Learning to detect the functional components of doorbell buttons using active exploration and multimodal correlation

This paper describes a large-scale experimental study, in which a humanoid robot learned to press and detect doorbell buttons autonomously. The models for action selection and visual detection were grounded in the robot's sensorimotor experience and learned without human intervention. Experiments were performed with seven doorbell buttons, which provided auditory feedback when pressed. The robot learned to predict the locations of the functional components of each button accurately. The trained visual model was also able to detect the functional components of novel buttons.