Published: 2012 Februar
Type: Research Article
Institution: AIFB, KIT
Erscheinungsort / Ort: Karlsruhe
The creation of mechanisms to control the behavior of mobile robots becomes increasingly complex with the complexity of the desired tasks and the environment involved. Evolutionary Robotics is a methodology for the automatic generation of robot control mechanisms using basic principles from natural evolution, i.e., mutation, recombination and selection. In this paper, a decentralized online-evolutionary model based on finite state machines (FSMs) is investigated with respect to its ability of evolving controllers for a swarm of mobile robots. The model originally has been proposed for simulation and is now implemented for a real-robot scenario. A new selection operator is proposed which implements a decentralized version of tournament selection with arbitrarily many parents. In a real-robot experimentation setup using "Wanda" robots it is shown that basic behaviors such as Collision Avoidance and Gate Passing can be evolved by the proposed model. Furthermore, the paper discusses some implementation details specific to real hardware; this includes particularly a specially designed communication protocol used for selection, an onboard fitness function based solely on sensoric information, and a recovery mechanism for the case of hardware failures.
Organic Computing Learning Robots Arena