An object-centric paradigm for robot programming by demonstration

Abstract

In robot programming by demonstration, we hypothesize that in a class of procedural tasks where the end goal primarily consists of the states of objects that are external to a task performer, one can significantly reduce complexity of a robot learner by not processing a human demonstrator’s motions at all. In this class of tasks, object behaviors are far more critical than human behaviors. Based on this virtual demonstrator hypothesis, this paper presents a paradigm where a human demonstrates an object manipulation task in a simulated world without any of the human demonstrator’s body parts being sensed by a robot learner. Based on the object movements alone, the robot learns to perform the same task in the physical world. These results provide strong support for the virtual demonstrator hypothesis.