Walking Condition Estimation Using Physical Reservoir Computing with External Echo State Network
Junyi Shen, Swaninda Ghosh, Tetsuro Miyazaki, Kenji Kawashima · 2024
The rapid development of power-assistive limb devices underscores the necessity of perceiving different walking conditions to enhance their adaptability. This research introduces a method to estimate the user's walking conditions using a computationally efficient model composed of a wearable pneumatic physical reservoir and an external Echo State Network (ESN). The system, focusing on wearability and energy efficiency, is achieved through a pneumatic physical reservoir of four interconnected Pneumatic Artificial Muscles (PAMs). Deformations in the PAMs, induced by the wearer's walking motions, generate dynamic changes in air pressures within the reservoir, which are then used for computation. The system incorporates an external ESN to counteract computational limitations from reduced hardware components and a linear compression layer to lessen the data amount engaged in model training. Our experiments, covering different walking conditions and reservoir configurations, demonstrate that the asymmetrical physical reservoir arrangement offers the most accurate estimations, albeit with practical limitations. Symmetrical reservoir configuration capturing both thigh and calf movements balances estimation accuracy and practical applicability. This study underlines the potential of integrating pneumatic physical reservoir computing into wearable assistive devices for walking condition detection to enhance their adaptability in diverse scenarios.