teamcraft/TeamCraft-Data-Cen

TeamCraft-Data-Cen数据集旨在开发Minecraft中的多模态、多代理协作。该数据集包含55,000个任务变体，每个任务变体都伴随着多模态提示和程序生成的专家演示。数据集的结构包括输入（多模态提示、观察和历史）和输出（动作空间）。数据集支持的研究领域包括多代理强化学习、多模态学习和任务规划以及协作环境中的嵌入式AI。数据集的大小约为127GB，包含超过一百万张图像和一个综合的训练JSON文件。

The TeamCraft dataset is designed to develop multi-modal, multi-agent collaboration in Minecraft. It features 55,000 task variants defined by multi-modal prompts and procedurally generated expert demonstrations. The dataset includes multi-modal prompts (language + images), observations, and actions. The dataset is structured with inputs including multi-modal prompts, observations, and history, and outputs consisting of high-level actions for all agents. The tasks covered include building, clearing, farming, and smelting. The dataset supports research in multi-agent reinforcement learning, multi-modal learning, and embodied AI in collaborative environments.