertical planes.

According to the model presented in [34], the angle of attack used for the horizontal plane (P*xy*) was defined between the velocity of the oar blade (<sup>→</sup> *Vxy*) and the longitudinal (transverse) axis of the oar blade, yielding α*xy*. Following a similar approach, the angle of attack used for the vertical plane (P*yz*) was defined between the <sup>→</sup> *Vyz* component of the oar blade velocity and the cross section (vertical axis) of the blade, resulting in α*yz*.

For the implementation of α*xy* and α*yz* calculations in a Matlab/Simulink® model, the following relationships were used (Figure 3):

$$
\alpha\_{xy} = \beta - \pi - \theta \tag{12}
$$

$$a\_{yz} = \gamma - \frac{\pi}{2} + \varphi\_{\prime} \tag{13}$$

where <sup>β</sup> is an angle that is defined between <sup>→</sup> *Vxy* and *Sx* in the horizontal plane (CCW around *Sz*), and γ is the angle between <sup>→</sup> *Vyz* and the *Sy* (CCW around *Sx*).

Subsequently, drag and lift coefficients for both the horizontal and vertical forces were defined as *Cxy <sup>D</sup>* = *fD* α*xy* , *Cxy <sup>L</sup>* = *fL* α*xy* , *Cyz <sup>D</sup>* = *fD* α*yz* , and *Cyz <sup>L</sup>* = *fL* α*yz* , where the functions *fD* and *fL* were implemented as a look-up table based on the presented experimental values [34].

In the final step, the projected areas of the oar blade in the horizontal and vertical planes were calculated as follows:

$$A^{xy} = A\_b p\_{inun} \varepsilon \alpha \varphi \tag{14}$$

$$A^{yz} = A\_b p\_{inum} \tag{15}$$

where *Ab* = 0.13 m<sup>2</sup> was chosen for a big blade used in sweep rowing. *pimm* is the proportion of the oar blade in the water, which was calculated as

$$p\_{imm} = \begin{cases} 0 & \text{for } \delta\_{Surfacc} < \delta \\ \left| \frac{\delta - \delta\_{Surfacc}}{\delta\_{InWater} - \delta\_{Surfacc}} \right| & \text{for } \delta\_{InWater} < \delta \le \delta\_{Surfacc} \\ & 1 & \text{for } \delta \le \delta\_{InWater} \end{cases} \tag{16}$$

where δ*Sur f ace* and δ*InWater* are the vertical oar angles at which the tip of the blade touches the water surface and the blade is fully immersed in the water, respectively. Placement of Equation (16) in Equations (14) and (15) allowed for a smooth transition of forces between the drive and recovery.

For the calculation of *Axy*, the projected area was assumed to be dependent only on the longitudinal oar angle ϕ since horizontal rotation of the oar θ was included in the calculation of horizontal angle of attack α*xy*, which would update the magnitude of drag and lift coefficients according to the fluid dynamic model [34]. Moreover, for the calculation of *Ayz*, we assumed that the projected area would only depend on the immersion ratio of the blade since the vertical angle of attack α*yz* was dependent on the longitudinal oar angle ϕ. In addition, a further trigonometric adjustment of the *Ayz* based on the horizontal angle θ resulted in low forces in the vertical plane during the pilot tests conducted on the simulator, which were not preferred by the experts. Therefore, we did not include any dependency of other oar angles to estimate *Ayz*.

Based on the previous calculations, the drag and lift forces of the model (LIAC) that resulted from the horizontal, vertical, and longitudinal rotation of the oar blade were calculated as follows:

$$F\_{Drag}^{xy} = \frac{1}{2} \mathcal{C}\_D^{xy} \rho A^{xy} \left(\overrightarrow{V}\_{xy}\right)^2 \tag{17}$$

$$\mathbf{F}\_{Lift}^{xy} = \frac{1}{2} \mathbf{C}\_{L}^{xy} \rho A^{xy} \left(\overrightarrow{V}\_{xy}\right)^{2} \tag{18}$$

$$F\_{Drag}^{yz} = \frac{1}{2} \mathcal{C}\_D^{yz} \rho A^{yz} \left(\overrightarrow{V}\_{yz}\right)^2 \tag{19}$$

$$F\_{Lift}^{yz} = \frac{1}{2} \mathcal{C}\_L^{yz} \rho A^{yz} \left(\overrightarrow{V}\_{yz}\right)^2. \tag{20}$$

To calculate the emerging horizontal (*F*χ*<sup>B</sup> <sup>H</sup>* ) and vertical (*F*χ*<sup>B</sup> <sup>V</sup>* ) plane forces acting on the oar blade in the boat-fixed coordinate system (χ*Boat*), the drag and lift forces resulting from Equations (17)–(20) were transformed in the horizontal and vertical planes:

$$\overrightarrow{F}\_{H}^{\chi\_{B}} = \prescript{\text{Bant}}{}{\mathcal{R}}\_{\text{Straum}}(-\pi)^{\text{Straum}} \mathcal{R}\_{xy}(\boldsymbol{\beta}) \left[ \begin{array}{cc} F\_{\text{Drag}}^{xy} & F\_{\text{Lift}}^{xy} & 0 \end{array} \right]^{T} \tag{21}$$

$$\stackrel{\rightarrow\_{\mathcal{K}}}{F}^{\chi\_{B}} = ^{\text{Box}}\mathcal{R}\_{\text{Stram}}(-\pi)^{\text{Stram}}\mathcal{R}\_{yz}(\chi) \left[ \begin{array}{c} 0 & F\_{Drag}^{yz} & F\_{Lift}^{yz} \end{array} \right]^{T} \tag{22}$$

where *Boat*R *Stream*(θ*rot* = −π) was used to define the χ*Stream* in the boat-fixed coordinate system (χ*Boat*), and *Stream*R *xy*(β) and *Stream*R *yz*(γ) were the transformation matrices in the horizontal and vertical planes, respectively.

