This paper explores the suitability of commonly employed classification methods to action-selection tasks in robotics, and argues that a classifier’s introspective capacity is a vital but as yet largely under-appreciated attribute. As illustration we propose an active learning framework for semantic mapping in mobile robotics and demonstrate it in the context of autonomous driving. In this framework, data are selected for label disambiguation by a human supervisor using uncertainty sam- pling. Intuitively, an introspective classification framework – i.e. one which mod- erates its predictions by an estimate of how well it is placed to make a call in a particular situation – is particularly well suited to this task. To achieve an efficient implementation we extend the notion of introspection to a particular sparse Gaussian Process Classifier, the Informative Vector Machine (IVM). Furthermore, we leverage the information-theoretic nature of the IVM to formulate a principled mechanism for forgetting stale data, thereby bounding memory use and resulting in a truly life-long learning system. Our evaluation on a publicly available dataset shows that an introspective active learner asks more informative questions compared to a more traditional non-introspective approach like a Support Vector Machine (SVM) and in so doing, outperforms the SVM in terms of learning rate while retaining efficiency for practical use.