Rehabilitation is critical for restoring motor function in individuals affected by neurological disorders such as stroke, which often results in upper-limb impairments and reduced independence in daily living. In recent years, robotic rehabilitation systems have gained significant traction for their ability to deliver repetitive, high-intensity, and task-specific therapy—factors known to promote neuroplasticity and functional recovery (Kabir et al., 2022; Gopura et al., 2023). These systems offer distinct advantages over conventional therapy, including precise kinematic feedback, repeatable movement trajectories, and objective performance metrics (Carbone and Gonçalves, 2022; Han et al., 2023).

Various robotic modalities have emerged, including active-assisted devices, active-constrained robots, and adaptive exercise platforms, each tailored to patient-specific needs. Despite these advances, robot-only systems often lack the neuromuscular engagement necessary to restore voluntary motor control. To overcome this, hybrid rehabilitation strategies that combine robotic assistance with Functional Electrical Stimulation (FES) have been explored, leveraging the benefits of both modalities (Niu et al., 2022; Montoya et al., 2022).

Real-time control is a critical requirement in robotics, particularly in medical and rehabilitation applications where precision, safety, and responsiveness are paramount. The Robot Operating System 2 (ROS2) has emerged as a robust middleware framework designed to address real-time constraints and enhance modularity in robotic systems (Open Robotics). Despite its increasing adoption across various robotics domains, the application of ROS2 in rehabilitation robotics remains relatively limited, positioning our work among the pioneering efforts in this field. Research has demonstrated ROS2’s potential in rehabilitation applications, such as trajectory tracking for knee rehabilitation robots (Arcos et al., 2022) and brain–computer interface (BCI)-controlled robotic arms for assistive technology (Rivas et al., 2024). However, the broader implementation of ROS2 in rehabilitation exoskeletons and motor recovery systems is still emerging.

Recent advancements, including bioinspired hierarchical electronic architectures for robotic locomotion assistance (Delgado-Oleas et al., 2023) and the integration of digital twins with exoskeletons for telerehabilitation (Falkowski and Gwardecki, 2024), highlight the growing interest in intelligent, adaptable rehabilitation systems. Furthermore, IoT-enabled humanoid robotics for motor rehabilitation (Moghbelan et al., 2024) and multi-layered assessment approaches for hand spasticity using exoskeletons (Yu et al., 2025) reinforce the necessity of real-time, adaptive control frameworks. By leveraging ROS2’s enhanced communication reliability, reduced latency, and dynamic adaptability, our research contributes to expanding its applicability in medical robotics, demonstrating its potential to revolutionize patient-centered robotic interventions.

FES delivers electrical impulses to targeted muscles, inducing contractions that mimic voluntary movement. This approach has demonstrated efficacy in promoting motor relearning and upper-limb strength (Niu et al., 2019; Sousa et al., 2022; Tefertiller et al., 2022). However, traditional FES protocols are typically open-loop and static, often leading to overstimulation, patient discomfort, and premature muscle fatigue—particularly in patients with intact sensory pathways (Peckham and Knutson, 2005; Doucet et al., 2012). Moreover, most systems lack the ability to adapt stimulation parameters in real time based on muscle condition (Sharififar et al., 2018).

Machine learning (ML) has increasingly been applied to enhance adaptive neuromuscular electrical stimulation (NMES) for rehabilitation, offering personalized and precise control over functional electrical stimulation (FES). Recent advances in deep learning and ML-based FES controllers have demonstrated significant improvements in system adaptability, robustness, and patient-specific tuning. Researchers in Arcolezi et al. (2021) introduced an ML-driven approach to optimize the tuning of a robust integral of the sign of the error (RISE) controller for lower-limb rehabilitation, leveraging system identification techniques to improve NMES effectiveness. Similarly, research by Sierotowicz and Castellini (2023) explored robot-inspired human impedance control via functional electrical stimulation, applying ML-based techniques to adapt stimulation patterns dynamically based on user responses. These works underscore the growing role of intelligent controllers in rehabilitation, allowing for real-time adjustments and improved neuromuscular recovery. By integrating deep learning models into NMES control, future advancements can further enhance the precision and adaptability of rehabilitation protocols, ensuring optimal stimulation levels tailored to individual patients’ needs.