According to the model presented by [34], the resulting plane forces of <sup>→</sup> *F* χ*B <sup>H</sup>* and <sup>→</sup> *F* χ*B <sup>V</sup>* define two distinct 2D models in the horizontal (P*xy*) and vertical (P*yz*) planes, respectively. In order to generate a 3D force (*FOar*) from two 2D models (<sup>→</sup> *F* χ*B <sup>H</sup>* and <sup>→</sup> *F* χ*B <sup>V</sup>* ), we employed a combination factor, which would satisfy the following constraints:

$$F\_{\rm Car} = f\left(V\_{\rm xyz}^2\right)\mathbb{C}\_1 f\left(V\_{\rm xy}^2\right) + \mathbb{C}\_2 f\left(V\_{\rm xy}^2\right), \text{ where } \mathbb{C}\_1, \mathbb{C}\_2 \le 1 \text{ and } \mathbb{C}\_1 + \mathbb{C}\_2 = 1. \tag{23}$$

*C*<sup>1</sup> and *C*<sup>2</sup> are the combination factors for the horizontal and vertical planes, respectively. The underlying assumption for the relationship in Equation (23) was that the blade forces calculated in each plane, i.e., *f V*2 *xy* and *f V*2 *yz* , were dependent on the square of blade velocity that was projected onto the horizontal and vertical planes, respectively. Thus, to satisfy the relationship in Equation (23), *C*<sup>1</sup> and *C*<sup>2</sup> were defined as follows:

$$\mathcal{C}\_1 = \frac{V\_{xy}^2}{V\_{xy}^2 + V\_{yz}^2} \tag{24}$$

$$\text{C2} = \frac{V\_{yz}^2}{V\_{xy}^2 + V\_{yz}^2}. \tag{25}$$

Therefore, the 3D force resulting from Equations (23)–(25) was defined as

$$
\overrightarrow{F}\_{\text{Our}} = \mathbb{C}\_1 \overrightarrow{F}\_H^{\chi\_B} + \mathbb{C}\_2 \overrightarrow{F}\_V^{\chi\_B}.\tag{26}
$$

#### 2.3.6. Rendering of Forces on the Rowing Simulator

To calculate the forces, rowing models were implemented in Matlab/Simulink® (r2013b, The MathWorks, Inc., Natick, MA, USA). To render the rowing forces at the end-effector, the actuators driving each rope were simultaneously controlled by the Matlab/Simulink® model running on an xPC-target at 1 kHz. The implemented rowing models required all oar angles (θ, δ, ϕ), the seat position *xs*, and the seat acceleration .. *xs* as inputs to calculate the forces. Longitudinal oar angle ϕ and both of the seat parameters were measured with respective sensors in the instrumented boat. To retrieve the values of horizontal θ and vertical δ oar angles, the end-effector's position was initially calculated from the forward kinematics of the robot, i.e., change of rope lengths. Subsequently, a coordinate transformation was applied to the end-effector's position to calculate the horizontal θ and vertical δ oar angles. The longitudinal oar angle ϕ was measured with the potentiometers on the oar.

#### 2.3.7. Control of Rope Tension Forces during Recovery

In the presented rowing models, the resistance forces resulting from air friction during recovery were not explicitly modeled. Instead, the resistance was provided with the residual forces acting at the oarlock, which results from tensioning the ropes of the tendon based parallel robot to control the translational motion of the end-effector, i.e., shortened oar.

A technical challenge regarding the use of a tendon based parallel robot is that rendering of high rowing forces at the end-effector due to rapid oar movements can result in unwanted transversal vibrations on the ropes and resonance on the robot. Since the vibrations mainly occur when the ropes are flexed, the stiffness control of the robot plays an important role in the haptic rendering of the rowing models on the simulator. Therefore, the natural (eigen) frequency of the system needs to be determined to understand the vibrational characteristics and assure the stiffness of the robot.

Assuming that the ropes are modeled as taut strings in our robot, the wave equation that governs the natural transversal vibration of the rope can be described as follows [38]:

$$
\omega\_n = \frac{n\pi}{l\_i} \sqrt{\frac{\tau\_i}{\zeta}} \tag{27}
$$

where ω*<sup>i</sup> <sup>n</sup>* is the eigen frequency of the *i*th rope, *n* is *n*th frequency, *li* is the length of the *i*th rope, τ*<sup>i</sup>* is the tension force on the *i*th rope, and ζ is the mass per unit rope length.

According to Equation (27), the natural frequency for transversal vibration of the rope is proportional to the square root of the rope tension and inversely dependent on the length of the rope. Confirming the relationship presented in Equation (27), three out of five ropes of the tendon based parallel robot are lengthened during catch-to-drive transition in the simulator, which reduces ω*<sup>n</sup>* and thus results in increased transversal vibrations on the ropes. However, the vibration on the ropes can be suppressed if the eigenfrequency ω*<sup>n</sup>* is increased. Although increasing the tension on the ropes would yield a stiffer tendon-based parallel robot, increased stiffness would also yield higher friction forces at the oarlock, which may cause unrealistically high resistive forces during recovery. Therefore, we could not use fixed, high tension forces on the ropes during whole rowing cycle to counteract the vibrations resulting from high rowing forces during catch and drive.

In the previous studies on the rowing simulator, the vibration problem could be suppressed by setting the minimum rope tension force (τ*min rope*) to a fixed value of 50 N , since the rendered forces at the end-effector were low due to the fact that body–arm rowing was used as the rowing task [24–27]. However, this minimum rope tension of 50 N yielded neither realistic recovery forces nor eliminated the transversal vibration issue for the rapid, full-body rowing movement. Therefore, we implemented a position-dependent τ*min rope* adjustment in the LIAC model. Three different thresholds for τ*min rope* were defined dependent on the vertical movement of the oar (δ) (Figure 4): τ*min rec* for the rope tension during recovery, τ*min transit* during catch and release, and <sup>τ</sup>*min drive* during drive.

**Figure 4.** Illustration of the vertical oar rotation-dependent rope tension adjustment for the tendon based parallel robot. A typical rowing cycle is illustrated with the thick elliptic line. Four arrows show the direction of the oar blade within the cycle.

The linear adjustment of the rope tension forces were defined as follows:

$$
\tau\_{\rm rops}^{\rm min} = \begin{cases}
\tau\_{\rm rcc}^{\rm min} & \text{for } \delta\_{\rm Tramit} < \delta \\
\tau\_{\rm rcc}^{\rm min} + \left(\tau\_{\rm rmsit}^{\rm min} - \tau\_{\rm rcc}^{\rm min}\right) \left(\frac{\delta\_{\rm Tramit} - \delta}{\delta\_{\rm Tramit} - \delta\_{\rm Surf}}\right) & \text{for } \delta\_{\rm Surnic} < \delta \le \delta\_{\rm Tmmit} \\
\tau\_{\rm rmsit}^{\rm min} + p\_{\rm imm} \left(\tau\_{\rm drive}^{\rm min} - \tau\_{\rm rmsit}^{\rm min}\right) & \text{for } \delta \le \delta\_{\rm Surnic}
\end{cases} \tag{28}
$$

where δ*Transit* is an empirically set value (1◦) for the vertical oar angle above the simulated water surface. In Equation (28), initially set values for the parameters were τ*min transit* <sup>=</sup> 50 N, <sup>τ</sup>*min drive* = 60 N.

