Kinematic Modeling for Biped Robot Gait Trajectory Using Machine Learning Techniques

This paper presents the predictive models for biped robot trajectory generation. Predictive models are parametrizing as a continuous function of joint angle trajectories. In a previous work, the authors had collected the human locomotion dataset at RAMAN Lab, MNIT, Jaipur, India. The MNIT gait dataset consists of walking data on a plane surface of 120 human subjects from different age groups and genders. Thirty-two machine learning models (linear, support vector, k-nearest neighbor, ensemble, probabilistic, and deep learning) trained using the collected dataset. In addition, two types of mapping, (a) one-to-one and (b) many-to-one, are presented for each model. These mapping models act as a reference policy for the control of joints and prediction of state for the next time instant in advance if the onboard sensor fails. Results show that the deep learning and probabilistic learning models perform better for both types of mappings. Also, the probabilistic model outperforms the deep learning-based models in terms of maximum error, because the probabilistic model effectively deals with the prediction uncertainty. In addition, many-to-one outperforms the one-to-one mapping because it captures the correlation between knee, hip, and ankle trajectories. Therefore, this study suggests a many-to-one mapping using the probabilistic model for biped robot trajectory generation.