Smart Textiles that Teach: Fabric-Based Haptic Device Improves the Rate of Motor Learning

People learn motor activities best when they are conscious of their errors and make a concerted effort to correct them. While haptic interfaces can facilitate motor training, existing interfaces are often bulky and do not always ensure post-training skill retention. Here, we describe a programmable haptic sleeve composed of textile-based electroadhesive clutches for skill acquisition and retention. We show its functionality in a motor learning study where users control a drone's movement using elbow joint rotation. Haptic feedback is used to restrain elbow motion and make users aware of their errors. This helps users consciously learn to avoid errors from occurring. While all subjects exhibited similar performance during the baseline phase of motor learning, those subjects who received haptic feedback from the haptic sleeve committed 23.5% fewer errors than subjects in the control group during the evaluation phase. The results show that the sleeve helps users retain and transfer motor skills better than visual feedback alone. This work shows the potential for fabric-based haptic interfaces as a training aid for motor tasks in the fields of rehabilitation and teleoperation.


Introduction
Over the past decade, robotic teaching aids have been developed to train people in a variety of motor activities by providing sensory feedback [1,2,3,4,5,6,7]. Motor learning is an error-driven process and the rate of learning depends on how sensory feedback is provided during training. Typically, people learn to rectify their errors based on a combination of visual, auditory, and haptic feedback [2]. Motor training by haptic feedback is of particular interest to researchers because it can be applied directly to the part of the body where corrective action is needed [8,9]. This effectiveness hinges on two critical and intimately linked factors -the haptic interface and the training method [10,11,12].
Haptic interfaces should provide reliable, intuitive, and clear feedback when required and be unobtrusive when they are not. Within the scope of motor training, haptic interfaces are generally employed in two fields -rehabilitation and teleoperation [5,13]. Existing haptic interfaces are mostly grounded i.e., they are fixed to a rigid base, which limits their applicability to tasks that do not require much user displacement, such as object manipulation [14]. With the miniaturisation of electronic components, such as integrated circuits and batteries, wearable interfaces have become a viable alternative because they allow users a greater degree of mobility [15]. This increased mobility has been demonstrated through multiple exoskeletons for gait rehabilitation and robot teleoperation that improve user performance in the specific tasks for which they are designed [16,17,18,19,20,21,22,23,24,25,26,27]. However, these interfaces are bulky because of the heavy actuators that they use, such as motors and pumps, which cause user fatigue over prolonged use. In fact, actuators are not always necessary to help users perform motor tasks better. A recent study showed how a simple, unpowered clutch and spring ankle exoskeleton can increase human walking efficiency by re-purposing their expended energy [28]. Indeed, new lightweight, fabric-based haptic interfaces have been developed in recent years to circumvent the problems associated with using heavy actuators and promote user comfort for continuous usage [29,9,30,22,31,32,33,34,35,36,37,38] ( Table 1). Amongst these fabric-based haptic interfaces, certain interfaces use electrostatic adhesive (EA) clutches to apply kinesthetic feedback through movement braking and passive springs [39,36,32,40,41]. They operate at low power (∼ 1 mW) and are easily integrated into other textile-based wearables, such as clothing. These new types of fabric-based, lowpower-consuming haptic interfaces are less complex in mechanical design compared to existing interfaces, which are composed of rigid components. The absence of actuators compels users to rely on the feedback to both identify and correct their errors, which helps them learn the motor task faster. Furthermore, unlike existing interfaces, these interfaces are not designed for any one specific task alone. Rather, they can be re-purposed to help train users in a variety of motor tasks.
A successful haptic training method ensures that users are immersed in learning the motor activity and are provided timely feedback to improve their performance over the course of 2 Advanced Intelligent Systems, Preprint Version, Published (July 2021) Low et al. [22] Bianchi et al. [30] Textile haptic devices Year Feedback type Power source Wearable type Tethering Ramachandran et al. [32] Culbertson et al. [31] Wu et al. [34] Park et al. [33] Hinchet et al. [36] Carpenter et al. [35] Lee et al. [37] This study Zhu et al. [38]   training. Existing haptic-based training methods can be broadly divided into two categories -haptic guidance and error amplification [42,43]. In the former, the haptic system physically guides users to minimize errors they commit during training and accomplish a task. In the latter, the haptic system amplifies user errors to intentionally increase the difficulty of the task [44,45]. Studies show that haptic guidance increases user performance during training compared to baseline conditions, but that performance levels precipitate when the guidance is not provided [46]. Some have posited that the guidance overtly increases user dependency [47]. This dependency curtails both skill retention and skill transfer post-training. On the other hand, error-amplifying systems deliver longer periods of skill retention at the expense of longer training periods [48]. However, some studies state that the comparative effects of haptic guidance and error amplification on motor learning cannot be generalised, because they are subject to the type of motor activity [42,49]. Nonetheless, there is some consensus that the method of feedback provision and the resulting outcome is dependent on the user skill level [10,50]. Accordingly, novice users benefit more from haptic guidance, whereas expert users gain more from error amplification. There is also recent evidence that suggests, allowing users to select the type and magnitude of haptic feedback can accelerate the rate of motor learning [6,51]. These studies conclusively show that users must actively utilize haptic feedback to rectify their errors and acquire motor skills. However, so far, this active motor learning has only been demonstrated with haptic interfaces made with rigid and bulky devices [6].
Here, we present a novel haptic interface that promotes active user involvement in recti-fying movement errors during motor learning and is composed of only soft, fabric-based components. The interface consists of a programmable elbow sleeve that comprises multiple fabric-based EA clutches, which can rapidly restrict the joint movements of a wearer. In this study, the haptic sleeve is programmed to afford users a margin of error, but provides a motion-blocking feedback to 1) make users aware of their error, and 2) prevent these errors from growing. The haptic feedback operates like a wall that confines the motion range.
Through heightened awareness of their errors, users can consciously avoid them in future iterations of the task. In this study, we experimentally show that the proposed haptic interface increases the training success to learn, retain, and transfer motor skills in a drone teleoperation task.

