Goal-Conditioned RL

Goal-conditioned reinforcement learning studies how an agent can use a desired state or outcome to choose actions, so one learned policy can handle many tasks. Common approaches include hindsight experience replay, contrastive and self-supervised learning, and offline learning from recorded trajectories; hierarchical methods add intermediate goals for long-horizon tasks. Researchers evaluate these methods on maze navigation, locomotion, and manipulation benchmarks such as OGBench and D4RL, with some work extending to physical robots. A central challenge is making goals and value estimates generalize to difficult or unseen targets without losing reliable progress in long-horizon, sparse-reward, or contact-rich settings.