Intelligent upper-limb exoskeleton integrated with soft bioelectronics and deep learning for intention-driven augmentation

Neuromotor disorders from stroke and age-related declines in musculoskeletal strength limit upper-limb function, affecting independence and incurring substantial healthcare costs. Existing exoskeletons mechanically augment strength but often fail to predict real-time user intention due to absent sensory feedback and AI, rely on stationary setups with bulky hardware and wiring, and typically assist only single joint movements. Haptic/physiological sensing is crucial; EMG can directly reflect muscle activation, yet many systems omit it or use it only for relative comparisons rather than closed-loop control. The authors posit that an ideal upper-limb exoskeleton should accurately predict intention, be portable, lightweight, easy to use, support multiple joints, and incorporate high-fidelity sensory feedback. This study reports an intent-driven upper-limb exoskeleton integrating soft bioelectronic EMG sensors, cloud-based deep learning for motion prediction, and soft pneumatic artificial muscles for multi-joint assistance.

The paper surveys prior upper-extremity exoskeletons, identifying key gaps: lack of real-time intention prediction and adaptive model updating; stationary, non-portable designs with complex, rigid hardware and wiring; limited to single-joint assistance (e.g., elbow or shoulder only); and inadequate integration of physiological sensors, with EMG often used for relative muscle activity rather than control. Table 1 compares this work against multiple references, highlighting that earlier systems commonly used rigid frames with gel electrodes, lacked wireless sensing, cloud computing, and deep-learning-based real-time classification, and supported fewer motions. Some devices provided assistance without sensors or with limited EMG integration, and many were not portable. The presented system uniquely combines wireless soft dry-electrode EMG sensing, cloud-based deep learning with real-time prediction and model updates, multi-joint assistance (shoulder flexion/extension and elbow flexion/extension), and portability.

System architecture: Four wireless soft EMG sensors adhere to the skin over biceps brachii, triceps brachii, medial deltoid, and latissimus dorsi to capture muscle activation. EMG data streams via BLE to a tablet and then to Google Cloud for real-time processing and deep-learning-based motion classification. Predicted motions are wirelessly sent to a wearable exoskeleton that actuates soft pneumatic artificial muscles (PAMs) to augment user-intended movements. Overall assistance latency is 500–550 ms (cloud response 200–250 ms, PAM actuation ~100 ms, EMG window 200 ms). Hardware: Three PAMs located in a backpack provide actuation, transmitting force via cables to a lightweight carbon-fiber exoskeleton frame. Each PAM comprises a silicone bladder within a polyester mesh sleeve, pressurized via compressor and solenoid valves; a pressure sensor enables real-time monitoring. Force–contraction characterization from 10–80 psi shows approximately linear force vs. contraction at fixed pressures, with maximum observed force ~897 N and displacement 87 mm at 80 psi for a 34 cm bladder. For safety and longevity, operating pressure is set to 10–60 psi with a 70 psi relief valve. Each PAM weighs 104 g; the exoskeleton frame weighs ~670 g; the fully integrated wearable system (battery, cables, PCB, PAMs, backpack) weighs ~4.7 kg. The exoskeleton uses telescoping carbon-fiber tubing and adjustable arm mounts (3D printed and thermoformed) for multiple body sizes; ball joints allow natural movement. Electronics in the backpack control valves and compressor and interface with the cloud GUI for monitoring and manual control. Soft EMG sensors: Sensors incorporate stretchable gold nanomembrane dry electrodes on a silicone adhesive patch, a flexible circuit (nRF52832 MCU, ADS1292 AFE/ADC, BLE), IMU, and a rechargeable battery with magnetic charging. Serpentine electrode design and soft encapsulation enhance skin conformality and reduce motion artifacts. Finite element analysis indicates mechanical stability under 30% strain; cyclic stretching (30% strain, 300 cycles) yields minimal resistance drift. The flexible circuit maintains wireless data integrity after repeated bending. Comparative measurements during elbow flexion show EMG signals overlapping those from a commercial gel-electrode system, with comparable SNR and improved skin compatibility (no irritation after 12 h wear vs. rash from gel electrodes). Battery life is ~5.1 h. Signal processing and cloud computing: Raw EMG is bandpass filtered (10–250 Hz) with 60 Hz notch (59–61 Hz), segmented into 1 s windows with 250 ms overlap, and rectified. Standard scaler normalization is applied. The cloud application (Python/Keras/TensorFlow) hosts a CNN+LSTM model per muscle; outputs are fused to determine motion class. Valve control GPIOs respond within ~50 ms; pressure feedback supports closed-loop actuation. A cloud-based GUI displays sensor/actuator status, real-time classifications, pressures, and allows manual valve control and emergency pause. Deep-learning model: For each muscle (biceps, triceps, medial deltoid, latissimus dorsi), a CNN+LSTM architecture processes 1D EMG segments using two 1D convolutional layers with batch normalization, max pooling, Leaky ReLU, dropout (0.3), followed by LSTM and fully connected layers to classify three states (rest, onset, activation). Optimization uses ADAM (learning rate 0.0001) with cross-entropy loss; early stopping and learning-rate annealing employed. Training uses data collected from five human subjects, with 50 EMG sets per movement. Data splits reported include 80% training/20% testing (main text) and 60%/20%/20% (Methods) for training/validation/testing. Control logic: The system detects onset activity in agonist muscles to initiate assistance and uses antagonist activation to pause or reverse/return to rest (e.g., biceps onset triggers elbow flexion; triceps activation pauses/vents; medial deltoid onset triggers shoulder flexion; latissimus dorsi activation pauses/vents). Combined motions (e.g., sequential elbow and shoulder flexion) are supported. GUI provides real-time feedback and control. PAM and sensor fabrication/characterization details: PAM components (silicone tubing, polyester mesh, aluminum end fittings, PEX-B clamps) are assembled and characterized on a motorized test stand across pressures 10–80 psi. Soft sensors are fabricated via spin-coated silicone adhesive layers, flexible PCB assembly, elastomer encapsulation, and laser-patterned Au/Cr serpentine electrodes transferred to the adhesive substrate; mechanical testing uses cyclic stretch and bending; skin-electrode impedance and SNR are measured against gel electrodes. Human studies followed IRB approval (Georgia Tech H21214).

