2.1. Passive shoulder movement assistance

Stretching machines are used to improve muscle flexibility and strength. The Chest Spread (Park Corp.), a stretching machine, is designed to stretch chest and back muscles (Chest Spread, 2020). It consists of a simple mechanism that does not require the control of human joint angles. With the aid of this machine, the user can perform thoracic stretching without the therapist’s support. However, since this machine was designed to adjust its intensity based on the lower-limb movement (e.g., pedaling), its use is challenging for older adults with limited lower-limb mobility.

Passive mechanisms are used in industrial exosuits to assist shoulder motion (Tian et al., Reference Tian, Wei, Luo, Yang, Chen, Liu, Feng, Li, Zhu and Yi2024). However, many mechanisms use gravity-compensation mechanisms to support long-duration tasks, making it difficult to apply them to stretching scapulae.

2.2. Robotic movement assistance

Upper-limb exoskeletons are used for medical and rehabilitation purposes such as neurorehabilitation, musculoskeletal rehabilitation, and motion assessment. For shoulder or scapular rehabilitation purposes, these exoskeletons provide high DoF control of the shoulder. Gopura et al. (Reference Gopura, Kiguchi and Li2009) developed a 7-DoF upper-limb exoskeleton robot that can assist shoulder internal/external rotation. The robot is capable of supporting shoulder rotation through a mechanism that adjusts to the shoulder rotation center displacement. In addition, B Kim and Deshpande (Reference Kim and Deshpande2017) developed an exoskeleton system, “Harmony,” that can support glenohumeral joint and upper arm movements. The robot uses a 2-DoF mechanism to control the position of the humeral head to adjust to the displacement of the center of rotation of the shoulder. However, exoskeletons that support shoulder joint movements for bilateral configurations such as “Harmony” are often bulky due to their complex mechanisms to support movements with many DoFs (Rosen et al., Reference Rosen, Milutinović, Miller, Simkins and Kim Hand Li2014; Kim and Deshpande, Reference Kim and Deshpande2017; Gull et al., Reference Gull, Bai and Bak2020; Dhatrak et al., Reference Dhatrak, Durge, Dwivedi, Pradhan and Kolke2024).

Several exoskeletons or exosuits for assisting activities of daily living (ADL) for occupational purposes are more compact than exoskeletons for rehabilitating patients with paralysis (Jalal et al., Reference Jalal, Harith, Hasan, Salim and Lin2024; Tian et al., Reference Tian, Wei, Luo, Yang, Chen, Liu, Feng, Li, Zhu and Yi2024). For example, Kwok and Yu (Reference Kwok and Yu2024) developed a lightweight bilateral underactuated upper-limb exoskeleton for bimanual ADL. In addition, Kim et al. (Reference Kim, Xiloyannis, Accoto and Masia2018) developed a soft exosuit for industrial applications. However, these exoskeletons only provide assistance or gravity compensation for the shoulder abduction/adduction and shoulder extension/flexion and thus cannot achieve scapular adduction/abduction. They achieve compactness by treating humeral head displacement as an acceptable or negligible misalignment.

Continuous passive motion (CPM) machine is a mechanical device that supports a joint and can be set to move slowly through a designated ROM to promote controlled movement in the operated joint (Intervention applications, 2013). For example, the Kinex Shoulder CPM machine (Kinex Medical Company, 2018) is designed to move the patient’s shoulder joint through a prescribed ROM over a specified duration. However, most CPM machines do not consider changes in the position of the humeral head.

Some robots for respiratory rehabilitation assistance have also been studied (Zhu et al., Reference Zhu, Cong, Liu, Liu, Yi and Inoue2015; Lee et al., Reference Lee, Hahn, Beom, Park, Cho, Kang and Cho2022; Zhang et al., Reference Zhang, Ge, Wang, Qin, Wu, Wang, Shi, Xue, Guo, Zhang, Wang and Wang2024; Zhou et al., Reference Zhou, Pu, Yuan and Fei2025). Zhou et al. (Reference Zhou, Pu, Yuan and Fei2025) have proposed a bedside assistive robot using pneumatic actuators. This robot is a soft robot and is more comfortable than rigid robots. However, because it uses pneumatics, it requires the use of air compressors and other equipment, making it difficult to use at home.

2.3. Cable-driven wearable approach

Among the robotic assistance approaches, cable-driven wearable robots for the shoulder are suitable for home use due to their lightweight and compact design. Several cable-driven mechanisms are designed to support upper-limb movement, including the displacement of the humeral head (Samper-Escudero et al., Reference Samper-Escudero, Giménez-Fernandez, Sánchez-Urán and Ferre2020; Georgarakis et al., Reference Georgarakis, Xiloyannis, Wolf and Riener2022). However, most methods are intended to assist with daily life activities and reduce the burden of work. Therefore, these cable-driven exoskeletons are designed to support the arm-raising motion based on the scapulohumeral rhythm; hence, they are not sufficient to assist the scapular adduction/abduction for thoracic stretching (Thalman and Artemiadis, Reference Thalman and Artemiadis2020).

