2.1. Requirements definition and concept selection

PDS establish quantitative targets derived from clinical needs and biomechanical analysis. Table I summarizes key specifications organized into functional, performance, dimensional, and system categories.

Table I.

Product design specifications for coordinated shoulder-elbow cable-driven exosuit.

Table I. Long description

The table presents the design specifications for a coordinated shoulder-elbow cable-driven exosuit, organized into four main categories: Functional, Performance, Dimensional, and System. It has four rows and three columns. The columns are labeled Category, Parameter, and Specification. Row 1: Category, Functional; Parameter, Target joints; Specification, Shoulder (2 DOF: flex/ext, abd/add); Elbow (1 DOF: flex/ext). Row 2: Category, Functional; Parameter, Assistance type; Specification, Admittance control with biomechanical modeling. Row 3: Category, Performance; Parameter, Range of motion; Specification, 0-90 degrees all DOF (preserve functional ROM). Row 4: Category, Performance; Parameter, Torque; Specification, Shoulder: 4-6 Nm; Elbow: 3-5 Nm. Row 5: Category, Performance; Parameter, Workspace coverage; Specification, 95 percent functional ADL volume. Row 6: Category, Dimensional; Parameter, Total mass; Specification, less than or equal to 5 kg (battery, actuators, structure). Row 7: Category, Dimensional; Parameter, Form factor; Specification, Wearable, portable, clothing-compatible. Row 8: Category, System; Parameter, Actuation; Specification, Cable-driven, centralized motor unit. Row 9: Category, System; Parameter, Sensing; Specification, Force sensors, IMU (100 Hz). Row 10: Category, System; Parameter, Portability; Specification, Battery-powered, 2 h operation.

The torque specifications (shoulder 4–6 Nm, elbow 3–5 Nm) represent the target assistive torque range at the joint anchor points, derived from ADL biomechanical requirements for the 5th–95th percentile population. These differ from the worst-case gravitational torques computed in Section 2.2 (7.8 Nm shoulder, 2.6 Nm elbow), which represent maximum loading at $90^\circ$ elevation with full arm weight. The PDS torque targets correspond to the desired assistive fraction (gravity compensation assistance level) rather than the total gravitational load, consistent with the admittance control strategy that provides partial unloading proportional to user effort.

Figure 1.

Systematic design methodology framework.

The 3-DOF requirement (2-DOF shoulder + 1-DOF elbow) responds to biomechanical evidence that 85–90% of ADL tasks require coordinated shoulder-elbow movements [Reference Murray, Buchanan and Delp10], while single-joint assistance creates compensatory muscle activation in adjacent joints. The portability constraint ( $\lt$ 5 kg) addresses clinical adoption barriers where stationary devices remain confined to institutional settings. Range of motion specifications (0– $90^\circ$ all DOF) derive from functional workspace analysis, where reaching, feeding, and manipulation tasks concentrate within this operational envelope.

State-of-art analysis identified representative portable cable-driven systems demonstrating effective single-joint assistance: the FALCO system provides 2-DOF shoulder assistance (flexion/extension, abduction/adduction) through anchor-point force sensing [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22], while the elbow-focused device by Lotti et al. achieves 1-DOF elbow support using force-based control [Reference Lotti, Xiloyannis, Missiroli, Bokranz, Chiaradia, Frisoli, Riener and Masia23]. Both systems exemplify portable cable-driven actuation with demonstrated clinical efficacy for isolated joint movements. However, critical analysis reveals fundamental limitation: neither system supports coordinated shoulder–elbow motion essential for functional ADL execution. Anchor-point force sensing strategies, while effective for single-joint configurations, impose spatial constraints during simultaneous multi-joint movements due to sensor positioning requirements at distributed cable attachment points on adjacent body segments, fundamentally limiting achievable workspace in portable multi-joint configurations.

This identified gap motivated exploration of alternative sensing architectures enabling 3-DOF coordination while preserving cable-driven portability advantages. The proposed motor-proximal sensing configuration addresses this limitation through strategic relocation of force sensors from distributed anchor points to motor-adjacent locations combined with geometric compensation algorithms, preserving force measurement capability while eliminating spatial constraints that restrict coordinated movement workspace.

Table II.

MCDM evaluation comparing cable-driven alternatives for multi-joint upper limb assistance. Weights ( $W_{C_i}$ ) derive from PDS priorities; ratings ( $R_{iy}$ ) calculated using explicit formulas detailed in text.

Table II. Long description

The table compares three alternatives for cable-driven multi-joint upper limb assistance across twelve criteria. It has three rows labeled as FALCO, Lotti et al., and Proposed device, and thirteen columns labeled as C1 to C12 and Score. Each row contains numerical values representing weights and ratings for each criterion, with the final score calculated for each alternative. Panel A: The first row, FALCO, shows values 1.67, 1, 2, 0.75, 3, 3, 1, 1, 0, 0, 0.67, 2, and a score of 106.7. Panel B: The second row, Lotti et al., shows values 2.94, 1, 2, 0.85, 3, 3, 0, 0, 1, 0, 0.33, 2, and a score of 96.6. Panel C: The third row, Proposed device, shows values 1.43, 1, 2, 0.90, 3, 3, 1, 1, 1, 0, 1.00, 2, and a score of 122.6.

