Overview
- At the Conference on Robot Learning in Austin on Monday, NVIDIA released GR00T 2, Isaac Lab 3.0 and Isaac Lab‑Arena together with open evaluation suites DexBench and VABAR.
- GR00T 2 is billed as a world‑action vision‑language‑action model trained on diverse robot and human data including 100,000 hours of egocentric video, and NVIDIA reports it completed 40 of 50 tasks at least once with average success rates of 71% with one demo and 95% with five demos.
- Isaac Lab 3.0 is a modular, multi‑GPU simulator that lets developers swap physics engines and renderers and use deployment tools such as LEAPP, and NVIDIA says its Newton physics engine achieved up to 2.5x the throughput of PhysX on a vision‑based dexterous task.
- NVIDIA named early testers and partners — including Foxconn, Siemens, Nexuni, Agile Robots and Unitree — and emphasized industrial use cases, while cautioning that the reported task and throughput numbers are internal lab benchmarks not yet validated on factory floors.
- The stack pairs a foundation model with large‑scale simulation and open benchmarks to reduce dependence on scarce real training data and to reveal failure modes, a step that could speed practical automation in manufacturing and logistics if independent real‑world tests confirm the lab results.