Low Level Segmentation for Imitation Learning Using the Expectation Maximization Algorithm

摘要

Imagine a robot that is able to develop skills on its own, without being programmed directly. This robot would be invaluable in any business, factory, or laboratory. Unfortunately, this problem, known as inductive learning, is very difficult, and has several varieties. One such is imitation learning. The overall process of imitation learning begins with one robot observing another robot performing a task. The watcher then breaks down, or segments, the demonstrating robot's actions into basic actions called planning units. Next the observing robot uses the planning units to create a plan that accomplishes the required task. The execution of a successful plan demonstrates that the robot has correctly implemented an inductive learning process. The scope of this research does not allow the problem of imitation learning to be discussed in its entirety; however, it does investigate an important subset of the larger problem. This paper focuses on the segmentation of the data, specifically how to break it up into the steps that provide the building blocks of the robots ultimate plan.