XoSoft Exosuit: Modular Soft Robotics

• XoSoft Exosuit is a family of soft wearable robotic systems featuring compliant, modular designs for upper-limb augmentation, gait assistance, and rehabilitation.
• The systems employ innovative actuation methods—cable-sheath transmissions, clutch-spring mechanisms, and dynamic textile logic—to enhance natural movement and minimize user effort.
• Co-simulation and experimental studies demonstrate significant reductions in muscle torque and metabolic cost, emphasizing the importance of personalized anchor configurations and online adaptation.

Searching arXiv for XoSoft exosuit papers to ground the article in current literature. XoSoft Exosuit denotes a set of soft wearable robotic systems reported under the XoSoft name, spanning upper-limb augmentation, lower-limb gait assistance, and textile-integrated rehabilitation garments. Across these studies, the common design premise is the use of compliant, body-conformal structures rather than rigid exoskeleton frames, with assistance delivered through cable–sheath transmissions, clutch-spring series actuation, or embedded textile logic. The literature therefore presents XoSoft less as a single fixed device than as a research lineage with multiple embodiments, including a cable-driven upper-arm soft exosuit with adaptive gravity compensation, the quasi-passive lower-limb XoSoft Gamma, and a dynamic field programmable logic-driven fabric soft exosuit for rehabilitation tasks (Mukherjee et al., 2023, Lambranzi et al., 31 Jul 2025, Cleary et al., 2023, Mariani et al., 14 Nov 2025, Yadav et al., 2023).

1. Scope and system variants

The XoSoft literature covers three principal configurations: an upper-limb cable-driven exosuit, a lower-limb quasi-passive gait exosuit, and a fabric-based programmable rehabilitation exosuit. The reported systems differ in actuation, sensing, and task domain, but each uses soft interfaces, textile anchoring, or compliant transmission to reduce restriction of natural movement.

Variant	Core architecture	Primary research focus
Upper-limb cable-driven exosuit	Single-cable tendon–sheath mechanism; DC motor with gearbox and spool	Adaptive gravity compensation and online parameter learning
XoSoft Gamma lower-limb exosuit	Quasi-passive, cable-pulley and clutch-spring series actuator–based lower-limb exosuit	Energetics, biomechanics, anchor-point placement, and mental workload
Dynamic field programmable fabric exosuit	Three-layer textile panel with interchangeable logic-gate patches and embedded sensors	Motion-triggered reconfigurability for rehabilitation exercises

For the lower-limb platform, XoSoft Gamma is reported as having a total system mass of 4.4 kg for the backpack, electronics, and pneumatic system, with actuation modules based on clutch plus spring elements of stiffness $k = 1.6\ \mathrm{N/mm}$. For the upper-limb platform, the emphasis is instead on assistive elbow torque generation and adaptive estimation of human-arm and payload parameters. For the dynamic fabric prototype, the emphasis shifts from force augmentation to sensing, logic reconfiguration, and rehabilitation task encoding through textile circuitry (Lambranzi et al., 31 Jul 2025, Mukherjee et al., 2023, Cleary et al., 2023).

This diversity is significant because it shows that XoSoft research addresses not only actuation and control, but also attachment geometry, physiological burden, and garment-level programmability. A plausible implication is that “XoSoft” functions as a modular soft-exosuit research framework rather than a single immutable hardware artifact.

2. Upper-limb cable-driven architecture and transmission modeling

Mukherjee et al. describe the upper-arm XoSoft exosuit as using a single-cable tendon–sheath mechanism routed from a DC motor, with gearbox and spool of radius $R_m$, to two soft straps, one on the upper arm and one on the forearm. As the motor rotates by an angle $\theta_m$, it generates a cable tension

$F_1 = \frac{N\,\tau_m}{R_m},$

where $N$ is the gear ratio and $\tau_m$ the motor torque. Because the cable runs through a low-friction sheath of total curvature $\phi$, the tension immediately beyond the sheath is

$$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$

with $\mu$ the cable–sheath friction coefficient. The instantaneous moment arm is

$$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$

where the extension function $R_m$0 captures how cable length varies with elbow angle $R_m$1. The assistive elbow torque is then

$R_m$2

with

$R_m$3

This formulation makes the elbow-assist transmission explicitly state-dependent and exposes the role of sheath friction in reducing delivered torque (Mukherjee et al., 2023).

