What happened
NVIDIA’s November 9, 2026 announcement introduces GR00T 2, which it describes as an open frontier world action model built on NVIDIA Cosmos. The company says it was trained on diverse robot and human data, including 100,000 hours of NVIDIA open-source human egocentric data. GR00T 2 is intended to help robots learn dexterous manipulation, where tasks include aligning gears, connecting cables and assembling small components. NVIDIA reports that the model completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. With one demonstration per task, the reported average success rate was 71%; with five, it was 95%. The announcement also says AI agents can call GR00T 2’s API and supporting tools to sequence skills and recover from errors.
The announcement says GR00T 2 ranks among the top three models across vision-language-action and world-action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins. NVIDIA names several uses under evaluation or development: Nexuni is testing laundry handling, Foxconn is applying the model to factory automation for server manufacturing, and Siemens is evaluating it for industrial use cases. MicroAGI is using a version post-trained through reinforcement learning for Ethernet cable swapping with a YAM robotic arm. These are examples described by NVIDIA, not independently assessed deployment results.
Isaac Lab 3.0 is the latest version of NVIDIA’s open-source robot-learning framework. NVIDIA says researchers can use agents to set up training environments with simulation-ready assets, switch between its PhysX solver and the open-source Newton physics engine, select renderers for perception-based training and apply a reinforcement post-training library to foundation models such as GR00T. The company says closer Isaac Lab integration with Newton supports deformable objects such as cables and cloth, as well as articulated systems such as robotic hands. For a vision-based dexterous manipulation benchmark, NVIDIA reports up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. Its announcement does not give a price for the open tools.
NVIDIA also introduced a developer preview of version 0.4 of Isaac Lab-Arena, an open-source framework for composing repeatable policy evaluations and running them across thousands of variations in parallel. Two benchmarks are highlighted: DexBench, developed by RLWRLD with NVIDIA, covers 15 tasks across retail, logistics, manufacturing and hospitality scenes; VABAR, from UC Berkeley and NVIDIA, evaluates six industrial tasks with standardized variations, difficulty tiers and success metrics. DexBench early access is being evaluated by Genesis AI, Mimic Robotics and other partners. Teams can run it on Bifrost’s Manifold platform; NVIDIA says it includes failure analysis and free credits for initial evaluations.
What the archive adds
New Runtime’s earlier coverage described FLUX 3 Action as an attempt to make a world-action model a fast robot-control layer (October 1, 2026 note). The current announcement is a new event, not a restatement: NVIDIA is pairing a world-action model with simulation, post-training and evaluation tools. An earlier note also described Cosmos3-Nano as an omnimodal world model for Physical AI (November 9, 2026 note); that is related context, not evidence about GR00T 2’s performance.
Trend signal
This does not yet fit a pattern New Runtime tracks. NVIDIA’s announcement connects models, simulated training and task benchmarks, but its reported outcomes do not establish a broader shift in how production agent systems are built or evaluated.
Why now
NVIDIA frames the CoRL announcement around a shift in robotics research from locomotion toward manipulation requiring finer control, and around agentic robotics as a way to compose skills into programs. The event provides the company’s timing; it does not establish how quickly industrial adoption will follow.
Connections to other events
The GR00T 2 release meets the earlier world-action-model work covered in New Runtime’s October 1 note on FLUX 3 Action. Both concern using models to connect observations and actions for robot control, while NVIDIA’s announcement emphasizes dexterous skills, demonstrations and agent-accessible sequencing. New Runtime’s November 9 coverage of Cosmos3-Nano is adjacent because it concerns a world model for Physical AI, but the announcement does not state that Cosmos3-Nano is part of GR00T 2.
Why it matters
For teams building embodied systems, the practical contribution is not only another model checkpoint. NVIDIA is presenting a pipeline in which simulation environments, physics engines, policy post-training and evaluation benchmarks can be combined. That gives researchers a way to test whether a learned skill survives changes in objects, scenes and physical conditions before trying it on hardware. The reported Isaac Lab throughput comparison could matter where simulation time limits iteration, while DexBench and VABAR offer task structures for measuring more than a polished single demonstration.
The trade-off is that teams should treat the headline metrics as hypotheses to reproduce. Throughput depends on the task, solver, hardware and configuration; manipulation success depends on how tasks and failures are defined. The announcement does not provide enough detail to establish comparable costs or robustness in a team’s own environment. Builders should identify which parts are open and runnable, reproduce a small representative benchmark, and separately validate safety and transfer on their target robot before relying on the numbers for deployment decisions.
Limits and uncertainty
The performance figures in the announcement come from NVIDIA, and the company does not provide independent verification in the announcement. The reported task success rates do not, by themselves, specify the full evaluation protocol, task-level breakdowns or how repeated trials were counted. The Isaac Lab throughput result is qualified as “up to” 2.5 times and applies to a stated benchmark comparison; it should not be generalized to every workload. The announcement also does not establish the reliability, cost or safety of the cited industrial uses in production. Availability terms and practical requirements for each model, benchmark and tool may differ; teams should check the relevant project materials before planning an evaluation.
Skill to build now
Choose one manipulation task your team can measure, then make a minimal simulation-to-real evaluation plan: define the success signal, vary object position or physical properties, and record failure categories before changing the policy. If you evaluate GR00T 2 or Isaac Lab, preserve the task setup and solver configuration so your results can be compared with later runs rather than with a headline number alone.
Reader takeaway
NVIDIA is packaging robot control, simulation and evaluation as a more connected open workflow. The numbers make the release worth testing, especially for dexterous manipulation, but they are company-reported results. The useful next step is to reproduce a task that matters to your robot and measure where simulation performance does—and does not—transfer.
Open question
How well do GR00T 2’s reported manipulation results transfer to independently selected tasks and physical robots?