Nvidia GR00T robotics foundation-model rollout

Nvidia is rolling out GR00T robotics foundation models for humanoids, dexterous hands and grippers through world-action prediction and simulation.

Nvidia introduced GR00T 2 as a world-action model that predicts physical movement outcomes to generate robot actions. The model is designed for humanoids, dexterous hands and parallel grippers, and tests showed a 97% or higher completion rate in toy assembly. Nvidia has released robotics training data, hardware designs and supporting tools, while model weights and X-Adapt recipes are due in December.

The rollout matters for robotics developers, hardware makers and AI infrastructure buyers testing whether one foundation model can generalize across physical platforms.

• GR00T 2 is described as a world-action model. • Training used a 2,000-GPU cluster with GB300 NVL72, H100 and L40 systems. • Half of the GPU fleet was dedicated to pretraining. • The system runs on onboard robot compute for long-horizon tasks. • Nvidia released 100,000 hours of robotics training data.