- Cloud-based deep learning classified four upper-limb joint motions with high accuracy: 95.38% for biceps/triceps and 97.01% for medial deltoid/latissimus dorsi (average ~96.2%). - End-to-end response time for assistance was 500–550 ms (cloud response 200–250 ms, PAM actuation ~100 ms, EMG window 200 ms). - Soft PAMs generated up to ~897 N force and 87 mm displacement at 80 psi; operating range set to 10–60 psi with a 70 psi relief valve for safety. - The wearable exoskeleton frame weighed ~670 g; each PAM weighed 104 g; total integrated system weight ~4.7 kg. - EMG-based assistance substantially reduced muscle activation during unloaded movements: • Elbow flexion: MVC-normalized biceps activation 3.9-fold lower with assistance (mean ± SD: 2.43 ± 0.91 assisted vs. 9.36 ± 4.54 unassisted). • Shoulder flexion: MVC-normalized medial deltoid activation 3.5-fold lower with assistance (20.12 ± 4.23 assisted vs. 70.98 ± 8.52 unassisted). - Under load (6.8 kg/15 lb dumbbell), EMG reductions persisted: 1.4-fold (elbow flexion) and 1.6-fold (shoulder flexion) lower with assistance. - Static holding performance improved: assisted subjects could hold 6.8 kg for >3 min vs. <1 min without assistance. - Wireless soft dry-electrode EMG sensors provided signal quality comparable to commercial gel electrodes, with improved skin compatibility (no irritation over 12 h).

The study addresses the need for intention-driven, portable upper-limb assistance by integrating high-fidelity soft EMG sensing with cloud-based deep learning and soft pneumatic actuation. Real-time motion classification with ~96% accuracy and ~0.5 s response enables the exoskeleton to assist precisely at the onset of intended movements. The control logic using agonist/antagonist muscle signals allows users to initiate, pause, and combine joint actions naturally. High actuation force and displacement from PAMs, combined with a lightweight adjustable frame, translated into significant reductions in muscle activation during both unloaded and loaded tasks and extended endurance for static holds. The cloud architecture facilitates scalable model updates and uniform performance across devices, supporting adaptation to evolving user patterns and broader deployment. These results demonstrate a practical, user-friendly exoskeleton capable of multi-joint, intention-driven augmentation suitable for daily activities and with potential applicability to advanced prosthetics and assistive technologies.

This work demonstrates a fully integrated, intent-driven upper-limb exoskeleton that combines wireless soft EMG bioelectronics, cloud-based deep learning for real-time motion prediction, and soft pneumatic actuators to augment shoulder and elbow movements. The system achieves high classification accuracy (~96%), fast response (500–550 ms), and substantial strength assistance, reducing EMG activation by 3.5–3.9× during unloaded tasks and 1.4–1.6× under load, with improved endurance in static holds. The lightweight, adjustable design supports practical daily use, and the cloud platform enables scalable updates and maintenance. Future work will establish universal, generalized deep-learning models that accurately predict intended movements across multiple users to broaden applicability and support large-scale human-robot interaction studies.

- The deep-learning models currently target single-person scenarios; future work is needed to build universal, generalized models that predict intended movements across multiple users. - Fitment can present challenges when creating a comfortable interface between the exoskeleton and users, despite adjustable and customizable mounts. - The physiological impact of carrying a backpack load up to 10% of body weight remains disputable, although the system weight (~4.7 kg) is designed to be manageable. - Experimental datasets referenced include five human subjects and specific trained movements, indicating a limited dataset and motion scope for current evaluations.