NVIDIA introduces GR00T 2 world action model and open robot-training tools

NVIDIA introduced GR00T 2, which it calls an open frontier world action model for robots, at the Conference on Robot Learning (CoRL) held this week in Austin, Texas. It also announced open simulation tools, benchmarks and agentic workflows for training and testing robots. All performance figures below come from NVIDIA's own posts and have not been independently verified.

Is GR00T 2 new?

NVIDIA has previously previewed a model called GR00T N2. According to NVIDIA's announcement as reproduced by GamesBeat, CEO Jensen Huang "previewed GR00T N2" at the GTC keynote, describing a model based on DreamZero research with a new world action model architecture, due to be available by year's end. In that announcement NVIDIA claimed the model helps robots succeed more than twice as frequently as leading vision language action models when tasks and settings are new, and said it ranked first on MolmoSpaces and RoboArena among generalist robot policies; the page offered no supporting data. The CoRL post uses the name GR00T 2 and does not say whether it is the same model.

What does NVIDIA claim for GR00T 2?

NVIDIA says GR00T 2 is built on its Cosmos platform and learned from a mix of robot and human data, among it 100,000 hours of NVIDIA's open-source egocentric human data. According to the company, the model managed at least one successful run on 40 of 50 evaluated tasks with no task-specific demonstrations, and averaged 71% success across the 50 tasks with one demonstration per task, rising to 95% with five. NVIDIA also says AI agents can call the model through its programming interface and supporting tools to chain skills together and recover from mistakes.

It now says GR00T 2 places in the top three of all vision language action and world action models on four benchmarks, RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, with several leaderboard wins. The benchmark sets overlap only on MolmoSpaces, and the earlier claim (No. 1) differs from the new one (top three), so the two are not directly comparable. NVIDIA names early users: Nexuni is trying it for robotic laundry handling, Foxconn is using it in server-manufacturing automation, and Siemens is evaluating it across several industrial uses. MicroAGI uses a version post-trained with reinforcement learning to swap Ethernet cables with a YAM robotic arm, NVIDIA says.

What changed in simulation and evaluation?

NVIDIA says the latest release of its open-source robot-learning framework, Isaac Lab 3.0, lets researchers use agents to build training environments, toggle between the PhysX and Newton physics engines, pick renderers for perception-based training and run a new reinforcement post-training library on models such as GR00T. NVIDIA introduced Isaac Lab 3.0 in early access at its GTC keynote, according to the same announcement. Tighter integration with Newton, an open-source physics engine, lets it model deformable objects such as cables and cloth, and complex articulated systems such as robotic hands. In a vision-based dexterous manipulation benchmark, NVIDIA says Isaac Lab 3.0 with Newton reached a training throughput up to 2.5x that of Isaac Lab 2.3 on PhysX.

Partners are building on it, according to NVIDIA. Skild AI announced STEEL, a rigid-body physics solver built on Newton. A project called AlphaNIST, built in Isaac Lab, trains one policy to handle all 20 tasks of the NIST Assembly Task Board using only a yes-or-no success signal; NVIDIA reports training success above 98%, versus 81% for separate per-task policies, with roughly 39x less GPU time.

NVIDIA says its open-source Isaac Lab-Arena framework lets developers run policies across thousands of variations in parallel and analyze which conditions cause failures; the developer preview of version 0.4 arrives at the conference. Two new open benchmarks use it: DexBench, developed by RLWRLD with NVIDIA, which covers 15 tasks in four scenes: retail, logistics, manufacturing and hospitality, and VABAR, from UC Berkeley and NVIDIA, which evaluates robots across six industrial tasks.

What is agentic robotics?

NVIDIA highlights UC Berkeley's Graph-as-Policy (GaP) framework, which uses agents to connect individual robot skills into programs, tests them in NVIDIA's Isaac simulation technologies and refines them based on failures. NVIDIA says this keeps the programs interpretable so engineers can review them before deployment. Ken Goldberg, a UC Berkeley robotics professor who is a lead on the VABAR and GaP projects, is quoted in the post calling agentic robotics "the most exciting technological development since the internet in 1995." Ambi Robotics, Bosch Research and Siemens are exploring GaP, according to NVIDIA.