How to Train Your Posture: Haptic Feedback Can be Used for Postural Adaptation of the Trunk During Upper-Limb Motor Training

Poor trunk posture, especially during long periods of sitting, could lead to problems such as Low Back Pain (LBP) and Forward Head Posture (FHP). Typical solutions are based on visual or vibration-based feedback. However, these systems could lead to feedback being ignored by the user and phantom vibration syndrome, respectively. In this study, we propose using haptic feedback for postural adaptation. In this two-part study, twenty-four healthy participants (age 25.87 ± 2.17 years) adapted to three different postural targets in the anterior direction while performing a unimanual reaching task using a robotic device. Results suggest a strong adaptation to the desired postural targets. Mean anterior trunk bending after the intervention is significantly different compared to baseline measurements for all postural targets. Additional analysis of movement straightness and smoothness indicates an absence of any negative interference of posture-based feedback on the performance of reaching movement. Taken together, these results suggest that haptic feedback-based systems could be used for postural adaptation applications. Also, this type of postural adaptation system can be used during the rehabilitation of stroke patients to reduce trunk compensation in lieu of typical physical constraint-based methods.

Rakhi Agarwal is with the School of Mechanical and Aerospace Engineering, Nanyang Technological University, Singapore 639798, and also with the Department of Applied Mechanics, Indian Institute of Technology, Madras, Tamil Nadu 600036, India (e-mail: am18d004@smail.iitm.ac.in).
Asif Hussain  especially in this digital era [1]. Poor spinal posture has been associated with several musculoskeletal disorders and spinal deformities. Poor sitting posture has been linked to increased cases of Low Back Pain (LBP), Forward Head Posture (FHP), neck pain, and Upper Crossed Syndrome (UCS) [2], [3]. LBP is, in fact, one of the most prevalent musculoskeletal conditions, and treating it places a significant strain on healthcare systems. Therefore, it is of considerable interest to monitor and analyze the sitting posture, especially for people with desk jobs that include long periods of sitting, typically in front of computers. Most existing postural correction systems are based on visual or vibration feedback [4]. However, when used for long durations, these feedback systems have some disadvantages. Visual feedback systems can be easy to ignore, and visual feedback in the form of pop-ups can become irritating for the user [5]. Vibrationbased feedback systems suffer from some disadvantages too. They can lead to phantom vibration syndrome [6]. Moreover, constant vibrations can lead to adaptation to the vibration, with a reduction in sensitivity, thereby, reducing the user's capabilities in perceiving the vibration. Also, postural correction systems based on Kinesio taping have been evaluated for the treatment of Low Back Pain [7]. However, some studies have reported adverse effects of Kinesio taping, such as allergy to the tape and fatigue [8]. Therefore, it may be beneficial to explore other forms of feedback. There is a need for a postural correction system that can help adapt to a particular trunk posture while the user can simultaneously continue working, ideally without any visible decline in the work-performance.
Moreover, this kind of feedback system for postural adaptation could help during the rehabilitation of stroke patients. One of the major issues that arise during stroke rehabilitation is trunk compensation. During recovery, patients compensate for the impaired arm function by moving their trunk for practical gains [9]. However, this should not be confused with motor recovery. Practicing the movement without assessing and minimizing the compensatory strategies could lead to short-term 'functional' gains. However, it is associated with increased long-term problems such as pain, decreased range of motion, and learned non-use [10].
Therapists often try to decrease the compensatory trunk movement during reaching and grasping training. A widely used method for reducing trunk compensation is using a trunk restraint [11]. In this method, patients are physically strapped to the chair during upper-limb rehabilitation to minimize anterior trunk bending. However, this kind of restraint-based rehabilitation has many disadvantages. Firstly, it requires constant supervision of the patients. Hence, rehabilitation strategies requiring minimal therapist supervision are needed to reduce trunk compensation. Secondly, research has shown that manual guidance through a physical constraint could act as 'concurrent feedback,' and the feedback could become part of the learned motor task, thus enabling the patient to get dependent on the provided feedback [12]. Therefore, it is better to adopt a rehabilitation strategy that can control the amount and duration of feedback. A feedback system with a provision to easily 'switch on and off' as per requirement is desirable. Haptics can be used to reduce trunk compensation during upper-limb rehabilitation instead of giving feedback in the form of explicit information through visual cues, verbal instruction, or physical constraints [13].
There are many methods of providing haptic feedback in the literature. Various compact devices are available that can provide tactile haptic feedback [14], [15]. Such devices are broadly used for improving the haptic perception of virtual objects. Neuromuscular Electrical Stimulation (NMES) based methods are also described in the literature for rendering haptic force feedback [16], [17], [18]. These devices could have interesting applications as wearable haptic solutions. However, these systems are still in the embryonic stages and require extensive research to overcome the issues such as triggering of the skin receptors, and trade-off between the desired high-intensity stimulation and user comfort. Many robotic devices are also available that can provide the desired haptic feedback [19], [20], [21]. While all these methods can be evaluated for postural correction applications, this study focuses on using a robotic device for providing the required haptic feedback.
The overall purpose of this study is to examine whether providing haptic feedback about trunk orientation while performing a unimanual reaching task can lead to postural adaptation in neurologically intact participants. This study evaluated the adaptation to three arbitrary non-upright trunk postures in the anterior direction. Haptic feedback is provided in the form of damping at the end-effector such that if incorrect posture is detected, more force is required to move the handle. Postural adaptation is evaluated by analyzing whether participants retain the posture in the absence of haptic feedback.
The study is divided into 'Single postural target study' and 'Multiple postural targets study.' In 'Single postural target study', postural adaptation to a single postural target angle has been evaluated. As the participants move closer to the desired postural target, the damping of the end-effector is decreased exponentially. The purpose of this part of the study is to evaluate whether haptic feedback in the form of exponentially decaying damping can induce postural changes in healthy people. We hypothesized that participants would move to the desired postural target to minimize the forces at the end effector (Hypothesis 1). In 'Multiple postural target study', postural adaptation to two different target angles has been evaluated. In this part of the study, as participants moved closer to the desired target, damping at the end-effector was decreased in a sigmoidal manner. The purpose of this part of the study is to evaluate whether postural adaptation to different postures can be achieved by changing the parameters of the provided haptic feedback and whether sigmoidal decay of damping leads to better postural adaptation as compared to exponential decay. We hypothesized that during exponential decay, change in damping at low postural error might not be sufficient to be perceived by the participants. Hence, sigmoid decay might lead to better postural adaptation (Hypothesis 2). Specific haptic feedback during the two parts of the study has been discussed in detail in the methodology section.
This kind of feedback system can help improve people's sitting posture with job requirements of long sitting hours in front of a computer device. Moreover, this could be further explored to reduce trunk compensation in stroke patients. Regular unrestricted reaching movements made by healthy people are usually upright; adapting to a non-upright posture during reaching tasks can be considered unnatural. If haptics can induce postural changes to unnatural postures, it could prove helpful in stroke rehabilitation to reduce trunk compensation since maintaining an upright posture can be considered unnatural for these patients.

