Prediction of Undesired Situations based on Multi-Modal Representations