We present a computational model that can learn event tax-onomies online from the continuous sensorimotor informationflow perceived by an agent while interacting with its environ-ment. Our model implements two fundamental learning bi-ases. First, it learns probabilistic event models as temporal sen-sorimotor forward models and event transition models, whichpredict event model transitions given particular perceptual cir-cumstances. Second, learning is based on the principle of min-imizing free energy, which is further biased towards the detec-tion of free energy transients. As a result, the algorithm formsconceptual structures that encode events and event boundaries.We show that event taxonomies can emerge when the algo-rithm is run on multiple levels of precision. Moreover, weshow that generally any type of forward model can be used,as long as it learns sufficiently fast. Finally, we show that thedeveloped structures can be used to hierarchically plan goal-directed behavior by means of active inference.