The Variational Coupled Gaussian Process Dynamical Model

Dmytro Velychko, Benjamin Knopp, Dominik Endres

Abstract: We present a full variational treatment of the Coupled Gaussian Process Dynamical Model (CGPDM) with non-marginalized coupling mappings. The CGPDM generates high-dimensional trajectories from coupled low-dimensional latent dynamical models. The deterministic variational treatment obviates the need for sampling and facilitates the use of the CGPDM on larger data sets. The non-marginalized coupling mappings allow for a flexible exchange of the constituent dynamics models at run time. This exchange possibility is crucial for the construction of modular movement primitive models. We test the model against the marginalized CGPDM, dynamic movement primitives and temporal movement primitives, finding that the CGPDM generally outperforms the other models. Human observers can hardly distinguish CGPDM-generated movements from real human movements.