Design of the electroadhesive haptic sleeve
The electroadhesive haptic sleeve is a fabric-based exoskeleton that can be programmed to constrain elbow extension and flexion. It is composed of two electroadhesive clutches and three body attachments ( Figure 1A, Figure S1, Supporting Information). Each clutch is a parallel plate capacitor composed of overlapping dielectric-coated electrodes. The EA clutches used in this study are an improved version of our previously described fabric-based clutches to generate higher holding forces [32]. The higher forces are generated by replacing the fabric-electrodes of our previous clutches with metallized biaxially-oriented polyethylene terephthalate sheets. This was done to avoid dielectric cracking on the electrode surface, which was observed with the fabric electrode over prolonged usage. Furthermore, surface irregularities and wrinkling were prevented by replacing the fabric electrodes with the metallized plastic sheets.
When a voltage in the order of 100 V is applied across the overlapping electrode plates, they adhere to each other through electrostatic adhesion. As a result, the maximum holding force i.e., the tensile force needed to ply the plates apart longitudinally increases. In this engaged state, the magnitude of the holding force is dictated by the applied voltage ( Figure 1B).
When the clutch is disengaged by removing the high voltage, it recovers its initial mechanical properties within 40 ms ( Figure 1C). The elbow joint is a hinge-type joint. Hence, the to the forearm and upper arm body attachments to measure the elbow angle.

Motor learning to teleoperate a drone
We assessed the applicability of the EA haptic sleeve as a teaching aid for drone teleoperation tasks and used user errors as the performance metric to determine skill retention and transfer.
The experiments consisted of two drone teleoperation tasks -path following to examine the effect of haptic training on the retention of motor skills, and waypoint navigation to determine the transfer of those skills ( Figure 2A). Both tasks are frequently used methods in the context of drone teleoperation. Path following is often employed to navigate drones through long, narrow pipelines for maintenance and inspection or navigate through cluttered environments that are inaccessible to humans. Waypoint navigation is commonly used to help drones map areas of interest or carry out aerial monitoring [53]. For the path following motor task, subjects were asked to control the altitude of a drone flying through a cylindrical tube with multiple vertical bends while avoiding collisions with the walls. For the waypoint navigation motor task, subjects were asked to control the altitude of a drone through a series of rings positioned at different heights. The performance of the path following task was computed by measuring the altitude error with respect to the desired centreline of the tube throughout its length. The performance of the waypoint navigation task was computed by measuring the difference in height between the drone and the centre of each traversed ring. Waypoint navigation is chosen for the skill transfer test because it is similar to path following. Greater task similarity between the skill retention and transfer tests improves the chances of skill transfer. Despite their similarities though, waypoint navigation provides users more freedom in terms of drone movement than path following as the performance is only evaluated at discrete intervals. However, this additional freedom comes at the risk of increased movement errors if the user does not retain and transfer the necessary skills to control the drone.
There is a linear mapping between the elbow angle measured by the IMUs and the altitude of the simulated drone ( Figure

