Learning sparse models for task and motion planning

Abstract

In order for a robot to be able to act effectively in the complex environments we live in, it must be able to have an understanding of how its actions may effect the objects around it. In this work, we present a method for learning the relationships between some robotic action and the objects in the environment in order to predict which objects will be affected by that action. By leveraging the constraint that the effect of the robotic action is sparse, we are able to successfully learn these relationships, and apply them to make learning models of these actions more efficient, and also enable existing models to be more effective when used for task and motion planning.