Capturing Human Body Dynamics Using RNN Based on Persistent Excitation Data Generator

Human body walking movement involves both single and double support phases and is considered difficult to model. The aim of this study was to develop a method to capture human body dynamics during walking using Recurrent Neural Networks (RNN). In addition, a novel method using persistent excitation data generator is proposed to generate kinematic data to train the RNN in the absence of laboratory measurements. Kinematic data were applied to human body mathematical model to obtain required joint torques during bipedal walking. The RNN was used to approximate human body kinematics resulting from the joints torques for the walking movement. In order to test validity of the RNN model, model output was compared with human walking data captured in the laboratory. Simulation results show the model was able to approximate the joint angles during human walk with a low (10^-4 m) mean squared error for one stride.