英伟达在德克萨斯州奥斯汀发布机器人模型 GR00T 2：50 项新任务零样本平均成功率 40%，每项 5 次示范升至 95%
德克萨斯州奥斯汀市消息，英伟达于 11 月 9 日在机器人学习会议（CoRL）上发布 GR00T 2。英伟达称其为开放的前沿世界动作模型，与之同场公布的还有机器人学习框架 Isaac Lab 3.0 的应用进展与两项新的开放评测基准。
据英伟达公布，GR00T 2 基于其 Cosmos 世界模型构建，模型规模为 20 亿参数，可在同一个扩散式变换器中同时预测未来视频与机器人动作。模型代码与权重定于 12 月开放，英伟达研究页面已列出相关说明。
英伟达报告的评测结果显示，在目标机器人训练数据中未出现过的 50 项任务上，GR00T 2 零样本平均成功率为 40%。每项任务给 1 次示范后，平均成功率升至 71%；每项给 5 次示范后，升至 95%。
训练数据方面，英伟达称预训练使用了大规模人类第一人称视频与机器人遥操作数据，其中包含 10 万小时的开源第一人称数据集。模型可通过接口与配套工具按顺序执行技能，并在出错后恢复。
同场公布的 Isaac Lab 3.0 可在 PhysX 与 Newton 两种物理引擎之间切换，并新增强化学习后训练库。英伟达称，在一项视觉灵巧操作基准上，Isaac Lab 3.0 配合 Newton 的训练吞吐量最高达到 Isaac Lab 2.3 配合 PhysX 的 2.5 倍。
英伟达同时发布 Isaac Lab-Arena 0.4 开发者预览版，以及 DexBench 与 VABAR 两项开放基准，分别用于衡量机器人的灵巧程度与在工业任务中的适应能力。DexBench 由 RLWRLD 与英伟达合作开发，含 4 个场景共 15 项任务；VABAR 由加州大学伯克利分校与英伟达推出，含 6 项工业任务。
试用进展方面，英伟达称 Nexuni 正在测试 GR00T 2 处理衣物，富士康正将其用于服务器制造自动化，西门子正在评估工业应用场景。MicroAGI 则以强化学习后训练后的 GR00T 2 配合机械臂完成以太网线缆更换演示。
来源链接：The Robot Report 对英伟达 11 月 9 日发布的报道： https://www.therobotreport.com/nvidia-launches-gr00t-2-isaac-lab-3-0-corl/ ；英伟达开发者论坛官方发布摘要（2026-11-09）： https://forums.developer.nvidia.com/t/news-nvidia-at-corl-2026-advancing-robot-learning/388596
NVIDIA Unveils Robot Model GR00T 2 in Austin, Texas: 40 Percent Average Zero-Shot Success on 50 New Tasks, Rising to 95 Percent With Five Demonstrations Per Task
Austin, Texas — NVIDIA introduced GR00T 2 on November 9 at the Conference on Robot Learning, or CoRL. NVIDIA described it as an open frontier world action model, announced alongside further Isaac Lab 3.0 work and two new open evaluation benchmarks.
According to NVIDIA, GR00T 2 is built on its Cosmos world models and uses a 2-billion-parameter transformer that predicts future video and robot actions together in one diffusion transformer. NVIDIA's research page lists code and weights as available in December.
NVIDIA reported that GR00T 2 achieved 40 percent average success with no demonstrations on 50 tasks absent from the target robot's training data. With one demonstration per task, average success rose to 71 percent, and with five demonstrations per task it rose to 95 percent.
For training data, NVIDIA said pretraining drew on large volumes of first-person human video and robot teleoperation data, including a 100,000-hour open-source egocentric dataset. The model can be driven through an application programming interface and supporting tools to run skills in sequence and recover from errors.
Isaac Lab 3.0, presented at the same event, lets researchers switch between the PhysX and Newton physics solvers and apply a new reinforcement-learning post-training library. NVIDIA said that in a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reached up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.
NVIDIA also released the developer preview of Isaac Lab-Arena 0.4 and two open benchmarks, DexBench and VABAR, designed to measure robot dexterity and adaptability on industrial tasks. DexBench, developed by RLWRLD with NVIDIA, covers 15 tasks across four scenes, while VABAR, from UC Berkeley and NVIDIA, covers six industrial tasks.
On early use, NVIDIA said Nexuni is testing GR00T 2 for laundry handling, Foxconn is applying it to server-manufacturing automation and Siemens is evaluating industrial use cases. MicroAGI used a reinforcement-learning post-trained GR00T 2 to swap Ethernet cables with a robotic arm.
Sources: The Robot Report on NVIDIA's November 9 announcements: https://www.therobotreport.com/nvidia-launches-gr00t-2-isaac-lab-3-0-corl/ ; NVIDIA Developer Forums announcement summary, November 9, 2026: https://forums.developer.nvidia.com/t/news-nvidia-at-corl-2026-advancing-robot-learning/388596
发布：GSM·美国（发布链接）
图片来源：Wikipedia/Wikimedia Commons 词条主图：Nvidia
GSM Global · 本篇英文版 English edition of this story
