Shapelets are discriminative local patterns in time series, which maximally distinguish among different classes. Instead of considering full series, shapelet transformation considers the existence or absence of local shapelets, which leads to high classification accuracy, easy visualization and interpretability. One of the limitation of existing methods is robustness. For example, Search-based approaches select sample subsequences as shapelets and those methods intuitively may be not accurate and robust enough. Learning-based approaches learn shapelets by maximizing the discriminative ability. However, those methods may not preserve basic shape for visualization. In practice, shapelets are subjected to various geometric transformations, such as translation, scaling, and stretching, which may result in a confusion of shapelet judgement. In this paper, robust shapelet learning is proposed to solve above problems. By learning transform-invariant representative prototypes from all training time series, rather than just selecting samples from the sequences, each time series sample could be approximated by the combination of the transformations of those prototypes. Based on the combination, samples could be easily classified into different classes. Experiments on 16 UCR time series datasets showed that the performance of the proposed framework is comparable to the state-of-art methods, but could learn more representative shapelets for complex scenarios.