The newest generation of Nvidia’s world-action models demonstrated the ability to recover from mechanical errors to complete repetitive assembly trials during recent performance tests. Using its GR00T 2 system, the company’s robots successfully assembled a series of toys, maintaining a completion rate of 97% or higher across three separate tests by adjusting their planning in real time based on live camera feeds.
Nvidia will make the code for the foundation model available in December. The system is designed to be universal across hardware including parallel grippers and humanoids and was developed on a 2,000 GPU cluster of H100 and GB300 NVL72 chips using the SkyPilot compute platform for research experimentation.