Yadav et al. extend the transmission analysis to a 3D double tendon-sheath actuator system for upper-limb soft exosuit use. In that model, two tendons, agonist/flexor and antagonist/extensor, each run through a low-friction sheath, pass over a virtual pulley at the elbow, and wrap or unwrap on two motor-mounted spools of radii $R_m$4 and $R_m$5. Each tendon is pretensioned to $R_m$6 in the unloaded state and connected in series to a linear spring of stiffness $R_m$7. Slack is identified as a principal failure mode, and the study reports that slack can be effectively controlled by changing the pretension, spring constant, and size and geometry of the spool mounted on the axle of motor. Parameter sweeps with a 30 mm flexor spool, $R_m$8, $R_m$9, tendon diameter $\theta_m$0, $\theta_m$1, and elbow pulley radius $\theta_m$2 yielded recommended design ranges of $\theta_m$3, $\theta_m$4, and $\theta_m$5–0.9. The summary states that this configuration yields near-zero slack over the full $\theta_m$6–$\theta_m$7 elbow flexion under typical payloads up to 5 kg while maintaining motor torques $\theta_m$8 and user comfort (Yadav et al., 2023).

Together, these studies define the upper-limb XoSoft architecture as a soft-strapped, cable-mediated transmission whose performance is governed by friction, moment-arm geometry, pretension, and slack suppression. This suggests that control performance and transmission design are tightly coupled rather than separable design problems.

3. Arm dynamics and adaptive gravity compensation

For controller design, the wearer’s forearm plus any hand-held payload is modeled as a planar single-joint manipulator. The upper-limb dynamics are written as

$\theta_m$9

where $F_1 = \frac{N\,\tau_m}{R_m},$0 is forearm mass, $F_1 = \frac{N\,\tau_m}{R_m},$1 the center-of-gravity distance, $F_1 = \frac{N\,\tau_m}{R_m},$2 the moment of inertia, $F_1 = \frac{N\,\tau_m}{R_m},$3 the damping, $F_1 = \frac{N\,\tau_m}{R_m},$4 the payload located at distance $F_1 = \frac{N\,\tau_m}{R_m},$5 from the elbow axis, $F_1 = \frac{N\,\tau_m}{R_m},$6 the net human-generated torque, and $F_1 = \frac{N\,\tau_m}{R_m},$7 the exosuit assistive torque. In this 1-DOF planar model there are no Coriolis or centrifugal terms; the gravity term $F_1 = \frac{N\,\tau_m}{R_m},$8 is the only configuration-dependent nonlinear effect (Mukherjee et al., 2023).

The controller introduces the combined gravity parameter

$F_1 = \frac{N\,\tau_m}{R_m},$9

and the known regressor $N$0, so that the dynamics become

$N$1

If $N$2 were known, a pure gravity-compensation law would be

$N$3

which yields the gravity-free closed loop $N$4. The adaptive gravity compensation (AGC) controller replaces $N$5 with its time-varying estimate $N$6: $N$7 With parameter estimation error $N$8, the closed loop becomes

$N$9

The adaptive law stacks the unknowns into $\tau_m$0, with regressor $\tau_m$1, giving

$\tau_m$2

The prediction error is

$\tau_m$3

Using the Lyapunov candidate $\tau_m$4, $\tau_m$5, the summary reports $\tau_m$6, and the decoupled adaptive law

$\tau_m$7

This ensures $\tau_m$8 and all closed-loop signals remain bounded; under persistent excitation one can further show $\tau_m$9 (Mukherjee et al., 2023).

A distinctive feature of this formulation is that it does not assume knowledge of the anthropometric parameters of the wearer’s arm and the payload, and it is agnostic to the desired joint trajectory followed by the human arm. In the context of soft exosuits, that is methodologically important because anthropometry and payload can vary across users and tasks, while trajectory prescription can conflict with natural human motion.

