“A policy trained with Automatic Domain Randomization (ADR) achieves higher sim-to-sim transfer more quickly than policies trained with fixed domain randomization ranges from the start of training.”