Meta-World Offline Dataset

该数据集包含了5万个由随机策略在模拟环境中收集的剧集，这些剧集均带有2311条独特的自然语言指令标签。该数据集主要用于训练一个语言条件下的奖励函数，并评估机器人在各种基于视觉的操作任务上的性能表现。数据规模涉及5万个剧集和2311条独特的指令，任务重点在于基于视觉的操作任务。

This dataset contains 50,000 episodes collected via random policies in simulated environments, where each episode is associated with one out of 2311 unique natural language instruction labels. This dataset is primarily designed for training language-conditioned reward functions and evaluating robotic performance across a wide range of vision-based manipulation tasks. In terms of scale, the dataset covers 50,000 episodes and 2311 unique natural language instructions, with its core tasks focusing on vision-based manipulation tasks.