Several studies suggested wearable robots for resistance training that consider the movement of the humeral head. Park et al. developed a wearable fitness device for upper-limb exercise that constitutes cable-driven actuation to control the resistance profiles (Park et al., Reference Park, Kim, Hyung, Shin, Kim, Kim, Yoon, Ko, Hahm and Lee2023). In addition, Pyeon et al. (Reference Pyeon, Ahn, Kim, Cheon, Jeong and Cho2024) developed a cable-driven exosuit for upper-limb home fitness, capable of transmitting bidirectional resistance and applying focused loads to specific muscles utilizing cable routing. However, the primary objective of resistance training is strengthening muscle strength and muscular endurance, not to control the position or increase the ROM of the humeral head. Although many related studies have shown that the shoulder can assist with the ROM required for ADL, few studies have examined the effectiveness of robotic stretching for older adults in comparison with manual therapy.

We focus on a wearable robot with a cable-driven system to achieve stretching at home. A model is needed for movements involving scapular adduction/abduction. There is a musculoskeletal model for the scapula and shoulder, but it is difficult to simplify as much as the scapulohumeral rhythm, and the high DoF make the robot bulky. Therefore, the design is aimed only at supporting the movement of thoracic stretching.

A cable-driven robot for thoracic stretching has also been proposed by our group (Isobe et al., Reference Isobe, Hirokawa and Suzuki2023). This device was designed to support the adduction and abduction of scapulae by controlling the position of the humeral head using a 2-DoFs cable-driven mechanism. However, in the study, due to safety concerns and to align the intervention conditions, the same control parameters (e.g., pulling force) were applied to all participants regardless of their physical size and/or strength. As a result, the robotic stretching was unable to provide sufficient support regarding the ROM compared to the manual stretching by a therapist. Therefore, the pulling force needs to be adjusted to the individual to ensure that the robot provides a sufficient ROM. However, that study was limited to younger participants, leaving its effectiveness for older adults unverified. Therefore, verification is also needed for older adults.

To address these limitations and ensure clinical applicability, the current study incorporates the following specific advancements compared to our previous work (Isobe et al., Reference Isobe, Hirokawa and Suzuki2023):

• • an individualized target tension procedure based on user feedback;

• • a modified spine fixation mechanism to support the C7 reference point;

• • a mechanical characterization of Bowden cable friction effects; and

• • a feasibility study with older adults and comparison to a physical therapist.

2.4. Summary of the related works

Robots such as the upper-limb exoskeletons and exosuits in the previous studies mentioned above have not been investigated for the purpose of assisting thoracic stretching. To achieve thoracic stretching, controlling the humeral head position based on the understanding of the scapular motion is crucial. However, due to the high degree of freedom of the scapular movement, conventional exoskeleton approaches require a large mechanism. To make the system suitable for home use, it is essential to create a simpler mechanism. Exoskeletons and wearable robots designed for thoracic stretching lack performance verification of their effect on actual users’ range of movement, particularly for older adults.

Therefore, in this study, we propose a shoulder-wearable robot with a 2-DoF flexible cable-driven mechanism. We propose a control that varies the intensity of the stretching, which was fixed in our previous study (Isobe et al., Reference Isobe, Hirokawa and Suzuki2023), according to the individual, and to characterize the mechanical properties of the proposed mechanism. We also verify the feasibility of the proposed robot for scapular adduction/abduction assistance through experiments with younger and older adults.

3.1. Scapular adduction/abduction support robot

The overview of the proposed robot-assisted scapular stretching is shown in Figure 1. We propose a method to reproduce manual therapy shown in Figure 1(a) by using a wearable robot shown in Figure 1(c). Our method aims to reproduce manual therapy by modulating the pulling intensity based on shoulder reaction forces, as shown in Figure 1(b), providing robotic stretching for older adults. The proposed system dynamically supports scapular adduction by cable pulling while the elastic components of the shoulder brace passively facilitate the return motion to scapular abduction.

Figure 1.

(a) A physical therapist performing thoracic stretching. (b) Overview of the similarities between manual stretching and the proposed wearable robot to assist in thoracic stretching. The physical therapist grasps the shoulder and controls it based on the shoulder’s reaction force, while the proposed robot controls the shoulder, which is stabilized by a brace, based on the pulling force. (c) Overview of the proposed robot.

Figure 1. Long description

Panel one at the left depicts a therapist seated behind a patient, holding the patient’s shoulder for thoracic stretching. Panel two in the center is divided vertically; the left side is labeled Manual stretching and shows a hand grasping the shoulder, with arrows indicating control based on shoulder reaction force. The right side is labeled Robotic stretching and shows a shoulder brace, with arrows indicating control based on pulling force. Three numbered points below list: 1. Control based on shoulder reaction force, 2. Adjust individual flexibility, 3. Available for older adults. Panel three at the right illustrates the proposed robot, showing a person wearing a shoulder brace, cable guides, and linear actuators connected to the brace, with arrows indicating movement direction.

