NVIDIA发布GR00T 2世界行动模型
NVIDIA在CoRL推出开放世界行动模型GR00T 2。该模型从人类视频和机器人演示中学习,并同时预测未来观测与动作。作为一款20亿参数的Transformer,其训练数据包括402,730小时第一人称人类视频、10,184小时手持夹爪演示和23,500小时机器人遥操作数据。NVIDIA称,GR00T 2在未针对目标任务机器人进行演示训练的情况下,50项操作任务中有40项至少成功一次;每项任务展示一次后的平均成功率为71%,展示五次后为95%。不过,NVIDIA指出这些任务是从模型原本就容易学会的项目中挑选,因此结果的普适性有限。NVIDIA还展示了搭载Newton的Isaac Lab 3.0,并称其在视觉式灵巧操作基准中的训练吞吐量最高为使用PhysX的Isaac Lab 2.3的2.5倍。代码发布列定于12月;鸿海、西门子和Nexuni正在测试该模型,但均未宣布投入生产部署。

Nvidia unveiled GR00T 2, an open AI model that helps robots learn delicate hand skills, trained partly on 100,000 hours of human video. It reached 95% average success with five demonstrations per task in Nvidia’s tests. Foxconn, Siemens and Nexuni are testing it.
Nvidia has unveiled GR00T 2, an AI model that helps robots learn delicate hand skills partly by watching videos of people, the company said on Monday. It announced the model at the Conference on Robot Learning in Austin, Texas.
#NVIDIA#GR00T 2#Isaac Lab 3.0#Newton#Human-1M-100K#Cosmos-Predict2.5-2B