NVIDIA used the Conference on Robot Learning in Austin, Texas to announce a set of open robotics releases, described on the NVIDIA Blog. The common thread is training robots in simulation for delicate manipulation, such as aligning gears, connecting cables and assembling small parts, rather than only for walking or moving around.
GR00T 2
The headline release is GR00T 2, an open world action model built on NVIDIA Cosmos. It was trained on robot and human data, including 100,000 hours of NVIDIA's open egocentric human footage. NVIDIA reports these results:
- It completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations.
- With one demonstration per task it averaged 71% success across the 50 tasks, rising to 95% with five.
- AI agents can call its API and supporting tools to chain skills and recover from errors.
NVIDIA says it ranks in the top three on four public benchmarks: RoboCasa365, RoboLab, RoboDojo and MolmoSpaces. Early users named in the post include Nexuni, testing laundry handling, and Foxconn, applying it to server-manufacturing automation. Siemens is evaluating it, and MicroAGI uses a reinforcement-trained version for Ethernet cable swapping on a robotic arm.
Isaac Lab 3.0 and Newton
Isaac Lab 3.0 lets agents help set up training environments and adds a reinforcement post-training library for foundation models such as GR00T. Closer integration with the Newton physics engine lets it model deformable items like cables and cloth, and robotic hands with many joints. On one vision-based dexterous manipulation benchmark, NVIDIA reports up to 2.5x the training throughput of Isaac Lab 2.3 running PhysX.
A separate result, AlphaNIST, trains a single policy across all 20 tasks of the NIST Assembly Task Board. NVIDIA reports over 98% training success, against 81% for separate per-task policies, using roughly 39x less GPU time.
Evaluation and agents
The developer preview of Isaac Lab-Arena 0.4 arrives at the conference. Two new open benchmarks use it: DexBench from RLWRLD, with 15 tasks in four scenes covering retail, logistics, manufacturing and hospitality, and VABAR from UC Berkeley, with six industrial tasks. UC Berkeley's Graph-as-Policy framework uses agents to link individual skills into robot programs that engineers can read and check before deployment. Ambi Robotics, Bosch Research and Siemens are exploring it.
What is still unproven
All figures are NVIDIA's own or its partners' and come from simulation benchmarks and evaluations. The source names pilots and tests, not production deployments, and gives no pricing or licensing terms beyond calling the releases open.