Age-related and stroke-induced declines in musculoskeletal strength significantly impair upper extremity function, impacting daily life and incurring substantial healthcare costs. Existing robotic exoskeletons offer mechanical strength augmentation, but they lack the crucial ability to predict user intent in real-time due to the absence of effective sensory feedback and AI-driven prediction capabilities. Many existing exoskeletons rely on pre-programmed movements, limiting their practicality. Furthermore, their designs are often bulky, complex, and support only single-joint movements. This study addresses these limitations by developing a novel intelligent upper-limb exoskeleton that integrates soft bioelectronics, cloud-based deep learning, and a comfortable, lightweight design to achieve accurate real-time intention prediction and provide effective strength augmentation for multiple upper-limb joints. The goal is to create a user-friendly, portable exoskeleton that seamlessly integrates all the necessary elements to assist individuals with neuromotor disorders in performing everyday tasks.

Existing upper-extremity exoskeletons face several limitations. Many cannot predict user intent in real time, relying on pre-programmed movements instead of sensory feedback. Their designs are often bulky and complex, with extensive wiring and limited portability. Most assist only single-joint movements, insufficient for many daily tasks requiring combined elbow and shoulder motions. While strain sensors offer potential in motion recognition, they lack strength information, unlike electromyogram (EMG) signals. However, many exoskeletons incorporating EMG sensors only compare relative muscle involvement, rather than using it for direct sensory feedback. This work aims to overcome these limitations by creating a fully integrated system with accurate intent prediction, portability, lightweight design, multi-joint support, and high-fidelity sensory feedback.

This study developed an intelligent upper-limb exoskeleton system integrating soft wearable sensors, cloud-based deep learning, and soft pneumatic actuators. **Soft Pneumatic Artificial Muscles (PAMs):** Soft pneumatics were chosen for their compliance, low power consumption, lightweight nature, and inherent safety. The PAMs, designed to mimic human muscle contraction, consist of a silicone bladder, polyester mesh sleeving, aluminum end fittings, and pinch clamps. Characterized testing showed that a single PAM could generate 897 N of force with 87 mm displacement at 80 psi, while the exoskeleton utilizes three PAMs integrated into a lightweight (670g) carbon-fiber exoskeleton frame. The exoskeleton design allows for adjustable sizing to accommodate different users. A backpack houses the PAMs, compressor, solenoid valves, and electronics. **Soft Wearable EMG Sensors:** An array of soft, wireless EMG sensors with stretchable gold nanomembrane electrodes provides sensory feedback. These sensors are thin, lightweight, skin-conformable, and minimize motion artifacts. Finite element analysis confirmed their mechanical stability under strain. The sensors wirelessly transmit EMG signals to the cloud, eliminating the need for complex wiring. Performance testing showed comparable results to commercial sensors in terms of signal-to-noise ratio and reliability. **Cloud-Based Deep Learning:** A cloud-based deep-learning model uses CNN+LSTM architecture to classify muscle activation patterns and predict intended movement. The model receives 1-second-long filtered EMG signals as input, classifying muscle activation (rest, onset, activation) for four muscles (biceps, triceps, medial deltoid, latissimus dorsi). The model achieved high accuracy (95.38% for biceps/triceps and 97.01% for medial deltoid/latissimus dorsi), with a response time of 200-250ms. Hyperparameters were optimized using a random search method, with early stopping and dropout regularization to prevent overfitting. **System Integration and Control:** The system operates as follows: 1) Onset of muscle contraction; 2) EMG signals detected by sensors and transmitted to the cloud; 3) Cloud processes signals and predicts the intended movement; 4) Prediction sent to the exoskeleton; 5) PAMs actuate to assist movement. The entire process takes 500-550 ms. A custom Android app provides a user interface for monitoring, manual control of PAMs, and system status.

The developed intelligent exoskeleton system demonstrated high accuracy in predicting user intent and providing effective strength augmentation. The deep-learning model achieved an average accuracy of 96.2% in classifying four upper-limb movements (shoulder flexion/extension, elbow flexion/extension), with a response time of 500-550 ms. The soft pneumatic actuators generated sufficient force (897 N) and displacement (87 mm) to assist these movements. Real-life testing showed that the exoskeleton reduced muscle activity by 3.7 times on average compared to unassisted movements for both elbow and shoulder flexion. Even with a 6.8 kg weight, muscle activity was reduced by 1.4-1.6 times. The lightweight and comfortable design improved user experience. The cloud-based system facilitated real-time data analysis, model updates, and scalable deployment.

The results demonstrate the successful integration of advanced technologies to create a highly functional and user-friendly intent-driven exoskeleton. The high accuracy of the deep learning model, combined with the power and responsiveness of the soft pneumatic actuators, offers significant potential for improving upper extremity function in individuals with neuromotor disorders. The reduction in muscle activity suggests that the exoskeleton can effectively reduce fatigue and improve task performance. The cloud-based architecture allows for continuous improvement and personalization of the system, enhancing its long-term utility. This study's approach can be further applied to other assistive technologies and human-machine interface development.

This study presented a novel intent-driven upper-limb exoskeleton integrating soft bioelectronics, cloud-based deep learning, and soft pneumatic actuators. The system demonstrated high accuracy in movement prediction and substantial strength augmentation, significantly reducing muscle activity. The lightweight, comfortable design and cloud-based architecture enhance usability and scalability. Future work will focus on generalizing the deep learning model for broader applicability and exploring potential clinical applications.

The current study focused on a limited set of upper-limb movements and a small number of participants. Further research is needed to validate the system's effectiveness across a wider range of movements and a more diverse population. The reliance on cloud computing might pose limitations in environments with limited or unreliable internet connectivity. Long-term studies are necessary to assess the system's durability and potential long-term effects on users.