An adaptive framework of real-time continuous gait phase variable estimation for lower-limb wearable robots

Abstract

Phase-variable-based approaches are emerging in the control of lower-limb wearable robots, such as exoskeletons and prosthetic legs. However, real-time smooth estimation of the gait phase within each gait cycle remains an open problem. This paper presents a novel method for real-time continuous gait phase estimation during walking. The proposed framework consists of three subsystems: real-time kinematic data collection, gait phase variable estimation, and online adaptation of individual kinematics through backward data segmentation of completed gait strides. It is worth noting that we introduce an online learning mechanism for extracting and learning gait features from previous strides, in contrast with offline parameter tuning. The proposed basic gait model is initialized by human average data and is incrementally refined as a function of the individual gait features over different walking speeds. This provides a framework for long-term personalized control. Furthermore, the phase variable is constructed through the thigh angle measured by an inertial measurement unit. The resulting simple sensor system improves the usability of the proposed technique in wearable robotics. Validation experiments with seven healthy subjects, including treadmill walking and free level-ground walking, were conducted to evaluate the performance of the proposed method. In treadmill validation, the root-mean-square error (RMSE) of the phase estimator was 4.14 ± 1.68% for steady speeds, while it was 6.77 ± 2.29% for unsteady-speed walking. In level-ground validation, the average RMSE of the phase estimator was 4.59 ± 1.76%. Preliminary experiments were also conducted using a single-joint hip exoskeleton to demonstrate the usability of our method in lower-limb wearable robots.