Generalizable Imitation Learning Through Pre-Trained Representations