Compositional Automata Embeddings for Goal-Conditioned Reinforcement Learning