calicio/StarCraftMotion_sample

StarCraftMotion是一个基于《星际争霸II》回放构建的大规模基准数据集，用于在对抗性和部分可观测场景下的智能体模拟。每个样本是一个固定长度的场景窗口（145帧，16 FPS，约9秒），包含所有单位状态、动态地图层和每玩家经济时间序列。发布的数据集分割是“对抗性的”：场景被二次采样以增加交互密集窗口（相互可见和玩家间转换）的权重，使其成为部分可观测性下多智能体预测的压力测试。数据集包含469,187个场景，分为训练/验证/测试集：362,075 / 45,121 / 61,991。数据来源于64,327个回放级别的HDF5文件（Blizzard `3.16.1-Pack_1-fix`），涵盖多个地图，包括ID和OOD（仅测试）地图。数据集采用parquet格式存储，每行一个场景，所有每帧/每单位的数组列存储为类型化的Arrow `large_list`数组。数据集旨在用于对抗性部分可观测性下的多智能体模拟，以及雾化处理、ID与OOD地图泛化和交互密集场景的基准测试。

StarCraftMotion is a large-scale benchmark for agent simulation under adversarial and partial observability scenarios, built from StarCraft II replays. Each example is a fixed-length scenario window (`145` frames at `16 FPS`, ~9 seconds) containing all unit states, dynamic map layers, and per-player economy time series. The released split is adversarial: scenarios are subsampled to overweight interaction-heavy windows (mutual-visibility and inter-player transitions), making it a stress test for multi-agent prediction under partial observability. The dataset includes 469,187 scenarios, split into train/val/test sets: 362,075 / 45,121 / 61,991. It is sourced from 64,327 replay-level HDF5 files (Blizzard `3.16.1-Pack_1-fix`), covering multiple maps, including ID and OOD (test-only) maps. The data is stored in parquet format, with each row representing one scenario and all per-frame/per-unit array columns stored as typed Arrow `large_list` arrays. The dataset is intended for multi-agent simulation under adversarial partial observability and benchmarks for fog-of-war handling, ID vs OOD-map generalization, and interaction-heavy scenes.