#### 2.3.8. Configurable Parameters of LIAC for an Individualized Rowing Experience on the Simulator

A pilot study with three expert rowers, who were not included in main study, was conducted to assess and calibrate the newly developed LIAC model. The goal was to adjust the LIAC parameters, which would yield realistic forces for all sub-phases of the cycle and a corresponding boat velocity. Thus, we chose four parameters as follows that modulated the calculation of the forces: drag and lift adjustment coefficient (*Gdl*) for drive, curvature of the rowing blade (ϕ*bias*) for catch and release, minimum rope tension force (τ*min rec* ) for recovery, and boat drag coefficient (*CBdrag*) for the boat velocity.

For drive, three expert rowers reported that the resulting 3D model forces were greater than the forces felt during actual rowing. Thus, to calibrate the forces in the horizontal and vertical plane, we employed an additional gain, i.e., *Gdl*, for the adjustment of drag and lift coefficients:

$$\mathbf{C}\_{D'}^{xy} = \mathbf{G}\_{dl} f\_{\mathbf{D}} (\alpha\_{xy}) \tag{29}$$

$$\mathbf{C}\_{L'}^{xy} = \mathbf{G}\_{dl} f\_{\mathbf{L}}(\boldsymbol{\alpha}\_{xy}) \tag{30}$$

$$\mathcal{C}\_{D'}^{yz} = \mathcal{G}\_{dl} f\_{\mathcal{D}} \left( \alpha\_{yz} \right) \tag{31}$$

$$\mathbf{C}\_{L'}^{yz} = \mathbf{G}\_{dl} f\_{L}(\boldsymbol{\alpha}\_{yz}) \tag{32}$$

where *D* and *L* subscripts define the adjusted drag and lift coefficients, which were replaced in Equations (17)–(20). In the end of the pilot test, *Gdl* was set to 0.35 by taking the average of the preferred gains reported by three expert rowers.

Another critique of the pilot test rowers was that the vertical forces during the air–water transition were low. Thus, we increased the vertical forces in the catch and release by introducing a bias angle, i.e., ϕ*bias*, which was assumed to account for the lateral curvature of the rowing blade. Originally, the oar blade presented in the LIAC model was designed as a flat plate. Thus, introducing ϕ*bias* yielded:

$$
\varphi' = \varphi\_{\text{bias}} + \varphi\_{\text{act}} \tag{33}
$$

where ϕ*act* = ϕ is the actual longitudinal rotation of the oar blade and ϕ is the adjusted oar blade rotation used in the Matlab/Simulink® implementation. The initial value of ϕ*bias* was set to 2◦.

The minimum rope tension force in the tendon based parallel robot was adjusted for the rowing forces during recovery (τ*min rec* ). Since τ*min rope* was set to a fixed value of 50 N for the overall cycle in the BIFC model, the initial value of τ*min rec* in the LIAC model was also set to 50 N for the recovery (*pimm* = 0). All three rowers judged that the recovery force rendered by LIAC was unrealistically resistive; thus, an average value of 35 N was set for a smooth manipulation of the oar in the recovery.

Finally, pilot study rowers were asked to assess how well the perceived boat velocity matched the applied forces on the simulator. Based on the implementation of the rowing model, the boat velocity . *yboat*, was dependent on the driving forces applied by the rowers on the oar blades (*Fpropulsion*), inertia forces resulting from the movement of the rowers and the oar blades (*Finertia*), and drag forces acting on the boat (*FBoatDrag*).

$$
\dot{x}\_{\text{bat}}\ddot{y}\_{\text{bat}} = F\_{\text{Net}} = F\_{\text{propulsion}} - F\_{\text{inertia}} - F\_{\text{BantDrug}}\tag{34}
$$

where *mboat* is the combined mass of rowers and the boat, and .. *yboat* is the acceleration of the boat in *By*-axis [28]. In Equation (34), *FBoatDrag* is calculated as in [29] as follows:

$$F\_{B\text{out}Dng} = \mathcal{C}\_{Bdng} \left(\dot{y}\_{\text{bant}}\right)^2 \tag{35}$$

According to Equation (34) and Equation (35), a modulation of *CBdrag* affected the acceleration and velocity of the boat, which also altered the calculated drag and lift forces from Equations (17)–(20).

The initial value of *CBdrag* in the LIAC model was set to 4.94 <sup>N</sup> (m/s) <sup>2</sup> based on the calculations for a pair type of boat in [30]. Pilot study rowers judged that the perceived boat acceleration and velocity (from the visual display as well as resulting forces on the blade) felt realistic for the given initial value. Although *CBdrag* was not further modulated in the pilot test but depended on the type and size of the boat [29], *CBdrag* was implemented in the model to be configurable by the rowers.

In the pilot test, we observed that the three rowers individually preferred slightly different values for the presented parameters, yielding a variety of rowing model settings. Indeed, rowers may differ in many aspects, e.g., anthropometric measures, technique, strength, accustomed water conditions (river, lake, or sea), preferred boat rigging, etc. Therefore, we decided to present the previously chosen model parameters, i.e., *Gdl*, ϕ*bias*, τ*min rec* , and *CBdrag*, also to the main study expert rowers as a means of adjusting the rowing model to their individual preferences. Initial values of all the model parameters for the main study were determined by taking the average of preferred values of pilot study rowers.

#### *2.4. Experimental Protocol*

The study consisted of two parts. In the first part, five training blocks were conducted (Figure 5). Each of the blocks served to focus on a different aspect of the rowing movement. In each block, both the BIFC and LIAC models were presented twice in a randomized manner, yielding four trials within the block. The order of model presentation was based on Latin square randomization. The order of models presented within a training block was different in each block for a participant. Participants were not explicitly informed about the number of different models or the order of the model presentation. They were only instructed to focus on the evaluation of the movement aspect in each of the trials within the training block. Each trial with the presented model lasted for two minutes unless the participant decided to finish the trial and rate the model on the questionnaire.


**Figure 5.** Experimental protocol. Training 1 to 5 were run in the first part of the experiment. Training 6 was run in the second part. Conf. and Assm. stand for Configuration and Assessment, respectively.

Prior to the Training-1, participants were not provided with an additional familiarization trial with any of the models. After the completion of each training block (four trials), a two-minute break was given. The first part of the experiment ended with the completion of Training-5.

In the second part of the experiment, i.e., Training-6, the participant was only provided with the LIAC model. The participants were given a minute to get familiar (again) with the model. Thereafter, experts were free to configure the parameters that would adjust the relevant rowing forces within four to five minutes. Configuration of the LIAC model resulted in the 'individualized' model, i.e., IND. After the individual configuration, participants were given one more minute to row with the final set of configuration. Finally, participants were asked to re-evaluate each of the questions in the previously presented questionnaire. The final evaluation typically took five minutes.

