Robot self-preservation and adaptation to user preferences in game play, a preliminary study

It is expected that in a near future, personal robots will be endowed with enough autonomy to function and live in an individual´s home. This is while commercial robots are designed with default configuration and factory settings which may often be different to an individual´s operating preferences. This paper presents how reinforcement learning is applied and utilised towards personalisation of a robot´s behaviour. Two-level reinforcement learning has been implemented: first level is in charge of energy autonomy, i.e. how to survive, and second level is involved in adapting robot´s behaviour to user´s preferences. In both levels Q-learning algorithm has been applied. First level actions have been learnt in a simulated environment and then the results have been transferred to the real robot. Second level has been fully implemented in the real robot and learnt by human-robot interaction. Finally, experiments showing the performance of the system are presented.