Learning from Demonstration Using Variational Bayesian Inference

Abstract

Learning from demonstration (LFD) is an active area of research in robotics. There are many approaches to LFD. One of the most widely used approaches is the combination of Gaussian Mixture Model learning for modeling and Gaussian Mixture Regression for behavior generation (GMM/GMR) due to its advantages including easy learning using Expectation Maximization and the simplicity of serializing learned behaviors as well as the ability to model internal correlations and constraints within the task.

A critical parameter that affects the accuracy of learned behavior in GMM/GMR is the number of components in the mixture. A handful of approaches for selecting this number can be found in the literature including classical model selection methods like Bayesian Information Criteria and Akaik Information Criteria and more advanced methods including Dirichlet Process modeling. These approaches are either wasteful of computational resources or hard to implement. This paper introduces a LfD approach which uses GMM with a variational Bayesian Inference (VB) approach to select the number of Gaussians that best fit the data. The proposed method is compared to classical model selection approaches and a recently proposed symbolization based method and is shown to provide an appropriate balance between execution speed and model accuracy.