MCDM provided quantitative evaluation of actuation alternatives. The MCDM framework assigns weights $W_{C_i}$ to criteria based on relative importance, derived from PDS priorities. The overall score for alternative $y$ follows the weighted sum approach with portability constraint:

(1) \begin{equation} \mathrm{Score}_y = B_y \times \sum _{i=1}^{12} P_{iy} = B_y \times \sum _{i=1}^{12} (W_{C_i} \times R_{iy}) \end{equation}

where $B_y = R_{2y}$ is the portability boolean (nonportable devices receive zero score), $W_{C_i}$ is the weight of criterion $i$ , and $R_{iy}$ is the performance rating of alternative $y$ on criterion $i$ . Table II presents the complete evaluation matrix comparing the two representative single-joint systems against the proposed integrated design. The twelve criteria cover: C1 mass (weight 6), C2 portability (10), C3 self-wearability (5), C4 available torque (4), C5 fixing stability (8), C6 safety (10), C7–C10 DOF coverage boolean indicators for shoulder flex/ext, shoulder abd/add, elbow flex/ext, and wrist (weights 6, 4, 6, 3, respectively), C11 ROM compliance (10), and C12 control complexity (6). Rating calculations follow explicit, reproducible formulas. C1 (Mass): $R_{1y} = M_{max}/M_{dy}$ , where $M_{max}=5.0$ kg (PDS maximum) and $M_{dy}$ is device mass. Applied values: FALCO $M_d=3.0$ kg (shoulder-only configuration) $\to R_1=1.67$ ; Lotti et al. $M_d=1.7$ kg (elbow exosuit) $\to R_1=2.94$ ; Proposed $M_d=3.5$ kg (design target) $\to R_1=1.43$ . C2 (Portability): Boolean ( $R_2=1$ if battery-powered with no external infrastructure, 0 otherwise). All alternatives achieve $R_2=1$ . C3 (Self-wearability): Discrete scale (0: requires assistance; 1: difficult but achievable; 2: easily achievable). All cable-driven exosuit architectures achieve $R_3=2$ through adjustable textile interfaces. C4 (Available Torque): $R_{4y} = \min _{j \in \mathrm{DoF}_y}(T_{available,j}/T_{required,j})$ . Required torques from PDS: shoulder 6 Nm, elbow 5 Nm. FALCO: estimated shoulder capability $\approx$ 4.5 Nm $\to R_4=0.75$ ; Lotti et al.: elbow 4.25 Nm average $\to R_4=0.85$ ; Proposed: worst-case elbow 4.5 Nm $\to R_4=0.90$ . C5 (Stability of Fixing): Discrete scale (1–3). All cable-driven soft architectures employ multi-point textile interfaces achieving $R_5=3$ . C6 (Safety): Discrete scale (1–3). Cable-driven soft architectures inherently limit injury risk through compliant transmission: $R_6=3$ for all alternatives. C7–C10 (DOF coverage): Boolean indicators for shoulder flex/ext, shoulder abd/add, elbow flex/ext, wrist. FALCO: $R_7=1, R_8=1, R_9=0, R_{10}=0$ ; Lotti: $R_7=0, R_8=0, R_9=1, R_{10}=0$ ; Proposed: $R_7=1, R_8=1, R_9=1, R_{10}=0$ . C11 (ROM vs. Functional ROM compliance): $R_{11y} = \frac {1}{3}\sum _{l=1}^{3} x_{ly}$ , where $x_{ly} = R_{jy} \times (S_{ly}/S_{lf})$ represents DOF-specific workspace compliance ( $S_{lf}=90^\circ$ for all DOF from PDS). FALCO: $R_{11}=(1\times 1+1\times 1+0)/3=0.67$ ; Lotti: $R_{11}=(0+0+1\times 1)/3=0.33$ ; Proposed: $R_{11}=(1+1+1)/3=1.00$ . C12 (Control Complexity): Discrete scale (3: simple direct mapping; 2: moderate force/myoelectric; 1: high complexity adaptive). All evaluated alternatives employ force-based control: $R_{12}=2$ . The quantitative analysis reveals a systematic gap: FALCO (106.7) and Lotti (96.6) achieve incomplete functional DOF coverage, with scores primarily limited by C9 and C7/C8, respectively. The proposed architecture (122.6) achieves the highest score through complete 3-DOF coverage (C7–C9) and full ROM compliance (C11), at the cost of slightly higher mass. This analysis provided systematic guidance for the design direction, identifying coordinated multi-joint coverage as the primary differentiating criterion.

2.2. Component dimensioning

Systematic dimensioning establishes component specifications through worst-case biomechanical analysis, ensuring adequate performance across the target population (95th percentile: height H = 1750 mm, body mass M $_{body}$ = 75 kg). Anthropometric scaling follows established relationships [Reference Winter25]: upper arm mass m $_{ua}$ = 2.1 kg, forearm mass m $_{fa}$ = 1.2 kg, hand mass m $_h$ = 0.45 kg; upper arm length L $_{ua}$ = 325 mm, forearm length L $_{fa}$ = 255 mm.