A. Participants
A total of twenty-four healthy participants (13 male and 11 female) were recruited for this two-part study. Out of these twenty-four people, ten participants took part in the first study, and fourteen participants took part in the second study. None of the participants had a history of neuromuscular or musculoskeletal injury in the upper extremities. All participants had normal or corrected to normal vision. The mean (SD) age, height, and weight of participants were 25.87 (2.17) years, 168.08 (6.57) cm, and 64.15 (10.65) kg, respectively. Edinburgh Handedness Inventory was used to assess the handedness of the participants [22]. Twenty-two participants were right-handed, while the other two were left-handed. Participants performed the study task with their dominant hand.
All participants provided written informed consent before participating in the study. The study was approved by the Institutional Review Board of Nanyang Technological University, Singapore (IRB-2018-07-043).

B. Experiment Setup
The experiment task consisted of making upper limb reaching movements to the visual targets shown on the screen using a manipulandum. While making reaching movements, the trunk posture of the participants was monitored, and haptic feedback based on the posture was provided.
The participant was seated on a non-movable, non-revolving, height-adjustable chair in an upright position in front of a 24-inch computer screen (Dell Inspiron 24 All in One Desktop 11th Generation Core i5 -1135G7 Processor), which was used to display the experimental task, and a robotic manipulandum called H-Man was used to perform the reaching movements (as shown in Fig. 1). H-Man is a commercial device which is manufactured and distributed by ArtiCares Pte Ltd, Singapore. It is a backdrivable, two-degree-of-freedom robotic manipulandum used in motor control studies and upper limb neurorehabilitation [23], [24], [25]. Planar force fields can be programmed in H-Man based on the kinematics of the end-effector. The position and velocity data of the reaching movement were obtained from the H-Man at 1000 Hz.
The orientation of the trunk of the participant was measured using two Inertial Measurement Units (IMUs). Nine-axis IMUs were used for the study (MPU-9250, Invensense, CA, USA). The trunk orientation data were obtained from IMU at 100 Hz. Real-time feedback related to trunk posture was provided to the participants during training. Force on the end-effector was changed in the form of damping. Damping data was sent from the PC to H-Man via Transmission Control Protocol (TCP) with a delay of less than 200 ms. One IMU was attached to the participants' chest at the sternum (moving-IMU) using a custom-made strap (as shown in Fig. 1). The second IMU (not visible in Fig. 1) was fixed on the H-Man (fixed-IMU). Both IMUs were initially calibrated for hard iron distortions by rotating the sensor along each axis. The average of minimum and maximum magnetometer readings was recorded along each axis after rotation to get each axis's hard-iron distortion correction value. Also, the gyroscope and accelerometer noise were determined from the datasheet of the sensor. A nine-axis Kalman filter was used for sensor fusion and orientation estimation [26]. The orientation of the moving-IMU was measured with respect to the fixed-IMU to get the trunk posture.

C. Experiment Protocol
The study was divided into a "Single postural target study" and a "Multiple postural target study". In both studies, participants were asked to grasp and move the H-Man handle (end-effector of manipulandum) to perform reaching movements. The movement of the H-Man handle in the left-right or front-back direction was replicated on a computer screen by moving the cursor in the leftright or up-down direction, respectively. A blue circular cursor always showed the position of the handle on the PC screen.
Haptic feedback was provided as a resistive, velocitydependent force experienced on the end-effector in the form of linear damping. The force required to move the handle of the H-Man, when haptic feedback is enabled, is computed using the velocity of the end-effector and the damping value: where F b represents the planar force required to move the handle, b represents the damping value, and v is the velocity of the movement of the H-Man handle. Damping was determined based on the participant's posture, as explained in the following section.
The instructions provided to the participants regarding haptic feedback were similar in both studies. In both studies, participants were instructed that after a few reaching movements, 'starting training' would be displayed on the screen. During training, they would experience some resistance in moving the handle, depending on their upper-body posture. Participants were instructed that they could move their upper body to different postures to decrease the resistance and make the movement easier. Participants were not explicitly informed about the magnitude or direction of the desired posture and were required to find this information themselves through exploration. Also, the participants were advised that they could ask for breaks if needed. Since healthy people were recruited for this study, reaching movements performed by the healthy people in such a single-session study are in an almost upright posture. Therefore, to evaluate the hypothesis that haptic feedback can be used for postural adaptation to a particular trunk posture, adaptation to different non-upright postures was evaluated. The specific task for individual studies is described in detail in their respective sections.

1) Study 1 (Single Postural Target Study):
The objective of the first study was to evaluate if haptics-based feedback can lead to postural changes in healthy people. Ten participants took part in this study. Each participant was asked to perform a reaching movement using H-Man to move the cursor on the computer screen. The task was to move the cursor from the 'Visual target A', also called the home position, to the 'Visual target B', also called the end position shown on the screen in the form of black circles (as shown in Fig. 2(a)). The movement of the cursor from the home position to the end position constitutes a single trial. After each trial, the participant was required to bring the cursor back to the home position before starting the subsequent trial. The study consisted of three phases: The Baseline phase, the Training phase, and the Test phase. The Baseline and Test phase consisted of 20 trials each, and the Training phase consisted of 100 trials. Therefore, the entire study consisted of 140 trials.
Haptic feedback was disabled during the Baseline and Test phases, meaning no additional force was required to move the end-effector during these phases. During the Training phase, a virtual postural target was set at 20 o towards the anterior direction from the initial upright position. Participants were not explicitly informed about the specific postural target. Real-time feedback about the deviation of trunk posture from the desired postural target was provided in the form of haptic feedback. Haptic feedback was enabled during each trial of the Training Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. phase, which means that the resistance to move the end-effector depended on the participants' posture. The damping was maximum at the initial (i.e., upright) position. As the participant moves towards the postural target, the damping coefficient 'b' decreases exponentially: where b m represents the maximum damping value, p e represents the postural error, and p t represents the postural target. Here postural target, p t , is defined as 20 o ; postural error, p e , is the angle between the current trunk orientation and the postural target (as shown in Fig. 2(b)). To ensure an upper limit on the maximum force experienced by the participant, maximum damping, b m , is defined. The change in calculated damping values on changing the postural error for this task is shown in Fig. 2(c). Maximum damping, b m , is set as 120 Ns/m. Therefore, using (1), the range of force, F, required to move the handle will be: where v is the velocity at which the participant moves the handle.

