Grounding Aleatoric Uncertainty for Unsupervised Environment Design Michael Dennis Jack Parker-Holder Andrei Lupu UCL & Meta AI UC Berkeley University of Oxford MILA & Meta AI Heinrich Küttler

May-23-2025, 18:19:08 GMT–Neural Information Processing Systems 

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