#### *2.5. Task*

The goal of the experiment was to qualitatively assess the realism of the presented rowing models for different aspects of the rowing movement. Experts were asked to mainly perform a full-body rowing cycle by coordinating their body limbs. They could also try other types of movements, e.g., half-slide, body-arm, only-arms rowing, if they wanted. They were free to decide to row with a feathered, i.e., horizontal oar blade with respect to water surface, or a squared blade in the recovery. In Training-1 and Training-2, the rowers could also pause at a certain posture and insert the squared blade into virtual water or tap the feathered blade onto the virtual water surface.

Experts were instructed to consider the experiment on the rowing simulator as if it was a technical training on the water, i.e., rowing with low to medium stroke rates and 40–70% of maximum force application. To capture the realism of the rowing forces with both compared models, participants were not explicitly constrained to any stroke rate. However, due to our instruction, they were observed to row with a range between 12 and 28 strokes per minute.

In the first part of the study, expert rowers were asked to focus on specific aspects of the rowing movement. To familiarize the expert rowers with the simulator and compare the presented force models, rather secondary/inessential aspects of a rowing stroke were considered in the first two trainings. In Training-1, the focus was on the horizontal movement of the blade above the water surface. Rowers deliberately slid the feathered blade on the virtual water surface and tried braking with the squared blade, which is a typical move to stop the boat during water outings. In Training-2, vertical movement of the oar during air to water transition was considered. In this training, rowers tried to distinguish the vertical forces between the squared blade insertion into water and the feathered blade tapping onto water surface. In addition, they were requested to purposely square the oar blade near to the water surface to feel the vertical reaction forces from the water surface.

In Training-3, the focus was on the forces during drive and recovery. In Training-4, the focus was given to the oar blade forces during the air–water transition, i.e., catch and release. In Training-5, the rowers were asked to rate the overall perception of the rowing forces considering the whole cycle. Finally, they were asked to rate the realism of the flow (speed) of the visual scenario with respect to their stroke rate and the force they applied. Completion of every trial was followed by filling out a questionnaire.

In the second part of the study, i.e., Training-6, participants were asked which particular forces they would have liked to adjust for their individual preferences. Upon participants' requests, the study instructor set the respective parameters from the graphical user interface to update the relevant forces of the active rowing model. Participants were asked to give their verbal comments to fine tune the force model according to their rowing preferences. The configurable parameters were *Gdl*, ϕ*bias*, τ*min rec* , and *CBdrag*. Once the participants were satisfied with the final individualized parameter settings, they were asked to assess the realism again.

#### *2.6. Questionnaire*

Two questionnaires with the same questions were prepared for both parts of the experiment. The questionnaires were used for the qualitative comparison of the rowing models not only to each other but also to the real rowing scenario on the water.

The questionnaires consisted of one page for each training block. On top of every page, the sub-goal of the training block and an 11-point Likert-scale was presented. On the Likert-scale, '0 = Not realistic at all', '5 = Moderate (Neutral)', and '10 = Feels like on the water' statements were inserted below the respective values. In the middle of the page, two questions were stated for every training aspect. A written instruction and an illustration were present to help participants understand the questions. The designs of both questionnaires were the same except the trial counts within the blocks. For the first part, i.e., Training-1 to 5, Questionnaire-1 included four boxes (one for each trial) for the rating input for two questions at the bottom of the page. For the second part, Questionnaire-2 included two boxes for two questions: an upper box for the mean value of LIAC that was calculated by the study instructor after the first part, and a lower box for the new evaluation rating input.

Both questionnaires were prepared in Microsoft Word. The questionnaires were displayed on a monitor, which was placed to the front-left side of the participant. Participants were also provided a keypad input device, which was placed on the left side of the boat, to rate the respective aspect of the rowing movement. Participants entered numerical values from zero to ten using the keypad to fill the question boxes for each trial. Participants were left alone in the simulator, and the study instructor waited outside to avoid affecting the experts by any means during the model evaluation.

#### *2.7. Questionnaire Analysis (Statistics)*

For the statistical analysis of questionnaire evaluation, Matlab® 2017a and RStudio (Integrated Development Environment for R, version 1.1.463, RStudio Team, 2018) software were used. Separate analyses were conducted for each part of the study since the goal in each analysis was different.

For the first part, to check if the compared base rowing models (BIFC, LIAC) significantly differed in terms of realism based on the qualitative evaluation with the Questionnaire-1, the following linear mixed effect (*lme*) model was used.

$$Q\_i \sim \text{BaseModel} + (1 \mid \text{Expert}) \tag{lmc-1}$$

where *Qi* is the dependent variable that represents the questions in the questionnaire, and *i* = 1, ... , 10 is the index of each question. *BaseModel* is the categorically independent variable with two levels defined as BIFC and LIAC. Since the models were presented twice in the training blocks, both rating values were used for each rowing model for every *Qi* in the *lme-1*. *Expert* is defined as the random factor since ratings in both models belonged to one participant.

A separate analysis was conducted for the effect of individualization in the second part of the study. To examine if the individual adjustment of forces yielded a significant improvement in the realism ratings in Questionnaire-2 for the rowing on the simulator, *lme-2* was constructed as follows.

$$Q\_i \sim \text{Configration} + (1 \mid \text{Expert}) \tag{lmc\text{-}2}$$

where *Con figuration* is the categorical independent variable with two levels: the configurable base model, i.e., LIAC, and the individually configured model, i.e., IND. Since the configured model IND was rated once with Questionnaire-2 (after the individual adjustment in Training-6), the mean of two ratings for the LIAC model was retrieved from the first part of the study (Questionnaire-1) in *lme-2*.

For both lme models, "*lmerTest*" package was used to check the significance by retrieving the *p*-values. To check the main effect of *BaseModel* in *lme-1* and *Con figuration* in *lme-2,* the "*anova*" method was used. Normality of the residuals from both lme models were visually inspected using Q–Q plots. Resultant *p*-values below 0.05 were regarded to show significance. The *p*-values between 0.05 and 0.1 were considered as 'trending towards significance.

#### **3. Results**

For every participant, individual initial and final values of the LIAC model are listed in Table 1.



#### *3.1. Comparison of Base Rowing Models (BIFC vs. LIAC)*

#### 3.1.1. Aspect 1 and 2: Interaction between Oar Blade and Water Surface

For the horizontal and vertical blade movements above the water surface, the *lme-1* did not show any significant main effect of the *BaseModel* except the feathered blade sliding on the water surface (Table 2). For Question 1, LIAC was rated as more realistic than BIFC (*FBM*(1, 30) = 26.81; *p* < 0.001) with a slightly above-medium value (μ*Q*<sup>1</sup> *LIAC* = 5.95).