First, we measured the trajectory of the humeral head during manual therapy by a physical therapist. The following describes the approximation process for determining the pulling direction based on a model with C7 as the origin, as shown in Figure 2(a). The position of the humeral head $ {\boldsymbol{P}}_{HH}\left({x}_1,{y}_1,{z}_1\right) $ in 3-dimensional space represented by C7 as the origin is expressed by the following equation:

(3.1) $$ {\boldsymbol{P}}_{HH}\hskip0.5em =\left(\begin{array}{c}{x}_1\\ {}{y}_1\\ {}{z}_1\end{array}\right)=\left({\displaystyle \begin{array}{c}{l}_1\sin {\theta}_1\cos {\phi}_1\\ {}{l}_1\sin {\theta}_1\sin {\phi}_1\\ {}{l}_1\cos {\theta}_1\end{array}}\right) $$

$ {\hat{\boldsymbol{P}}}_{HH} $ represents the approximated position of $ {\boldsymbol{P}}_{HH} $ on a linear trajectory consisting of the normalized first principal components as follows:

(3.2) $$ {\hat{\boldsymbol{P}}}_{HH}={w}_1{\mathbf{a}}_1+\mathbf{u} $$

where $ {\mathbf{a}}_1 $ denotes the normalized first principal component, $ {w}_1 $ denotes the weight value obtained from the principal component analysis of $ {\boldsymbol{P}}_{HH} $ trajectory, and $ \mathbf{u} $ denotes the center of gravity of the trajectory, as shown in Figure 2(b). Assuming that the change in $ {l}_1 $ and $ {l}_s $ is proportional when the distance from C7 to $ {\boldsymbol{P}}_{HH} $ for another user is $ {l}_s $ , the humeral head position $ {\boldsymbol{P}}_{HH} $ can be expressed by the following equation:

(3.3) $$ {\hat{\boldsymbol{P}}}_{HH}=\frac{l_s}{l_1}\left({w}_1{\mathbf{a}}_1+\mathbf{u}\right) $$

The direction of this approximate line ( $ {\mathbf{a}}_1 $ ) is assumed to be independent of the user’s body size. The pulling direction is fixed by adjusting the width and height of the cable guides according to the $ {l}_s $ and $ {l}_1 $ . By controlling the cable pulling distance and tension with a cable-driven system using linear actuators, load cells, and Bowden cables, the 2-DoF robot can assist both scapular adduction/abduction simultaneously, as shown in Figure 3(a). When the distance from C7 to the cable guide is denoted as $ d $ , the distance between the cable guides, $ w $ , and the height from C7 to the cable guide, $ h $ , are expressed by the following equations:

(3.4) $$ w=2\left|\frac{l_s}{l_1}{u}_y+|\frac{l_s}{l_1}{u}_x-d|\frac{a_{1y}}{a_{1x}}\right|,h=\frac{l_s}{l_1}{u}_z+\mid \frac{l_s}{l_1}{u}_x-d\mid \frac{a_{1z}}{a_{1x}} $$

An example of the cable guide mounting positions is shown in Figure 3(b). The width and height of the cable guides are adjusted according to the C7 position. Bowden cables are also used to reduce friction between the cable and the cable guide. By using redundant wiring, the Bowden cable adapts to changes in the position of the cable guide. A flexible shoulder brace is used for connecting the body and the cable-driven system.

Figure 2.

(a) Shoulder link model. (b) Trajectory of the right humeral head and approximate line. $ {P}_{HH} $ : position of humeral head, $ {\hat{P}}_{HH} $ : approximated position of humeral head.

Figure 2. Long description

Panel a on the left shows a simplified human torso with arms extended, highlighting the right shoulder. The C7 and T7 vertebrae are labeled as spatial anchors. The humeral head is marked with a circle, and its position is denoted as P sub H H with coordinates x sub 1, y sub 1, z sub 1. The shoulder link l sub 1 connects C7 to the humeral head. Three rotation angles phi sub 1, theta sub 1, psi sub 1 are indicated at the shoulder joint. Panel b, to the right, overlays a three-dimensional coordinate system with axes x, y, z centered at O. The axis of thoracic stretching is shown as a dashed line. The position P sub H H is marked by a red dot, and a blue line connects O to P sub H H. An arrow labeled ‘Projection’ points rightward to a two-dimensional plot with axes a sub 1 and a sub 2. On this plot, the trajectory of the humeral head is shown as a curved orange line labeled ‘Trajectory of Humeral Head,’ while a blue line labeled ‘Approximate line’ runs close to it. The projected position, P hat sub H H, is marked by a red dot on the trajectory. The variable u is shown along the trajectory curve.

Figure 3.

(a) Proposed robot configuration and cable guides positioning parameters. (b) The position of the cable guides. By adjusting the height and width of the cable guides, calculated by the shoulder width of the user, the pull direction of the cable-driven system can be aligned with the direction of movement of the user’s humeral head.

Figure 3. Long description

The diagram contains two main sections. In the top row, the left panel shows a top view of a seated human model within a rectangular robot frame, with two colored cables (blue and red) extending from the frame corners toward the shoulders. The width (w) and depth (d) of the frame are labeled. The right panel presents a back view, labeling cable guides at shoulder height (h), Bowden cables, cables, load cells, and linear actuators along the spine and frame. Arrows indicate the cable routing from the actuators upward through the guides. In the bottom row, the left panel is a close-up perspective of the cable guide and Bowden cable positioned near the user’s shoulder, with labels pointing to each component. The right panel shows a wider perspective of the user seated in the robot, highlighting the cable path and guide placement relative to the shoulder. The spatial relationship between cable guides and user anatomy is emphasized, with cable direction aligned to the humeral head movement.