4. Co-simulation results and upper-limb performance

The upper-limb AGC controller was evaluated in a MATLAB + OpenSim co-simulation. OpenSim used the arm26 musculoskeletal model for human biomechanics and metabolic cost via a 5-term heat/work model, while MATLAB computed control $\phi$0 every time step and fed it into OpenSim’s external force interface. The elbow reference was a sinusoid,

$\phi$1

over 4 s, and four payload cases were examined: 0, 3, 5, and 10 kg. The reported performance metrics were the root-mean-square of human-generated torque $\phi$2, metabolic cost reduction $\phi$3 versus unassisted, and convergence of $\phi$4 (Mukherjee et al., 2023).

The key simulation findings are quantitative. Even with unknown $\phi$5, AGC quickly learns the gravity term: the 2-norm $\phi$6 decays in the first second. Human torque RMS reductions versus unassisted were reported as follows: no load, GC 35% and AGC 28%; 3 kg, GC 65% and AGC 50%; 5 kg, GC 67% and AGC 66%; 10 kg, GC 67% and AGC 66%. Metabolic cost drops were reported up to 67% for GC and 63% for AGC across increasing loads. The assistive torque $\phi$7 and motor torque $\phi$8 profiles for AGC rapidly match those of the ideal GC controller. The summary therefore states that AGC converges to GC performance after an initial learning transient (Mukherjee et al., 2023).

These results support two technical conclusions. First, online parameter learning can recover most of the performance of exact gravity compensation without prior anthropometry or payload information. Second, the dominant benefit reported in simulation is not merely reduction in actuator effort, but reduction in human effort in terms of human muscle torque and metabolic cost. The real-world implication stated in the summary is that the single-cable, soft-strap architecture minimally restricts natural arm kinematics while delivering up to two-thirds reduction in wearer’s effort.

5. Lower-limb XoSoft Gamma: actuation, anchor geometry, and gait effects

The lower-limb XoSoft Gamma is described as a quasi-passive, cable-pulley and clutch-spring series actuator–based lower-limb exosuit. Its total system mass is 4.4 kg, consisting of backpack, electronics, and pneumatic system. Actuation modules comprising clutch plus spring with $\phi$9 deliver hip flexion and extension assistance via elastic energy storage, and a finite-state machine control triggers clutch engagement and disengagement based on foot-contact events measured by FSRs in instrumented insoles. The cable path runs from the hip actuator, through a low-friction sheath, down to the knee loop, then up to the contralateral hip anchor to complete a force-transmission loop; soft textile straps maintain pre-tension and ensure the cable path follows a near-muscle line of action, specifically iliopsoas for hip flexion. A dog-bone aluminum frame of $$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$0 at each loop spreads load over approximately $$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$1 of textile, minimizing pressure peaks (Lambranzi et al., 31 Jul 2025).

Anchor points are explicitly defined relative to palpable bony landmarks: Hip–Anterior just above the anterior superior iliac spine on the frontal belt; Hip–Posterior aligned with the posterior superior iliac crest on the back of the backpack belt; Knee–Anterior-Central centered on the patellar tendon; Knee–Anterior-Lateral on the lateral side of the patella, approximately 2 cm lateral to Knee–Anterior-Central; and Knee–Posterior in the popliteal fossa, approximately 3 cm behind the knee joint line. Six 8-min treadmill trials at 5 km/h were performed with 11 healthy subjects under randomized conditions: baseline without exosuit and five anchor configurations, namely HA + KP, HA + KC, HA + KL, HP + KP, and HP + KC (Lambranzi et al., 31 Jul 2025).

The reported outcome measures include net metabolic rate, EMG-derived indices, and sagittal-plane hip, knee, and ankle kinematics from IMUs. Median NMR values were 2.673 W/kg for NoExo, 2.898 for HA + KP, 2.914 for HA + KC, 2.864 for HA + KL, 2.920 for HP + KP, and 2.940 for HP + KC. Friedman’s test gave $$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$2 across all six conditions and $$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$3 excluding NoExo. Pairwise significant increases versus NoExo were reported for HA + KP ($$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$4), HA + KC ($$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$5), HA + KL ($$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$6), HP + KP ($$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$7), and HP + KC ($$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$8). At the same time, subject-specific best savings reached $$F_2 = \begin{cases} F_1 e^{-\mu\phi}, & \dot\theta_m \ge 0,\ F_1 e^{\mu\phi}, & \dot\theta_m < 0, \end{cases}$$9 for Subject 10 in Configuration D and $\mu$0 for Subject 6 in Configuration E, while variability included increases up to 43.9% (Lambranzi et al., 31 Jul 2025).

