Active perception model extracting object features from unlabeled data

In this paper, we developed an active perception model for a robot. Active perception is defined as the recognition using categories based on robot's behaviors. The behavior for extracting the object feature is called an exploratory behavior. Several models which acquire exploratory behaviors by learning have been proposed. In the past, we also have proposed a model that is able to learn suitable exploratory behaviors by a reinforcement learning. However, these previous models need the information of the label to which the objects belong during learning. Then, it is difficult to apply these models to acquisition of the category of an unknown object. In this paper, we improve the model that we proposed previously. We develop the model that is able to learn exploratory behaviors without the object's label. Mobile robot simulations are performed to verify the effectiveness of the model. The results indicated that the model can not only acquire the exploratory behavior but also form the categories of the object's features.