Active learning of affordances for robot use of household objects

Learning to perform household tasks is a key step towards developing cognitive service robots. This requires that robots are capable of discovering how to use human-designed products. In this paper, we propose an active learning approach for acquiring object affordances and manipulation skills in a bottom-up manner. We address affordance learning in continuous state and action spaces without manual discretization of states or exploratory motor primitives. During exploration in the action space, the robot learns a forward model to predict action effects. It simultaneously updates the active exploration policy through reinforcement learning, whereby the prediction error serves as the intrinsic reward. By using the learned forward model, motor skills are obtained to achieve goal states of an object. We demonstrate through real-world experiments that a humanoid robot NAO is able to autonomously learn how to manipulate two types of garbage cans with lids that need to be opened and closed by different motor skills.