Reinforcement learning approach to cooperation problem in a homogeneous robot group

A distributed autonomous approach to adaptive system design is investigated through the cooperative carrying problem (CCP) using a homogeneous connected robot group. The task of carrying an object is supposed to be given only to the group of robots, for the purpose of putting the main interest on how to design online task decomposition mechanisms which should be autonomous and adaptive. The robot group dealt by this paper is comprised of same autonomous robots connected by a load. Reinforcement learning (RL) is adopted for a basic framework of the robot´s decision-making mechanism, so that quick online learning can be expected. However, since RL in a simple form is not effective in developing a stable cooperative behavior in a multi-agent environment, a novel decision-making mechanism is designed using two RL units, in which the first RL unit is for predicting its partners´ next states, and the other is for generating an action of its own. Several empirical experiments for three connected robots are conducted on a computer in order to investigate the effectiveness of the proposed mechanisms