Locomotion and Manipulation Environments

该数据集包含了4个运动环境（Walker、Cheetah、Quadruped和Stickman）以及1个操作环境（厨房），用于在没有奖励的情况下训练智能体，共涉及35项任务，这些任务通过文本提示来指导。数据集混合了结构化数据（重放缓冲区）和非结构化数据（探索数据），并将明确的奖励信息替换为简短的任务描述。这是一项大规模研究，旨在探究强化学习中使用视觉和语言提示进行多任务泛化的问题。

This dataset contains four locomotion environments (Walker, Cheetah, Quadruped, and Stickman) and one manipulation environment (Kitchen). It is developed for training AI agents in the absence of reward signals, encompassing a total of 35 tasks guided by textual prompts. The dataset integrates structured data (replay buffers) and unstructured data (exploration data), and substitutes explicit reward signals with concise task descriptions. This is a large-scale study aimed at investigating multi-task generalization in reinforcement learning using visual and language prompts.