Gaussian Process Based Model Predictive Controller for Imitation Learning

Humans still outperform robots in most manipulation and locomotion tasks. Research suggests that humans minimize a task specific cost function when performing movements. In this paper we present a Gaussian Process based method to learn the underlying cost function, without making assumptions on its structure, and reproduce the demonstrated movement on a robot using a linear model predictive control framework. We show that the learned cost function can be used to prioritize between tracking and additional cost functions based on exemplar variance, and satisfy task and joint space constraints. Tuning the weighting between learned position and velocity costs produces trajectories of the desired shape even in the presence of constraints. The approach is validated in simulation with a simple 2dof manipulator showing joint and task space tracking and with a 4dof manipulator reproducing trajectories based on a human handwriting dataset.