A Parallel, 2-DOF Exoskeleton for the Human Shoulder: Device Characterization and Preliminary Results on Healthy Subjects

flexion adduction, horizontal flexion and extension. It can drive the upper arm to any in between standard positions. At the core of the exoskeleton are four modular, pneumatic bending actuators with and replaceable inflation modules. Two antagonistic actuator pairs placed the humerus. One side of each pair is responsible for one direction of motion while the agonist drives the humerus the opposite direction. The exoskeleton is unique in that synchronized and simultaneous operation of the antagonistic pairs, additional flexibility by separated modules, movement

therapists and clinicians. Nevertheless, current robotic architectures are not optimized for the human 19 shoulder but are more apt for industrial environments. Pneumatically powered soft robotic actuators 20 present an attractive method to create shoulder exoskeletons due to their compliance and relatively low 21 mass. However, current actuators lack the necessary functions to provide support to the entire 22 shoulder's range of motion. 23

Methods 24
A modular, fabric pneumatic actuator was constructed. The actuator design allows it to perform three-25 dimensional (3-D) bends with minimal resistance. Four actuators were combined to create a soft 26 shoulder exoskeleton. Each actuator drives one direction of motion: elevation and depression, rotation 27 of the plane of elevation. The torque output of the actuator was measured using a customized two-axis 28 torque measurement system. Exoskeleton functionality was tested through surface electromyography 29 of relevant shoulder muscles. 10 healthy subjects were recruited and performed arm motions under the 30 assistance of the exoskeleton. 31

Results 32
The actuator can reach full bending (>360°) with low pressures (~10kPA). Its torque output is highly 33 dependent on its geometry. Moreover, torque output is reduced as the bending angles increase. The 34 actuators installed on the exoskeleton output 11.15N-m of torque at the neutral position, and 4.44 N-m 35 at 90° shoulder elevation. The test on healthy subjects showed that use of the exoskeleton reduces 36 muscle activation by up to 65% when performing shoulder elevation, and up to 34% when rotating the 37 plane of elevation. Use of the exoskeleton also resulted in a change in arm trajectory when performing 38 elevation and depression movements. 39

Conclusions 40
that an actuator's torque output will differ depending on the topology of the object to which it is 159 attached. This model simulates an actuator's attachment to the human shoulder and assumes that the 160 actuator is bent at the center, while the remaining modules remain straight. In their inflated state, the 161 modules can effectively be considered as an extruded ellipse (Figure 3-A, Figure 3-B) with the lengths 162 of the minor and major axes to be equal to half of the module's inflated height (Hi) and length (Li). 163 Figure 3-C shows the free-body diagrams upon which the rest of the model is based. In this 164 configuration, the actuator effectively generates torque with respect to the pivot point; it can be 165 effectively divided into two equal segments, with each segment generating equal but opposing torques 166 that effectively bring the entire system into static equilibrium. Each segment is further subdivided into 167 three sections: the angled section (middle pair, red), the edge section (first and last pair, yellow), and 168 the straight middle section (all other pairs, blue). Each section will have its own unique mechanical 169 behaviors; the subsequent subsections will concentrate on the n/2+1 to the nth module. The total torque 170 ( ) generated by the actuator is the sum of each module's generated torque ( ) [1]. 171

Force from module intersections ( ⃗ ⃗ ) 174
⃗ ⃗ is dependent on the nature of the contact area (AC) between two colliding modules. Due to the 175 elliptical shape of the modules, this model approximates the contact area as a rectangle. The contact 176 area can be determined by examining the cross-section as seen from the top-view (Figure 3-D). The 177 centers, Ci= (Ci,x,Ci,y) and Ci+1= (C i+1,x,C i+1,y) of the two ellipses are first established; in the proceeding 178 calculations, both centers lie on the x-axis while the origin coincides with C i+1. Points Mi= (MI,x,MIi,y) 179 and M i+1= (M i+1,x,M i+1,y) are the points of attachment on the fabric spine. Upon inflation, modules collide and rotationally displace by the bending angle (θj). They then trace a circular path centered 181 around point Oj; the arc length of this path is equal to module spacing (d). CI,x can then be calculated: 182 With the location of the centers known, the ellipses can now be mathematically defined with [4] and 186 [5] 187 188 = ( The points of intersection, Ij,1= (Ij,1,x, Ij,1,y) and I j, 2= (I j,2,x, I j,2,y), between the two ellipses are 193 At θ>-90°, the vertical component of ( ) no longer contributes to torque generation due to the 212 fact that vertical force is now directed away from the arm, and is transformed into tension on the 213 fabric spine, as opposed to generating a counter-acting force on the arm. This results in minor 214 compression of the actuator but does not significantly affect its performance. At θ>-90°, 2 +1 is 215 described by [9]. 216