3.1.2. Aspect 3 and 4: Interaction of Oar Blade and Virtual Water during Sub-Phases of the Rowing Cycle

For all four sub-phases of the rowing cycle, *lme-1* revealed a significant main effect of the *BaseModel* (Question 5, 6, 7, and 8, Table 2). The LIAC was rated more realistic than BIFC for the drive (Q5: *FBM*(1, 30) = 39.36; *p* < 0.001), recovery (Q6: *FBM*(1, 30) = 30.37; *p* < 0.001), catch (Q7: *FBM*(1, 30) = 24.86; *p* < 0.001), and release (Q8: *FBM*(1, 30) = 6.72; *p* = 0.015). For all the sub-phases of the cycle, rating of LIAC was above medium. However, BIFC was rated below medium for drive and recovery and medium for the catch and release.

3.1.3. Aspect 5: Overall Realism of the Rowing Forces and Synchronized Flow of Visual Scenario

The *lme-1* revealed a significant main effect of *BaseModel* for the overall realism of the rowing forces and a trending main effect for the flow of visual scenario (Q10: *FBM*(1, 30) = 3.91; *p* = 0.057). For Question 9, LIAC was rated more realistic than BIFC (Q9: *FBM*(1, 30) = 39.12; *p* < 0.001) that was rated below medium. For Question 10, both models were rated with a slightly above-medium value (Table 2).

#### *3.2. E*ff*ect of Individualized Configuration (LIAC vs. Individually Configured model (IND))*

#### 3.2.1. Aspect 1 and 2: Interaction between Oar Blade and Water Surface

The *lme-2* showed a main effect of *Con figuration* for the horizontal movements of the oar blade above the water surface (Q1: *FC*(1, 10) = 6.01; *p* = 0.034 and Q2: *FC*(1, 10) = 5.56; *p* = 0.041) (Table 2). The realism rating value for the forces of oar blade sliding (Q1) increased by 11% compared to the LIAC model (μ*Q*<sup>1</sup> *IND* = 6.60). Rating of the braking forces increased by 16%. Regarding the water surface interaction forces resulting from the vertical movement of the blade above surface, there was a trend for a main effect of the *Con figuration* for the Question 3 (Q3: *FC*(1, 10) = 4.86; *p* = 0.052). The rating value for the Question 3 increased by 10%. The *Configuration* did not yield any significance for the Question 4 although the rating value increased by 13%.

3.2.2. Aspect 3 and 4: Interaction of Oar Blade and Virtual Water during Sub-Phases of the Rowing Cycle

The *lme-2* showed a significant main effect of the *Con figuration* for all four sub-phases of the rowing stroke (for Question 5, 6, 7, and 8, Table 2). Compared to base LIAC model, the configured LIAC model, i.e., IND, was rated more realistic for the drive (Q5: *FC*(1, 10) = 6.13; *p* = 0.032), recovery (Q6: *FC*(1, 10) = 54.44; *p* < 0.001), catch (Q7: *FC*(1, 10) = 7.21; *p* = 0.023), and release (Q8: *FC*(1, 10) = 6.95; *p* = 0.025). The rating values of the drive, recovery, catch, and release were increased by 13%, 28%, 12%, and 17%, respectively.

3.2.3. Aspect 5: Overall Realism of the Rowing Forces and Synchronized Flow of Visual Scenario

The *lme-2* revealed a main effect of *Con figuration* only for the perception of overall rowing forces (Q9: *FC*(1, 10) = 16.74; *p* = 0.002) (Table 2). The rating of Question 9 increased by 18% after individual configuration (μ*Q*<sup>9</sup> *IND* = 7.90). Tuning of the rowing LIAC model did not significantly improve the perception of the visual scenario (Table 2), and the rating value increased by 6%.

#### **4. Discussion**

Robotic simulators provide an important environment for training of surgeons [39], pilots [40], motor learning of complex sports [41,42], and training of athletes for the performance improvement in real-life sports [22]. To benefit from simulator training in real-life tasks, simulators are required to realistically render the task aspects [43,44].

In this study, we propose a novel 3D rowing model to render the interaction forces between the oar blade and water for a rowing simulator. The implemented model was assessed by expert rowers in terms of the realism of the rendered rowing forces. The assessment of the realism was based on the subjective perception of the experts about how realistic the rowing forces were rendered. At first, two distinct 3D models were compared in numerous aspects of the rowing movement. We hypothesized that the new 3D rowing model would render the forces on the simulator more realistically than the previously implemented 3D model [21]. Secondly, we elaborated whether individualizing parameters of the new rowing model further increases the perceived realism of the interaction forces.

The consensus from the experts about the rowing simulator was that it provided a more realistic indoor training platform than any other rowing devices in the market due to incorporation of not only visual and auditory rendering of rowing aspects but also a realistic task force rendering with 3-DoF oar manipulation. Results from both parts of the study along with the impact of model parameters on the perceived realism are discussed in the following sub-sections.

#### *4.1. Comparison of Two 3D Rowing Models (BIFC vs. LIAC)*

Presented 3D rowing models that calculated drag and lift forces acting on the oar blade mainly differed in two parameters, i.e., use of angle of attack and the blade area in the water. In comparison to the BIFC model, the LIAC model not only incorporated the effect of angle of attack in horizontal plane

(α*xy*) but also in the vertical plane (α*yz*). In the LIAC model, the resultant drag and lift coefficients in horizontal and vertical planes were continuously adjusted according to the instantaneous angle of attack. On the contrary, a pre-determined value of *Cmax <sup>L</sup>* was used in the BIFC model for the forces in both planes and the vertical forces were not dependent on the angle of attack due to modeling assumptions. In addition, the BIFC model did not incorporate the effect of the oar-blade area in the water. However, forces from the LIAC model depended on the proportion of the blade area in the water.

We also investigated the effect of rope tension forces (τ*min rec* ) on the rendering of recovery forces. In real rowing, rowers push the oar handle with a minimal force to move the oar from release to catch position during recovery. Thus, to provide a realistic feeling on the simulator during recovery, the rendered forces on the oar blade should be the lowest compared to other sub-phases of the rowing cycle. Therefore, we initially set the τ*min rec* value for the LIAC model to 35 N based on pilot study and compared to the τ*min rec* value of the BIFC model that was kept at 50 N .

#### 4.1.1. Rowing Cycle and Sub-Phases

As hypothesized, the newly proposed 3D rowing LIAC model was rated as significantly more realistic than BIFC in drive (Q5), recovery (Q6), catch (Q7), release (Q8), and overall rowing cycle (Q9).

