Control Method in Coordinated Balance with the Human Body for Lower-Limb Exoskeleton Rehabilitation Robots.

Ground walking training using a floating-base lower-limb exoskeleton rehabilitation robot improves patients' dynamic balance function, thereby increasing their motor and daily life activity capabilities. We propose a balance-directed motion generator (BDMG) based on the principles of deep reinforcement learning. The reward function sub-components pertaining to physiological guidance and compliant assistance were designed to explore motion instructions that are harmoniously aligned with the human body's balance correction mechanisms. To address the sparse rewards resulting from the above design, we introduce a stepwise training method that adjusts the reward function to control the model's training direction and exploration difficulty. Based on the aforementioned generator, we construct a training and evaluation process database and design an abnormal command recognizer by extracting samples with diverse feature characteristics. Furthermore, we develop a sample generation optimizer to search for the optimal action combination within a closed space defined by abnormal commands and extremum points of physiological trajectories, thereby enabling the design of an abnormal instruction corrector. To validate the proposed approach, we implement a training simulation environment in MuJoCo and conduct experiments on the developed lower-limb exoskeleton system.