It should be noted that the kinematic modeling of the humeral head described in this section is used only for hardware placement – specifically, to identify the optimal spatial arrangement of the cable guides in relation to individual anatomical landmarks. Since the current prototype cannot perform real-time shoulder position sensing, this model remains static and is not integrated into the active dynamic control loop.

3.2. Shoulder brace

The shoulder brace serves multiple functions beyond simply connecting the shoulder to the cable-driven system. The shoulder brace supports and stabilizes the shoulder joint, and the shoulder brace generally plays a role in pulling the shoulders toward the torso and preventing them from dropping downwards. Its function of restricting shoulder movement is used to achieve combined control around the shoulder joint. In addition, the torque required for shoulder motion varies with the angle of rotation of the upper arm (Kuechle et al., Reference Kuechle, Newman, Itoi, Morrey and An1997). For example, Miyaguchi et al. (Reference Miyaguchi, Nojiri, Matsunaga and Kawaji2008) attempted to support shoulder motion by CPM based on upper arm rotation measurements. The proposed robot uses a shoulder brace to prevent excessive upper arm inner/outer rotation change.

Figure 4 shows a shoulder brace model and views of an older person wearing a brace. The shoulder brace is fixed with Velcro. Velcro is placed on the front of the body, making it easier for older adults to wear it alone. By pulling on the cable, the proposed robot can actively assist in scapular abduction, but releasing the cable does not actively assist in scapular abduction. Therefore, the braces on both sides are connected with elastic fabrics to passively support the movement of returning to the neutral posture during scapular abduction support.

Figure 4.

(a) Shoulder brace link model. The front elastic band passively supports scapular abduction. (b) An older person wearing a brace.

Figure 4. Long description

On the left, a line diagram shows a top view of a shoulder brace. At the center is a dashed circle representing the torso. Outward from the center, two arms extend, each with jointed segments. At the top, a zigzag line labeled elastic band connects the two arms. Arrows labeled scapular abduction point outward from the elastic band, while arrows labeled scapular adduction point inward along the arms. On the right, three photos display an older person wearing the brace. The first photo shows the front view, with the brace crossing the chest and shoulders. The second photo shows the back view, where the brace straps cross in an X shape over the upper back. The third photo shows the side view, highlighting the brace’s fit around the shoulder and upper arm.

3.3. Control

Review papers of the upper-limb control approach show that high-level interaction control strategies can be implemented in a range of methods, mainly including impedance/admittance-based strategies, adaptive control techniques, and physiological signal control (Rahman et al., Reference Rahman, Ouimet, Saad, Kenné and Archambault2010; Miao et al., Reference Miao, Zhang, Cao and Xie2018). The active motion of the scapula and the humeral head are modeled (FVD, Reference FVD1994), but the passive joint ROM is unclear, so an intuitive and simple control method is required.

Approaching the final ROM poses ethical safety concerns; therefore, our control strategy aims to apply high-intensity stretching tailored to the individual. Freitas et al. found a greater ROM in joint stretching with high-intensity loading than with low-intensity stretching (Freitas et al., Reference Freitas, Vaz, Bruno, Andrade and Mil-Homens2016). Therefore, we refer to the physical therapist’s direction of movement based on our previous study (Isobe et al., Reference Isobe, Hirokawa and Suzuki2023) and aim for stretching with force set at a high intensity to suit the individual. As described above in Section 1, the greater the amount of pull on the shoulder, the greater the load required. The tendency of change in load and distance changes significantly near the final ROM. Therefore, in this study, to achieve high-intensity stretching, the ROM is set by setting the maximum tension in advance, rather than by detecting the end-feel and setting the controlled ROM. Depending on the individual’s physique and flexibility, the maximum tension varies and is determined based on the user’s diagnosis. During practical stretching by the robot, the maximum tension is set as the target tension, and traction of the shoulder and adduction of the scapula are realized by position-based force control. When holding the stretching state after the target tension is reached, the difference between the target tension and the current tension is used as an input for position control to cope with fluctuations in shoulder reaction force. Since the robot is not equipped with a system for sensing the user’s shoulder position, the humeral head position used in the model described in Section 3.1 is not used for control. We adopted this position-based force control strategy because setting safety thresholds based on 3D shoulder posture is technically difficult due to the joint’s complex 3D kinematics. Furthermore, electromyography (EMG) was not used because it is difficult to distinguish the stretching state from passive tissue resistance using muscle activity alone.

The robot operation is divided into two stages: the preparation and control processes. The preparation process aims to determine the individualized target tension $ {F}_{\mathrm{ref}} $ based on user feedback. Following this, the control process executes a full stretching cycle consisting of three phases: loading, holding, and unloading.