For drive, expert rowers mentioned that the rendered forces with LIAC were realistically resistive. The development of the force during early drive and the reduction of the force in the late drive was also judged as realistic. Experts also reported a smoother force increase and decrease in the catch and release, respectively. Inclusion of parameter *pimm* in the blade area calculation accounted for a smooth transition of the oar blade during catch and release, which plays an important role in the exerted propulsion forces, continuity of the oar motion and elimination of unwanted decelerating blade drag forces [45].

On the other hand, the resistance provided by the BIFC model was reported to be low during the drive. It could be argued that a selection of a higher *Cmax <sup>L</sup>* value might have yielded a more realistically resistive force in drive. However, the force calculation in the BIFC model did not incorporate a parameter for the 'immersed blade area in the water' as in the case of LIAC. Thus, due to rapidly increased forces, the binary approach in the force modeling might have introduced a discrepancy during blade transition between air and water in the catch and release.

For recovery, superiority of the LIAC model over BIFC is not surprising because in the pilot tests, initially selected values of rope tension for LIAC (τ*min rec* = 50 N ) were also rated to be unrealistically high by three pilot experts. Nevertheless, we wanted to inspect and compare the overall ratings from main study expert rowers; thus, τ*min rec* forces in the BIFC model were maintained at 50 N . If we had lowered the τ*min rec* tension forces for the BIFC model, the observed significance regarding the realism of recovery forces of LIAC over BIFC would not have been found in our opinion.

#### 4.1.2. Auxiliary Aspects of Rowing Movement

During actual rowing outings, rowers usually may need to stop the boat. Thus, rowers immerse the squared blade into the water to apply a braking motion. In the simulator, expert rowers were also asked to simulate a similar scenario, in which they would immerse the squared blade into water to decelerate the boat (Question 2). In this aspect, the perceived horizontal braking forces were not perceived differently for BIFC and LIAC. Some rowers described that braking forces were higher than they expected, which also decelerated the simulated boat more compared to an actual boat. This observation may be explained by the fact that effects of the Froude number and the turbulence were omitted in both models. In real blade–water interaction, the immersed area of the oar blade in the water may reduce the drag and lift force coefficients due to a change of the Froude number [46]. In addition, the turbulence was also shown to reduce the drag and lift coefficients in a computational study [47]. Thus, both effects should be implemented to yield more realistic interaction forces.

Another simplification was the geometrical modeling of the oar blade. In both rowing models BIFC and LIAC, the oar blade was initially modeled as a flat rectangular plate, i.e., the longitudinal and lateral curvature of an original rowing blade [34] design was omitted. Based on the feedback from pilot study rowers regarding the vertical forces in the catch and release, ϕ*bias* was implemented in LIAC to account for a simplified form of a lateral curvature of the blade. However, due to the simplified blade geometry, the contact area between the blade and water surface increased. Thus, rendered horizontal and vertical forces above the water surface were generally perceived as higher for the simulator than a real rowing scenario.

In real rowing, due to rolling of the boat to one side, rowers may sometimes unintentionally touch and slide the blade on the water surface when moving the oar from release to catch with a feathered blade. In the simulator, participants were asked to make a deliberate attempt to slide the feathered blades to simulate this scenario (Question 1). Rowers reported that the BIFC model rendered horizontal forces that decelerated the boat when the blade was slid parallel to the water surface (feathered oar). According to the mathematical modeling of BIFC, the ideal forces along the *By*-axis should be zero. However, the lack of curved blade design in addition to the binary modeling of the oar immersion, i.e., either fully in or fully out of the water, might have facilitated the blade immersion during the feathered position, which may have induced the reported braking forces.

Although LIAC was rated more realistic than BIFC for the horizontal forces resulting from feathered blade sliding, the mean rating was only above average (μ*Q*<sup>1</sup> *LIAC* = 5.95). Braking forces were also reported for LIAC; however, the magnitude was perceived as lower in comparison to BIFC. The reason for the reported forces for LIAC was the introduction of ϕ*bias*, whose initial value was set to 2◦ to simulate a rotated blade instead of a (laterally) flat one. Thus, especially a rapid movement towards the catch while sliding the blade on the water surface resulted small resistive forces.

For both rowing models, the lack of curved blade design also resulted in excessive vertical contact forces. Naturally, the more parallel a blade is oriented with respect to surface, the greater the vertical forces become acting on the blade due to the water contact. In Question 3 and 4, experts were asked to rate the realism of the vertical forces resulting from water surface contact for a range of oar blade rotation (0◦ ≤ ϕ ≤ 90◦). For BIFC, experts judged that immersion or tapping with all ϕ orientations rendered higher vertical forces than real rowing. This observation may be attributed to the binary modeling of the oar immersion for the calculation of rowing forces. For LIAC, the squared blade insertion (ϕ = 0◦) into water was reported to render realistically low vertical forces. However, tapping with the feathered blade (ϕ = 90◦) rendered much higher vertical forces than the real rowing.

To encounter high vertical forces and ensure the safety of the users in the simulator, rendered vertical forces on the simulator were limited to 200 N in *Bz*-axis, which was mainly considered for the general rowing movement with the squared blade orientation. However, the selected force limit was perceived too high for rendering the vertical forces during feathered blade tapping onto the water surface. Thus, the vertical force limit should be lowered in the future implementation.

#### 4.1.3. Visual Scenario

Although the velocity model proposed in [28] was not further modified for the first part of the study, the resultant velocity of the simulated boat might have differed due to the design of distinct force models. In the simulator, the visual scenario shown on the screens was updated according to the boat speed ( . *yboat*) that was calculated through Equation (34). The result from Question 10 showed that the LIAC model was only marginally better than BIFC for the realism of the visual scenario flow. However, ratings for both models were only above medium. Discussions with the experts revealed that the design of the visual scenario itself could be improved for a more realistic virtual reality immersion, e.g., placement of the virtual buoys, rendering of the regatta lane, etc. Thus, although the goal was to judge how well the visual scenario updated with respect to the performed stroke rate and applied force, the majority of the experts were distracted by certain features in the visual scene such as the visual rendering of blade and buoys.

#### *4.2. E*ff*ect of Individualization on the Perceived Realism of a Rowing Simulator (LIAC vs. IND)*

In the second part of the study, experts were presented with a pool of model parameters, which can be individually configured to attain the most preferred rowing scenario on the simulator. The configurable parameters were chosen as *Gdl*, ϕ*bias*, τ*min rec* , and *CBdrag*.

