Re-using schematic grasping policies

Abstract

It can be difficult to generalize the solutions to grasping and manipulation problems because even small differences in problem context can require qualitatively different solutions. For example, small changes in the shape of an object to be grasped can necessitate different grasp strategies. In this paper, we introduce the action schema framework that represents generalized skills in a functional way such that all viable ways of accomplishing a task are represented as instantiations of the generalized skill. We also propose an on-line algorithm for learning how to instantiate the skill in a context-appropriate way. We test this approach with a robotic grocery bagging task where a dexterous humanoid robot learns to make correct qualitative decisions regarding how to grasp everyday grocery items and drop them in a paper bag