This project develops autonomous goal-conditioned reinforcement learning agents capable of selecting their own training goals while learning environment-agnostic representations. The work investigates methods for unsupervised goal selection based on flow-matching and discriminator-guided acquisition strategies, aiming to reduce dependence on task-specific supervision while maintaining strong performance across a diverse set of environments. Preliminary results obtained during ongoing work indicate that several proposed unsupervised methods achieve performance comparable to, and in some cases exceeding, supervised state-of-the-art baselines under stricter constraints. The requested resources will be used to complete statistical validation, hyperparameter studies, and ablation experiments across multiple reinforcement learning benchmark environments in preparation for publication submission during autumn 2026. The project constitutes a major part of a WASP-funded PhD project at Lund University and involves AI/ML method development with no sensitive data.