Generation of search behavior of robots by an extended probabilistic flow control

Abstract

The probabilistic flow control method (PFC) can generates behaviors of robots that compensate information uncertainty. We improve PFC and verify it on an actual robot in this paper. The original PFC is a decision making method for POMDPs (partiall observable Markov decision processes). It is a modified method of the Q_MDP value method, which makes a robot control so as to maximize an expected value of an evaluation function when the state of the robot is only known as a probability distribution.
PFC biases the expected value "optimistically," and it makes a robot behave as if searching a goal of a task. In this paper, we make the intensity of bias adjustable. With appropriate parameters of the intensity, we find that robots behave more effectively than the original PFC method. Then the improved PFC method is implemented on an actual mobile robot that has poor self-localization ability. The robot shows goal search behavior that compensates the uncertainty of self-localization with the improved method.