2) Study 2 (Multiple Postural Targets Study):
Fourteen participants took part in the second part of the study. Participants recruited in the second part of the study were different than the first part of the study, and therefore, were naïve to the protocol. The task consisted of moving the cursor from the 'Visual target A' to the 'Visual target B' (as shown in Fig. 2(a)), which constituted a trial, and then back to the 'Visual target A', which formed the subsequent trial. In this part of the study, postural adaptation was trained and tested for two virtual postural targets at angles 15 o and 30 o in the anterior direction from the initial upright posture, and hence there were two Training and Test phases. Therefore, the study consisted of 5 phases: Baseline phase, During the Baseline and both Test phases, haptic feedback was disabled. During both Training phases, haptic feedback was enabled based on the sigmoid function, with maximum damping at the initial (upright) posture and decreasing value of damping coefficient 'b' as the participant moves towards the correct postural target: where b m represents the maximum damping value; postural target, p t , is defined as 15 o for one of the Training and Test phases and 30 o for another Training and Test phases; postural error, p e , is defined as the angle between the current trunk orientation and the postural target. Change in the calculated damping value on changing the postural error for this study is shown in Fig. 2(d). Maximum damping, b m , is set as 200 Ns/m. Therefore, using (1), the range of force, F, required to move the handle will be: where v is the velocity at which the participant moves the handle.

D. Data Analysis
The primary variable of interest was anterior trunk bending which represented the angular deviation of the trunk from the upright posture in the anterior direction. Anterior and posterior bending of the trunk in the sagittal plane was considered positive and negative, respectively, and 0 degrees was deemed vertical to the ground, i.e., upright posture. Anterior trunk bending was averaged across participants to compute mean anterior trunk bending, and it was plotted during different phases of training for visualization. Anterior trunk bending before and after the intervention was compared to assess if the haptic feedback during training resulted in the adaptation to the desired postural target. Secondary variables related to the reaching movement were computed to check the impact of postural training on the performance of the reaching movement. The raw kinematic data (position and velocity) obtained from H-Man were filtered using a low pass filter (Butterworth: 6 th order; cut-off frequency, F c : 20 Hz; sampling rate, F s : 1000 Hz). The filtered data was used in offline data processing to evaluate reaching task performance. The velocity of movement of the end-effector and straightness of reaching movement were analyzed. Straightness of movement was evaluated by calculating the path-to-length ratio, i.e., the ratio of the total length of path traversed by the end-effector and the shortest straight-line distance between the start and endpoint for each trial [27]. Therefore, the straightness value can be greater than or equal to 1, where a value of 1 indicates a perfect straight-line movement. Also, some additional parameters were analyzed (discussed in the Appendix).
The two studies (single postural target study and multiple postural target study) differed in the postural target and task conditions. The percentage of adaptation was computed for each postural target in both studies to assess the outcome of the change in task conditions on the level of adaptation. The percentage of adaptation was calculated: where A represents the percentage adaptation; p t represents the postural target, defined at the beginning of the study; and p e represents the postural error, defined as the angle between the current trunk orientation and the postural target during the Test phase. As anterior trunk bending varies from the upright posture to the target posture, the percentage of adaptation, A, can vary between 0-100%, where a value of 100% represents a complete adaptation to the defined postural target.

E. Statistics
For both studies, a priori power analysis was performed before the commencement of the main study using G-Power tool. Pilot data were collected from four participants for Study 1 (Single postural target study). Power analysis was performed using the pilot data, which yielded a power of 96% for N = 6 participants. Then pilot data were collected from four participants for Study 2 (Multiple postural targets study). Power analysis was performed using pilot data, which yielded a power of 96% for N = 10 participants. Results of power analysis were used to determine the lower limit of sample size for each study.
Statistical analysis was performed using SPSS (IBM SPSS Statistics, Version 28.0). In both studies, Friedman ANOVA was performed on the anterior trunk bending from the upright posture to investigate the overall differences between Baseline, Training, and Test phases. For pairwise post-hoc analysis to compute the difference between phases, the Wilcoxon-signed rank test was used. TOST ('Two one-sided tests') equivalence testing was done to evaluate if there is an order effect of different postural targets (15 o   in both studies. A 0.05 significance level (α) was used for all statistical tests, and Bonferroni correction was applied in the case of multiple comparisons.

