aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time AI score47

Nvidia releases GR00T 2, using more human video to train robotsMachine translation

AI introduction

GR00T 2 is an open world-action model for robot arms, hands and humanoids. Nvidia pretrained it on 402,730 hours of human head-camera video, 10,184 hours of handheld-gripper demonstrations and 23,500 hours of teleoperated robot data. Ordinary human video lacks motor commands, which previously made it difficult for such models to use.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

英伟达发布了 GR00T 2,这是一款可控制机械臂、机器人手和人形机器人的开放世界动作模型。

机器人通常通过录像学习,录像将每一帧摄像头画面与精确的电机指令配对,而这些数据只能由人操控真实机器人获得。

普通的人类活动视频成本低得多,数量也丰富得多,但其中没有电机指令,因此早期模型几乎无法利用它。

GR00T 2 通过使用 402,730 小时的人类头戴摄像头视频和 10,184 小时的手持夹爪操作演示进行预训练,缓解了这一问题,而遥控操作机器人的数据仅有 23,500 小时。

Source: Rohan Paul · x.com

Model
Found an error? Send a correction