EMG results show that the best net rectus femoris assist was Configuration B, HA + KP, with Overall Interaction Index $\mu$1, while the best biceps femoris assist was Configuration D with OII $\mu$2, the only increase reported for BF. Kinematic changes were not confined to the assisted joint: hip flexion peaks decreased in all exo conditions versus NoExo, with the maximum reduction in Configuration F of $\mu$3 from $\mu$4 to $\mu$5, and knee and ankle were also altered despite hip-only actuation. Dynamic Time Warping distances indicated the largest hip deviation in B and E and the largest knee and ankle deviations in C and F. The summary therefore emphasizes that no single configuration was optimal across all subjects and that a personalized approach is necessary to transmit the assistance forces optimally (Lambranzi et al., 31 Jul 2025).

This body of results directly counters two common simplifications. The first is that anchor-point placement is a secondary implementation detail; the reported data treat it as a crucial determinant of energetics, muscle activation, and kinematics. The second is that a hip-assist exosuit affects only the hip; the reported gait alterations at the knee and ankle show system-level biomechanical coupling.

6. Mental workload and physiological assessment during assisted walking

Mental workload has been evaluated on the lower-limb XoSoft platform using subjective and objective measures. Mariani et al. studied 18 healthy adults walking on a treadmill at 3.5 km/h for 2 min per trial under three device modes—No Exo, Exo OFF, and Exo ON—and three task conditions: simple walking, an “Easy” counting-down task, and a “Hard” subtract-3 task. The exosuit configuration consisted of two soft leggings with adjustable straps for the hip and ankle, a backpack of 4.4 kg housing electronics, vacuum pump, valves, and pneumatic plumbing, silicone insoles with FSRs for gait-phase detection, and a quasi-passive series actuator using pneumatic clutch, steel spring, and elastic band. Assistance was delivered symmetrically at hip flexion/extension and ankle dorsiflexion, and gait was segmented by a finite-state machine based on FSR thresholds (Mariani et al., 14 Nov 2025).

Three mental-workload metrics were compared: the raw NASA-TLX questionnaire, the average percentage change in pupil size (APCPS), and the Baevsky stress index. APCPS was based on

$\mu$6

where $\mu$7 is current pupil diameter and $\mu$8 the baseline mean pupil diameter, with APCPS defined as the mean of PCPS over each 2-min trial. The stress index was defined from the RR-interval histogram as

$\mu$9

and converted to decibels by

$$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$0

Tobii Pro Glasses 3 provided binocular eye tracking at 100 Hz, while an Empatica E4 wristband provided BVP at 64 Hz and heart rate at 1 Hz (Mariani et al., 14 Nov 2025).

The reported findings show that overall NASA-TLX increased with task difficulty in all three exo modes. For simple walking, Exo OFF versus No Exo yielded $$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$1 and Exo ON versus No Exo $$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$2; for dual tasks, no significant NASA-TLX differences across exo modes were reported. Median APCPS increased from Simple to Easy to Hard across modes, with the right eye more sensitive: for Exo ON, Easy versus Hard for the right eye gave $$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$3, while the left eye showed no pairwise significance. Approximate mean APCPS for the right eye increased from about 0.3% in No Exo Simple to about 1.0% in No Exo Hard, and from about 0.5% in Exo ON Simple to about 1.3% in Exo ON Hard. By contrast, $$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$4 showed no statistically significant changes with task or exo mode and stayed approximately 20–22 dB across all conditions. Correlations between APCPS and NASA-TLX subscales were stronger for the right eye, reaching $$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$5 for Mental Demand, $$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$6 for Effort, and $$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$7 for Frustration (Mariani et al., 14 Nov 2025).

The interpretation advanced in the summary is that pupillometry, particularly right-eye APCPS, offers an objective correlate to the gold-standard NASA-TLX in assessing mental workload during exosuit-assisted walking, whereas cardiac-based SI was less sensitive under locomotion. The reported discussion also notes that backpack weight contributed to perceived Physical Demand and recommends quieter pneumatic systems and adaptive assistance modulated by real-time pupillometry. This suggests that lower-limb exosuit evaluation cannot be reduced to metabolic or kinematic benefit alone; cognitive load and perceived burden are also design-relevant system outputs.