Gravitational torque analysis determines actuator requirements. For shoulder assistance accounting for upper arm, forearm, and hand contributions, maximum loading occurs at $90^\circ$ shoulder flexion with elbow extended:

(2) \begin{equation} \tau _{shoulder,gravity}(\theta _{y1}) = \left [m_{ua} \cdot \frac {L_{ua}}{2} + (m_{fa} + m_{hand}) \cdot L_{ua}\right ] \cdot g \cdot \sin (\theta _{y1}) \end{equation}

yielding $\tau _{shoulder,max}$ = 7.8 Nm at $90^\circ$ elevation. For elbow assistance accounting for forearm and hand:

(3) \begin{equation} \tau _{elbow,gravity}(\theta _{y1}, \theta _{y2}) = \left [m_{fa} \cdot \frac {L_{fa}}{2} + m_{hand} \cdot L_{fa}\right ] \cdot g \cdot \sin (\theta _{y1} + \theta _{y2}) \end{equation}

where $\theta _{y1} + \theta _{y2}$ gives the absolute forearm angle, yielding $\tau _{elbow,max}$ = 2.6 Nm when the forearm is horizontal. Note that these worst-case values reflect the maximum gravitational loading at $90^\circ$ elevation with fully extended arm, and exceed the PDS torque targets (4–6 Nm shoulder, 3–5 Nm elbow): the PDS specifications define the desired assistive torque fraction for partial gravity compensation, not the total gravitational load.

Figure 2 presents gravitational torques across operational range, showing shoulder torque reaching 7.8 Nm at $90^\circ$ flexion and elbow torque reaching 2.6 Nm maximum.

Figure 2.

Gravitational torques during $0^\circ$ – $90^\circ$ operational range for representative user (H=1750 mm, M $_{body}$ =75 kg). Shoulder torque (blue) reaches maximum 7.8 Nm at $90^\circ$ flexion; elbow torque (red) reaches maximum 2.6 Nm when forearm is horizontal.

Cable force requirements incorporate variable moment arm geometry and transmission efficiency. MATLAB kinematic simulation confirms moment arms vary 55–95 mm shoulder (mean 75 mm) and 25–78 mm elbow (mean 51 mm). Accounting for Polytetrafluoroethylene (PTFE) Bowden conduit transmission efficiency ( $\eta _{trans}$ = 0.88):

(4) \begin{equation} F_{cable,required} = \frac {\tau _{gravity}}{\eta _{trans} \cdot r(\mathbf{q})} \end{equation}

where $r(\mathbf{q})$ denotes the configuration-dependent moment arm. Peak forces occur at approximately $73^\circ$ shoulder flexion (118 n) and $61^\circ$ elbow flexion (58 N), shown in Figure 3.

Figure 3.

Required cable forces during $0^\circ$ – $90^\circ$ operational range. Peak forces 118 N (shoulder, blue) and 58 N (elbow, red) occur where ratio $\tau _{gravity}/r(\mathbf{q})$ is maximum.

Motor torque requirements with direct pulley drive (r $_p$ = 20 mm) and safety factor (SF = 1.5) for dynamic loads yield $\tau _{motor,shoulder}$ = 3.5 Nm and $\tau _{motor,elbow}$ = 1.7 Nm required. Selected actuator: AK60-6 quasi-direct drive motor (6 Nm continuous) provides adequate capacity with conservative margins. Force sensors: DYMH 103 S-type load cells (200 N, $\pm$ 0.5% FS). Cables: Kevlar braided Ø1.5 mm, breaking load $\gt$ 500 N. Bowden: PTFE liner Ø4 mm.

Table III summarizes realized system specifications demonstrating PDS compliance.

Table III.

System specifications demonstrating PDS compliance.

Table III. Long description

The table has four rows and three columns. The columns are labeled Category, Component/parameter, and Value. Row 1: Category, Mass; Component/parameter, Total system mass; Value, 3.3 kg. Row 2: Category, Actuation; Component/parameter, Motor torque (rated); Value, 6 Nm. Row 3: Category, Actuation; Component/parameter, Cable tension (max shoulder/elbow); Value, 118/58 N. Row 4: Category, Actuation; Component/parameter, Transmission efficiency; Value, 0.88. Row 5: Category, Sensing; Component/parameter, Load cells (3 DYMH 103); Value, 200 N, 0.5% FS. Row 6: Category, Sensing; Component/parameter, IMUs (2 BMI085); Value, 6-axis, 100 Hz. Row 7: Category, Performance; Component/parameter, Operational ROM (all DOF); Value, 0.

Systematic dimensioning achieves 3.3 kg total mass below the 5 kg target while providing specified torque capabilities across full operational workspace. Figure 4 presents complete computer-aided design (CAD) assembly integrating commercial components with custom 3D-printed structures.

Figure 4.

Complete CAD assembly: (1) Bowden cable; (2) Kevlar wire; (3) mounting shoulder; (4) cantilever mechanism; (5) shoulder slider guide; (6) shoulder anchor point; (7) elbow anchor point; (8) actuation unit with three motors and control electronics.

3.1. Sensing architecture and design rationale

Current cable-driven exoskeletons position force sensors at anchor points to measure actual applied forces after transmission losses. This configuration provides direct force measurement advantageous for single-joint systems where cable routing can be optimized for one degree of freedom. However, coordinated multi-joint assistance introduces critical spatial constraints.

