Back to news
launchNVIDIA2026-11-09

NVIDIA unveiled the GR00T 2 world action model at CoRL on November 9, with code and weights to open in December

GR00T 2 is a two billion parameter model built on Cosmos-Predict2.5-2B, completing 40 of 50 tasks outside its training data zero-shot, with average success of 71 percent from one demonstration and 95 percent from five

On November 9, at the Conference on Robot Learning in Austin, Texas, NVIDIA introduced GR00T 2, an open frontier world action model built on NVIDIA Cosmos. GR00T 2 uses a two billion parameter transformer that combines recent camera views, robot state and a written instruction to generate expected future views and the next action sequence, and a single pretrained checkpoint covers parallel gripper arms, dexterous hands and humanoids.

NVIDIA reported that the model completed 40 of 50 evaluated tasks zero-shot, giving a 40 percent success rate, on tasks absent from the target robot training data. On the Dexter evaluation, average success reached 71 percent with one demonstration per task and 95 percent with five, where five demonstrations supply under three minutes of robot data per task followed by an approximately 30 minute adapter fine-tune on a single RTX 5090. The technical whitepaper qualifies the figure, noting that evaluated tasks were selected among those the model can learn with five episodes at a success rate of at least 80 percent.

On training data, pretraining combines 402730 hours of first person human video, 10184 hours of handheld gripper demonstrations across 2897 tasks, and 23500 hours of robot teleoperation. NVIDIA also released the Human-1M-100K dataset under a CC BY 4.0 license, opening 100000 hours of human video with language annotations and head and hand motion data for commercial and noncommercial use. Although the model is introduced as open, its research page states that code will be available in December.

GR00T 2NVIDIACoRL世界动作模型机器人开源