kantine/BotFails

BotFails是一个专为机器人操作中通用故障检测研究设计的新型多模态数据集。该数据集通过主臂遥操作收集，包含视觉、本体感觉和自然语言任务指令等多模态观察数据，涵盖了10个不同的操作任务，分为家庭和工业两个领域。每个任务都包含多个异常类型的注释，包括非故障异常和真实故障。数据集结构包括标签、正常训练和测试三个主要目录，采用LeRobot数据集格式存储观测数据、视频、元数据和标签。

BotFails is a novel dataset specifically designed to support research on general failure detection in robotic manipulation. It provides multimodal observations — including vision, proprioception, and natural language task instructions — collected across a semantically diverse set of manipulation scenarios. The dataset covers 10 distinct manipulation tasks spanning domestic and industrial domains, with episode-level annotations identifying the presence and type of anomaly. The dataset is organized into labels, normal_train, and test directories, following the LeRobot dataset format for storing observations, videos, metadata, and labels.