Preliminary Study in Motion Assistance of Soft Exoskeleton Robot based on Data-driven Kinematics Model Learning

Abstract

Exoskeleton is widely used to enhance human mobility. Especially in recent years, the soft exoskeleton robots have developed rapidly, which could realize natural human-machine physiological coupling. However, the motion patterns and physiological parameters are significant various between different subjects. The parameters of the soft exoskeletons change differently during motion. In this paper, we proposed kinematics model based on data-driven model learning. The proposed model learning method not only has the fast learning ability of model-based controller, but also has the adaptability of sensor-based controller. Firstly, we use the data of the rigid model to pre-train the kinematics model network, which can make the output of the network consistent with the kinematics model. Then, we use the sensors to collect the actual motion data and send the motion data into the pre-trained neural network model. By increasing the iteration times of training, the network model outputs model parameters that are consistent with the trend of the simulation model. Through the training and learning of the bionic motion platform, the speed of learning adaptation in human body can be accelerated.