At the 2026 Conference on Robot Learning in Austin, held from November 9 to 12, Nvidia presented GR00T 2, its new robot foundation model, along with Isaac Lab 3.0 and an open benchmark for evaluating robot policies in simulation. GR00T 2 follows GR00T N1.6 and N1.7, which were based on a 3-billion-parameter vision-language-action model running on the Cosmos backend. Isaac Lab 3.0, released in late October, uses a modular architecture that supports the Newton and PhysX physics engines and enables scalable training across multiple GPUs. Nvidia is combining these tools with reinforcement learning and imitation learning to reduce reliance on real-world training data, while also providing deployment support through products such as Jetson Thor. Partners including Agile Robots and Unitree are positioned to build on these releases as Nvidia describes the sector as approaching a turning point for physical AI.
Source: Read the original article