During coordinated movements combining shoulder elevation with elbow flexion, sensor positioning at anchor points on adjacent segments (upper arm and forearm) limits range of motion through mutual physical interference. The implemented architecture addresses this limitation through strategic load cell repositioning from anchor points to actuation unit locations. Relocating force sensors from cable anchor points to positions adjacent to the motors enables elbow flexion range of 0– $90^\circ$ while maintaining shoulder elevation range of 0– $90^\circ$ , meeting PDS specifications. This configuration eliminates mechanical constraints at distributed anchor points while requiring geometric compensation to translate motor-proximal measurements into effective anchor-point forces.

Figure 5 illustrates the actuation unit integrating motors, motor-proximal load cells, control electronics, and pulley system within compact posterior-mounted configuration.

Figure 5.

Actuation unit with motor-proximal sensing. Internal configuration shows brushless motors (1), dual microcontroller system (2), strategically repositioned load cells (3), and integrated power management (battery 14.8V, 3,700 mAh) (4).

Figure 5. Long description

A diagram of an actuation unit with various labeled components. Panel A: The diagram shows an actuation unit with four main components. The brushless motors are labeled as 1 and there are three of them. The dual microcontroller system is labeled as 2. The strategically repositioned load cells are labeled as 3, with two visible. The integrated power management, which includes a battery of 14.8V and 3,700 mAh, is labeled as 4.

3.2. Geometric compensation

Motor-proximal load cell positioning measures force at motor-side pulleys rather than anchor points, requiring geometric compensation to estimate effective cable tension throughout the workspace. The pulley mounts directly on the load cell body, with cable routing through the pulley and extending through Bowden conduits to anchor points on limb segments. In an ideal cable system with negligible bending stiffness, cable tension remains approximately uniform along its length. The load cell measures the resultant force from both cable segments entering and exiting the pulley. For cable under uniform tension $T$ routing through the pulley:

(5) \begin{equation} F_{pulley} = 2T \end{equation}

where the factor of 2 accounts for both cable segments contributing to the measured force. Effective cable tension transmitted to the anchor point, accounting for Bowden conduit transmission losses, is

(6) \begin{equation} T_{anchor} = \eta _{trans} \cdot \frac {F_{pulley}}{2} \end{equation}

where $\eta _{trans}=0.88$ was established through experimental characterization of the PTFE-lined Bowden conduits in Section 2.2.

The geometric compensation algorithm translates the estimated anchor tension $T_{anchor}$ into the effective assistive torque at the joint, accounting for the configuration-dependent moment arm. Following the model established in ref. [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22] for shoulder cantilever geometry (Figure 6), the cable length $k(\alpha )$ from the cantilever to the arm anchor as a function of shoulder elevation angle $\alpha$ is computed via the law of cosines:

(7) \begin{equation} k(\alpha ) = \sqrt {k_1^{*2} + k_2^{*2} + 2k_1^* k_2^* \cos (\gamma ^* + \beta ^* + \alpha )} \end{equation}

where $k_1^*$ , $k_2^*$ are fixed geometric parameters derived from anthropometric measurements ( $l_b^*$ , $h_b^*$ , $l_a^*$ , $h_a^*$ ) and $\gamma ^*$ , $\beta ^*$ are fixed angles of the cantilever configuration. The Jacobian relating cable displacement to joint angle gives the effective moment arm:

(8) \begin{equation} J(\alpha ) = \frac {\partial k}{\partial \alpha } = \frac {-k_1^* k_2^* \sin (\gamma ^* + \beta ^* + \alpha )}{\sqrt {k_1^{*2} + k_2^{*2} + 2k_1^* k_2^* \cos (\gamma ^* + \beta ^* + \alpha )}} \end{equation}

The estimated assistive torque at the shoulder joint is then:

(9) \begin{equation} \tau _i = J(\alpha ) \cdot T_{anchor} \end{equation}

This torque estimate $\tau _i$ is compared with the reference gravity-compensation torque $\tau _r$ (Section 3.3) to compute the tracking error for the admittance controller.

The geometric compensation algorithm continuously updates calculations at 100 Hz based on real-time joint configuration from distributed IMU sensors (BMI085, 6-axis). Quaternion-based orientation representation avoids gimbal lock during large-amplitude coordinated movements. IMU fusion integrates accelerometer and gyroscope data to determine shoulder elevation angle $\alpha$ (0– $90^\circ$ ) and elbow flexion angle $\theta$ (0– $90^\circ$ ), enabling continuous Jacobian update throughout the operational envelope.

Figure 6 illustrates the geometric configuration, including the cantilever mechanism for shoulder assistance and distributed anchor points for elbow support.

Figure 6.

Geometric configuration showing shoulder assistance with cantilever mechanism and elbow assistance with cable routing from shoulder anchor to forearm attachment. Motor-proximal load cells measure cable tensions while preserving extended range of motion. Parameters $k_1^*$ , $k_2^*$ , $\gamma ^*$ , $\beta ^*$ are the fixed cantilever geometry constants used in the compensation algorithm.