Acquisition and retention of motor skills
Subject performance errors were measured for both teleoperation tasks and the temporal changes in their performance described the motor learning characteristics of each group (Fig-ure 3A). Differences in performance errors between groups for the same task phases revealed the effects of differentiated feedback provision. The performance errors of the path following task were compared to understand how each of the groups attempted to acquire and retain motor skills specific to this task ( Figure 3C). Subjects in groups 2 and 3 committed fewer errors (32.89% and 37.56% respectively) in the evaluation phase compared to the baseline phase for the path following task. On the other hand, subjects in group 1 committed 12.65% more errors during the evaluation phase of path following compared to the baseline phase.
A paired t-test showed that the performance of subjects in group 1 (FPV) deteriorated from baseline to evaluation for the path following task indicating that the training phase between the baseline and evaluation phases did not benefit the subjects and may have even been deleterious to them (t =-2.535, P <0.05). On the other hand, subjects in group 2 (FPV + arrows) and group 3 (FPV + EA clutches) performed significantly better in the evaluation phase compared to the baseline, which implies that they were able to learn the task over  Figure 3F). For a given amplitude or wavelength, each comparison was made between groups using a one-way ANOVA followed by a post-hoc Holm-Sidak corrected t-test. Groups 2 and 3, which received augmented visual and haptic feedback, respectively, performed signif- Different drone speeds might have produced more noticeable differences.

Transfer of motor skills
Subject had to utilize the motor skills garnered during the path following task to improve upon their baseline performance in the waypoint navigation transfer task ( Figure 3B).

Subjective assessment of drone teleoperation and sleeve comfort
We assessed each subject's comfort in learning to control the drone while wearing the haptic sleeve using questionnaires. Since differentiated feedback was only provided during the path following phase, subjects filled three questionnaires, each corresponding to the three phases of path following -baseline, training, and evaluation at the conclusion of each respective phase. Subjects were asked to assess the degree to which they were in control of the drone's movement, their level of comfort with the haptic interface, and a self-evaluation of their performance improvement over the recently concluded phase by indicating a grade on a 7-point Likert scale (0-6). For all groups, at least 80 % of the subjects rated both their degree of control over the drone's movement and their level of comfort with the wearable system between 4 and 6 (both included) for all three phases (Figure 4). Subjects of groups 2 and 3 were asked to rate the helpfulness of the additional sensory feedback (arrows and EA clutches) in performing the tasks as part of the training questionnaire. Despite the quantitative results which indicated that subjects in these two groups improved upon their baseline performance, only 50 % of them found the additional sensory feedback (arrows or EA clutches) to be qualitatively helpful. It is interesting that participants do not perceive the feedback as qualitatively helpful and yet, they tend to perform better. The rather counter intuitive response from the subjects may be due to the additional feedback creating sensory   overload, but this claim cannot be verified as part of our study. For the self-evaluation of performance improvement subjects could answer "Yes" or "No". For group 1, the percentage of subjects who responded to their self-evaluation of performance improvement as "Yes" increased (70% -"Baseline"; 80% -"Training"; 90% -"Evaluation"). On the other hand, the percentage reduced for group 2 (90% -"Baseline"; 70% -"Training" and "Evaluation"). The percentage of affirmative responses increased for group 3 over the phases (80% -"Baseline" and "Training"; 90% -"Evaluation"). In the evaluation questionnaire for all groups, subjects were also asked to rate their overall experience in participating in the experimental study. All subjects rated their overall experience between 4 and 6 in participating in the experiments.