A. Adaptation to Single Postural Target
The anterior trunk bending of each participant during each phase is shown in Fig. 3(a). For this study, the virtual postural target was set as 20 o in the anterior direction, and accordingly, haptic feedback was enabled for the Training phase. Fig. 3(a) shows that anterior trunk bending increased during the Training phase for most participants when haptic feedback was enabled, except for participant number 7 (P7). The mean anterior bending of the trunk of all the participants at the end of the reaching movements is shown in Fig. 4 for the Single postural target study.
We found that the mean anterior bending of the trunk increased on the application of haptic-based feedback. The trunk posture was significantly different between the Baseline, Training, and Test phases of the single postural target study (Friedman ANOVA: p = 0.007, W = 0.490, where effect size estimate W is Kendall's W value ranging from 0 to 1). The participants moved the trunk towards the predefined postural target when haptic feedback was enabled during the Training phase. This was supported by post hoc analysis using the Wilcoxon signed-rank test with Bonferroni correction (Bonferroni adjusted α = 0.017). We found a significant difference between Baseline and Training phases (p = 0.004, r = 0.854, where r is effect size ranging from 0 to 1), and a significant difference between Baseline and Test phases (p = 0.004, r = 0.854). The change in posture was retained after the removal of haptic feedback during the Test phase. Post-hoc analysis revealed no significant difference between the Training and Test phases (p = 0.185, r = 0.42). Fig. 3(b) shows the distribution of mean trunk orientation during different phases for all participants for the single postural target study. The average value of all trials during a phase for each participant is plotted to form the box plot in Fig. 3(b). Therefore, haptic-based feedback resulted in a change in trunk posture from the initial upright posture.  Anterior trunk bending increased during both training phases for most participants, and posture is maintained during both Test phases. (b) Box plot for anterior trunk bending from the upright posture for multiple postural targets study: The horizontal bar, small square box, and small triangles represent the median, mean, and outliers, respectively. The double and triple asterisks signify the p-value < 0.01 and p-value < 0.001, respectively. Trunk bending during both Training and Test phases is significantly greater than in the Baseline phase.

B. Adaptation to Multiple Postural Targets
anterior trunk bending increased during the Training phase for both targets for all the participants. Mean anterior bending of the trunk of all the participants at the end of the reaching movements is shown in Fig. 6 for multiple postural targets study. We found that the mean anterior trunk bending increased on the application of haptic feedback. The trunk posture was significantly different for different phases of the multiple postural target study (Friedman ANOVA: p < 0.0001, W = 0.754). The participants moved their upper body forward when haptic feedback was applied for both predefined postural targets. This was supported by the post hoc analysis using Wilcoxon signed-rank test with Bonferroni correction (Bonferroni adjusted α = 0.00625). We found a significant difference between the Baseline phase and Training Anterior trunk bending for the 30 o postural target was greater than the 15 o postural target. This was supported by post hoc analysis which revealed a significant difference between Training phases for both targets (p < 0.0005, r = 0.872) and between Test phases of both targets (p < 0.0005, r = 0.872). Varying the haptic feedback parameters resulted in postural adaptation to different postural targets. Fig. 5(b) shows the distribution of trunk orientation for all participants for the multiple postural target study. The average value of all trials during a phase for each participant is plotted to form the box plot in Fig. 5(b).
In this study, the order of Training of postural target was varied between participants (15 o target followed by 30 o target, or 30 o target followed by 15 o target). Both these conditions resulted in equivalent adaptation (TOST equivalence test: p < 0.0001).
There is no effect of order in which different postural targets are trained using haptic feedback.

