Self-Reconfiguring Modular Robot Learning for Lower-Cost Space Applications

Abstract

The applications for self-reconfiguring modular robots are far reaching. Their advantage comes from their ability to form into many different shapes which allows them to perform a multitude of tasks. This is especially advantageous for space applications, due to the associated high cost of launching equipment into space. Self-reconfiguring robots are not a panacea, however. A trade-off between flexibility and capability exists. A modular robot that configures itselfto do a specific task may be unable to perform as well as a robot that was designed specifically for that task. This disadvantage can be mitigated by having the units form into optimal configurations for the associated tasks. This is difficult to pre-define for tasks that are not known a-priori. To this end, the robots are designed to learn and choose the most efficient ways of reorganizing into the desired configuration. In this paper, the use of machine learning for commanding a self-reconfiguring modular robot system is investigated. A method is proposed under which the robot system would be able to learn through a trial and error process to form itself into different configurations, more optimally.