Mapping Model for Genesis of Joint Trajectory using Human Gait Dataset

Abstract

Walking gait trajectory genesis for the biped robot is a cumbersome task because of many degrees of freedom. In this paper, the authors have proposed machine learning models for kinematic modeling of human locomotion data. The human locomotion gait dataset has been taken from the MNIT gait dataset which is collected in previous work. The Gait dataset have contains the walking data of 120 subjects from different age-group. The machine learning models can become biased due to overfitting/underfitting. Thus, the K-fold cross-validation technique is employed for the training of machine learning models for mitigating biasing. In addition, two types of mappings have been developed i.e., one-to-one and many-to-one. One-to-one mapping has been used to map the knee, hip, and ankle trajectory to knee, hip, and ankle trajectory respectively. While many-to-one mapping has been used to map the combined trajectory of the knee, hip, and ankle to individual knee, hip, and ankle trajectory. The advantage of many-to-one is that it captures the connection between the knee, hip, and ankle efficiently. The accuracy of developed machine learning mapping is evaluated in terms of average error, maximum error, and root mean square error. The result shows that the Lasso family is performing the best among the developed models and also the many-to-one mapping outperforms the one-to-one mapping. Finally, an open discussion is presented for future research direction for gait generation and applications.