Exploratory data for Offline RL (ExORL)

该数据集采取以数据为中心的离线强化学习方法，通过无监督的、无需奖励的探索生成数据，并用下游任务的奖励对其进行重标记，以训练策略。此外，该数据集使得传统的非策略强化学习算法在下游任务上的表现能够超越或媲美最先进的离线强化学习算法。该数据集包含了一百万次交互，这些交互是在多个剧集过程中收集的，适用于多任务离线强化学习任务。

This dataset adopts a data-centric offline reinforcement learning approach, generating data through unsupervised, reward-free exploration and relabeling it with rewards from downstream tasks to train policies. Furthermore, this dataset enables traditional off-policy reinforcement learning algorithms to outperform or match the performance of state-of-the-art offline reinforcement learning algorithms on downstream tasks. This dataset contains one million interactions collected across multiple episodes, which is suitable for multi-task offline reinforcement learning tasks.