Towards semi-supervised learning of semantic spatial concepts for mobile robots