Treball realitat mitjançant programa de mobilitat. Norges teknisk-naturvitenskapelige universitet. Institutt for teknisk kybernetikk

Developing a control strategy for a wheeled snake robot can be difficult given the number of parameters involved. In this thesis we have studied the use of a reinforcement learning framework to develop a control strategy that allows a wheeled snake to lift its head as much as possible. The learning process has been done using a simulator developed for SINTEF's pipe inspection robot PIKo. The reinforcement learning methodology used has been CACLA with an RBF network as function approximator. Various alternatives have been proposed and used for the action space in simulations showing positive results. Issues with the simulator have been detected and workarounds proposed for them.