Near-Optimal Second-Order Guarantees for Model-Based Adversarial Imitation Learning