Researchers propose a new approach called QDOS (Quality-Diversity Offline Skill learning) for improving the performance of reinforcement learning policies. QDOS combines offline and online data to extract diverse and high-value skills from pre-collected datasets. This is achieved through an Advantage-Weighted Quality-Diversity pretraining objective, which estimates the advantage of each trajectory segment. The approach also incorporates a dual dataset reuse strategy, where offline data is used for both skill pretraining and populating the online replay buffer via pseudo-labeling. Experiments show that QDOS outperforms strong baselines in various tasks, including structured manipulation and unstructured locomotion. --- Why it matters: This matters to researchers because it provides a new method for improving reinforcement learning policies using offline data, which can be particularly useful when dealing with sparse rewards or limited online data. The approach's ability to extract diverse and high-value skills can accelerate exploration and improve final returns in challenging domains. Source: https://arxiv.org/abs/2608.19684