Arie Rachmad Syulistyo, Yuichiro Tanaka, Ninnart Fuengfusin, Dinda Pramanta, Hakaru Tamukoh
Upper-limb movement intention recognition is a fundamental task to intuitive human-robot interaction, prosthetic control, and rehabilitation. Decoding intention in real time can improve assistance and effectiveness in such areas. Existing approaches face practical trade-offs: conventional machine learning classifiers are computationally efficient, but they typically require handcrafted and feature engineering to preprocess multivariate time series signals. On the other hand, deep recurrent neural networks can learn temporal representations directly from data, however they often impose substantial training costs that are challenging for wearable edge devices. This study investigates reservoir computing, specifically echo state networks (ESNs), that provides temporal processing with lightweight training by adjusting only a linear readout layer. We conducted a comparative evaluation using seven upper-limb movement tasks (Rest, Flex, Grasp, Lift, Return, Release, and Extend) with synchronized EEG, EMG, and IMU signals from the multimodal upperlimb movement intent detection challenge dataset. Our results demonstrate that ESNs offer a compelling balance between temporal modeling capability and computational efficiency for assistive and rehabilitation applications. © 2026 IEEE.
Graduate School Of Life Science And Systems Engineering, Kyushu Institute Of Technology, Kitakyushu, Japan; State Polytechnic Of Malang, Department Of Information Technology, Malang, Indonesia; Research Center For Neuromorphic Ai Hardware, Kyushu Institute Of Technology, Kitakyushu, Japan; Advanced Mobility Research Institute, Kanazawa University, Kanazawa, Japan; Kyushu Institute Of Information Sciences, Department Of Information And Network Sciences, Dazaifu, Japan