Probabilistic Modeling of Human Locomotion for Biped Robot Trajectory Generation

The wheel-type robot has found numerous applications in hospitals, restaurants, entertainment, the automation industry, etc., and shows its applicability in solving the tasks efficiently. However, it failed to achieve the same efficiency in an unstructured environment that is mostly found in the real world. Thus, a biped robot can replace the wheel-type robot for better performance. The biped robot has many joints which make it a complex higher degree of freedom system. Hence, the designing of the controller, reference trajectory generation, state estimation and, filter design for feedback signal is a very cumbersome task. This paper focuses on the generation of the reference trajectories. Since human locomotion is optimal naturally, therefore, the human data is used for this study, which is collected at Robotics and Machine ANalytics (RAMAN) Lab, MNIT, Jaipur, India. In the literature, various authors have implemented model-based learning methods to develop a model based on data. However, these models suffer from model bias i.e., it is assumed that learned model accurately define the real system. Therefore, in this paper, the authors have proposed probabilistic models to model the human locomotion data. The reference trajectory is generated using the Bayesian ridge regression, Automatic relevance determination regression, and Gaussian process regression. The performance evaluation of developed models are based on average error, maximum error, root mean square error, and percentage normalized root mean square error.