| Title | A Developmental Approach to Goal-Based Imitation Learning in Robots |
| Publication Type | Report |
| Year of Publication | 2013 |
| Authors | Chung MJae-Yoon, Lei J, Gupta A, Fox D, Meltzoff AN, Rao RPN |
| Series Title | University of Washington Technical Reports |
| Document Number | UW-CSE-13-11-04 |
| Date or Month Published | 2013-11-04 |
| Institution | University of Washington |
| City | Seattle |
| Abstract | We propose a new developmental approach to goal-based imitation learning that allows a robot to: (1) learn probabilistic models of actions through self-discovery and experience, (2) utilize these learned models for inferring the goals of human demonstrations, and (3) perform goal-based imitation for human- robot collaboration. Our approach is based on Meltzoff’s “Like- me” hypothesis in developmental science, which states that children use self-experience to bootstrap the process of intention recognition and imitation. Such an approach allows a robot to leverage its increasing repertoire of learned behaviors to interpret increasingly complex human actions, even when the robot has very different actuators from humans. We present preliminary results illustrating our approach using a simple robotic tabletop organization task. We show that the robot can learn a probabilistic model of its actions on a small set of objects, and use this model for both goal inference and goal-based imitation of human actions. We also present results demonstrating that the robot can use its learned probabilistic model to seek human assistance whenever it recognizes that its inferred actions are too uncertain, risky, or impossible to perform.
|
| Citation Key | 9700 |