Conclusion
We described a wearable haptic sleeve that uses fabric-based EA clutches to impart kinesthetic feedback by blocking body joint movement and experimentally showed its functionality as a teaching aid for motor activities in drone teleoperation tasks. This study examined and compared the effects of providing haptic feedback for motor training with different forms of visual feedback.
The results show that subjects in the control group, who received FPV visual feedback did not acquire and retain sufficient motor skills to improve path following performance and subsequently, were unable to perform better in the transfer waypoint navigation task. Instead, subjects who received either augmented visual feedback in the form of arrows to correct elbow movement or haptic feedback from the electroadhesive haptic sleeve to physically block elbow rotation, displayed performance improvement from baseline to evaluation for both path following and waypoint navigation. This performance improvement could be attributed to users relying on both types of feedback to determine their errors and consciously learning to to avoid them. This conclusion is reinforced by the observable improvement in performance level with respect to the control group immediately after receiving additional haptic/visual feedback. While this additional feedback assists subjects in identifying and avoiding errors during the training phase, their newly acquired motor skills are retained even after the training phase. This show that the subjects did not become overtly dependent on the additional feedback.
The increase in performance errors with the amplitude of tube reference trajectories for all subject groups can be ascribed to the difficulty in rapidly alternating between forearm extension and flexion, especially when approaching the maximum and minimum elbow angle limits. Indeed, the performance errors of specific training sessions on average were higher for all groups than individual sessions within the baseline and evaluation phase due to the training tubes having larger amplitudes. The relatively small differences in performance for all groups with changing tube wavelength may be due to the chosen drone speed. For higher drone speeds, we might expect to see greater performance differences with fewer errors being committed for larger tube wavelengths.
The comparable beneficial effects of haptic feedback and of corrective arrow displays are noteworthy because augmented visual feedback is generally accepted as a benchmark in feedback-based training. Indeed, there are no observable statistically significant differences between the two forms of additional feedback based on the data collected during the subject studies. While the haptic feedback is just as effective as a teaching aid as augmented visual feedback, haptic feedback could be qualitatively more helpful than augmented visual feedback when provided directly to the part of the body responsible for erroneous motion. This is because concentrating the relay of augmented sensory feedback through a single feedback channel (vision) can severely increase the risk of sensory overload over prolonged periods of robot teleoperation [54]. Furthermore, haptic feedback could be used instead of vision when visual feedback from the robot is occluded [54,55]. Certainly, visual occlusions are not uncommon when operators need to inspect infrastructure, for example maintenance after the occurrence of a natural disaster. In these instances, first-person visual perspective may not reveal concealed structures in the robot's periphery due to a variety of reasons, including insufficient lighting. Another possibility is that visual feedback may be cut off from the operator intermittently due to poor transmission. In addition, haptic feedback can be provided to the visually impaired. In this work, the haptic feedback was binary i.e., either it behaved as a compliant cloth or it blocked the range of motion. In the future, lightweight variable stiffness technologies which use low melting point materials and shape memory materials that can provide a range of blocking forces could also be employed for different training tasks.
Furthermore, extension of this technology to other joints, such as the wrist and fingers, could be used for more complex teleoperation and rehabilitation tasks.

