Frontier

NVIDIA GR00T 2 Hits 95% Success With Five Demos

NVIDIA unveiled GR00T 2, its world action model pretrained on hundreds of thousands of hours of human video and robot data. The model reaches 95% average success on 50 tasks after five demonstrations per task and a 30-minute fine-tune on one RTX 5090, with code due in December 2026.

NVIDIA GR00T 2 Hits 95% Success With Five Demos
Photo: NVIDIA (company image)

ZeroGantry analysis

GR00T 2's 30-minute RTX 5090 adaptation and 95% few-shot result on a 22-DoF bimanual platform cut per-task data collection by roughly an order of magnitude versus prior GR00T releases. Operators running mixed fleets of Unitree G1 and custom dexterous arms can now share one pretrained checkpoint across bodies with minimal new teleop. Ship for research teams validating on controlled tasks; watch for production until December weights and independent long-horizon benchmarks appear.

NVIDIA GR00T 2 Launches as Open World Action Model

NVIDIA introduced GR00T 2 on November 9, 2026, at the Conference on Robot Learning in Austin, Texas. On [NVIDIA's GR00T 2 research page](https://research.nvidia.com/labs/gear/gr00t-2/), the company reports that the model reaches 95% success after a 30-minute fine-tune with five demonstrations per task, up from 40% zero-shot, on 50 tasks never seen on the target robot. The model predicts future video frames alongside robot actions in a single pass, distinguishing it from prior vision-language-action approaches. Built on the Cosmos-Predict2.5-2B backbone, GR00T 2 runs across parallel grippers, dexterous hands, and humanoids without embodiment-specific retraining from scratch. Partners including Foxconn for server manufacturing, Siemens, and Nexuni for laundry tasks are already evaluating it in real settings.

Pretraining drew from 402,730 hours of egocentric human video, 10,184 hours of handheld UMI gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation on platforms such as YAM, Franka, Sharpa, Wuji hands, and the Unitree G1 humanoid. Of the human video, 100,000 hours are openly released as Human-1M-100K on Hugging Face under CC BY 4.0. This mixture lets the model absorb skills demonstrated only by people or on other machines and transfer them zero-shot to a new target robot.

Performance Metrics on Unseen Tasks

NVIDIA evaluated GR00T 2 on 50 manipulation tasks never performed by the target bimanual robot with two 22-DoF Sharpa hands. Zero-shot success averaged 40 percent across trials, with at least one success on 40 of the 50 tasks. One demonstration per task raised average success to 71 percent. Five demonstrations, totaling under three minutes of robot data, lifted performance to 95 percent average success after a 30-minute LoRA fine-tune on a single RTX 5090, with 37 tasks succeeding in every trial. Most zero-shot failures stemmed from execution precision rather than incorrect task decomposition.

The same checkpoint delivered strong results on contact-rich dexterous work previously out of reach for earlier models. Seating GB300 memory modules reached 85 percent success, while assembling a toy truck hit 65 percent. Reinforcement learning on the real robot further boosted three insertion tasks to 88–96 percent. These outcomes rely on the model's ability to forecast scene changes, which helps with deformable objects like cables or cloth where small early errors compound.

Adaptation Speed and Hardware Requirements

Fine-tuning occurs via 30-minute LoRA updates on one RTX 5090, lowering the barrier for labs without large GPU clusters. The approach contrasts with earlier GR00T versions that required more robot-specific data collection. NVIDIA also released Isaac Lab 3.0 in late October 2026, which delivers up to 2.5 times the training throughput of version 2.3 on vision-based dexterity benchmarks using PhysX. New benchmarks such as DexBench and VABAR accompany the framework to standardize evaluation.

Code for GR00T 2 is scheduled for December 2026, while the dataset and Isaac Lab 3.0 are already available. Model weights and a full license remain pending, limiting immediate commercial redistribution. Researchers note that the 50-task set was drawn from tasks already capable of reaching at least 80 percent success after five demonstrations, so broader generalization claims require external validation on production SKUs.

Implications for Humanoid Training Pipelines

GR00T 2 directly supports faster iteration for platforms such as Tesla Optimus and Figure 02 by transferring skills learned from human video or other robots. [NVIDIA Isaac GR00T N1.7 VLA Model Ships with Human Video Pretraining](/articles/nvidia-isaac-gr00t-n1-7-vla-model-ships-with-human-video-pretraining) showed the prior generation's reliance on more robot data; the new world-action formulation reduces that requirement. [NVIDIA GR00T N2 Preview at GTC 2026 Advances Physical AI](/articles/nvidia-gr00t-n2-preview-at-gtc-2026-advances-physical-ai) outlined the architectural progression toward joint video and action prediction.

Fleet operators gain a practical path to cross-embodiment reuse. A policy trained on Unitree G1 teleoperation data can adapt to a Sharpa-handed bimanual setup with minutes of new demonstrations rather than hours. This matters for cost-sensitive deployments where collecting fresh teleop on every variant is prohibitive. [NVIDIA Cosmos Framework Boosts GR00T World Model Training](/articles/nvidia-cosmos-framework-boosts-gr00t-world-model-training) details how the underlying video backbone accelerates this transfer.

Remaining Questions on Reproducibility and Scope

Independent benchmarks such as RoboDojo report real-world policy success below 13 percent even for top entries, highlighting the gap between curated 50-task suites and open-ended factory variability. NVIDIA's own results include confidence intervals that widen with fewer rollouts, and the 95 percent figure reflects averages across selected tasks rather than end-to-end uptime. External teams have posted lower numbers on harder unseen verbs and longer horizons.

The December code release will clarify whether weights carry an open license or usage restrictions that affect commercial humanoid programs. [NVIDIA SONIC Scales Humanoid Control Beyond Task-Specific Limits](/articles/nvidia-sonic-scales-humanoid-control-beyond-task-specific-limits) explored scaling laws for control; GR00T 2's few-shot results test whether world prediction extends those laws to dexterous, multi-embodiment settings. Production teams should request robot-specific pick-rate and cycle-time data on their own parts before committing.

Path Forward for Physical AI Research

GR00T 2 demonstrates that heavy pretraining on human video can compress the robot data needed for new tasks to under three minutes per skill. Combined with Isaac Lab 3.0 throughput gains, the stack lowers the compute and collection

Sources

Topics

Related articles

AI-derived: drafted with AI assistance and checked against cited sources. How we label articles