Robust Reinforcement Learning via Adversarial Kernel Approximation