Manufacturing of electroadhesive haptic clutches
Each clutch consisted of three pairs of dielectric-coated electrodes that were interleaved in an interdigitated architecture. The electrodes (150 mm × 35 mm) were 15 µm biaxiallyoriented polyethylene terephthalate that were metallised on one surface. The metallised surface was coated with a 20 µm layer of high-κ dielectric ink, a ferroelectric composition of Barium Titanate and Titanium Dioxide (Luxprint 8153, DuPont). The dielectric was oven-cured at 140°C for 60 min. Post-curing, the solid dielectric that remained was 10 µm thick. The non-metallised surface of each electrode was bonded to a 120 µm sheet of polymethyl methacrylate (PMMA). Each of the three pairs of interleaved electrodes overlapped with their dielectric surfaces in contact. By virtue of the interdigitated architecture, one electrode of each pair was maintained at high voltage with respect to its paired electrode that was grounded when the clutch was engaged. Therefore, the two sets (high voltage and ground) of three electrodes that were maintained at the same voltage were bonded together To determine the clutch disengagement time, the clutch was initially fixed between the vices of the tensile tester and then engaged at 250 V. The clutch was loaded in tension until it reached the maximum holding force at which point the clutch is disengaged by setting the applied voltage from the board to zero. At the same time, a serial command was sent to the tensile tester to register the time of disengagement. The disengagement characteristics were reported by averaging the precipitating holding force measurements for three trials. The disengagement time was measured by computing the 90 % drop from the initially measured maximum holding force corresponding to the applied voltage.
Simulation framework for drone teleoperation The drone flight tasks were simulated using Gazebo, an open source robotic simulator [56]. Gazebo provided 3D visual scene rendering and simulated on-board sensors such as RGB cameras. The simulation environment and drone model were based on our previous work on machine-learning-based multi-drone control [57]. For the flight tasks, single-integrator dynamics were used to move the drone through the environment. The drone flight was restricted within the x-z plane and moved forward at a constant speed of 5 m s −1 . For each of the three baseline and evaluation sessions of waypoint navigation, a set of 20 rings of 1m diameter each were generated in the environment. The rings were spaced at equal intervals of 4 m along the x-axis with the openings facing the drone's direction of travel. For each session, the reference trajectory passing through the ring centres was created using cubic spline, such that spline slope was zero at the ring centres. Furthermore, the z-position of the ring centres were constrained to remain between 9 m and 11 m above the ground. For the baseline and evaluation sessions of path following, 1 m diameter tubes measuring 76 m in horizontal length were rendered in the same x-z plane as the drone, appearing 4 m in front of it. The centreline of each tube, which was the drone's reference trajectory was also produced using cubic splines. The trough and the peak of the trajectory were constrained to 9 m and 11 m respectively. For the training sessions, 1m diameter tubes measuring 76 m in horizontal length were produced with tube centreline reference trajectories as sine curves. To generate nine distinct tubes, nine distinct combinations of amplitude and wavelength were used. The three amplitudes values were 0.5 m, 1 m, and 1.5 m, and the three wavelength values were 8 m, 10 m, and 12 m.

Human subject study
A total of 30 adult subjects (ages between 23 and 42, mean = 29.86, standard deviation = 4.32) were recruited, primarily from the university. Each of the subjects were healthy, had normal hearing and normal or corrected-to-normal vision. None of the subjects had prior knowledge of the tasks they were expected to perform. They were randomly assigned to the three groups with 10 subjects each (6 men, 4 women in each group). The subjects provided written informed consent, and the study was approved by the EPFL Human Research Ethics Commission. Before the motor learning studies were performed, preliminary psychophysical measurements were made to ascertain the Just Noticeable Difference in terms of the holding force needed to constrain their elbow flexion and extension. The applied DC voltage was directly proportional to the maximum holding force. A voltage value of 300 V corresponding to the results of the Just Noticeable Difference was used as the operating voltage for the engagement of the clutches during the experiments. The dimensions of the haptic devices were determined a priori by using the anthropometric data collected from 10 individuals, specifically, their forearm length, their upper arm length, and the changes in length on the ventral and dorsal arm faces associated with forearm extension and flexion about the upper arm respectively. While the haptic devices are robust to repeated usage, the elastic fishing line used as the low-stiffness spring can become brittle at the knots, as is often the case with some types of vulcanized rubber. To prevent any deleterious effect as a result of this changing mechanical property, the springs were replaced after every 5 subjects. The clutch plates were operated at a voltage that would not cause electrical shorting. To prevent the accumulation of surface charges on each clutch plate due to the DC voltage operation, the polarity of the clutch plates was reversed for each subject after every session of each experimental phase.

Statistics
All analysis on the data gathered from the simulated flight tasks was carried out in MATLAB 2019b (Mathworks, MA, US). The paired t-test is used when comparing a pair of normally distributed sets of data subjected a single condition, here the effect of the training condition for each group. To compare three or more groups of normally distributed data obtained from different sample populations, an Analysis of Variance (ANOVA) needs to be performed