Natural Actor-Critic for Robust Reinforcement Learning with Function Approximation