NVIDIA Says GR00T 2 Reaches 95% on Unseen Tasks After 30-Minute Fine-Tune
Summary
NVIDIA introduces GR00T 2, a world-action model that predicts future video and robot actions together, pretrained on 402,730 hours of human egocentric video plus 23,500 hours of robot teleoperation. NVIDIA reports 40% zero-shot success on 50 unseen tasks, rising to 95% after a 30-minute fine-tune with five demonstrations. Code is due in December.
Key Points
- NVIDIA reports a two-handed robot lines a lab funnel with filter paper in 65% of trials after learning only from human video.
- GR00T 2 Turbo predicts motion in 124 ms on the Unitree G1's onboard Jetson Thor, NVIDIA says.
- NVIDIA releases 100,000 hours of its Human-1M dataset as Human-1M-100K on Hugging Face.