편마비 환자를 위한 자유에너지 원리 기반 무릎 보조 로봇 제어의 고도화

Abstract

This study presents a customized control framework for a knee-wearable exoskeleton designed to assist hemiplegic stroke patients by combining real-time gait-phase estimation with a free-energy principle (FEP)-based active inference controller. Conventional exoskeletons often rely on time-based estimators or adaptive oscillators to synchronize assistance, but these methods suffer from phase delays and poor adaptability when the walking speed varies. To address this issue, we developed a lightweight neural estimator using a temporal convolutional network coupled with a Transformer encoder to infer the user’s gait phase from inertial measurement unit data. The model, pretrained on healthy gait datasets and fine-tuned on each patient’s data using a learning-without-forgetting strategy, achieves submillisecond inference latency and maintains phase errors below 3 %. Building on this, we formulate an FEP-based controller that unifies sensor filtering, state estimation, and torque computation within a variational inference framework. By modeling sensory observations and latent gait states probabilistically, the free energy can be approximated as a precision-weighted squared prediction error. Minimizing this free energy drives both the internal state update and control torque generation, enabling the robot to actively reduce discrepancies between expected and measured sensor signals. The controller coexists with a conventional proportional-derivative (PD) assistive mode, allowing direct performance comparisons. Experimental findings based on the analysis of the results obtained from seven male hemiplegic subjects demonstrate that the proposed FEP control restores posture gait symmetry in comparison to PD control. These results suggest that fusing individualized gait-phase estimation with active inference control can enhance rehabilitation outcomes and support intuitive human–robot cooperation for users with asymmetric gait patterns.