MotIF-1K

该数据集名为MotIF-1K，包含了653个人类演示和369个机器人演示，覆盖了13个任务类别，旨在为机器人动作理解的视觉-语言模型提供基准测试和精细调整。该数据集捕捉了各种上下文相关的动作，其中包括不同的中间轨迹，确保了对视觉-语言模型（VLMs）的全面挑战，使其在成功检测时考虑整个轨迹。该数据集规模涵盖了1022个演示，跨越了13项任务，其核心任务是机器人动作理解和成功检测。

This dataset, named MotIF-1K, consists of 653 human demonstrations and 369 robot demonstrations spanning 13 task categories. It is designed to provide benchmarks and fine-tuning resources for vision-language models (VLMs) targeting robotic action understanding. This dataset captures various context-aware actions including diverse intermediate trajectories, which poses comprehensive challenges to VLMs, requiring them to consider the full trajectory when conducting successful action detection. With a total of 1022 demonstrations across 13 tasks, the core objectives of this dataset are robotic action understanding and successful action detection.