Adjustment of *Gdl* was mainly proposed for the magnitude of the drag/lift forces in drive, catch and release; however, it also influenced the forces resulting from blade and water surface interaction. Seven out of ten rowers preferred a different value of *Gdl* (range: 0.32–0.40, Table 1) after testing the simulator with a variety of *Gdl* settings. The modulation of *Gdl* resulted in an analogous effect as the damper setting on an ergometer. Four out of five male rowers preferred a higher setting while three females settled on a lower setting than the initial value of 0.35.

The bias factor ϕ*bias* was originally introduced to modulate the vertical forces during catch and release. Change of ϕ*bias* affected the value of the projected blade area for both the horizontal and vertical plane forces. In Training-6, eight out of ten rowers did not prefer the previously introduced ϕ*bias* for rendering higher vertical forces in the catch and release (Table 1). ϕ*bias* was set back to 0◦ for these participants. Thus, while rowing with a squared blade, decreased value of ϕ*bias* resulted in lower vertical forces.

Configuration of τ*min rec* was offered for the resistive forces during recovery. Overall, nine experts decided to configure the rope tension forces (τ*min rec* ) to a lower value than the initial setting (initial: 35 N , range after configuration: 30–35 N, Table 1).

Finally, modification of *CBdrag* was proposed to alter the boat velocity and thus, the resultant drag forces. Originally, the parametrization of *CBdrag* was suggested during pilot test to account for different rowing boat types since the rowers may be more familiar with training on a certain type of boat. However, during the main study, some rowers criticized the decelerating drag forces acting on the blade at the water entry and exit. The decelerating drag forces were caused by the stream velocity (negative of calculated boat velocity), which acted in the +*By*-direction. Therefore, *CBdrag* was the only relevant configurable parameter available to us, which could influence the magnitude of decelerating drag forces through boat velocity. For example, increment of *CBdrag* resulted in a reduced boat velocity (Equation (34)), which reduced the braking forces in +*By*-direction at the water entry and exit.

#### 4.2.1. Rowing Cycle and Sub-Phases

The individualization of the model parameters significantly improved the perceived realism of forces in drive, recovery, catch, release, and overall rowing cycle.

For drive, individualization of *Gdl* was the key factor for the increased realism of resistive forces in the water. All participants except one were content with the magnitude of resistance, the peak force, and the change of forces in the water (μ*Q*<sup>5</sup> *IND* = 8.20). One participant reported that he felt constrained by the simulator when he tried to apply high propelling forces. This observation can be attributed to the fact that the rendering of the horizontal forces was also bounded with a safety limit of 450 N in the *By*-axis. For the experiment, rowers were instructed to row with 40–70% of maximum force application and within low to medium stroke rates to assess the general capability of the simulator. Since the participants were not explicitly informed about the safety limit, their expected force peak was not rendered for the rowers who applied higher propelling forces. Therefore, the simulator will be tested in future for a higher force rendering to account for higher propelling forces by the rowers.

For recovery, individualization of rope tension forces significantly improved the perceived realism of forces (μ*Q*<sup>6</sup> *IND* = 8.00). Although a setting of 30 N rendered reduced recovery forces, some rowers preferred to experience the slight resistance, referring to the real rowing in a windy weather. Few rowers actually preferred to set even lower values than 30 N for τ*min rec* stating that the residual resistance could still be perceived. However, due to design criteria of a tendon based parallel robot, the ropes must always be under tension to control the end-effector movement. Additionally, the selected value for τ*min rec* should ensure a smooth transition from τ*min rec* to τ*min drive* without causing disruptive tension–force jumps on the ropes. Thus, for control and safety reasons the lowest limit of τ*min rec* was set to 30 N . Nevertheless, the selected value was based on the initial trials on the simulator, and further experiments may yield a safe and less resistive τ*min rec* tension forces on the ropes.

Overall, individualization of ϕ*bias*, *Gdl*, and *CBdrag* led to significant improvement in the ratings of catch and release (μ*Q*<sup>7</sup> *IND* <sup>=</sup> 8.00, <sup>μ</sup>*Q*<sup>8</sup> *IND* = 7.80). A final setting of ϕ*bias* = 0◦ was mainly reported to have rendered more realistic vertical forces during blade transition. In addition, four rowers who preferred a higher value of *CBdrag* judged that the increment encountered the previously criticized high decelerating drag forces at the initial and final blade contact with the water. However, other rowers who preferred the same or lower *CBdrag* values notified us that lower braking forces would be preferred. If the rendering of high braking forces at the water entry and exit had been reported in the pilot study, implementation of Froude number and turbulence effects to reduce the drag/lift coefficients might have been a more convenient approach than adjusting *CBdrag*.

#### 4.2.2. Auxiliary Aspects of Rowing Movement

Overall, the individualization of the *Gdl*, ϕ*bias*, and *CBdrag* significantly improved the perceived realism of the horizontal forces resulting from blade contact with the water. However, due to previously presented blade modeling simplifications and safety limits, the increment on the rating of vertical contact forces were insignificant.

Setting ϕ*bias* back to 0◦, as preferred by the majority of experts, led to a significant increase in the realism perception of the horizontal forces when sliding the blade (Question 1). Although experts were more consistent with the rating values for the configured model IND, on average the realism was only above average (μ*Q*<sup>1</sup> *IND* = 6.60). This result may be explained with the lack of longitudinal curvature in the modeling of the blade surface. In a real rowing case, inserting a feathered blade into the water requires a deliberate attempt to push the oar handle upwards due to the curved design of the blade. However, due to the flat-plate modeling of the oar blade, rowers were still able to insert the feathered blade into the simulated water, which produced slight decelerating drag forces.

Significant improvement in the realism of the braking forces (in Question 2) may be attributed to the rowers who settled on a lower value for *Gdl* and/or *CBdrag* during individualization, which reduced the high braking forces encountered in the first part of the study. Nonetheless, the average rating for the realism of braking forces were only above average (μ*Q*<sup>2</sup> *IND* = 6.80). The result can be attributed to the rowers, who preferred higher values for the both drag coefficients. Thus, the previously reported issue regarding the excessive braking forces was repeated.

For the vertical forces acting on the blade due to contact with water surface, individualization of *Gdl* only resulted in a marginally significant increment (μ*Q*<sup>3</sup> *IND* = 6.30). The majority of the rowers, who preferred ϕ*bias* = 0◦, mentioned that the minor reduction of the vertical forces were more realistic compared to ϕ*bias* = 2◦. However, for the tapping with feathered blade onto the water surface, previously reported unrealistically high vertical force rendering was stated again in Question 3 and 4. Repetition of the issue was an expected outcome due to the lack of realistic blade modeling and used safety limit of 200 N . in *Bz*-axis, which were not among the configurable parameters.

