NVIDIA GR00T 2 reaches 40% zero-shot success on 50 unseen robot tasks using human video

NVIDIA introduced GR00T 2 on November 9, 2026, at the Conference on Robot Learning (CoRL) in Austin. The company reported 40% zero-shot success on 50 tasks absent from the target robot's training data, for a model trained partly on human video. The details come from a short summary published by Unite.AI.

Architecture

GR00T 2 is a 2-billion-parameter transformer built on Cosmos-Predict2.5-2B. The model takes three inputs: recent camera views, robot state and a written instruction. From these it generates expected future views and the next action sequence.

Key figures

• Parameters: 2 billion
• Base model: Cosmos-Predict2.5-2B
• Reported result: 40% zero-shot success on 50 tasks not present in the target robot's training data
• Training data: includes human video, according to NVIDIA's reported result
• Context window, pricing and benchmark comparisons against other models: not yet disclosed

What this means

The reported 40% figure measures zero-shot performance on tasks the target robot's training data did not contain. That makes human video a data source for robot policies, rather than relying only on data collected from the robot itself. The design pairs predicted future views with action output in a single 2-billion-parameter model built on Cosmos-Predict2.5-2B. The 40% result is NVIDIA's own report, and it covers 50 tasks.