C. Percentage of Adaptation for Different Postural Targets
The percentage of adaptation was analyzed during the Test phase for all postural targets (15 o , 20 o , and 30 o targets). The  Fig. 7 shows the distribution of the percentage of adaptation during the Test phase for different postural targets. We found that the percentage of adaptation was distinct for different postural targets (Kruskal-Wallis ANOVA: p = 0.0001, r = 0.69). Task parameters varied between the two studies. We found that the percentage of adaptation for the multiple postural target study Moreover, for the single postural target study, there was no significant difference between the velocity at the Baseline phase and the Test phase (p = 0.48, r = 0.23). Similarly, for the multiple postural target study, there was no significant difference between the velocity at the Baseline phase and velocity at the last Test phase (p = 0.17, r = 0.37). Therefore, there was no overall change in velocity for both studies.
Furthermore, in the multiple postural target study, participants were trained in random order for two different postural targets. Therefore, we compared velocity at the end of the first Training phase with the velocity at the beginning of the second Training phase for both orders. We found that velocity at the beginning of the second Training phase decreased as compared to the velocity at the end of the first Training phase for both orders (for the first training target of 15°followed by the second training target of 30 o : p = 0.046, r = 0.73; for first training target of 30°followed by second training target of 15 o : p = 0.022, r = 0.86). Therefore, the increase in velocity was not continuous across the entire study. Velocity increased at the end of the first Training phase, then again decreased at the beginning of the subsequent Training phase before increasing again at the end.

E. Straightness of Reaching Movement
We found that there was no change in the straightness of reaching movement before and after the haptic intervention for both the single postural target study (Friedman ANOVA: p = 0.206, W = 0.4) as well as the multiple postural target study (Friedman ANOVA: p = 0.396, W = 0.07). Therefore, applying haptic feedback related to trunk posture did not significantly change the straightness of reaching movement for any postural target (15 o , 20 o , and 30 o targets).