The preparation process is an empirical protocol for determining the individualized target tension $ {F}_{\mathrm{ref}} $ for each user. As explained in Section 1, the force increases as the pulling distance increases according to the force–displacement properties of a joint. Therefore, we estimate the maximum pulling force by gradually pulling the brace while the user wears the device. The force at a distance considered appropriate by the user is set as the target tension $ {F}_{\mathrm{ref}} $ .

The robot’s control process strategy is mathematically described as follows. The control process performs stretching support based on the $ {F}_{\mathrm{ref}} $ estimated in the preparation process. The robot uses a discrete-time, position-based force control method to ensure stable, safe stretching. The system updates the position command $ {x}_{\mathrm{ref}} $ at each sampling interval based on the tension $ F\left[k\right] $ measured by load cells. At the start of the control process, the control logic for the next step $ k+1 $ is expressed as follows:

(3.5) $$ {x}_{\mathrm{ref}}\left[k+1\right]=\left\{\begin{array}{ll}{x}_{\mathrm{lim}}& F\left[k\right]<{F}_{\mathrm{ref}}\\ {}x\left[k\right]& {F}_{\mathrm{ref}}\le F\left[k\right]<{F}_{\mathrm{max}}\\ {}{x}_0& {F}_{\mathrm{max}}\le F\left[k\right]\end{array}\right. $$

where $ k $ is the discrete time step; $ {x}_{\mathrm{ref}}\left[k+1\right] $ and $ x\left[k\right] $ are the next position command (mm) and current measured position (mm), respectively; $ F\left[k\right] $ and $ {F}_{\mathrm{ref}} $ are the measured and target tensions (N), respectively; $ {F}_{\mathrm{max}} $ is the safety limit tension (N); and $ {x}_{\mathrm{lim}} $ and $ {x}_0 $ are the displacement limit and initial position (mm), respectively.

When the tension reaches the user-specific target $ {F}_{\mathrm{ref}} $ , the system switches to the holding phase. At this transition, the position command $ {x}_{\mathrm{ref}} $ is set to the current measured position $ x\left[k\right] $ to ensure a smooth transition to force maintenance. In this phase, the system uses a proportional control to maintain the force, where the position command is updated as follows:

(3.6) $$ {x}_{\mathrm{ref}}\left[k+1\right]=\left\{\begin{array}{ll}x\left[k\right]+{K}_p\left({F}_{\mathrm{ref}}-F\left[k\right]\right)\Delta t& F\left[k\right]<{F}_{\mathrm{max}}\\ {}{x}_0& {F}_{\mathrm{max}}\le F\left[k\right]\end{array}\right. $$

where $ {K}_p $ is the proportional gain (mm/(N $ \cdot $ s)). Finally, in the unloading phase, the position command is reset to $ {x}_0 $ . This allows the cable to loosen, enabling the scapulae to return to the neutral position, passively assisted by the elastic components of the shoulder brace. Implementation details are described in Section 4.2.

4.1. Mechanical design

Figure 5 shows the robot used in this experiment. The mechanism of this robot has several modifications from the previous robot (Isobe et al., Reference Isobe, Hirokawa and Suzuki2023). First, the shape of the back cushion was changed to support the position of C7 (seventh cervical vertebra) since this robot models shoulder motion using C7 as a reference point. The upper thoracic spine is also fixed, as it is considered important to fix the spine to support thoracic spine joint mobilization.

Figure 5.

The developed robot: (a) Back view, (b) Side view.

Referring to the movements of physical therapists, they perform one dynamic stretch in about 2–3 s. For safety reasons, it is not necessary to control the actuator as fast as they do, but an actuator with a speed approaching that of a therapist’s movements is desirable. For force, the mass of the upper limb should be considered, and an actuator with a force greater than the traction force of the mass should be selected. The specifications of the selected actuators for the proposed robot are shown in Table 1.

Table 1.

Specifications of the proposed robot

Table 1. Long description

From top to bottom, the table has two columns labeled List and Value. The first row under List is Max speed (no load) with Value 18 millimeters per second. The next row is Max force lifted (lowered) with Value 90 newtons. The following row is Back drive force with Value 200 newtons. The final row is Max static force with Value 500 newtons.

4.2. Robot behavior control

According to Section 3.3, the robot realizes position-based force control. Figure 6 shows the block diagram of the proposed system. The linear actuators and motor drivers receive position control commands from a microcontroller called ESP32, which receives target value commands from ROS (Stanford Artificial Intelligence Laboratory et al., 2020). The target tension (represented by $ {F}_{\mathrm{ref}} $ ) and repetition count are decided by the user. To enhance user intuition and ease of use, the target tension is set in units of kgf. The control frequency is about 40 Hz, and the position command $ {x}_{\mathrm{ref}} $ is updated at a sampling interval $ \Delta t=0.025 $ s. This method allows the system to maintain the target tension even if the user moves slightly. For example, if a user suddenly changes posture or resists, causing $ F\left[k\right] $ to increase above $ {F}_{\mathrm{ref}} $ , the error becomes negative. As a result, $ {x}_{\mathrm{ref}}\left[k+1\right] $ becomes smaller than the current $ x\left[k\right] $ , and the cable tension is automatically released within 25 ms. If the tension reaches $ {F}_{max} $ , the target is immediately reset to $ {x}_0 $ to loosen the cables and ensure user safety, and the robot terminates the stretching. In an emergency, the actuators can be stopped or extended to their maximum length by pressing the emergency stop button attached to the robot or the reset button on the operation control panel. In this study, safety is ensured by having the experimenter stay within 2 m of the robot when the user is using it and by immediately stopping the robot in case of an emergency.

