Adaptive Learning Control in Evolutionary Design of Mobile Robots,

International audience This is a study of an application of neuraltechnics to the learning of control laws withinthe framework of the evolutionary design of roboticssystems. The present paper proposes the replacementof the evolutionary synthesis of the individual’scontrol law by its learning. The learning of neuralcontroller is carried out on-line when the robot undergoesevaluation tests. Thus, a robot that is apriori inadequate to solve a task can, thanks to thetraining it goes through, improve its performance. Itparticipates then to the global improvement of thepopulation while it would have been eliminated withoutlearning. A mobile robot that could be equippedwith up to 4 independent driving wheels and thatmust attain a given configuration will be taken asan example. The whole unit uses a simulation ofthe robot and its environment in which all dynamiceffects are taken into account. Results show the accuracyand strength of the method since even thestructures which would have been in fact eliminatedto carry out this kind of task, are controlled withreasonable efficiency.