IV. DISCUSSION
The main objective of this study was to evaluate the utility of haptic feedback to promote adaptation to a particular trunk posture during a unimanual reaching task. The results show that this kind of intervention can be used for postural adaptation applications. The force feedback about the trunk posture at the robot end-effector was used to guide the participant to the desired pose.
From the plot of the Baseline phase in Fig. 4, it is evident that the upper body has some anterior bending movement in the absence of any feedback. This non-zero trunk bending can be attributed to the trunk being recruited as part of the kinematic chain for reaching and grasping with the upper limb, even in healthy people [28]. However, as shown in Fig. 4, the mean anterior trunk bending during the Training phase is significantly larger than the Baseline phase, suggesting that haptic feedback can lead to postural changes. This posture is maintained even after switching off the haptic feedback during the Test phase.
It may be argued that the anterior bending of the trunk in the single postural target study may be a generic response to counter the force experienced at the end-effector, irrespective of the angle of the defined postural target. We conducted another study with two different postural targets to ensure that the postural changes are a direct result of the provided haptic feedback and to determine if postural adaptation to different angles can be induced by changing the feedback parameters. In the Multiple Postural Target study, participants adapted to both postural targets. Mean anterior trunk bending after the haptic intervention is greater than the Baseline value. Fig. 6 shows the change in mean anterior trunk bending in the Multiple postural target study. Therefore, participants adapted to different angles by changing the feedback parameters according to different postural targets, and postural adaptation using haptics-based feedback is not limited to any single postural target angle. Participants maintained the posture during the Test phase after turning off the feedback.
Moreover, based on the difference between trunk bending at Baseline and Training phases in our first study, we noticed that though there is a postural adaptation, the posture at the end of the training was not the exact postural target selected for the study (as seen in Fig. 4). At the final pose, the trunk bending angle was less than the chosen postural target. This implies that the haptic feedback guided the participants towards the desired postural target, only to some degree, and no strong adaptation was obtained. We speculated that this could be due to many factors explained below.
The haptic feedback in this study is in the form of damping of the robot end-effector. In the single postural target study, damping decreased exponentially as the participant's trunk moved towards the desired postural target in the first study. Since participants were not explicitly instructed on even the direction of the desired postural target, exponential decay was chosen so that they could recognize the feedback as soon as they started moving towards the target during exploration at the beginning of the training. This kind of exponential curve was chosen in the hope that even a slight movement in the correct direction would result in haptic feedback above the 'Just Noticeable Difference (JND)' threshold. JND is defined as the minimum difference required between the magnitude of two successive stimuli for the difference to be reliably perceived as different stimuli by the participants [29]. However, as seen in Fig. 2(c), following a steep decay in damping, as the postural error decreases, there is minimal change in the damping. Therefore, it may be the case that feedback was not perceived by the participant due to the low value of damping and minimal change in damping at lower postural error. In the second study (multiple postural target study), a sigmoid decay of damping was chosen instead of an exponentially decaying damping ( Fig. 2(d)). Sigmoid decay ensured a steep decline at high postural error values to guide the participants at the beginning of training towards the correct direction of the postural target and sufficient damping at medium-to-low postural error values to help the participants "fine-tune" their posture to the desired postural target. Moreover, during the Training phase in the first study (single postural target study), haptic feedback was disabled after every trial, during which the participant brought the cursor back to the initial position before starting the subsequent trial. This kind of discrete nature of feedback could have led to "resetting" of trunk posture after every trial, disabling the continuous training of participants' posture in this kind of short-term study. The experiment task was slightly modified in the second study such that moving the cursor from the 'Visual target A' (initial position) to the 'Visual target B' constituted one trial and bringing the cursor back to the 'Visual target A' represented the ensuing trial, and feedback was always enabled during the training. Therefore, the participant was provided continuous feedback during the Training phase.
These changes in the second study resulted in better adaptation to the desired postural targets. Percentage adaptation in the Multiple Postural Target study (15 o and 30 o targets) was higher than in the Single Postural Target study (20 o target), indicating that the task parameter modifications engaged in the second study resulted in a better adaptation. Other results from previous motor learning studies could be explored for postural adaptation application in the future to improve the percentage of adaptation, such as continuous task, fading feedback frequency [30], variable practice schedules with contextual interference [31], combining haptic feedback with other feedbacks such as visual or auditory cues [32], reinforcement feedback [33].
