This study describes the design, simulation, and development process of a rehabilitation glove driven by soft pneumatic actuators. A new, innovative finger soft actuator design has been developed through detailed kinematic and workspace analysis of anatomical fingers and their actuators. The actuator design combines cylindrical and ribbed geometries with a reinforcing element—a thicker, less extensible structure—resulting in an asymmetric cylindrical bellow actuator driven by positive pressure. The performance of the newly designed actuator for the rehabilitation glove was validated through numerical simulation in open-source software. The simulation results indicate actuators’ compatibility with human finger trajectories. Additionally, a rehabilitation glove was 3D-printed from soft materials, and the actuator’s flexibility and airtightness were analyzed across different wall thicknesses. The 0.8 mm wall thickness and thermoplastic polyurethane (TPU) material were chosen for the final design. Experiments confirmed a strong linear relationship between bending angle and pressure variations, as well as joint elongation and pressure changes. Next, pseudo-rigid kinematic models were developed for the index and little finger soft actuators, based solely on pressure and link lengths. The workspace of the soft actuator, derived through forward kinematics, was visually compared to that of the anatomical finger and experimentally recorded data. Finally, an ergonomic assessment of the complete rehabilitation glove in interaction with the human hand was conducted.

Loss of hand function due to conditions like stroke or multiple sclerosis impacts daily activities. Robotic rehabilitation provides tools to restore hand function, while surface electromyography (sEMG) enables the adaptation of the device’s force output to the user’s condition, thus enhancing rehabilitation outcomes. This study focuses on accurately predicting grip force during medium wrap grasps using a single sEMG sensor pair, addressing the challenge of escalating sensor requirements. We conducted sEMG measurements on 13 subjects at two forearm positions, validating results with a hand dynamometer. Established flexible signal processing steps achieved high peak cross-correlations between the processed sEMG signal and grip force. Influential parameters were subsequently identified through sensitivity analysis. Leveraging a novel data-driven Koopman-based approach and problem-specific data lifting, we devised a method for the estimation and short-term prediction of grip force from processed sEMG signals. The method achieved a weighted mean absolute percentage error (wMAPE) of ~5.5% for grip force estimation and ~17.9% for 0.5-second predictions. The methodology proved robust regarding precise electrode positioning, as the effect of sensing position on error metrics was non-significant. The algorithm executes exceptionally fast, processing, estimating, and predicting a 0.5-second sEMG signal batch in just ~30 ms, facilitating real-time implementation.

Loss of hand functions, often manifesting in the form of weakness or spasticity from conditions like stroke or multiple sclerosis, poses challenges in performing activities of daily living (ADLs). The broad area of rehabilitation robotics provides the tools and knowledge necessary for implementing efficient restorative therapies. These therapies aim to improve hand functionality with minimal therapist intervention. However, the human hand evolved for various precision and power gripping tasks, with its intricate anatomy featuring a large number of degrees of freedom—up to 31—which hinder its modeling in many rehabilitation scenarios. In the process of designing prosthetic devices, instrumented gloves, and rehabilitation devices, there is a clear need to obtain simplified rehabilitation-oriented hand models without compromising their representativeness across the population. This is where the concept of kinematic reduction, focusing on specific grasps, becomes essential. Thus, the objective of this study is to uncover the intra-finger dependencies during finger flexion/extension by analyzing a comprehensive database containing recorded trajectories for 23 different functional movements related to ADLs, involving 77 test subjects. The initial phase involves data wrangling, followed by correlation analysis aimed at selecting 116 dependency-movement relationships across all grasps. A regularized generalized linear model is then applied to select uncorrelated predictors, while a linear mixed-effect model, with reductions based on both predictor significance and effect size, is used for modeling the dependencies. As a final step, agglomerative clustering of models is performed to further facilitate flexibility in tradeoffs in hand model accuracy/reduction, allowing the modeling of finger flexion extensions using 5–15 degrees of freedom only.

Soft pneumatic glove actuators for hand rehabilitation require compact, accurate models that can be evaluated in real time. At the same time, high-fidelity finite element (FE) simulations are too slow for iterative design and control. We develop a finite element-based calibration pipeline that combines a dependency-constrained human finger kinematic model with a segmented pseudo-rigid-body (PRB) description of ribbed-bellow soft pneumatic actuators sized to individual fingers. FE models with symmetry and contact generate pressure–pose data for the MCP, PIP, and DIP spans, from which we extract per-segment bending angles and axial elongations, fit simple pressure–kinematics relations, and identify PRB parameters using basin-hopping global optimization. The calibrated PRB reproduces FE flexion–extension trajectories for index and little finger actuators with millimetric accuracy (mean segment positioning errors of approximately 2.3 mm and 0.7 mm), preserves finger-like bending localized in the bellows, and maintains negligible compression of inter-joint links (below 1.2%). The pressure–bend and pressure–elongation maps achieve near-unity adjusted $R^2$, and the PRB forward kinematics evaluates complete pressure trajectories in less than half a millisecond, compared with several hours for the corresponding FE simulations. This pipeline provides a practical route from detailed FE models to controller-ready reduced-order surrogates for design-space exploration and patient-specific control of soft rehabilitation actuators.

Pneumatic artificial muscles, including McKibben-type actuators, offer high power-to-weight ratio, compliance, and inherent safety, but their nonlinear pressure–contraction behavior, hysteresis, saturation, and load-dependent dynamics make accurate position control challenging. This study develops a practical data-driven workflow that derives a branchwise feedforward compensator and an LQI or PID controller from one open-loop characterization experiment. Quasi-static characterization first identifies a conservative control-ready voltage window. A bounded random excitation within this window is replayed with $4\mathrm{s}$ holds to expose terminal and transient behavior. The same experiment supplies branchwise discrete plant models and a feedforward lookup. Two open-loop-derived transient layers, voltage creep compensation and dynamic pressure referencing, are applied to the raw lookup before simulation. Four controller variants are compared on a common simulated closed-loop benchmark built from the identified plant: a feedforward-only baseline, a branchwise proportional–integral–derivative (PID) baseline, a base linear quadratic integral (LQI) controller with displacement and pressure feedback, and a velocity-state LQI extension with a filtered velocity estimate. A multi-metric optimization score balances tracking RMS, settled oscillation, command activity, saturation, and gain magnitude. The score selects the base LQI within the LQI family. The selected gains and transient layers are deployed in a real-time implementation with manually reduced position gains. The controllers are then evaluated on a common reference stream against the physical actuator. Although simulation metrics cannot be transferred directly to the real system, the combined-metric ranking of the controllers remains unchanged.