Intelligent Upper-Limb Exoskeleton with DL-Augmented Strength

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Feb 15 2024

In an article published in the journal Nature, researchers presented an intelligent upper-limb exoskeleton system utilizing deep learning (DL) to predict and augment human strength. Soft wearable sensors collected real-time muscle activities, enabling cloud-based DL to predict upper-limb joint motions with 96.2% accuracy.

The system, responding to human intention in 500–550 milliseconds, incorporated soft pneumatics, generating 897 newtons of force to assist movements. Compared to an unassisted exoskeleton, the intent-driven exoskeleton reduced human muscle activities by 3.7 times on average, addressing age and stroke-related musculoskeletal strength decline.

Background

Individuals with neuromotor disorders resulting from stroke-induced and age-related declines in musculoskeletal strength face challenges in performing daily tasks. Stroke affects one in four adults over 25, with 12.2 million cases annually, leading to neuromotor disorders in 20–40% of victims. This hampers functional independence and incurs significant healthcare costs, amounting to $65 billion annually in the United States. Existing robotic exoskeletons address upper-extremity strength but lack real-time intention prediction, limiting their practicality. Most exoskeletons are stationary, bulky, and focus on a single joint movement, hindering mobility and adaptability.

This research introduced a groundbreaking human-intent-driven robotic exoskeleton. Leveraging DL on a cloud platform, it accurately predicted user movements in real time. Soft bioelectronics, including pneumatic artificial muscles (PAM) and wireless electromyogram (EMG) sensors, provided high-fidelity sensory feedback. This exoskeleton was portable, lightweight, user-friendly, and supported multiple upper-limb joint movements. Addressing the gaps in previous work, this integrated solution seamlessly combined accurate intention prediction, portability, and sensory feedback, presenting a comprehensive approach for assisting individuals with neuromotor disorders in performing daily tasks. The novel features distinguished this work from existing exoskeletons, marking a significant advancement in the field.

Method

The authors introduced a comprehensive system comprising a PAM actuator, an exoskeleton, and a soft sensor array for human upper-extremity strength augmentation. The PAM, constructed from silicone tubing, polyester mesh, aluminum fittings, and stainless steel clamps, underwent quasistatic testing to characterize its mechanical properties. The exoskeleton, predominantly made of carbon fiber with aluminum connectors, allowed natural body movement through a super swivel ball joint and adjustable telescoping tubing. Three-dimensional (3D)-printed arm mounts enhanced user comfort, and PAM provided force assistance for multiple joint movements.

A soft sensor system with fabric substrate, flexible circuits, and nanomembrane electrodes was developed. The fabric substrate, created using Silbione, provided a base for the circuit, while the circuits, encapsulated in elastomers, contained a lithium polymer battery assembly. Gold/chromium electrodes, deposited by E-beam evaporation, form a stretchable serpentine pattern on a polymer film, facilitating skin conformability. Mechanical and electrical tests demonstrated the sensor's durability, with a cyclic stretching test and circuit bending test.

The study employed a wireless EMG recording circuit with an nRF52832 microcontroller and Bluetooth system-on-chip. Data from biceps, triceps, and medial deltoid muscles were transmitted to Google Cloud, where a convolutional neural networks (CNN) + long short-term memory ( LSTM) algorithm predicted motion classes in real-time, achieving a response rate of 200–250 milliseconds. The cloud communicated with the exoskeleton driver, which utilized a microcontroller and pressure feedback system for valve control, providing user strength augmentation.

Demonstrations showed effective strength augmentation through EMG signal analysis, comparing conditions with and without exoskeleton assistance. The study included a human subject study with volunteers aged 18 or older, conducted following Institutional Review Board protocols at the Georgia Institute of Technology.

Results

The researchers presented an innovative intent-driven upper-limb exoskeleton system integrating wearable bioelectronics and cloud-based DL to augment human strength. The exoskeleton utilized soft PAMs and soft sensors to provide immediate assistance based on the prediction of user intention for four key upper-limb movements: shoulder flexion, shoulder extension, elbow flexion, and elbow extension. The soft wearable sensors, placed on the skin, collected real-time muscle signals through EMG and transmitted data to a cloud server.

The cloud-based DL algorithm processed the signals, predicted the intended movement, and wirelessly transferred the output to the exoskeleton for real-time assistance. The PAMs, designed to mimic human muscles, operated within a safe pressure range. The exoskeleton's lightweight carbon fiber frame, adjustable through 3D-printed mounts, enhanced user comfort and convenience. The developed soft EMG sensors demonstrated compatibility, reliability, and a high signal-to-noise ratio, offering effective muscle activity monitoring. The cloud-based DL model achieved accuracy rates of 95.38% for biceps/triceps and 97.01% for medial deltoid/latissimus dorsi muscle activation classification.

Real-life demonstrations showcased a substantial reduction in EMG signals during assisted joint movements, confirming the exoskeleton's ability to augment human strength. The system's cloud infrastructure allowed for scalability, updates, and data collection across diverse users. The integration of soft bioelectronics, cloud-based DL, and a user-friendly exoskeleton positioned this system as a promising advancement in human-robot interaction for strength augmentation in daily activities.

Discussion

The authors presented a groundbreaking integration of a cloud-based DL platform, a robotic exoskeleton for strength augmentation, and soft bioelectronics for sensory feedback, resulting in an intent-driven robotic exoskeleton. The system utilized soft PAMs for lightweight yet powerful assistance in multiple upper-extremity joint movements. Soft bioelectronic sensors monitored EMG signals, processed through cloud-based DL, enabling accurate user intent prediction.

The exoskeleton exhibited a rapid response time of 500–550 milliseconds, significantly reducing EMG activities during movements. With a high test accuracy for motion classification, the exoskeleton demonstrated a 6.9-fold reduction in EMG during elbow flexion and a 3.4-fold reduction during shoulder flexion. This innovation marked a transformative leap in robotic exoskeleton technology, offering potential applications for individuals with neuromotor disorders in real-life scenarios.

Conclusion

In conclusion, the researchers introduced an intelligent upper-limb exoskeleton, integrating DL for real-time motion prediction and muscle activity augmentation. The system, incorporating soft wearable sensors and cloud-based DL, achieved a remarkable 96.2% accuracy in predicting upper-limb joint motions. The lightweight exoskeleton, driven by soft PAMs, significantly reduced muscle activities by 3.7 times on average, addressing age and stroke-related strength decline. This innovative, user-friendly solution marks a significant advance in assisting individuals with neuromotor disorders in daily tasks, setting a new standard in robotic exoskeleton technology.