3.3. Control system implementation

The control system implements personalized biomechanical modeling using XYZ Euler angle parametrization [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22] to provide coordinated assistance across shoulder and elbow joints. The shoulder joint is modeled as a 2-DOF revolute system within the control framework (elevation $\alpha$ and longitude $\beta$ ), consistent with the 2-DOF actuation design. The parametrization follows [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22]:

(10) \begin{equation} \alpha = \arccos \bigl (c(\phi _2)\,c(\phi _1)\bigr ) \end{equation}

(11) \begin{equation} \beta = \arctan \!\left (-\frac {s(\phi _2)}{c(\phi _2)\,s(\phi _1)}\right ) \end{equation}

where $\phi _1$ , $\phi _2$ are the XYZ Euler angles extracted from IMU quaternions, $c(\cdot )$ denotes cosine, and $s(\cdot )$ denotes sine.

Under quasi-static conditions (joint velocities and accelerations are low), the gravitational potential energy of the upper limb kinematic chain is

(12) \begin{equation} U = -g\bigl \{\cos (\alpha )\bigl [K_{ul} + K_l\cos (\theta )\bigr ] - \sin (\alpha )\cos (\beta )\bigl [K_l\sin (\theta )\bigr ]\bigr \} \end{equation}

where $K_{ul} = m_u l_{Gu} + m_l l_u$ and $K_l = m_l l_{Gl}$ are anthropometric constants ( $m_u$ , $m_l$ : upper arm and forearm+hand masses; $l_{Gu}$ , $l_{Gl}$ : distances of respective centers of mass from proximal joint), $\alpha$ is the shoulder elevation, $\beta$ is the arm longitude, and $\theta$ is the elbow flexion. This expression follows directly from the z-coordinates of the arm segment centers of mass in the world frame, as derived in ref. [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22].

The reference gravity-compensation torque at the shoulder is

(13) \begin{equation} \tau _r = g\bigl \{\sin (\alpha )\bigl [K_{ul} + K_l\cos (\theta )\bigr ] + \cos (\alpha )\cos (\beta )\bigl [K_l\sin (\theta )\bigr ]\bigr \} + D_1\dot {\alpha } + D_2\dot {\theta } \end{equation}

where $D_1$ and $D_2$ are positive damping coefficients added to model biological joint damping and ensure smooth controller behavior [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22]. These coefficients were determined empirically through iterative experimental tuning using the Ziegler–Nichols method as initial estimate, following the same procedure reported in ref. [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22] for the FALCO exosuit. Specifically, the tuning procedure minimized torque tracking error during controlled sinusoidal reference trajectories while maintaining stability across the five participants. The resulting values were consistent across participants ( $D_1 = 0.15 \pm 0.03$ Nm $\cdot$ s/rad, $D_2 = 0.08 \pm 0.02$ Nm $\cdot$ s/rad), confirming robustness of the empirical identification.

The low-level admittance controller converts the torque tracking error $\tau _{error} = \tau _r - \tau _i$ into motor velocity commands. In the Laplace domain [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22, Reference Keemink, van der Kooij and Stienen30]:

(14) \begin{equation} \frac {\dot{\!x}(s)}{\tau (s)} = R(s) = B_a + \frac {D_a}{s} + M_a s \end{equation}

where $B_a$ , $D_a$ , and $M_a$ are the admittance proportional, integral, and derivative gains, respectively, tuned empirically for stable torque tracking.

The system implements two parallel synchronized control loops. Shoulder tensioning control incorporates the high-level gravity compensation model Eq. (13) with the low-level admittance controller Eq. (14), translating torque error into motor speed commands. Shoulder translating control utilizes a secondary motor to reposition the anchor point on the slider, aligning the cable direction with the highest elevation point of the arc (tracking $\beta$ through a PID position controller with adaptive fading filter). Elbow control implements an independent admittance loop with geometric compensation adapted to the elbow cable routing. All three loops share IMU kinematic data.

Figure 7 presents the CAD representation of the coordinated control architecture.

Figure 7.

CAD representation of coordinated control architecture showing three parallel loops: elbow assistance (red), shoulder tensioning (blue, high-level feedforward gravity compensation + low-level admittance feedback), and shoulder translating (blue, PID position control of anchor geometry). Each subsystem operates independently while sharing IMU sensor information.

Figure 7. Long description

A CAD representation of a coordinated control architecture showing three parallel loops: elbow assistance, shoulder tensioning, and shoulder translating. The elbow assistance control is highlighted in red, the shoulder tensioning control is highlighted in blue with high-level feedforward gravity compensation and low-level admittance feedback, and the shoulder translating control is highlighted in blue with PID position control of anchor geometry. Each subsystem operates independently while sharing IMU sensor information.

Figure 8 presents the detailed control flow. The feedforward pathway (gravity compensation model) provides the reference torque $\tau _r$ ; the feedback pathway (admittance controller) closes the loop using the interaction torque $\tau _i$ estimated from motor-proximal load cells via the geometric compensation Eqs. (5–9).

Figure 8.

Control flow diagram. The shoulder tensioning loop (magenta) combines feedforward gravity compensation (reference torque $\tau _r$ from biomechanical model, Eq. (13) with feedback admittance control (interaction torque $\tau _i$ from motor-proximal load cells via geometric compensation, Eq. (9). The translating loop (green) tracks the longitude angle $\beta$ via PID position control. The elbow loop (orange) implements independent admittance control with geometric compensation of motor-proximal force measurements. All subsystems share joint angle estimates ( $\alpha$ , $\beta$ , $\theta$ ) from IMU sensors, where $\omega _{ref}$ denotes the motor velocity command for the translating actuator.