#### 4.2.3. Visual Scenario

During model individualization in Training-6, rowers were instructed to focus on the configuration of the rowing forces. When the participants settled on their preferred rowing forces on the simulator, they were asked to evaluate the realism of the simulated boat speed from the visual scenario displayed on the screens.

Results showed that although a slight improvement was observed in the rating, force model configuration failed to significantly improve the perception regarding the flow of visual scenario (μ*Q*<sup>9</sup> *IND* = 6.40). Based on the rowers' feedback, the design of the visual scenario was the main factor that might have distracted the experts. Rowers especially stated that the virtual buoy placement in the visual scenario was perceived narrower than a normal regatta lane, which prevented them from comparing the simulator to the real life rowing. A few rowers also mentioned that they did not have adequate experience on a pair (2-) boat. These rowers reported that they were accustomed to rowing in bigger boats such as four (4-) and eight (8+), which are faster than a pair (2-) according to regatta results [48]. Thus, the resultant simulated boat velocity might have been slower for these rowers.

#### *4.3. Practical Implications*

To experience a realistic rowing training scenario in a simulator, expert rowers demand well-designed and synchronized rendering of visual, auditory, and haptic aspects of the task. In our simulator, forces rendered by the implemented rowing model and the synchronous rendering of visual and auditory displays were generally found to be realistic by the expert rowers. However, simplifications during modeling and technical limitations of the hardware setup might have led to perception of unnatural forces or unrealistic auditory and visual features, which may have confounded the overall realism of rowing scenario in the simulator. Therefore, in cases where residual technical problems cannot be readily resolved due to practical or safety reasons, individualization of the model parameters might be an appropriate approach to maintain or increase the perceived realism of the simulator.

Indeed, in our study, individualization of the model parameters was observed to yield a significant improvement in the acceptance of training on the simulator. Thus, expert rowers were also asked which other parameters of the rowing model and features to enhance the indoor training they would prefer to have on the simulator.

The majority of the experts proposed the parametrization of the inboard and outboard length of the oar, which could be virtually configured based on the rigging preferences of the rower. In real rowing, the preferred inboard and outboard oar lengths are usually found after a trial and testing process in the trainings. The required length of the oar also needs to be adjusted for the weather conditions. Thus, our rowing simulator can provide an easy and fast rigging within a training session.

In addition, expert rowers were interested in visualizing the errors related to the kinematic and dynamic outcomes of their rowing strokes in the simulator. Although commercial products, which can provide propulsion forces and oar kinematics while rowing on a boat, are available, these devices are not yet capable of providing feedback about the resulting errors, i.e., difference between desired and actual performance [49,50]. Therefore, as previously studied with beginner rowers, three modalities rendered on the rowing simulator could also be used to present feedback to the expert rowers for performance enhancement [26,51,52].

Some expert rowers were also interested in the visual display of an 'opponent boat', against which the rower in the simulator can train or race to simulate a regatta scenario. According to this idea, the 'opponent boat' could be set to propel with a desired speed setting, and the rower in the simulator could train to match the set goals for the training. Therefore, realistic rendering of the rowing forces can further be employed to augment the simulator to train professional athletes [53].

Finally, thanks to the achieved haptic realism of a sweep rowing application, potential sources of injuries associated with the asymmetries unique to the sweep rowing can be systematically assessed on our virtual reality simulator [54–56].

#### *4.4. Limitations*

Although rolling motion is important in real rowing for training the balance of the boat, the actuation of rolling angle on the rowing simulator was not included for safety reasons and controlling issues of the tendon based parallel robot. As reported by the participants, fixation of the boat for rolling motion helped to focus more on the relevant features and forces of rowing motion, although on the other hand, it lessened the overall haptic realism of the simulator compared to a training with an actual rowing boat on water. This technical drawback may be counteracted by introducing augmented reality headsets that can combine real life view of the rowing simulator and virtual rendering of boat rolling on the simulated water.

Another technical limitation was the use of fixed oarlock height for all the participants. The oarlock height was not adjustable due to the attachment of the end-effector (shortened oar) with the tendon based parallel robot. To operate the simulator, the oar and the boat had to be kept at a pre-defined position for the calibration of rope tension forces in the tendon based parallel robot. Since changing the end-effector position (by changing oarlock height) affected the calibration process of the robot, the oarlock was not adjusted. Therefore, some participants could not row with their most preferred rigging setting, which might also reflect the overall rating of the force models.

Due to the scope of the study, assessment of the realism of the rendered rowing forces were based on the subjective perception of the expert rowers. Therefore, future studies are required to objectively assess the realism by comparing the rendered rowing forces on the simulator to the on-water force measurements.

#### **5. Conclusions**

In this study, we presented a new rowing model that computed the blade–water interaction forces based on three rotational movements of the oar and implemented the model in our rowing simulator. We asked expert rowers to qualitatively assess the realism of the resulting forces for various rowing aspects on our rowing simulator. Based on the evaluation of experts, we found that the newly proposed 3D rowing model, which incorporated continuously adjusted drag/lift coefficients and immersed area of the blade, was highly realistic for the sub-phases and overall rowing cycle. We also showed that when the rowing model parameters were configured according to individual preferences of expert rowers, the perceived realism significantly increased.

The implementation of the proposed 3D rowing model for the haptics and the audiovisual rendering of the rowing scenario on the virtual reality simulator provide a realistic indoor training opportunity to rowers with different expertise levels. Thanks to the parametric design of the model and the real-time operation of the simulator, rowers and coaches can choose and immediately test the inquired settings to increase the rowing performance. The pool of configurable parameters can easily be extended to include the boat and oar rigging, which may further help rowers to find their most preferred rowing configuration in a boat. Therefore, our rowing simulator does not only provide an indoor rowing training facility but also a platform, which assists rowers and coaches to find the most optimal rowing settings.

**Author Contributions:** E.B. contributed to the rowing model development, vibration control of the parallel robot, experimental protocol design, questionnaire preparation, conducting of the experiments, data acquisition, and statistical analysis and prepared the manuscript. P.B. conducted his MSc. semester thesis, in which he contributed to the rowing model development and vibration control of the parallel robot. N.G. and G.R. contributed to the rowing model conceptualization, solution of technical problems encountered during model development, and they revised the manuscript. P.W. participated in the experimental protocol design and questionnaire preparation, contributed to discussion of the statistical analysis results, and revised the manuscript. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Acknowledgments:** We would like to thank the rowers who took part in testing the models during the development phase and pilot study. We also express our gratitude to Michael Herold-Nadig for his technical support and safety testing of the rowing simulator.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **References**


© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

#### *Article*
