As people age, declines in muscle strength, coordination, and sensory function often make activities of daily living, such as doing laundry, getting dressed, cooking, carrying objects, and using tools, increasingly difficult. These difficulties not only reduce convenience in daily life but also limit older adults’ independence. To meet the growing need for elderly assistance, assistive technologies such as electric wheelchairs, prostheses, and soft exoskeletons have continued to develop, while enabling these devices to naturally understand human motor intent has become a key issue in improving human–robot collaboration. Electromyography signals reflect muscle activity and neuromuscular control and usually appear before visible limb movement, making them an intuitive and natural interface for assistive robot control. However, bringing EMG-based control into real-time daily use still faces several challenges. First, reliable mapping between EMG signals and motor intent depends on accurate labeling, whereas traditional manual annotation or external motion-capture systems increase system complexity and workload. Second, many existing methods still focus on single-joint or independently modeled joint recognition, making it difficult to capture coordinated multijoint movements such as those involving the hand and elbow in daily activities. Third, EMG signals vary across tasks, postures, and application scenarios, so models often struggle to generalize to new tasks and may forget previously learned capabilities. Therefore, for real-world elderly assistance, there is a clear need for a unified intelligent EMG-control framework that can jointly address physiological signal labeling, functional multijoint intent decoding, and behavioral task adaptation. “Therefore, for real-world elderly assistance, there is a clear need for a unified intelligent EMG-control framework that can jointly address physiological signal labeling, functional multijoint intent decoding, and behavioral task adaptation.” said the author Jiaqi Xue, a researcher at City University of Hong Kong.
This study proposed a 3-level AI-driven EMG control framework for upper-limb assistance in older adults, translating EMG signals into robotic assistive actions across physiological, functional, and behavioral levels. The researchers collected upper-limb EMG data from younger and older participants, focusing on 4 muscles related to hand and elbow movement: the triceps, biceps, extensor digitorum, and flexor digitorum superficialis. Based on different muscle activation characteristics, they designed selective active labeling and contextual labeling methods to reduce reliance on manual annotation and external motion-capture devices. At the functional level, a one-dimensional convolutional neural network was built to simultaneously predict hand states, including relaxed, open, and closed, and elbow states, including relaxed, flexed, and extended, from 4-channel EMG signals, with a voting mechanism used to stabilize real-time outputs. At the behavioral level, knowledge distillation was introduced so that the model could learn new daily tasks, such as cooking, lifting a bag, and pulling a grocery cart, while retaining previously learned basic joint-movement knowledge. Finally, the framework was deployed in a real-time robotic control system combining EMG acquisition, edge computing, and a 6-axis collaborative robotic arm to validate its feasibility in daily assistance scenarios.
The results showed that the 3-level AI-driven EMG control framework effectively improved EMG labeling, multijoint intent recognition, and adaptation to daily tasks. At the physiological level, selective active labeling was more suitable for hand opening and closing, where EMG signals showed short and strong fluctuations, while contextual labeling was more suitable for elbow flexion and extension, where muscle activation was more continuous. Both strategies reduced manual annotation burden while maintaining consistency with actual movement states. At the functional level, the one-dimensional convolutional neural network achieved stable hand and elbow state prediction, with offline accuracies of 95.42% and 93.97%, respectively. In real-time multijoint coordination, the model reached an overall accuracy of 95.34% across 9 hand–elbow motion combinations and successfully drove smooth coordinated robotic assistance. At the behavioral level, knowledge distillation helped the model adapt to more complex daily activities, such as cooking, lifting a bag, and pulling a grocery cart. In the cooking task, the student model significantly improved hand and elbow prediction compared with the teacher model, with a maximum improvement of 11.25%, while retaining basic joint-movement knowledge. Real-time robotic experiments further showed that the system could control the robotic arm according to users’ EMG intentions to pick up, move, and put down a wok, validating its stability and practical potential for elderly daily assistance.
The significance of this work lies in providing a complete technical pathway from physiological EMG signal processing to functional multijoint intent decoding and behavioral adaptation in daily activities. This allows elderly-assistance robots to move beyond recognizing isolated movements and better address the real-world need for coordinated hand–elbow actions and changing task contexts. By designing physiologically informed labeling methods, the study reduced reliance on manual annotation and external sensing devices; by building a multi-output one-dimensional convolutional model at the functional level, it achieved accurate multijoint coordination recognition; and by introducing knowledge distillation at the behavioral level, the model could learn new daily tasks while retaining previously acquired basic abilities, improving adaptability and scalability for long-term use. Overall, the framework shows that EMG-driven assistance should not be viewed only as an intention-recognition problem, but as a stable, transferable, and task-adaptive human–robot collaboration mechanism. “In the future, we will combine flexible sensors to enhance wearing comfort, optimize model lightweighting and system latency, and introduce more user feedback and interaction mechanisms to promote the practical application of such systems in home assistance, rehabilitation training, and healthy aging scenarios.” said Jiaqi Xue.
Authors of the paper include Jiaqi Xue, Ziqi Li, Xiaoyang Zou, Zijia Qu, Shengjie Yang, Colin Pak Yu Chan, Yanchen Liu, Zhou Zhao, Jing Zhang, Clio Yuen Man Cheng, Haiyang Wang, Kehan Zou, Yafei Zhao, Vivian Weiqun Lou, Ning Xi, and King Wai Chiu Lai.
This work was partially supported by grants from the Research Grant Council of the Hong Kong Special Administrative Region Government (TBRS Grant T42-717/20-R and CRF Grant C7100-22G).
The paper, “Artificial-Intelligence-Driven Electromyography Adaptation for Elderly Assistance at Physiological, Functional, and Behavioral Levels” was published in the journal Cyborg and Bionic Systems on Jul 15, 2026, at https://doi.org/10.34133/cbsystems.0638.
Cyborg and Bionic Systems