The middle section (all other modules) 218
It can be seen in arm. These modules slightly bend and effectively act as straight beams. Such a seamless system could 245 easily be extended to other joints in the arm, such as the elbow and the wrist. However, a system for 246 the entire arm is beyond the scope of this paper. The exoskeleton is controlled by a pneumatic control 247 system; exoskeleton output is controlled by regulating the pressure supplied to the actuators. The 248 details of the control system are available in additional file 1. 249 The operation of the actuators allows the humerus to trace a natural trajectory. From the neutral 250 position, the humerus can perform shoulder abduction and adduction by activating the elevation or 251 depression actuators (Figure 4-A). The shoulder can also be rotated along the plane of elevation by 252 sequentially activating the elevation and steering actuators (Figure 4-B). Simultaneous activation of 253 the elevation and steering actuators also grants the ability to perform shoulder flexion or extension 254 (Figure 4-C). A video of these movements is made available to supplement this paper (Additional file 255 2). In addition to performing the basic anatomical movements, the controlled activation of both steering 256 actuators and the elevation actuators allows the performance of reaching actions (Figure 4-D, Figure  257 4-E). 258

Actuator Experiments 259
In order to verify the accuracy of the physical model, three actuator variants were constructed, with 260 each variant having a total of eight modules installed. The variants differ in the size and pattern of the 261 modules installed. Their geometric parameters are listed in Table 1. Module widths were based on the 262 lower 5th percentile of female upper arm diameters for B, and the lower 5th percentile of male upper 263 arm diameters for C (27). Module spacing (d) was set at 25mm in order to minimize the discontinuity 264 in the curvature profile of the actuator. While tighter spacing is desired, the size of the 3-D printed 265 locking structures presents a practical minimum. Alternate module patterning (i.e. ABAB, ACAC) was 266 chosen since a homogenous pattern induced buckling in preliminary tests. This pattern involves the 267 installation of alternating module sizes; specifically, module A was installed between each B and C 268 module. The module lengths were then set to 65mm and 90mm to ensure sufficient spatial interference 269 between adjacent modules. 270 The actuator' static free bending output in response to a pressure input was measured by hanging the 271 actuator vertically. The detailed setup is available in additional file 1. 272

Module Width (W)
The installation arrangement of modules on the variants

5-A).
As an inflatable structure, an inflated actuator will always attempt to remain in a 2-D pose by 274 generating a straightening torque when acted upon by an external force that drives it into a 3-D pose. 275 While the actuator is only capable of generating torque along its primary axis, it is important to quantify 276 how much unwanted straightening torque the actuators generate when forced into a 3-D pose. As shown 277 in All experiments were repeated three times, and the samples were dismounted and remounted before 282 each repetition in order to minimize the effects of any actuator movement that may have occurred 283 during a measurement. 284

Healthy Subject Testing Protocol 285
Ten healthy subjects were recruited into the study; the test was reviewed and approved by NUS movements. At the peak of each movement, the subject was instructed to pause in order to ensure that 324 each movement was isolated from the proceeding movement. Once the sensors were applied, the 325 subjects donned a neoprene sleeve to ensure that the sEMG sensors remain stationary and to minimize 326 the presence of motion artifacts. 327 The test was conducted in three phases. The first phase (free phase, F) serves as a baseline; the subjects 328 were asked to perform arm motions while only wearing an empty neoprene sleeve (Figure 6-E). In the 329 second phase (unpowered phase ,U), the neoprene sleeve was removed, and the subjects wore an 330 unpowered exoskeleton (Figure 6-F). At this stage, the exoskeleton was disconnected from the control

Actuator Static Characteristics 344
Each actuator, when unloaded, was able to achieve full bending (>360°) when excited with a pressure 345 of 10kPa. An increase in pressure did not have any effect on its pose. 346  (Figure 7-B). The unfolding nature of the modules ensure that the contact area is nearly constant when the actuator 363 is positioned between 0° to 90°. They also suggest that torque output has a linear response to increases 364 in pressure.
An increased reduction in available contact area occurs when the A-A' angle and B-B' angle change 366 simultaneously. The effects of this behavior on variant D2 are shown in Figure 7-C. In this 367 measurement, the A-A' angle was set to -90° and the B-B' angle was varied from 45° to 0° at 15° 368 increments. This range of motion corresponds to that required of the majority of ADLs (19). The 369 actuator was then pressurized to 80kPa. An overall decrease in the torque output along the A-A' axis 370 was found as the B-B' angle was increased. However, the platform was not able to detect any 371 significant torque output (>0.5N-m) along the B-B' axis. This can be attributed to the ability of each 372 individual module to translate in 3-D as well as the aspect ratio of the modules; they show minimal 373 surface area along the B-B' axis. Effectively, the modules reposition themselves such that there is 374 minimal resistance during operation, resulting in small B-B' torques. 375

