Solving Complex Tasks Hierarchically from Demonstrations

Abstract

Robot learning from demonstration is an approach to enable non-expert users to program a robot for new tasks effectively. In this paper, we consider the problem of how to learn complex tasks from demonstrations and how to use learned knowledge to solve a new task. A two-layer hierarchical framework is proposed. In the bottom layer, a Bayesian non-parametric learning algorithm is used to segment the trajectory of the whole task into subtasks or skills which are consequently trained as probabilistic movement primitives. In the top layer, a linear temporal logic specification is used in the synthesis framework that generalizes to new tasks. Our proposed approach is tested and validated with a coffee refill experiment on a Baxter humanoid robot.