IMU sensors positioned on upper arm and forearm provide three-dimensional orientation processed through quaternion representation at 100 Hz via BLE UART protocol [Reference Wang, Zhang, Kong, Su, Yuan and Zhao27].

5.1. Phase 1: Sensing accuracy assessment

Before proceeding to full system human trials, the motor-proximal sensing approach required experimental validation confirming geometric compensation algorithms maintain adequate control accuracy across extended ROM necessary for functional multi-joint coordination.

Comprehensive validation employed extensive data collection across complete operational workspace (shoulder elevation 0– $90^\circ$ , elbow flexion 0– $90^\circ$ ) during human subject trials. Data were aggregated across all five participants and all experimental conditions, resulting in measurement samples spanning the complete functional range.

Kinematic data were collected at 100 Hz. For each movement configuration, the system recorded: (1) theoretical gravitational torque $\tau _{theory}$ computed from the biomechanical model Eq. (12), (2) motor-proximal measured forces $F_{pulley}$ , and (3) geometrically compensated torque estimates $\tau _{estimated}$ via Eqs. (6–9), and (4) estimation error $\epsilon = \tau _{estimated} - \tau _{theory}$ .

Aggregate analysis across the complete workspace revealed R $^2$ = 0.75, RMSE = 1.51 Nm, and near-zero mean error (0.13 Nm) indicating absence of systematic bias. However, workspace-dependent analysis reveals critical spatial variation:

• • 0– $30^\circ$ elevation: R $^2$ = 0.66, RMSE = 1.08 Nm – moderate correlation with acceptable error for admittance control.

• • 30– $60^\circ$ elevation: R $^2$ = 0.22, RMSE = 2.26 Nm – degrading correlation reflecting increasing geometric sensitivity.

• • 60– $90^\circ$ elevation: R $^2$ = 0.06, RMSE = 3.40 Nm – poor correlation due to near-zero Jacobian $J(\alpha )$ Eq. (8) at elevated angles, where small changes in cable geometry produce large estimation uncertainty.

This workspace-dependent degradation is a fundamental consequence of the motor-proximal architecture: at high shoulder elevation, the effective moment arm approaches its minimum value, amplifying measurement uncertainty. The 0– $60^\circ$ range encompasses the primary ADL workspace (reaching, feeding, and object manipulation), where sensing accuracy is adequate for functional admittance control. The 60– $90^\circ$ range, relevant for overhead tasks, presents a known limitation requiring future architectural refinement (discussed in Section 6).

Figure 11 presents overall validation results demonstrating correlation between theoretical and measured torques with unbiased error distribution.

Figure 11.

Gravity compensation validation: (a) Correlation between theoretical and measured torques (R $^2$ = 0.75, RMSE=1.51 nm); (b) error distribution showing near-zero bias (mean=0.13 nm, std=1.51 nm).

Figure 11. Long description

Panel A: A scatter plot shows the correlation between theoretical gravity torque and estimated torque. The x-axis represents theoretical gravity torque in Newton meters (Nm), and the y-axis represents estimated torque in Newton meters (Nm). The plot includes a red dashed line indicating perfect agreement. The R-squared value is 0.747, and the root mean square error (RMSE) is 1.51 Nm. The data points are scattered around the line, showing a general trend of increasing estimated torque with increasing theoretical gravity torque. Panel B: A histogram displays the error distribution between theoretical and estimated torques. The x-axis represents the error in Newton meters (Nm), and the y-axis represents the frequency. The histogram shows a near-zero bias with a mean error of 0.13 Nm and a standard deviation of 1.51 Nm. The majority of the errors are concentrated around zero, indicating a good agreement between theoretical and estimated torques.

5.2. Phase 2: Human subject trials

5.2.1. Experimental methods

Five healthy male participants (age 29 $\pm$ 5 y, height 1.8 $\pm$ 0.1 m, body mass 73 $\pm$ 9 kg) were recruited through convenience sampling. The homogeneous demographic profile was intentionally selected to minimize anthropometric confounding variables in this proof-of-concept study. The limitation of this choice for generalizability is acknowledged in Section 6.

Participants met selection requirements: (1) anthropometric measurements compatible with exosuit parameters (height 1.65–1.95 m, body mass 60–90 kg), (2) normal joint ROM across all upper limb articulations, (3) sufficient muscle strength to complete testing protocol without assistance, and (4) absence of neurological, musculoskeletal, or cardiovascular conditions.

All experimental procedures were conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants. Equipment contacting participants was thoroughly disinfected between users using standard sterilization protocols.

Experimental protocol employed within-subjects repeated-measures design. Participants completed all tasks under two randomized conditions: Exo On (full exosuit assistance with active gravity compensation) and No Exo (sensor-only configuration with IMUs worn, no mechanical assistance).

Controlled ROM tasks consisted of three primary movements performed at $90^\circ$ amplitude with 5 repetitions $\times$ 3 series each (Figure 12): (a) elbow flexion–extension, (b) shoulder flexion–extension, and (c) shoulder abduction-adduction. Reference trajectories were sinusoidal profiles at approximately 0.2 Hz. Functional tasks included a standardized drinking sequence (Figure 13).

Figure 12.

