RetroAct
收藏资源简介:
RetroAct数据集是一个包含974个虚拟环境的集合,这些环境基于Stable Retro框架。数据集通过注释每个环境的行为和控制来组织,以便于生成具有相似行为的跨环境的大规模交互数据集。数据集中的环境主要是各种风格的复古游戏,共分为483个具有相似控制的行为类别,其中最常见的是平台游戏。通过随机代理在环境中自动收集数据,创建了两个子数据集:Platformers-200和Platformers-50,分别用于预训练和微调。
The RetroAct dataset is a collection of 974 virtual environments built upon the Stable Retro framework. The dataset is structured by annotating the behaviors and control schemes of each individual environment, to facilitate the generation of large-scale cross-environment interaction datasets that exhibit similar behavioral characteristics. The environments in the dataset are primarily retro games across various styles, which are categorized into 483 behavioral categories with consistent control mechanisms, among which platform games are the most common. Data was automatically collected via random agents interacting within these environments, leading to the creation of two subdatasets: Platformers-200 and Platformers-50, which are respectively intended for pre-training and fine-tuning.

- 1Exploration-Driven Generative Interactive Environments索菲亚大学St. Kliment Ohridski学院INSAIT · 2025年