Figure 6.

Block diagram of the proposed system.

In the loading phase of the control process, the target is set to $ {x}_{\mathrm{lim}} $ at first. For safety, the robot is configured to reach $ {F}_{\mathrm{ref}} $ in approximately 8 s, slower than the 2 s typical of manual therapy. To prevent prolonged strain if the target tension is not reached, a maximum duration of 12 s is set for this phase. Upon reaching $ {F}_{\mathrm{ref}} $ , the system enters the holding phase and maintains the target tension for a maximum of 5 s.

In addition, the safety limit $ {F}_{\mathrm{max}} $ is set to $ {F}_{\mathrm{ref}}+19.6 $ N. If $ F\left[k\right] $ exceeds $ {F}_{\mathrm{max}} $ during any phase, the system immediately resets to $ {x}_0 $ and terminates the operation within 25 ms.

5.1. Evaluation of the proposed mechanism

To evaluate mechanical properties and ensure safety, we evaluated the capability of the proposed cable-driven mechanism for scapular stretching. The objective was to verify that excessive tension is not applied when using the system on a human shoulder. Instead of using a human participant, we employed a stiff spring to approximate the force–displacement properties of the shoulder joint. This spring simulates the passive torque of a human shoulder, considering the relationship that tension increases proportionally with traction. The mechanism is shown in Figure 7. The model number of the spring used was AWY20-175 (MISUMI), with an initial tension of 12.75 N, a spring constant of .39 N/mm and maximum deflection of 167.5 mm. To evaluate the tracking performance of the end-effector during traction, a spring was selected that could accommodate the maximum displacement required for scapular adduction, simulating the compliance of the shoulder. However, determining an appropriate target tension is difficult due to limited consensus on the actual passive torque of the shoulder joint complex, particularly regarding data from scapular adduction movements used in this protocol. Therefore, to ensure safety in the user study, the maximum tension was set to 49.0 N (5.0 kgf) in this experiment, assuming users with a low maximum tolerable load. Traction started when the wire tension exceeded 1.2 N. The system was controlled to perform the stretching behavior as described in Section 3.3 for 10 times. The positions of the end-effectors of linear actuators were measured by potentiometers mounted on the actuators, and the tensions of the system were measured by the load cells. Only load cell 1 is used for control, and load cell 2 is only used to measure the force on the spring instead of the shoulder. Load cell 2 was not used for measurement in the user study because of the effect on measurement due to contact with the human body.

Figure 7.

Experiment system in Section 5.1. (a) The mechanism of the system. (b) The appearance of the experimental system.

The mean values and standard deviation ranges for 10 trials are shown in Figure 8. The mean values are depicted with a bold line, with the standard deviation ranges shown as a shaded area. In addition, the target tension in the pulling and holding phase is depicted with a gray line. According to Figure 8, observation of the average tension in load cell 1 and the actuator’s pulling distance confirmed that traction ceased once the target of 49 N was reached. However, the tension on load cell 2 did not attain 49 N, and the average tension in load cell 2 reached a maximum of 35.3 N. Furthermore, it was observed that while load cell 1 exhibited oscillations after reaching the target tension, load cell 2 remained stable.

Figure 8.

Changes in the mean and standard deviation of the tension of the two load cells and the pulling distance of the actuator over 10 trials.

Figure 8. Long description

The top panel displays tension in newtons on the y-axis versus step percentage on the x-axis. Loadcell 1 shows a sharp rise to about 45 newtons at 20 percent, then fluctuates between 40 and 50 newtons until 80 percent, followed by a steep decline. Loadcell 2 rises more gradually, peaking near 35 newtons at 20 percent, then plateaus around 30 newtons until 80 percent, before decreasing. Shaded regions indicate standard deviation over 10 trials. The legend at the top right identifies loadcell 1 in blue and loadcell 2 in red. The bottom panel shows actuator position in millimeters on the y-axis versus step percentage on the x-axis. The position increases linearly to about 80 millimeters at 20 percent, remains constant until 80 percent, then decreases linearly back to zero.

5.2. User study with younger adults

To evaluate the robot’s performance on younger adults, we measured the relative surface ROM of the scapula using motion capture while the participants wore the robot. Although surface-based measurements – such as the motion capture markers used in this experiment and the palpation used in the subsequent older adult study – are susceptible to skin motion artifacts and soft-tissue deformation, they provide practical indicators of the robot’s effect. Accordingly, the results presented in this section and Section 5.3 are framed as an assessment of relative surface range-of-motion rather than absolute skeletal kinematics.

