We present a computational model that can learn event taxonomies online from the continuous sensorimotor information flow perceived by an agent while interacting with its environment. Our model implements two fundamental learning biases. First, it learns probabilistic event models as temporal sensorimotor forward models and event transition models, which predict event model transitions given particular perceptual circumstances. Second, learning is based on the principle of minimizing free energy, which is further biased towards the detection of free energy transients. As a result, the algorithm forms conceptual structures that encode events and event boundaries. We show that event taxonomies can emerge when the algorithm is run on multiple levels of precision. Moreover, we show that generally any type of forward model can be used, as long as it learns sufficiently fast. Finally, we show that the developed structures can be used to hierarchically plan goal-directed behavior by means of active inference.