Model-Learning-Based Partitioned Control of a Human-Powered Augmentation Lower Exoskeleton

This paper presents a new model-learning-based partitioned control strategy of a wearable lower exoskeleton. Here, the dynamics of the coupled human-exoskeleton system along with the corresponding resulting interaction torques are learned based on nonparametric regression technique and then incorporated in the control system for swing phase. This promising combination of partitioned control scheme and incremental model learning has provided the exoskeleton with the ability to adapt to various dynamics of human operators, to reduce the physical interaction between the operator and the exoskeleton, and minimize the sensory system used in the system. In this method, movement data containing the information of dynamics and interaction was collected in a number of walking cycles, and then training and prediction procedure were performed to aid the controller. We have demonstrated the feasibility of the proposed method through an exoskeleton prototype that employs walking sessions on a bench-testing over different ranges of walking speeds (0.8–1.2 m/s) with various subjects. In the simulation results, the control performance of the proposed algorithm was qualitatively compared to other fundamental controllers including a classical impedance control and a Rigid-Body-Dynamics model-based control. The resulting interaction torque reduced to greater than 32%. These results were re-evaluated in the real system and similar performance was achieved.