Learning from Experience in Manipulation Planning: Setting the Right Goals

Abstract

In this paper, we describe a method of improving trajectory optimization based on predicting good initial guesses from previous experiences. In order to generalize to new situations, we propose a paradigm shift: predicting qualitative attributes of the trajectory that place the initial guess in the basin of attraction of a low-cost solution. We start with a key such attribute, the choice of a goal within a goal set that describes the task, and show the generalization capabilities of our method in extensive experiments on a personal robotics platform.