At the Conference on Robot Learning in Austin, Nvidia introduced GR00T 2, an open frontier world action model for robots built on its Cosmos platform and trained on diverse robot and human data, including 100,000 hours of open-source egocentric human video. The headline claim: GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations at all, achieved 71 per cent average success across 50 tasks from a single demonstration each, and rose to 95 per cent with five demonstrations per task.
The Robot Report’s coverage of the November 9 announcement confirms the model already ranks in the top three across all vision-language-action and world action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins. What distinguishes GR00T 2 from its predecessor is the shift from locomotion to dexterous manipulation — the model transfers learned skills between different robot bodies without requiring robot demonstrations, and AI agents can call its API to chain skills in sequence and recover from errors.
The adoption list is what moves this from research demo to industrial signal. Nexuni is testing GR00T 2 for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, Siemens is evaluating it across industrial use cases, and MicroAGI post-trained the model with reinforcement learning to perform Ethernet cable swapping on a robotic arm.
The simulator underneath matters as much as the model
The second half of Nvidia’s announcement is Isaac Lab 3.0, the open-source robot learning framework, now integrated with the Newton physics engine so that simulated objects deform realistically — cables bend, cloth drapes, a strawberry squishes under a robot’s grip. That physics fidelity is what makes skills trained in simulation transfer to real hardware, and Nvidia says the combination delivers up to 2.5 times the training throughput of the previous Isaac Lab generation. Two new open benchmarks, DexBench and VABAR, measure dexterity and adaptability across retail, logistics, manufacturing and hospitality scenes rather than tabletop toy problems.
Why this round matters
The competitive context is that Google’s DeepMind reported a 96 per cent success rate for Gemini Robotics 2.5 in a Mercedes-Benz factory pilot only days earlier — the two announcements now bracket the physical-AI race as a two-platform contest with open and closed poles. Nvidia is running the open-platform play: model, simulator, benchmarks and data pipeline all public, which is exactly the strategy we analysed when Google Intrinsic and Nvidia Isaac went head-to-head over robotics platforms.
The data flywheel is also the point. Nvidia’s new open Video-to-Data workflow turns ordinary human video into simulated robot demonstrations — training data without robots — and its EgoHand model for reconstructing 3D hand motion cleared 40,000 Hugging Face downloads in two weeks. That connects to the 100,000-hour egocentric dataset behind LG and Nvidia’s Seoul-based robot training data project: the labs that industrialise video-to-skill pipelines will not need robot fleets the size of their rivals’, and GR00T 2’s one-demonstration learning rate is the strongest evidence yet that the approach generalises.
For a NZ angle: manufacturers watching the Foxconn server-automation trial should note that GR00T 2 is open-weight and runs on commodity GPUs, so the barrier to piloting is integration effort rather than licensing fees — the same dynamic that made cloud AI adoption a procurement decision rather than a research project.