Arm Trajectory 376
Trajectories extracted from a single subject are shown in Figure 8. Each subject was able to reach the 377 90° arm position for all motions as observed during the test. Application of the exoskeleton increased 378 the starting angle of elevation from 0° (i.e. neutral position/pendulum position) to ~15. However, it is 379 interesting to note that each subject was able to return to true neutral position (i.e. 0°) after the first 380 cycle. This is attributed to the forced movement of the subject, allowing each subject to overcome the 381 resistance introduced by the exoskeleton and forcibly return to true neutral. A closer visual inspection 382 of the individual trajectories shows an observable difference in the paths among the three phases. 383 The full tabular data of the change in RMS of the arm trajectories are found in Additional file 1 ( Table  384   S2 and Table S3) (Table S4 and Table S5). 405 The results show a decrease in the mean RMS of muscle activation in the population for fast PvF 406 comparisons. Statistically significant results were obtained from multiple muscles during fast 407 exercises: the lateral deltoid and infraspinatus during abduction and adduction, the anterior deltoid 408 during forward flexion and extension, the pectoralis major during horizontal flexion, and the pectoralis 409 major and posterior deltoid during horizontal extension. The exoskeleton was able to reduce deltoid 410 activation by 65% (p<0.01) when performing abduction and adduction, and by 45% (p<0.01) and 25% (p<0.05) when performing forward flexion and extension, respectively. The activation of the pectoralis 412 major was reduced by 34% (p<0.05) when rotating the arm's angle of plane of elevation. While 413 reductions in some muscles were also observed during slow movements, the reduction magnitude was 414 considerably lower as compared to those acquired during fast movements. Statistically significant 415 results were only obtained for the anterior deltoid and lateral deltoid during slow forward flexion and 416 extension, and the anterior deltoid for horizontal extension. The slow movements only apply minimal 417 loads to the shoulder muscles that result in low magnitudes in the measured EMG voltages, which 418 explains the high variance in non-significant findings. The assistance provided by the exoskeleton 419 would not be apparent in the voltage waveform due to the inherent, relatively low signal-to-noise ratio 420 of EMG (31). Nevertheless, the statistically significant results for forward flexion and extension 421 indicate that the exoskeleton is capable of providing support during reaching movements. increased by 25% (p>0.05) and 28% (p>0.05) while performing slow abduction and adduction, 432 respectively, while infraspinatus activation increased by 10% (p>0.05) when performing slow 433 abduction and adduction. Other muscles did not feature statistically significant readings. Most notably, 434 forward flexion and extension did not exhibit any change in muscle activation for both speeds. The high variance in the data is attributed to the various strategies employed by the subjects in order to defy 436 the mechanical resistance added by the exoskeleton, with some strategies more successful than others. 437 Moreover, the fit of the sleeve to a participant may also play a role in the amount of mechanical 438 resistance it would impart on the user. as evidenced by its ability to trace its full range of motion with minimal input. This feature, along with 447 its soft robotic nature, ensured that the shoulder sleeve will have maximum mechanical transparency. 448 Results from various tests indicate that the exoskeleton will be able to provide torque assistance 449 regardless of the position of the actuators or the pressure supplied. Nevertheless, the positions of the 450 actuators influence their torque outputs. Load bearing actuators must operate from the -180°to -90° A-451 A' angle range in order to maximize their utility. This behavior is consistent with that from our previous 452 work(27,28). These actuators would supply approximately 7%, 24.6% and 25.8% of the torque 453 necessary to maintain arm elevation of 90° for a typical, stretched, male arm with a mass of 3.5kg (32). 454 Users can still benefit from the high peak torques at low elevation angles (i.e. -180°to -90° A-A') when 455 performing high velocity, dynamic movements. 456 The final design of the exoskeleton is intended to capitalize on the capabilities of the modular actuators. 457 When comparing the torque curves of the D2 and D3 variants, it can be seen that the difference is negligible. This suggests that minor buckling occurs with the D3 actuator and that a significant portion The mean RMS values show that activation of shoulder muscles during fast unpowered movements 507 increased by up to 29%, as compared to free movements, and slow movements showed an increase of 508 up to 48%. However, it must be noted that the absolute change in activation is negligible since the 509 overall load on the arm is similarly smaller during slow movements. Nevertheless, both PvF and PvU 510 comparisons yielded the same amount of muscle activation reduction, indicating that the exoskeleton 511 is not only capable of overcoming its own mechanical resistance, but also capable of effectively 512 negating it. 513 The unique soft structure of the exoskeleton also allowed subjects to retain their shoulder's natural 514 range of motion regardless of the prevailing condition (i.e. free, unpowered or powered phase). While 515 their relaxed, neutral position was elevated by approximately 15 o when wearing the exoskeleton, each 516 subject was still able to reach the true neutral position (i.e. 0 o elevation) when effort was applied. While 517 this required the application of force on the arm-either by the user, or the exoskeleton-the results 518 show that the exoskeleton was able to compensate, and overall muscular activation was still reduced. 519 Meanwhile, the statistical results for fast forward flexion and extension, fast abduction, and slow 520 abduction seem to suggest that use of the exoskeleton affects elevation movements. These nevertheless 521 indicate that the use of the exoskeleton, in its current form, will reduce the kinematic accuracy of the 522 user's arm. It is important to note that subjects were only given one day to use the exoskeleton, and 523 further training and familiarization may allow them to regain movement accuracy (34) . 524 The results highlight the positive benefits of the exoskeleton to the user by measuring sEMG signals 525 as way to ascertain its effects on muscle activation. A comprehensive comparison of the results of the 526 exoskeleton with that found in the literature is currently not possible due to a lack of a standard 527 evaluation protocol. By using an unsuited state as a baseline, we were able to remove the mechanical 528 resistance imparted by the actuators from our measurements. This is not always the case when 529 exoskeletons are evaluated. Nevertheless, sEMG measurements have also been used to assess exoskeleton performance. O'Neill et al was able to achieve 62.75% of muscle reduction on the lateral 531 deltoid when performing abduction and 16.94% on the pectoralis major when performing horizontal 532 flexion (35). Simpson et al also collected sEMG signals, but their data processing method was different 533 from the one presented here (23). Tiseni et al achieved a mean reduction of 56.9% on the lateral deltoid 534 using a hybrid soft-hard exoskeleton on five healthy subjects (36). The vast other majority of functional 535 tests with exoskeletons directly deal with subjects with shoulder impairments. They typically 536 implement clinical tests which are designed to gauge the level of shoulder functionality. 537 It is important to note that the exoskeleton lacks a controller that involves human input in the activation 538 of the exoskeleton's functions. Admittance control has proven to be effective in predicting the onset of 539 user motion and has been implemented in a number of exoskeletons (37,38). Continuum-based 540 actuators impart forces on a large surface area, making it impractical to measure forces through 541 conventional sensors. While the addition of an admittance controller would improve the performance 542 of the exoskeleton, it is difficult to implement. We believe that the observed changes in arm trajectory 543 when wearing the powered exoskeleton is also a product of a lack of user input in the control system. 544 The inclusion of intent sensing would allow the exoskeleton to make on-the-fly adjustments to its 545 motion and will therefore reduce its effect on the user's kinematics. These changes are therefore a 546 product of the lack of synchronization between the user and the exoskeleton, and not of the mechanical 547 design of the exoskeleton. The build quality of the exoskeleton in the current study also limited possible 548 performance. The exoskeleton was only capable of withstanding 80kPa of continuous, sustained 549 pressure and 150kPa of intermittent pressure. Improving the build quality delays, the onset of device 550 failure, which could lead to the ability to handle higher intermittent pressures and the generation of 551 stronger torques. While there is evidence of mechanical resistance throughout the exoskeleton's range 552 of motion, the resistance of the steering actuators must be significantly reduced. Natural, free, shoulder 553 movement in this direction only presents minimal load to the shoulder muscles and the exoskeleton 554 should therefore present the relatively same amount of low resistance. Reduced mechanical resistance 555 would also simplify the application of an admittance-based controller for backdrivability. 556