Controlled range of motion tasks performed at standardized $90^\circ$ amplitude with fixed repetition structure enabling systematic comparison across assistance conditions.

Figure 13.

Functional drinking task demonstrating complex multi-joint coordination: (a) initial reach, (b) precision grasping, (c) coordinated lifting and drinking motion, (d) controlled return to rest position.

Each participant completed a single 2-hour session: (1) 10-min familiarization, (2) baseline measurements (maximal voluntary contraction (MVC) collection for electromyography (EMG) normalization, anthropometric calibration), (3) randomized task execution with 5-min rest periods between conditions, and (4) subjective feedback collection. Figure 14 illustrates the complete data acquisition infrastructure deployed during experimental trials.

Figure 14.

Data acquisition setup with dual-pathway architecture: kinematic processing station and wireless EMG system, both providing real-time signal visualization during trials.

5.2.2. Functional outcomes

Surface EMG data were acquired using a wireless Delsys system targeting three key muscles: biceps brachii (elbow flexion), anterior deltoid (shoulder flexion), and medial deltoid (shoulder abduction). Electrode placement followed SENIAM guidelines (Figure 15).

Figure 15.

EMG electrode placement following SENIAM guidelines: (a) biceps brachii and anterior deltoid positioning; (b) medial deltoid placement optimized to avoid interference with cable routing during multi-joint movements.

EMG preprocessing: bandpass filtering (15–450 Hz, 4th-order Butterworth), full-wave rectification, low-pass filtering (6 Hz cutoff), normalization to individual MVC values collected at the start of each session following standardized isometric contraction protocols. Root mean square (RMS) values were computed over complete movement cycles. The reported RMS values are expressed as fractions of MVC (dimensionless, where 1.0 = 100% MVC). Values above 1.0 observed in some muscles reflect the higher activation demands of the multi-joint coordinated tasks relative to the isolated MVC reference contractions, which were collected for each muscle individually at a fixed posture and may not capture peak activation across all task postures. Kinematics: Joint angles derived from IMU quaternions recorded at 100 Hz.

Statistical Analysis: Paired $t$ -tests compared conditions. With three muscles tested, Bonferroni-corrected significance threshold is $\alpha _{corr} = 0.05/3 = 0.017$ . Effect sizes reported as Cohen’s $d$ (small: 0.2, medium: 0.5, large: 0.8). Degrees of freedom df = 4 (n $-$ 1).

Table IV presents aggregate EMG analysis across all experimental tasks.

Table IV.

Aggregate muscle activation analysis. RMS values represent normalized activation (fraction of MVC) averaged across tasks. Bonferroni-corrected significance threshold $\alpha _{corr}=0.017$ (3 comparisons). $\dagger$ denotes significance after Bonferroni correction.

Table IV. Long description

The table presents data on muscle activation with and without an exoskeleton. It has five rows and six columns. The columns are labeled Muscle, No exo (RMS±SD), Exo on (RMS±SD), Reduction (percent), p-value, and Cohen’s d. The rows are labeled Biceps Brachii, Anterior Deltoid, Medial Deltoid, and Average. Row 1: Biceps Brachii, No exo 1.294±0.265, Exo on 1.059±0.162, Reduction 18.2 percent, p-value 0.031, Cohen’s d 0.9. Row 2: Anterior Deltoid, No exo 1.802±0.335, Exo on 1.636±0.285, Reduction 9.2 percent, p-value 0.089, Cohen’s d 0.6. Row 3: Medial Deltoid, No exo 2.037±0.321, Exo on 1.473±0.249, Reduction 27.7 percent, p-value 0.012, Cohen’s d 1.2. Row 4: Average, No exo, Exo on, Reduction 18.4 percent ± 9.3. The table shows the reduction in muscle activation when using an exoskeleton, with notable reductions in the Medial Deltoid and Biceps Brachii muscles.

After Bonferroni correction ( $\alpha _{corr}=0.017$ ), medial deltoid reduction (27.7%, p = 0.012, d = 1.2) reaches statistical significance, representing a large effect consistent with effective shoulder abduction assistance. Biceps brachii reduction (18.2%, p = 0.031, d = 0.9) shows a large effect size but does not survive Bonferroni correction, and should be interpreted as exploratory. Anterior deltoid (9.2%, p = 0.089, d = 0.6) approaches a medium effect but is non-significant. The 18.4% average reduction across muscles is reported as indicative of clinically meaningful assistance: reductions of this magnitude are consistent with thresholds associated with measurable fatigue reduction in endurance-based rehabilitation protocols reported in the cable-driven exosuit literature [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22, Reference Lotti, Xiloyannis, Missiroli, Bokranz, Chiaradia, Frisoli, Riener and Masia23], although direct clinical significance thresholds vary by application and patient population.

Figure 16 presents representative electromyographic activation patterns during the pick-and-place task.

Figure 16.

Electromyographic analysis during pick-and-place task: (a) temporal EMG activity showing activation patterns throughout movement cycles; (b) RMS values normalized to MVC comparing no exo (blue) and exo on (green) conditions.

