Nvidia’s GR00T 2 targets scalable robot skill learning

Nvidia shifts focus toward transferable robot skills

Nvidia’s GR00T 2 launch points toward reducing the task-specific demonstrations needed to train robots for manipulation. The company said the model completed 40 of 50 evaluated tasks at least once without task-specific demonstrations and averaged 71% success across those tasks with one demonstration per task, rising to 95% with five. These are Nvidia-reported benchmark results rather than independently established performance across deployed systems.

The model was introduced at the Conference on Robot Learning in Austin, Texas. Unlike earlier vision-language-action models, GR00T 2 is a world action model that predicts future video frames together with robot actions before executing an action chunk. It was built on Nvidia Cosmos and trained on robot and human data, including 100,000 hours of Nvidia open-source human egocentric video.

Simulation throughput addresses training bottlenecks

Alongside GR00T 2, Nvidia released Isaac Lab 3.0, an open-source robot-learning framework with tighter integration between the PhysX and Newton physics solvers. In Nvidia’s vision-based dexterous-manipulation benchmarks, Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.

The framework also supports simulated robots, perception-oriented renderers and reinforcement-learning post-training for models such as GR00T. Its practical significance is a potential reduction in the simulation work required to train and refine manipulation policies, although the throughput comparison is limited to the cited benchmarks.

On the newly introduced AlphaNIST benchmark, a single policy covering 20 tasks achieved 98% training success, compared with 81% for separate per-task policies, while using roughly 39 times less GPU time. Nvidia also introduced DexBench, covering 15 tasks across four industrial scenes, and VABAR, which evaluates six industrial tasks with standardized variations. These figures describe training performance on Nvidia’s benchmarks and do not establish commercial adoption.

Open tools extend beyond the model

Nvidia introduced an open Video to Data workflow intended to turn human video into simulated robot demonstrations and presented EgoHand, a model for reconstructing three-dimensional hand motion from monocular egocentric video. Noitom Robotics said its HiPHI-Hand dataset—11 hours of synchronized dual-fisheye video, headset-referenced camera extrinsics and optical-motion-capture joint positions—had been processed into Nvidia’s SOMA-Hand representation for use in EgoHand’s training corpus.

Noitom planned a live EgoHand demonstration at CoRL and has integrated the model into one of its egocentric-data pipelines for evaluation as a hand-annotation stage. The distinction matters: HiPHI-Hand was released openly, while EgoHand’s broader training corpus and the reported 14 million-image training set do not represent unrestricted access to every underlying asset.

Deployment remains at testing and evaluation stages

Nvidia identified Foxconn as applying GR00T 2 to factory automation for server manufacturing, Siemens as evaluating it across industrial use cases, Nexuni as testing robotic laundry handling and MicroAGI as using a post-trained model for Ethernet-cable swapping. These disclosures show active technical evaluation, not confirmed orders, production deployments or guaranteed procurement.

GR00T 2 and a lighter Turbo variant are planned for open-source release in December, according to Humanoids Daily’s report of Nvidia’s CoRL presentation. The reviewed reporting conflicts over what was already open: some tools and datasets were available at the conference, but the principal model weights and cross-embodiment training recipe were still scheduled for December. The next material checkpoint is therefore that planned release, followed by broader reproducibility and real-world performance testing.

Industry and product implications