6
Conclusion 557 The exoskeleton presented demonstrated the feasibility of creating a fabric-based shoulder 558 exoskeleton, powered by flexible pneumatic actuators. The torque response and free bending response 559 of the underlying actuator were explored. The exoskeleton was tested on healthy subjects in order to 560 measure its efficacy. Our results quantitatively demonstrated the efficacy of our exoskeleton in 561 imparting torque to the shoulder joint. While there are limitations to our study, the results suggest that 562 the unique exoskeleton configuration is effective in providing support to the human shoulder 563 throughout its entire range of motion. We were able to create an exoskeleton configuration with two 564 degrees-of-freedom. It is able to support not only abduction and adduction, and horizontal flexion and 565 extension, but also reaching motions (i.e. forward flexion and extension). Reaching movements allow 566 the exoskeleton to assist the users in a wide variety of ADLs in a seamless manner. While these ADLs 567 may possibly be accomplished using sequential humeral elevation and rotation of the plane of 568 elevation, this series of motions is undoubtedly unnatural. Moreover, the exoskeleton was able to 569 provide support throughout the entire range of motion of the shoulder and this mainly attributed to the 570 minimal mechanical resistance of the actuator coupled with the parallel actuation configuration of the 571 exoskeleton. 572

7
List of Abbreviations and Symbols 573