5.2.2. Procedure

Before the main experiment, we determined the individualized target tension for each participant using the robot. To determine the individualized target tension, the experimenter manually pulled the cable to gradually increase the stretching intensity. The participants were given the following specific instruction: “We will gradually increase the stretching intensity. Please tell us when it reaches an intensity that feels ‘just right’ without pain. However, if you feel any pain during the adjustment or the experiment, please let us know immediately, and we will stop the device.” The tension value recorded at the point where the participant signaled was set as the target tension for the robotic intervention. This procedure was consistent for both younger and older adult groups. In this experiment, the target tension was set at over 49.0 N(5.0 kgf) to suit the individual. Note that, as indicated in the mechanical evaluation (section 5.1), the actual pulling force applied to the participant’s shoulder was lower than this set value due to friction loss in the power transmission.

Following this preparation, the experiment consisted of the following three stretching conditions:

• Self-stretching: Used to assess neutral posture. Participants received prior instructions on the method of active stretching movements.

• Manual stretching: Conducted by the experimenter (referred to as Manual).

• Robotic stretching: Conducted by the proposed robot (referred to as Robot).

Stretching was performed in the following order: Self-stretching $ \to $ Manual $ \to $ Robot. Although the experimenter was not a physical therapist, he received instruction from a physical therapist regarding the scapular stretching. A minimum rest period of 10 min was implemented between each condition to prevent residual effects and order effects.

To measure scapular motion, the positions of five markers attached to the back were measured using an optical motion capture system. The markers were attached at the positions of the ISB-recommended anatomical landmarks angulus acromialis (AA), trigonum spinae (TS), and angulus inferior (AI) (Wu et al, Reference Wu, van der, Veeger, Makhsous, Van Roy, Anglin, Nagels, Karduna, McQuade, Wang, Werner and Buchholz2005), indicated as blue dots in Figure 9. However, in some cases, these markers were often occluded by the shoulder brace during the experiment. Therefore, referring to the fact that the distance between the AI and the spine is a common metric for evaluating scapular adduction, we attached markers at positions easily visible from the camera on a straight line connecting the AI and the TS, indicated as red dots in Figure 9. Then, the distance between these markers was calculated as the interscapular distance to evaluate scapular adduction. It should be noted that this metric relies on surface landmarks. Consequently, the measurements inevitably include the effects of soft tissue deformation and skin motion artifacts, which may differ slightly from the underlying skeletal kinematics. However, this method was considered sufficient for evaluating the relative changes in ROM within the scope of this study. Markers were attached to the skin over the scapula, ensuring the right and left markers were at the same height. Note that the position of the markers, the initial distance between the scapulae, and the position of the band of the brace differed among the participants due to the individual differences in shape and size of the scapula. Moreover, we attached markers in C7 and T7 as reference points to check for misalignment of the other markers.

Figure 9.

The position of retroreflective marker in the experiment of Section 5.2.

Figure 9. Long description

At the center, two red markers are placed on the scapular plane, connected by a horizontal orange arrow labeled interscapular distance. Above, C 7 is marked at the base of the neck. Moving outward and upward, T S and A A are labeled at the upper scapular region, with blue dots marking their positions. Below, T 7 is labeled at the lower thoracic spine, and A I is marked at the lower scapular angle. The schematic uses dashed lines to connect labels to anatomical points, and shaded areas indicate the scapulae.

An example of the change in interscapular distance for each stretching condition is shown in Figure 10. The minimum value of the interscapular distance in each condition was used to estimate the effect of stretching on scapular adduction.

Figure 10.

Interscapular distance of a participant in each condition.

5.2.4. Result

The box plot of the interscapular distances for all participants in the Neutral, Manual, and Robot conditions is shown in Figure 11. Significant differences were confirmed between manual stretching and neutral posture and between robotic stretching and neutral posture ( $ W=0 $ , adjusted $ p<0.05 $ ). Wilcoxon signed-rank test showed a significant difference between robotic stretching and manual stretching ( $ W=0 $ , adjusted $ p=0.023<0.05 $ ). However, the mean difference between the two conditions was 5.1 mm (130.9 mm for the robotic stretching and 125.8 mm for the manual stretching). This difference was only 7.1% of the mean difference observed between neutral posture and manual stretching. In addition, no one complained of pain or discomfort during the stretching.

Figure 11.

Box plot of the interscapular distance of participants in each condition, along with the results of statistical analysis *: $ p<.05 $ ).

5.3. User study with older adults

To evaluate the robot’s performance and short-term effect on older adults, scapula–spine distances and respiratory function were measured with the robot worn by the older adults. The purpose of this experiment was to clarify the feasibility of robotic stretching for older adults compared to manual stretching by the physical therapist and its effect on the immediate improvement of respiratory function. The target tension was set at over 49.0 N (5 kgf) to suit the individual.

5.3.2. Procedure

The procedure for determining the individualized target tension was consistent with that used in the younger adults study (Section 5.2). Based on the participants’ feedback regarding the maximum comfortable intensity, the target tension settings for the older adults ranged from 98.0 N (10.0 kgf) to 137.3 N (14.0 kgf) at the actuator side. Note that, as indicated in the mechanical evaluation (Section 5.1), the actual pulling force applied to the participant’s shoulder was lower than these set values due to friction loss in the power transmission.