7. Textile logic, rehabilitation functions, and cross-cutting design implications

Cleary et al. report a distinct XoSoft configuration: a dynamic field programmable logic-driven fabric soft exosuit. Its core architecture is a three-layer textile panel mounted on the back of a shrug-style jacket. Layer 1 provides power and ground buses embroidered using stainless-steel conductive thread; Layer 2 is an interchangeable conductive-thread interconnect panel; and Layer 3 consists of modular logic-gate patches, specifically AND, OR, and NOT, built from conductive nylon fabric, SMD 0805 resistors of 1 k$$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$8 and 10 k$$J_f(\theta) = \frac{\partial h(\theta)}{\partial \theta},$$9, 2N2222 NPN transistors, and 9 mm snap fasteners. Because gates and interconnect panels are physically removable, the wearer or clinician can swap logic functions without unstitching or rewiring the garment (Cleary et al., 2023).

The sensing architecture uses six momentary textile “push” switches as motion detectors. On the control arm, wrist flexion closes clusters 2 and 4 to activate Circuit 1, while elbow flexion closes clusters 1, 3, and 5 to activate Circuit 2. On the exercise arm, one sensor each is located at the wrist, elbow, four-finger pad, and thumb pad. Both circuits implement

$R_m$00

with truth table

$R_m$01

and rehabilitation mappings such as $R_m$02 wrist flexion or finger flexion, $R_m$03 elbow flexion or thumb flexion, and $R_m$04 both. Sensors are sewn into a Lycra gauntlet with zig-zag stitches, and velcro straps over the elbow region allow exact placement on each patient’s anatomy (Cleary et al., 2023).

Experimental validation included logic-gate characterization and motion-based exercise tests. Each gate’s output voltage was measured three times per input combination, with threshold $R_m$05 interpreted as logic 1. Example Circuit 1 readings for input $R_m$06 were: NOT$R_m$07 output 4.88 V, NOT$R_m$08 0.12 V, AND$R_m$09 4.02 V, AND$R_m$10 0.12 V, and OR 3.23 V. Range-of-motion tests recorded activation at wrist flexion 52°, elbow flexion 107°, thumb flexion 69°, and finger flexion 88°. Repetitive trials of 10 cycles per exercise showed consistent ON-state voltages, such as 2.80–2.91 V for Circuit 1 wrist and 2.85–3.03 V for Circuit 1 elbow, while reliability exceeded 95% correct LED indication over 100 flexion/extension cycles. Overall system switching was reported as less than 1 ms, dominated by momentary-switch bounce and LED turn-on (Cleary et al., 2023).

Viewed alongside the upper-limb AGC controller and the lower-limb XoSoft Gamma studies, this textile-logic prototype broadens the meaning of XoSoft beyond mechanical assistance. It demonstrates that the XoSoft research space also includes reconfigurable wearable computation, with future enhancements proposed as scaling to more than 8 inputs and outputs, replacing SMD parts with all-textile passive components, adding pressure, temperature, or biopotential sensors, and conducting human-subject trials to correlate exosuit data with Fugl–Meyer and other clinical outcomes. Across the XoSoft literature more generally, the stated future directions are hardware validation of AGC, sensing of $R_m$11, optimization of transmission $R_m$12 for reduced frictional losses, lightweight redesign of the 4.4 kg backpack, adaptive clutch timing, integration of subject anthropometry into pre-fitting simulation, and validation in elderly or clinical populations (Mukherjee et al., 2023, Lambranzi et al., 31 Jul 2025, Cleary et al., 2023).

Taken together, these studies characterize XoSoft as a soft-exosuit research program centered on compliant actuation, subject-specific attachment, online adaptation, and wearable sensing or logic. The persistent technical theme is not merely softness as a material property, but softness as a systems strategy: compliant transmission, soft anchoring, low-profile sensing, and reconfigurable textile interfaces are all used to couple assistance or task monitoring to the body while preserving natural movement.