minecraft-vla-videos
收藏资源简介:
HEXTANT Minecraft Rollout Videos (VPT ranged combat) 是一个用于强化学习和具身人工智能研究的公开数据集,专门针对Minecraft游戏环境。该数据集包含使用VPT(MineStudio_VPT.rl_for_shoot_animals_2x)策略进行远程弓箭战斗的rollout游戏过程录制视频。数据集仅包含视频文件,训练用的parquet格式文本动作数据和标签存储于单独的私有仓库。视频内容展现了高度的环境多样性,在kite_ranged场景下包含7种生物群落(热带草原、沼泽、平原、沙漠、针叶林、森林、随机)、13种生物类型(牛、骷髅、羊、鸡、尸壳、猪、苦力怕、僵尸猪灵、溺尸、僵尸、流浪者、蜘蛛、洞穴蜘蛛)、3种天气条件(晴朗、下雨、雷暴)以及3个时间阶段(白天、夜晚、午夜)。数据集规模约为3200个视频文件,对应约330个任务变体,视频格式为MP4,采用第一人称视角,分辨率为128x128,帧率为20fps。根据可测量的击杀数据,该数据集中智能体的远程击杀成功率约为45%。视频附带简化的质量标签,采用三级分类:远程击杀(2分)、远程命中(1分)、无效射击或未射击(0分)。该数据集适用于视觉语言动作模型训练、强化学习策略评估、Minecraft智能体行为分析等研究任务。
HEXTANT Minecraft Rollout Videos (VPT ranged combat) is a public dataset for reinforcement learning and embodied AI research, specifically designed for the Minecraft game environment. It contains rollout video recordings of long-range bow combat using the VPT (MineStudio_VPT.rl_for_shoot_animals_2x) strategy. The dataset includes only video files, with training parquet-format text action data and labels stored in a separate private repository. The video content exhibits high environmental diversity, featuring 7 biomes (savanna, swamp, plains, desert, taiga, forest, random), 13 creature types (cow, skeleton, sheep, chicken, husk, pig, creeper, zombified_piglin, drowned, zombie, stray, spider, cave_spider), 3 weather conditions (clear, rain, thunderstorm), and 3 time phases (day, night, midnight) in the kite_ranged scenario. The dataset comprises approximately 3200 video files, corresponding to about 330 task variants, in MP4 format with a first-person perspective, 128x128 resolution, and 20fps frame rate. Based on measurable kill data, the agents long-range kill success rate in this dataset is approximately 45%. Videos come with simplified quality labels in a three-level classification: long-range kill (2 points), long-range hit (1 point), and ineffective shot or no shot (0 points). This dataset is suitable for research tasks such as vision-language-action model training, reinforcement learning policy evaluation, and Minecraft agent behavior analysis.
数据集概述
数据集名称:HEXTANT Minecraft VLA Rollout Videos (VPT ranged combat)
许可证:MIT
任务类别:强化学习
标签:Minecraft、具身AI、VLA、VPT
数据集规模:约3200个视频文件,覆盖约330个任务变体。
数据内容
- 本数据集仅包含视频文件,不包含用于训练的parquet格式文本动作标签或标签文件。
- 视频来源于使用VPT策略(
MineStudio_VPT.rl_for_shoot_animals_2x)在Minecraft中进行远程(弓箭)战斗的游戏录像。 - 视频格式:POV视角,分辨率128×128,帧率20fps。
- 文件组织方式:
<task>/<stamp>/*.mp4。
环境多样性(kite_ranged场景)
- 生物群系(7种):savanna、swamp、plains、desert、taiga、forest、lottery
- 生物种类(13种):cow、skeleton、sheep、chicken、husk、pig、creeper、zombified_piglin、drowned、zombie、stray、spider、cave_spider
- 天气(3种):clear、rain、thunder
- 时段(3种):day、night、midnight
质量标签(简要)
- 远程击杀率(基于可测量击杀)约为45%。
- 质量等级划分:
ranged_kill(2):远程击杀ranged_hit(1):远程命中fired_ineffective/no_fire(0):无效射击或未射击
- 详细的分类本体和附加信息请参考相关研究仓库。
其他说明
该数据集是OpenHA/HEXTANT研究项目的一部分,公开内容仅包含视频。





