Exploring the Edges of Latent State Clusters for Goal-Conditioned Reinforcement Learning