Nvidia has introduced an open robotics model, a major simulation update and new benchmarks aimed at training robots for delicate manipulation rather than basic movement. The releases include the GR00T 2 world action model and Isaac Lab 3.0, announced around the Conference on Robot Learning.
GR00T 2 was trained on robot data and 100,000 hours of human first-person video. Nvidia says it completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. Average success reached 71% with one demonstration per task and 95% with five, although those figures come from Nvidia’s evaluation.
Isaac Lab 3.0 lets researchers choose between PhysX and the open Newton physics engine. Tighter Newton integration adds simulation of deformable objects such as cloth and cables, along with articulated robotic hands. Nvidia reports up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX on one dexterous-manipulation benchmark.
The release also includes Isaac Lab-Arena for testing policies across repeated variations, plus DexBench and VABAR for industrial tasks. The goal is to expose the conditions that cause failures before deployment. Several manufacturers and robotics companies are evaluating parts of the stack, but simulated and benchmark performance does not by itself establish reliability in production environments.