DMCA
Learning From Observation and Practice Using Behavioral Primitives: Marble Maze (2004)
| Citations: | 3 - 0 self |
BibTeX
@MISC{Bentivegna04learningfrom,
author = {Darrin C. Bentivegna and Christopher G. Atkeson and Gordon Cheng},
title = {Learning From Observation and Practice Using Behavioral Primitives: Marble Maze},
year = {2004}
}
Share
 |
 |
 |
 |
||
OpenURL
Abstract
We explore how to enable robots to rapidly learn from watching a human or robot perform a task, and from practicing the task itself. A key component of our approach is to use small units of behavior, which we refer to as behavioral primitives. Another key component is to use the observed human behavior to define the space to be explored during learning from practice. In this paper we manually define task appropriate primitives by programming how to find them in the training data. We describe memory-based approaches to learning how to select, provide subgoals for, and perform behavioral primitives. We demonstrate both learning from observation and learning from practice on a marble maze task, Labyrinth. Using behavioral primitives greatly speeds up learning relative to learning using a direct mapping from states to actions. 1