Figure 16. Long description

The image contains two sets of graphs. Panel A shows three line graphs depicting temporal EMG activity for the medial deltoid, anterior deltoid, and biceps muscles. The x-axis represents time in seconds, and the y-axis represents EMG activity as a percentage of maximum voluntary contraction (percent MVC). The green lines indicate the exoskeleton on condition, while the blue lines indicate the no exoskeleton condition. Panel B shows three bar graphs comparing the root mean square (RMS) values normalized to MVC for the same muscles. The x-axis is labeled RMS, and the y-axis represents the RMS values. The green bars represent the exoskeleton on condition, and the blue bars represent the no exoskeleton condition. The graphs illustrate differences in muscle activation patterns and RMS values between the two conditions.

Movement quality assessment demonstrated preserved joint coordination during assisted conditions. Trajectory tracking: elbow flexion–extension R $^2$ = 0.82 $\pm$ 0.071, shoulder flexion–extension R $^2$ = 0.64 $\pm$ 0.14, shoulder abduction-adduction R $^2$ = 0.65 $\pm$ 0.15.

The functional drinking task provided comprehensive evaluation of naturalistic movement patterns. Figure 17 shows representative kinematic profiles, demonstrating preserved coordination across multiple cycles. The temporal offset in EMG and kinematic peak timing between Exo On and No Exo conditions (most visible in Figure 18) reflects the dynamic response of the admittance controller: the exosuit assistance introduces a compliance-like behavior that slightly shifts the timing of peak muscle recruitment and joint angles relative to the unassisted condition. This effect is intrinsic to admittance control and has been reported in comparable cable-driven systems [Reference Missiroli, Ciaramella, Radaelli, Rambaldi, Frisoli, Casadio and Masia22, Reference Keemink, van der Kooij and Stienen30]. The angular differences between conditions in shoulder kinematics during the drinking task arise primarily from the altered torque balance at the shoulder joint: with active gravity compensation, the effective gravitational load perceived by the muscles is reduced, leading to modified trajectory profiles as the neuromuscular system adapts to the assistive input.

Movement precision via RMSE: elbow $13.04 \pm 1.31^\circ$ , shoulder flexion $16.36 \pm 4.16^\circ$ , shoulder abduction 13.37 $\pm 3.32^\circ$ .

Range of motion analysis revealed minimal kinematic constraints. Shoulder flexion: 87.2 $\pm 3.4^\circ$ (Exo On) vs. 89.8 $\pm \,2.9^\circ$ (No Exo), 2.9% reduction. Shoulder abduction: 85.9 $\pm 3.8^\circ$ vs. 88.7 $\pm 3.2^\circ$ , 3.2% reduction. Elbow flexion: 88.9 $\pm 2.7^\circ$ vs. 90.1 $\pm 2.4^\circ$ , minimal constraint. All DOF achieved $\gt 85^\circ$ ROM under assisted conditions.

5.3. Comparative positioning

Table V presents systematic comparison establishing the system’s position within the technological landscape. Heterogeneous metrics, different participant populations, and different task protocols preclude direct performance ranking; comparison focuses on capability positioning.

Table V.

Representative upper limb cable-driven assistance systems. Heterogeneous metrics, tasks, and participant populations preclude direct performance comparison.

Table V. Long description

A table comparing upper limb cable-driven assistance systems. The table has five rows and five columns. Column headers are System, DOF assisted, Mass (kg), Sensing, and Primary outcome. Row 1: FALCO, 2 (shoulder), 3.0, Anchor point, 31% avg. activation reduction. Row 2: Lotti et al., 1 (elbow), 1.7, Anchor point, 35% fatigue reduction. Row 3: Kim et al., 2 (rigid), 7.5, Force, 26% overhead task. Row 4: Proposed system, 3 (coord.), 3.3, Motor-prox., 18% multi-joint avg.

$^\dagger$ Across coordinated shoulder-elbow tasks; only medial deltoid reduction survives Bonferroni correction.

Figure 17.

Kinematic tracking performance during standardized movement tasks for representative participant. Green: Exo On condition, blue: No Exo condition, red dashed: sinusoidal reference trajectory ( $90^\circ$ amplitude). Similar tracking patterns observed across all participants.

Figure 17. Long description

Three line graphs depict elbow flexion-extension, shoulder flexion-extension, and shoulder abduction-adduction angles over time. Panel A: The line graph shows elbow flexion-extension angles in degrees over time in seconds. The green line represents the Exo On condition, the blue line represents the No Exo condition, and the red dashed line represents the sinusoidal reference trajectory. Panel B: The line graph shows shoulder flexion-extension angles in degrees over time in seconds. The green line represents the Exo On condition, the blue line represents the No Exo condition, and the red dashed line represents the sinusoidal reference trajectory. Panel C: The line graph shows shoulder abduction-adduction angles in degrees over time in seconds. The green line represents the Exo On condition, the blue line represents the No Exo condition, and the red dashed line represents the sinusoidal reference trajectory.

Figure 18.

Kinematic analysis during functional drinking task showing coordinated shoulder and elbow joint angles. Green: Exo On, blue: No Exo. Preserved coordination across movement cycles demonstrates system maintains natural multi-joint coordination patterns during complex functional activities.

The proposed system achieves multi-joint coordination unavailable in prior portable systems. Lower muscle activation reduction relative to FALCO and Lotti reflects task complexity rather than inferior performance: single-joint systems concentrate all actuation on one DOF, while coordinated assistance distributes torque across three DOF simultaneously. Direct performance comparison is further complicated by different experimental tasks and participant populations.