Evolution of a Control Architecture for a Mobile Robot

Most work in evolutionary robotics used a neural net approach for control of a mobile robot. Genetic programming has mostly been used for computer simulations. We wanted to see if genetic programming is capable to evolve a hierarchical control architecture for simple reactive navigation on a large physical mobile robot. First, we evolved hierarchical control algorithms for a mobile robot using computer simulations. Then we repeated one of the experiments with a large physical mobile robot. The results achieved are summarized in this paper. 1 Motivation Currently mobile robots are usually programmed by hand. For many real-world applications this introduces considerable diiculties due to the complexity of the task 10]. Instead of programming the robots directly Darwinian evolution may be used to automate this process 2]. Thus one would reduce the problem to nding an appropriate tness function which describes how well a particular individual solves the task. The diierent approaches in evolutionary robotics are described by Noll et al. 21]. A number of researchers have already evolved control architectures for mobile robots using computer simulations. Few have worked with physical mobile robots. Brooks 4] emphasized the importance of using physical mobile robots as opposed to computer simulations. To our knowledge , those who used physical mobile robots only experimented with miniature robots such as the Khepera or a gantry-robot. Most work in evolutionary robotics was done using neural net In contrast to this work, we wanted to see if genetic programming 13, 15] can be used to evolve a hierarchical control architecture for a simple reactive navigation task on a large physical mobile robot. If evolution is carried out on a large physical mobile robot safety measures have to be taken such that possible damage to the environment or to the robot is avoided. As Matari c and Clii noted 17], most research in evolutionary robotics fo-cused on the evolution of simple tasks with long evaluation times. Recently, Noll 20] evolved a garbage collecting behavior for a mobile robot using accurate computer simulations of the Khepera and a neural net control architecture which also executed successfully on the real robot. Evaluation times may be reduced using computer simulations especially if more diicult tasks are addressed. Therefore it is also important to see if results from computer simulations can be transferred into the real-world.