Grounding Aleatoric Uncertainty for Unsupervised Environment Design

Adaptive curricula in reinforcement learning (RL) have proven effective for producing policies robust to discrepancies between the train and test environment.