RoboCupSimData
收藏资源简介:
通过这个项目,我们提供了必要的软件,以使用现有团队(二进制文件)创建用于在 RoboCup 足球模拟中学习的数据集。足球模拟器通常会在比赛进行时记录比赛(即真实数据),而此模拟中每个球员获得的本地、嘈杂和受限信息通常会丢失。我们的模拟器补丁增加了记录每个玩家的个人感知的能力,还增加了将这些记录转换为 CSV 文件的工具。 我们还提供了来自 RoboCup 足球模拟联赛 (2D) 中一些顶级球队(2016 年和 2017 年)比赛的大型数据集,其中 11 个机器人(“代理”)组成的球队相互竞争。总的来说,我们使用了 10 支不同的球队进行比赛,产生了 45 对独特的配对。对于每个配对,我们进行了 25 场比赛(10 分钟),导致 1125 场比赛或超过 180 小时的游戏时间。生成的 CSV 文件是 17GB 数据(压缩)或 229GB(解压缩)。 该数据集是独一无二的,因为它既包含“地面实况”数据(场上所有对象的全局、完整、无噪声信息),也包含每个机器人的嘈杂、局部和不完整的感知。这些数据以 CSV 文件以及原始足球模拟器格式提供。
Through this project, we provide the necessary software to create datasets for machine learning in RoboCup Soccer Simulation using existing teams (binary executables). Standard soccer simulators typically record matches in real-time during gameplay, which constitutes global ground truth data, yet the local, noisy, and bounded perception information obtained by each individual player during the simulation is usually lost. Our simulator patch adds the capability to record each player's personal perceptual data, alongside tools to convert these records into CSV files. We also release a large-scale dataset sourced from matches held by some top-performing teams in the RoboCup Soccer Simulation League (2D) between 2016 and 2017, where teams composed of 11 robotic "agents" compete against each other. In total, we utilized 10 distinct teams, resulting in 45 unique team pairings. For each pairing, we conducted 25 matches (each lasting 10 minutes), yielding a total of 1,125 matches or over 180 hours of gameplay. The generated CSV files occupy 17 GB when compressed and 229 GB when decompressed. This dataset is unique in that it contains both ground truth data—global, complete, noise-free information about all objects on the field—and the noisy, partial, and incomplete perceptions of each robotic agent. The data is provided in both CSV format and the original soccer simulator native format.