In both studies, the primary task is to make reaching movements to the 'Visual targets' shown on the screen using the upper limb, and the secondary task is to adapt to a particular posture. During training, feedback about the secondary task is provided as a force on the robot's end-effector controlled by the upper limb, which performs the primary task. So, the task is essentially a trade-off between applying extra force by the upper limb (to counter the damping force on the end-effector) or moving to an unnatural and uncomfortable posture (forward bending of the upper body). Based on the changes in posture before and after the intervention, it can be said that people moved to an uncomfortable posture rather than applying extra force on the end-effector. This kind of intervention could prove helpful for stroke survivors during rehabilitation to reduce trunk compensation since maintaining an upright posture is an uncomfortable posture for these patients. This trade-off could be the reason for incomplete adaptation to the correct target posture too. It could be the case that the force was not challenging enough at low postural error values. Perception of this kind of intervention depends on the upper body physical strength of every individual as well. For some people with relatively more physical strength in their upper body, force may not be challenging enough to motivate them to adapt to this unnatural posture. This could also be the reason for the absence of adaptation for some participants (P7 in Study 1 as seen in Fig. 3(a)) and minimal change in posture after intervention for some participants (P1 and P4 in Study 1 as seen in Fig. 3(a), P7 and P12 in Study 2 as seen in Fig. 5(a)). It remains to be seen whether adaptation to any postural target will increase if the maximum damping force is customized based on an individual's strength.
In this study, since the feedback was not related to the primary task, there was a possibility that the feedback could negatively interact with the primary task, and the performance at the primary task of reaching movement could decline during training, despite the improvements in the secondary task of postural adaptation. Previous studies have shown that dual tasks can interfere with each other due to many reasons, such as concurrent tasks requiring activation of overlapping areas of the cerebral cortex [34], [35], and an increase in working memory demands while performing two tasks simultaneously [36]. In contrast, other studies have shown the benefits of dual-task motor practice in motor learning [37], [38]. In this study, there was no significant change in the velocity of reaching movements at the beginning and the end of the studies. Moreover, there was no significant difference in the straightness of the reaching movements before and after the intervention, implying that the performance at the secondary task improved while there was no visible decline in the performance at the primary task. This kind of feedback strategy could help monitor and improve the posture of people prone to undesirable or non-upright postures while sitting for long durations. No negative interference with the primary task suggests that this kind of feedback can be used for postural correction without affecting work performance. Further longterm studies can be done to test the application of haptic-based feedback in such a population. No significant difference also means that there was no improvement in the performance at the primary task. However, this could be attributed to the fact that since this study was conducted on healthy participants, the primary task of simple reaching movements was not challenging enough, and there was no room for any significant improvement in the primary task performance. It remains to be seen how the primary and secondary tasks interact for the people with subpar upper-limb motor performance, such as older adults and patients with neuromuscular disorders.
We noticed that the velocity of reaching movements after training was significantly higher than the velocity at the beginning of the training for all postural targets. This increase in velocity could be due to continuous practice of the reaching movements during training [39]. However, there was no overall improvement in velocity at the end of the study as compared to the beginning of the study for both single and multiple postural target study. Moreover, in the case of the multiple postural target study, we found that the velocity of reaching movement at the beginning of the second Training phase was significantly less than the velocity at the end of the previous Training phase. Therefore, an increase in velocity at the end of Training phases for all postural targets could be because of the fact that initially, there is high damping at the end-effector when the haptic feedback is enabled due to higher postural error at the beginning of training. Therefore, more effort is required to move the handle to counter the damping, resulting in slower movements. As people adapt to the postural targets, damping at the end-effector decreases, resulting in faster-reaching movements.
During both studies, participants were only instructed that if it gets difficult to move the H-Man handle, they can adjust their posture to make the task easier. With such minimal instructions, participants were able to move towards the correct posture and adapt to it. No visual feedback was provided about their posture either. This kind of haptic feedback-based postural adaptation could play a crucial role in stroke rehabilitation applications where such implicit adaptation is needed instead of explicitly strapping the patients. A limitation of this kind of haptic feedback based solution for postural correction is that it requires an active device capable of providing the required haptic feedback. Future studies can explore inexpensive and portable options for providing haptic feedback to ensure widespread use of the proposed solution in working environments.

