原帖引用 · AI 翻译
@shawnshenjx 真的是一次重大发布。🫡
@rohanpaul_ai 谢谢 Rohan!从摄像头到 JSON,正是这个意思。谁都不该为了找到一次抓取动作而在数小时的视频里来回翻找。
原文
@rohanpaul_ai Thanks Rohan! Camera to JSON is exactly it. Nobody should be scrubbing through hours of video to find one grasp
根据双手正在做的动作来搜索视频,是目前机器人数据领域最有用的想法之一。
从工人头上的摄像头,到训练运行中的 JSON 文件。
来自 https://t.co/yaP1XdpqxR 的 Memories Upload,会为第一人称影像中的每个时刻生成附带时间戳的双手各自的描述,以便按动作查询影像档案。
机器人通过观察人类来学习手部技能,因此 https://t.co/yaP1XdpqxR 付钱请世界各地的人们录下自己工作时的双手,然后对这些影像进行清理、标注,并出售给研发机器人的实验室。
原文
Searching video by what the hands are doing is one of the most useful ideas in robotics data right now. from a camera on a worker’s head to a JSON file in a training run. Memories Upload, from https://t.co/yaP1XdpqxR, gives every moment of first person footage a time-stamped caption for each hand, so the archive can be queried by action. Robots learn hand skills by watching humans, so https://t.co/yaP1XdpqxR pays people worldwide to record their hands at work, then cleans, labels and sells that footage to labs building robots.