Four conditions of scapula–spine distance were measured. (1) neutral posture (Neutral), (2) self-stretching, (3) manual stretching by a physical therapist (Manual), and (4) robotic stretching (Robot). Considering that the ROM of the scapula is generally increased by the physical therapist’s treatment or robotic stretching, self-stretching was measured at first. Self-stretching was performed to confirm the participant’s ROM and to confirm that it was safe to perform the robotic stretching. The procedure of this experiment is shown in Figure 12. The distances of Neutral conditions were measured before self-stretching. Five participants performed the robotic stretching first, followed by manual stretching, and the remaining four participants received manual stretching and then received robotic stretching. Participants performed stretching 10 times for each condition. In addition, the participants in the experiment measured FVC (forced vital capacity), FEV1.0 (Forced Expiratory Volume in 1 s), and the FEV1.0% (FEV1/FVC) using a spirometer (CHESTGRAPH HI-301 U, CHEST M.I., Inc.) before and after the two conditions (Manual and Robot) mentioned above.

Figure 12.

Measurement procedure in experiment of Section 5.3.

Figure 12. Long description

At the top, the first sequence labeled N equals 5 shows four boxes in a row: Instruction, Self, Manual, Robot, separated by three vertical blue bars representing 10 minutes rest. Below, the second sequence labeled N equals 4 shows four boxes: Instruction, Self, Robot, Manual, also separated by three blue bars. From the Robot box, a dashed line branches downward to a sequence of three boxes: F V C, Stretching, F V C, with F V C and Stretching defined in the legend below. The legend at the bottom clarifies that F V C means F V C Measurement, the blue bar means 10 minutes rest, and Stretching means Stretching 10 times.

Figure 13 shows the scapula–spine distance used as an assessment parameter. We used an optical motion capture system in Section 5.2, but in this experiment, we used a digital caliper because the marker occlusion by the user’s and therapist’s bodies and waiting bare-chested is a burden for older adults. Although this metric provides a one-dimensional measurement of three-dimensional scapular motion, measuring the scapula–spine distance via palpation is a standard clinical practice in manual therapy for assessing scapular adduction (Johnson, Reference Johnson and Johnson2012). Therefore, we considered this method valid and sufficient for verifying the feasibility of the robot-assisted stretching in this study. Using a skin marker pen, draw a line at the location of the spine and a dot at the location of the AI. Since the dots were written on the surface of the skin, the position of the dots (AI) can be displaced from the actual position of the scapula during stretching. To ensure consistency and accuracy, all measurements were performed by a single experimenter who had been trained in palpation techniques by a licensed physical therapist. The stretching was performed 10 times. The first five trials were excluded to allow for the convergence of initial shoulder brace displacement and to enable the experimenter to confirm stable robot operation. Consequently, the latter five times were measured. To ensure a steady state during the measurement in manual stretching, the physical therapist was instructed to hold the scapulae in the maximally adducted position for 5 s. The average value was calculated as the scapula–spine distance.

Figure 13.

Measurement position in experiment of Section 5.3.

For respiratory function evaluation, FEV1.0 and FEV1.0%, which are commonly used for COPD assessment, were adopted for FVC measurement. Although there is a study showing significant differences in FVC and FEV1.0 values even in immediate manual therapy for COPD patients (Yilmaz Yelvar et al., Reference Yilmaz Yelvar, Çirak, Demir, Dalkilinç and Bozkurt2016), it should be noted that this study was conducted on healthy older adults. The participants were instructed by a physical therapist on how to use the stretching device, practiced using it, and then rested enough before the measurement.

5.3.4. Result

Figure 14 summarizes the average scapula–spine distance of all participants for Neutral, Manual, and Robot conditions as box plots. There were significant differences between the condition of neutral posture and the condition of robotic stretching ( $ W=0 $ , adjusted $ \mathrm{p}<.05 $ ) and between the condition of the neutral posture and the condition of manual stretching ( $ W=0 $ , adjusted $ \mathrm{p}<.05 $ ). There was no significant difference between the manual stretching and the robotic stretching. In addition, the physical therapist interviewed the patients, but no one complained of pain or discomfort from stretching.

Figure 14.

Box plot of the scapula–spine distance of participants in each condition, along with the results of statistical analysis in Section 5.3 (*:adjusted $ \mathrm{p}<.05 $ ; n.s.: not significant).

The results of the pre-and post-FEV1.0 changes for each condition are shown in the box plots in Figure 15. Due to missing data, we analyzed data from eight participants for the Robot condition and nine for the Manual condition. Due to a technical error in the data logging process, data for one participant in the Robot condition were lost. Consequently, we analyzed data from eight participants for the Robot condition and nine for the Manual condition. It is noted that this data loss occurred randomly and was unrelated to the participant’s performance or fatigue. Friedman tests were conducted for the precondition in all stretching conditions. No significant differences were found when the significance level was set at 5%. Wilcoxon signed-rank sum tests were conducted on the FEV, FEV1.0, and FEV1.0% values pre- and post-stretching. No differences were found in all conditions when the significance level was set at 5%.

Figure 15.

Box plot of pre-and post-FEV1.0 for each condition in Section 5.3 (n.s.: not significant).