EgoExoLearn

EgoExoLearn是一个大规模数据集，旨在模拟人类通过观察他人演示视频来学习执行任务的过程。该数据集包含120小时的日常场景和专业实验室环境下的第一人称和演示视频数据。数据集不仅记录了视频，还提供了高质量的注视数据和详细的多种模态标注，为建模人类跨视角异步程序动作能力提供了平台。EgoExoLearn特别关注日常辅助和专业支持的潜在应用，通过引入如跨视角关联、跨视角动作规划和跨视角参考技能评估等基准，期望成为创建能够无缝学习人类在现实世界中演示的AI代理的重要资源。

EgoExoLearn is a large-scale dataset developed to simulate the human process of learning to perform tasks by observing others' demonstration videos. This dataset contains 120 hours of first-person and demonstration video data collected in daily scenarios and professional laboratory environments. In addition to recording videos, it also provides high-quality gaze data and detailed multi-modal annotations, serving as a platform for modeling humans' cross-view asynchronous procedural action capabilities. EgoExoLearn specifically focuses on the potential applications of daily assistance and professional support. By introducing benchmarks such as cross-view association, cross-view action planning, and cross-view reference skill evaluation, it is expected to become an important resource for creating AI agents that can seamlessly learn from human demonstrations in the real world.