A Bayesian Method for Learning POMDP Observation Parameters for Robot Interaction Management Systems

Abstract

Technology has allowed robots to enter more personal settings in our society, appearing in environments alongside humans. These new situations provide a new set of problems, including the interaction and control of the robot by untrained humans, as well as adapting to an unconstrained world designed for humans. In this paper, we address the issue of robot learning in these environments while taking advantage of a user working alongside the robot. We present a framework for gradually learning a model of the user through a parametric observation function. This type of framework allows us to begin with a rough model of the world and adjust it from experience. By relying on an oracle providing optimal policy information, we are able to learn the observation model and adjust the robot’s behavior to match that of the oracle. We address the problems of learning and modifications necessary to handle the observation function and learning for rare events. We demonstrate the feasibilty of the algorithm on a robot-interaction domain and compare against a model-free method for action-selection.