To this end, we introduce an unsupervised active pre-training algorithm for diverse behavior induction (APD). We explicitly characterize the behavior variables with a state-dependent sampling method, ...