NVIDIA Unveils GR00T 2 at CoRL Conference to Enable Robots to Learn from Human Video
NVIDIA 推出 GR00T 2,让机器人通过人类视频学习
On November 9, 2026, NVIDIA introduced GR00T 2 at the CoRL conference in Austin. The model, built on a 2-billion-parameter Transformer based on Cosmos-Predict2.5-2B, achieves a 40% zero-shot success rate on 50 unseen tasks using pretraining data including 402,730 hours of first-person human video. While NVIDIA describes it as an open model with code scheduled for release in December, industrial applications remain in testing phases. The system supports parallel gripper arms, dexterous hands, and humanoid robots, leveraging techniques like success-guided sampling to improve training efficiency.
Source: Chinese robotics — Dataset and collection discovery · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
human first-person video hours
402,730 hours
View original evidence
预训练数据包括 402,730 小时的第一视角人类视频、10,184 小时涵盖 2,897 项任务的手持夹爪演示,以及 23,500 小时的机器人遥操作数据。
Open source S7hand-held gripper demonstration hours
10,184 hours
View original evidence
预训练数据包括 402,730 小时的第一视角人类视频、10,184 小时涵盖 2,897 项任务的手持夹爪演示,以及 23,500 小时的机器人遥操作数据。
Open source S7robot teleoperation hours
23,500 hours
View original evidence
预训练数据包括 402,730 小时的第一视角人类视频、10,184 小时涵盖 2,897 项任务的手持夹爪演示,以及 23,500 小时的机器人遥操作数据。
Open source S7Reported success rate
65%
Reported trials
View original evidence
NVIDIA 报告称,通过人类演示学会的滤纸布置任务,在没有针对该任务进行遥操作的情况下,成功率达到 65%;从手持夹爪迁移到 YAM 机械臂的玩具自卸卡车组装任务,成功率达到 60%。
Open source S10Reported success rate
60%
Reported trials
View original evidence
NVIDIA 报告称,通过人类演示学会的滤纸布置任务,在没有针对该任务进行遥操作的情况下,成功率达到 65%;从手持夹爪迁移到 YAM 机械臂的玩具自卸卡车组装任务,成功率达到 60%。
Open source S10Reported success rate
95%
Reported trials
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: Reported pending Open source S8
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Human-1M-100K 数据集 开放其中 100,000 小时的人类数据,包括视频、语言标注以及头部和手部运动数据,并采用 CC BY 4.0 许可,允许商业和非商业用途。
Open source S8
预训练数据包括 402,730 小时的第一视角人类视频、10,184 小时涵盖 2,897 项任务的手持夹爪演示,以及 23,500 小时的机器人遥操作数据。
Open source S7
NVIDIA 报告称,这款部分通过人类视频训练的模型,在目标机器人的训练数据中未出现的 50 项任务上,实现了 40% 的零样本成功率。
Open source S3
NVIDIA 报告称,通过人类演示学会的滤纸布置任务,在没有针对该任务进行遥操作的情况下,成功率达到 65%;从手持夹爪迁移到 YAM 机械臂的玩具自卸卡车组装任务,成功率达到 60%。
Open source S10
每项任务提供一次演示时,平均成功率达到 71%;提供五次演示时,平均成功率达到 95%。
Open source S11
Source:Chinese robotics — Dataset and collection discovery · unite.ai