Robotic Manipulation Datasets for Offline Compositional Reinforcement Learning

本数据集由宾夕法尼亚大学的研究团队创建，旨在支持离线组合强化学习技术的发展。数据集包含2.56亿条转移数据，每条数据记录了模拟机器人操作的一个步骤，包括状态、动作、奖励等信息。数据集通过不同的机器人代理收集，每个代理在256个不同的任务上执行操作，以模拟不同的性能水平。这些数据可用于训练和评估机器人学习算法，特别是在处理新任务时的泛化能力。数据集的应用领域包括机器人控制、自动化和人工智能，旨在解决如何有效利用预先收集的大规模数据来提高机器人学习效率的问题。

This dataset was developed by a research team at the University of Pennsylvania to support the advancement of offline combinatorial reinforcement learning techniques. It contains 256 million transition records, with each entry documenting a single step of simulated robotic manipulation, including key information such as state, action, reward, and other relevant details. The dataset was collected using multiple robotic agents, where each agent performs operations across 256 distinct tasks to simulate varying performance levels. This dataset can be used to train and evaluate robotic learning algorithms, particularly their generalization capabilities when tackling novel unseen tasks. Its application domains include robotic control, automation, and artificial intelligence, and it aims to address the challenge of effectively leveraging pre-collected large-scale data to enhance the efficiency of robotic learning systems.