Self-Organized Construction by Population Coding

Abstract

The automatic generation of robot controllers by machine learning or evolutionary computation is still challenging and even more so for collective robotics. We follow the recently proposed paradigm of 'population coding' to compose robot swarms for collective construction. We define a controller template as finite state machine, enumerate a finite number of specified robot controller types to choose from, and use evolutionary robotics to evolve effective homogeneous and heterogeneous compositions of robot swarms using selections of these controllers. Besides an objective for solving the actual construction task we also add objectives for subtasks, and to minimize the number of different chosen robot types. For three variants of a collective construction task we find effective solutions with both homogeneous and heterogeneous swarms.