英伟达发布GR00T 2：从人类视频学习机器人动作
原标题：Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot
AI 摘要
英伟达发布机器人基础模型GR00T 2，采用“世界-动作模型”架构，在预测下一步动作的同时生成短期未来视频，从而将人类视频中的动作迁移到机械臂，减少传统遥操作训练成本。该模型使用超过1000块GB300芯片训练，包含约40万小时头戴摄像头拍摄的人类操作视频；实验室中五指机器人未经专门遥操作即可完成点蜡烛、切香蕉等任务，成功率10%至80%不等。英伟达还推出可在Jetson Thor上运行的轻量版GR00T 2 Turbo，并指出触觉传感器缺失、高层语言模型拖慢响应等落地障碍。
正文节选
Nvidia's model GR00T 2 watches humans on video, then imagines itself as the robot To teach robots new tasks, humans usually have to laboriously teleoperate them during training. Nvidia's new AI model GR00T 2 is designed to cut that effort by transferring human movements from videos to robotic arms. To train a robot for new tasks, engineers usually spend hours remotely controlling the machines. This so-called teleoperation is expensive, slow, and needs its own data for every robot model. In a new