Improved Robustness and Safety for Pre-Adaptation of Meta Reinforcement Learning with Prior Regularization