To address these limitations, we propose a closed-loop, AI-enhanced hybrid rehabilitation system that integrates robotic assistance with real-time, EMG-driven neuromuscular electrical stimulation, as illustrated in Figure 1. Central to our approach is a Support Vector Machine (SVM)-based model that continuously estimates muscle fatigue from EMG signals, enabling adaptive, patient-specific stimulation protocols. A symmetrical, dual-arm robotic platform was developed to support bilateral upper-limb therapy (Bouteraa et al., 2023; Bouteraa et al., 2020), featuring automatic limb-length adjustments to accommodate anatomical variations. The system is implemented using a ROS2-based control architecture, enabling real-time processing, modular extensibility, and remote therapist supervision.

This closed-loop configuration facilitates adaptive therapy sessions that dynamically respond to patient needs, ensuring both safety and effectiveness.

The main contributions of this work are as follows:

• Development of a dual-arm rehabilitation robot with real-time anatomical adaptation and bilateral therapy support.

The remainder of this paper is structured as follows: Section 2 describes the system architecture and methodology. Section 3 presents experimental results and clinical analysis. Section 4 concludes with key findings and discusses future directions.

The proposed hybrid rehabilitation system was evaluated on three post-stroke patients over 30 therapy sessions conducted across a 6-week period. Quantitative assessments were carried out before and after the intervention, focusing on three clinical metrics: range of motion (RoM), active torque, and passive torque at the targeted joint. As shown in Figures 6–8, the system enabled substantial improvements across all three patients. On average:

• Range of motion increased by 44%, reflecting enhanced joint flexibility.

To further validate system performance at an individualized level, automated reports were generated by the system’s ROS2 interface, summarizing patient-specific progress over time. These are presented in Figures 6–8, respectively.

The summarized outcomes were:

• Patient 1: Final RoM of 82°, RPF = 9.1/10, active torque increased from 0.91 Nm to 1.42 Nm, passive torque reduced from 0.39 Nm to 0.23 Nm.

These individualized reports suggest the system’s potential to support measurable functional improvements while adapting stimulation to real-time physiological states. Although the improvements in range of motion, active torque, and passive torque were consistent across all patients, the small sample size (n = 3) limited statistical power. Wilcoxon signed-rank tests were performed and yielded p-values of 0.25 across all three metrics. While these values did not reach conventional significance thresholds, the observed trends suggest meaningful clinical improvement and motivate larger-scale studies for future validation.

Figure 9 presents the PCA scatter plot of the EMG features. As shown, there is a clear separation between the Fatigue and Non-Fatigue classes in the space of the first two principal components, validating the discriminative quality of the features despite the preprocessing steps.

The results of this pilot study suggest that the integration of a symmetrical dual-arm robotic platform with adaptive, EMG-driven neuromuscular stimulation can effectively enhance upper-limb rehabilitation outcomes in post-stroke patients. Specifically, we observed measurable improvements in range of motion, increased active torque, and decreased passive resistance across all participants. These gains can be attributed to three key factors: (1) the use of bilateral symmetry training, which leverages mirrored motion to promote inter-limb coordination and neuroplasticity; (2) the implementation of a fatigue-aware stimulation strategy, which maintains engagement while preventing overstimulation; and (3) the adoption of a real-time, ROS2-based control system, which ensures responsive adaptation and continuous clinician oversight. Together, these components create a synergistic framework that enhances therapy personalization and efficacy. Improvements in range of motion, active torque, and passive torque highlight both neuromuscular reactivation and reduction of joint stiffness. The range of motion (ROM) in the joints of the three participants increased by approximately 44% from their initial states, demonstrating the effectiveness of the rehabilitation robot in enhancing joint mobility. These results are broadly consistent with recent trends reported in the literature (Rouse et al., 2020; Mashayekhi and Moghaddam, 2022), where comparable systems have achieved 30%–40% improvements in RoM over similar rehabilitation durations. While our observed ∼44% improvement is encouraging, caution is warranted in making direct comparisons due to differences in sample size, patient characteristics, and evaluation protocols.

