Evolutionary Design of Moving Objects

The aim of this work is to implement a system for automatic evolutionary design of virtual robot controllers. In particular, Linear Genetic Programming representation combined with a steady-state genetic algorithm will be used to find a suitable program that will lead a given virtual robot across a sequence of points denoting a predefined trajectory. The MuJoCo physics engine is applied to allow the user to specify the robot shape and to evaluate its behavior according to candidate programs generated by the genetic algorithm. The goal is to train the robot to follow the given path by optimizing the distance of the robot from the given points during the simulation. The optimization is performed by evolving the programs for a given number of generation of the genetic algorithm. Several sets of experiments will be presented and obtained results will be evaluated