Real-Time Motion Intent Prediction with Minimal Sensors and Neural Networks for Industrial Assistive Robots

초록

Accurately inferring user intent remains a central challenge to improving the efficiency of assistive wearable robots in industrial settings. Many existing approaches rely on costly or body-worn sensors, which raise production costs and reduce wearability. Using data from low-cost, non-body-worn sensors, this study infers user intent with a neural network and, guided by neurophysiological timing evidence, specifies and validates a control-period target that minimizes intent-motion mismatch. The evaluation focuses on real-time intent prediction during the frequently occurring industrial action of lift-and-put-down. To address the delay inherent in joint-encoder signals-which reflect motion only after muscle-driven movement occurs-minimal plantar-pressure and load-cell sensors were incorporated to capture earlier cues of body-weight shift without direct skin attachment. Results indicate that a control period of approximately 15 ms-derived from neurophysiological timing of afferent conduction and early cortical responses-is advantageous for closed-loop operation, and the proposed data, network, and post-processing configuration achieved an accuracy of 99.39 % (+/- 1.36 %), with a 95 % confidence interval of [97.71 %, 100.00 %]. These findings provide initial evidence that minimal-sensor intent inference can alleviate cost and wearability barriers to the commercialization of assistive wearable robots.

키워드