Similarly, joint torque assessments revealed a 45% increase in active torque and a 36% reduction in passive torque, highlighting the system’s ability to improve neuromuscular function. These outcomes represent a significant advancement compared to our previous work using exoskeletons alone (Bouteraa et al., 2023) and traditional electrical stimulation techniques that did not consider muscle fatigue (Bouteraa et al., 2020). Related studies have combined exoskeletons with functional electrical stimulation (Inoue et al., 2022; Abdallah and Bouteraa, 2024), achieving moderate improvements in motor function. However, the integration of real-time muscle fatigue classification in our proposed protocol enables more adaptive and personalized interventions, leading to superior rehabilitation outcomes.

While recent research has investigated muscle fatigue estimation during rehabilitation training (Mashayekhi and Moghaddam, 2022; Zasadzka et al., 2022; Maffiuletti et al., 2018), integrating fatigue information directly into the primary control system remains a considerable challenge. In most existing systems, patient safety is managed manually by pausing rehabilitation sessions or modifying stimulation settings when signs of fatigue are observed.

This study addresses that limitation by implementing a robust Support Vector Machine (SVM) classifier capable of detecting muscle fatigue with a mean accuracy of 95%, outperforming previous approaches (Zasadzka et al., 2022) where fatigue detection accuracy was limited to 90% using linear models.

Unlike traditional open-loop stimulation systems that apply fixed parameters throughout therapy, our approach adapts stimulation in real time based on muscle fatigue levels estimated via EMG signals and SVM classification. This closed-loop mechanism allows the system to optimize therapeutic intensity without exceeding fatigue thresholds—improving comfort and engagement. Similar adaptive strategies have been explored in exoskeleton control (Montoya et al., 2022), but few have combined fatigue adaptation with bilateral robotic training and ROS2 integration.

The system’s ability to generate patient-specific reports—including real-time tracking of torque and mobility parameters—offers a significant advantage for clinical monitoring and decision-making. The inclusion of the RPF adds a high-level indicator of recovery progress, useful for both therapists and automated systems.

Bilateral mirrored motion, achieved through the symmetrical robot design, provides consistent proprioceptive feedback and leverages inter-limb coordination to stimulate neuroplasticity. Coupled with adaptive FES, this synergy supports the re-establishment of functional motor pathways (Niu et al., 2022).

Despite promising outcomes, the study is limited by a small sample size. Future work will include a larger clinical cohort, integration of cognitive intent (via EEG), and long-term outcome tracking. Nevertheless, the results demonstrate the viability and therapeutic potential of a real-time, AI-enhanced hybrid rehabilitation system grounded in physiological data. Another limitation of this pilot study is the absence of standardized clinical outcome assessments, such as the Action Research Arm Test (ARAT) or post-intervention Fugl-Meyer scores. Although objective biomechanical measures (e.g., RoM, torque) provided valuable insights into neuromuscular improvements, the inclusion of validated clinical scales would enhance the clinical relevance and comparability of results. Future studies will integrate such tools to provide a more comprehensive evaluation of functional recovery.

This study presented a novel AI-driven hybrid rehabilitation system that integrates symmetrical dual-arm robotic assistance with adaptive neuromuscular electrical stimulation, guided by real-time EMG analysis and machine learning. The proposed framework leverages a Support Vector Machine classifier to detect muscle fatigue and dynamically adjust stimulation intensity, ensuring patient specific, fatigue-aware therapy.

Preliminary testing on three post-stroke patients demonstrated measurable improvements in range of motion, active torque, and reduction in joint stiffness. In addition to group-level gains, individualized reports highlighted the system’s ability to adaptively track and optimize progress in real time. The ROS2-based control architecture enabled seamless integration of sensing, processing, actuation, and reporting—facilitating modularity, clinical usability, and remote supervision. These promising findings motivate further investigation with larger clinical trials to robustly evaluate efficacy and generalizability.

By addressing key limitations of conventional FES and robot-only systems, this work contributes a scalable, closed-loop solution for next-generation neurorehabilitation. Future efforts will focus on expanding clinical trials, incorporating EEG-based cognitive intent detection, and further enhancing the system’s autonomy and personalization.