V. CONCLUSION
Our study shows that the proposed solution for posture adaptation is promising. Participants managed to adapt to the correct posture with minimal instruction. This kind of posture adaptation in healthy participants can be explored further for different applications. Future studies can explore various immersive tasks, including 3-dimensional tasks, instead of a simple discrete reaching task. Also, force feedback parameters could be varied between participants depending on the individual participant's strength. Further, it remains to be seen how valuable this kind of intervention is in the case of neurorehabilitation, for example, in stroke rehabilitation.

A. The smoothness of Reaching Movements
The smoothness of reaching movements was evaluated using Spectral Arc Length (SPARC) (defined in [40]). We did not find any change in the smoothness of reaching movements before and after the intervention in both the single postural target study and the multiple postural target study. Friedman ANOVA revealed no significant difference before and after the haptic feedback intervention for both single postural target (p = 0.206, W = 0.41) and multiple postural target (p = 0.168, W = 0.128) studies. Therefore, haptic feedback related to trunk posture did not improve or worsen the smoothness of reaching movement in this study.

B. Force at the Beginning and end of the Training
Range of force required to move the handle were as follows: 3.48N ≤ F 20 ≤ 9.52N , 1.92N ≤ F 15 ≤ 12.38N , and 3.72N ≤ F 30 ≤ 11.68N ; where F 20 represents the force required to move the handle for 20 o postural target, F 15 represents the force required to move the handle for 15 o postural target, and F 30 represents the force required to move the handle for 30 o postural target. We found that the force required to move the end-effector decreased at the end of training compared to the beginning of training for all postural targets (as shown in Fig. 9). This was supported by Wilcoxon signed-rank test (for 20 o target: p = 0.014, r = 0.77; for 15 o target: p = 0.002, r = 0.84; and for 30 o target: p = 0.012, r = 0.67). Force at the end effector is dependent on both damping and velocity value.
Anterior trunk bending increased as the participants moved towards the postural targets during training (as shown in Figs. 4 and 6), and therefore, damping decreased accordingly (as shown in Fig. 2(c) and Fig. 2(d)). Moreover, velocity increased towards the end of the training phase for all postural targets (as shown in Fig. 8). Therefore, an overall decline in force at the end-effector indicates that people optimize trunk posture to minimize the force experienced at the end-effector.

C. Velocity of Reaching Movement for Each Participant
The velocity of reaching movement for all participants at the beginning and end of the training phase for different postural targets is shown in Fig. 10. Fig. 10(a) shows the velocity of reaching movement at the beginning and end of the training phase of the 20 o postural target. For most participants, velocity increased at the end of the training phase as compared to the beginning of the training phase (except participant 3 (P3)). Similarly, for 15 o postural target (as shown in Fig. 10(b), velocity at the end of training phase as compared to the beginning (except for participant 5 (P5)). Moreover, for 30 o postural target (Fig. 10(c)), similar trend in velocity was observed and velocity increased at the end of the training phase for all participants (exceptions are P2, P4, P13). Therefore, overall trend was maintained by most of the participants. Increase in velocity can be attributed to the fact that at the beginning of the training phase, there was no adaptation to the desired postural target. Therefore, high postural error resulted in high damping value. Hence, more force was required to move the end-effector. As the participants started adapting to the desired postural targets towards the end of the training phase, postural error decreased resulting in a decrease in the damping value. Therefore, less force was required to move the handle, which could be a reason for the increase in velocity at the end of the training phase.