Enhancing generalization in autonomous driving through track-agnostic reinforcement learning