Nvidia GR00T 2 Hits 71% One-Shot Success Across 50 Tasks

Nvidia officially launched GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas, introducing a 2-billion-parameter vision-language-action foundation model that averaged 71% success across 50 robotics tasks with a single demonstration per task. The model, built on Cosmos-Predict2.5 and trained on 100,000 hours of open-source human egocentric video, reached 95% success when given five demonstrations and completed 40 of 50 tasks without any task-specific training.

Release Date and Availability

Officially launched on November 10, 2026, GR00T 2 represents Nvidia's latest entry into robotics foundation models. The company announced at CoRL 2026 that model weights and the full 100,000-hour training dataset will become available in December 2026. No specific licensing terms were disclosed in the initial announcement, though Nvidia described the training data as open-source human egocentric footage.

Zero-Shot and Few-Shot Performance Benchmarks

GR00T 2 demonstrated strong generalization across industrial robotics tasks. The model completed 40 of 50 evaluated tasks at least once without any task-specific demonstrations, showcasing robust zero-shot capabilities. With one demonstration per task, success rates averaged 71% across all 50 tasks. Performance jumped to 95% when the model received five demonstrations per task, indicating efficient few-shot learning.

Nvidia positioned GR00T 2 among the top three models across both vision-language-action and world action model categories. The company tested the model against two new industrial benchmarks: DexBench, featuring 15 tasks across four industrial scenes, and VABAR, which evaluates six industrial tasks with standardized variations.

Isaac Lab 3.0 Training Infrastructure

Alongside GR00T 2, Nvidia released Isaac Lab 3.0 with the Newton physics solver. In vision-based dexterous manipulation benchmarks, Isaac Lab 3.0 delivered up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. The AlphaNIST benchmark revealed additional efficiency gains: a single 20-task policy reached 98% training success compared to 81% for separate per-task policies, while consuming approximately 39 times less GPU time.

The training infrastructure improvements address a core challenge in robotics: the computational cost of learning diverse manipulation skills. By enabling multi-task policies to outperform specialized models while using dramatically less compute, Isaac Lab 3.0 lowers the barrier for industrial deployment.

Industry Collaborations and Expert Reaction

Nvidia highlighted partnerships with Ambi Robotics, Bosch Research, Flexion, FieldAI, and Siemens in the GR00T 2 announcement. Ken Goldberg, a UC Berkeley robotics professor leading the VABAR benchmark work, stated: "To me, agentic robotics is the most exciting technological development since the internet in 1995."

The release arrives during an active period in robot learning: 191 robotics stories have been published in the last 90 days, including Nature Will's 50-billion-parameter model for robot hand manipulation announced one day earlier. The clustering of foundation model releases signals accelerating investment in generalist robot learning systems.

What This Means

GR00T 2's combination of strong zero-shot performance and efficient few-shot adaptation positions it as a practical tool for industrial robotics deployments where task-specific training is expensive or impractical. The December 2026 release of weights and training data will enable independent benchmarking and potential fine-tuning for specialized applications. Isaac Lab 3.0's training efficiency gains could accelerate the development cycle for custom robotics policies, particularly in manufacturing and logistics environments where rapid adaptation to new tasks drives competitive advantage.