Learning Motion Trajectories of Differential Drive Mobile Robot Using Gaussian Mixtures and Hidden Markov Model

Abstract

In this paper we focus on learning of motion trajectories for differential drive mobile robot. Learning from Demonstration (LfD) is applied to enable mobile robot to learn and reproduce trajectories which are hard to model due to unknown dynamics and uncertainty. To solve this problem, we build the probabilistic model of the mobile robot motion in two steps. Firstly, Gaussian Mixture Model (GMM) of incremental robot motions is built, while in the second step the hidden Markov Model (HMM) is applied to extract transition matrix and most likely sequence of GMMs for each training trajectory. The final step assumes estimating the optimal sequence of incremental motions for each of N different training trajectories. To summarize the main idea: each of N hidden Markov models model the desired robot motion given as a sequence of M low level robot motions. In contrast to conventional mixture models, hidden Markov model captures temporal dependencies between mixtures, which is i a significant advantage of HMM over conventional mixture models. Experimental results demonstrate applicability and optimal performance of the proposed learning algorithm.