How is pricing calculated?

Pricing is based on the hourly rate of operators in your selected region. You only pay for the time operators spend controlling your robots.

What robots are supported?

We support 9+ robot arms including SO-100, Koch, Franka FR3, Franka Panda, WidowX 250, ViperX 300, ViperX 300s, and ALOHA.

What is the latency for teleoperation?

Our platform provides ultra-low latency of 30-70ms worldwide, enabling real-time responsive robot control.

NVIDIA Isaac GR00T N1

NVIDIA, United States · March 2025

Foundation VLAOpen weights, NVIDIA OneWay NCAlso written GR00T N1, GR00T-N1-2B, Isaac GR00T N1

Parameters: 2.2 B; 2,190,019,826 parameters in BF16 according to the safetensors metadata of nvidia/GR00T-N1-2B on Hugging Face.
GPU memory: 4.1 GB; weights at bf16, computed. Fits an 8 GB card
Inference latency: 63.9 ms; NVIDIA L40 GPU in bf16, for sampling one chunk of 16 actions.
Weights: NVIDIA OneWay NC; NVIDIA OneWay Noncommercial License for the weights.

What it is

GR00T N1 is NVIDIA's opening move in open humanoid foundation models and the paper that the whole N1.x line still cites as its architecture reference. It splits the policy into an Eagle-2 VLM running at 10 Hz for interpretation and a flow-matching diffusion transformer that emits 16-action chunks in 63.9 ms on an L40. The distinctive contribution is the data pyramid, which converts action-less human video into a trainable pseudo-embodiment through latent actions and an inverse dynamics model. Absolute scores are modest, 32.1 percent on RoboCasa with 100 demos per task, and the weights are noncommercial only. Anyone building a product should look at N1.7 instead, since it is the first version in the family whose license permits commercial use.

Architecture

Backbone: NVIDIA Eagle-2 VLM as the System 2 reasoning module, running at 10 Hz on an L40 GPU
Action head: diffusion transformer (DiT variant) trained with action flow matching as the System 1 module, with denoising-step conditioning via adaptive layer normalisation, cross-attending to the VLM output tokens, plus embodiment-specific MLP encoders and decoders for variable state and action dimensions
Action chunk: 16 steps per forward pass
Parameters: 2,190,019,826 parameters in BF16 according to the safetensors metadata of nvidia/GR00T-N1-2B on Hugging Face. The model card states 2B.
Pretraining data: Organised as a data pyramid. Real robot layer: 88 hours of in-house teleoperation at 20 Hz with head-mounted camera plus fine-grained and coarse-grained action annotations. Middle layer: the 88 hours were augmented roughly 10x to 827 hours of neural-generated video via fine-tuned image-to-video models, plus simulation trajectories from DexMimicGen and RoboCasa. Base layer: human egocentric video datasets without action labels, converted into pseudo-embodiments through a learned latent-action codebook and a trained inverse dynamics model. The public sim dataset nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim contains 1.34M examples.
Embodiments: Fourier GR-1 humanoid with Fourier dexterous hands, Franka Emika Panda (RoboCasa), bimanual arm setups from DexMimicGen, simulated humanoid and tabletop configurations

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 2.2 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
Precision	Weights	Note
fp32	8.2 GB	Training master weights, rarely used for inference
bf16 or fp16	4.1 GB	The usual way these checkpoints are served
int8	2.0 GB	Quantised, expect some loss of precision on fine motions
int4	1.0 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference latency, reported	63.9 ms on NVIDIA L40 GPU in bf16, for sampling one chunk of 16 actions. The paper separately states the System 2 VLM runs at 10 Hz on an L40 and the System 1 DiT generates motor actions at 120 Hz.

No VRAM table was published for N1. The paper notes that compute-constrained fine-tuning was tested on a single A6000 GPU, and the model card lists Ampere, Hopper, Lovelace, Blackwell and Jetson as supported microarchitectures.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

DexMimicGen (DexMG) 9 tasks, post-trained with 100 demonstrations per task
66.5%
success rateFirst-party. Baselines: BC Transformer 53.9 percent, Diffusion Policy 56.1 percent.source
VLABench Overall score reported on the model card
39.7 score
overall scoreFirst-party, taken from the Hugging Face model card. The card does not describe the evaluation protocol.source
RoboCasa Kitchen 24 atomic tasks, Franka Emika Panda, post-trained with 100 demonstrations per task, average success rate over 100 trials
32.1%
success rateFirst-party. Baselines on the same split: BC Transformer 26.3 percent, Diffusion Policy 25.6 percent.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

Real-world GR-1 humanoid evaluation Average over pick-and-place, articulated, industrial and coordination tasks, full teleoperation dataset
76.8%
success rateFirst-party. Diffusion Policy on the same full dataset reached 46.4 percent. With only 10 percent of the data GR00T N1 still reached 42.6 percent, within 3.8 points of Diffusion Policy trained on everything.source
Simulated GR-1 tabletop (Digital Cousin suite) 24 tasks, GR-1 humanoid with Fourier dexterous hands, 100 demonstrations per task
50%
success rateFirst-party. Baselines: BC Transformer 16.1 percent, Diffusion Policy 32.7 percent. This is the split with the largest margin over baselines.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Deployed on the Fourier GR-1 humanoid with Fourier dexterous hands for bimanual manipulation, evaluated across pick-and-place, articulated-object, industrial and coordination task groups. In a coordinated multi-robot grocery packing setting the pre-trained GR00T-N1-2B reached 76.6 percent (11.5 of 15) in the first setting and 73.3 percent (11 of 15) in a second setting involving novel objects, using five different objects with three trials each. Teleoperation for data collection ran at 20 Hz.

Fine tuning it yourself

Fine-tune via the Isaac-GR00T repository using GR00T LeRobot format data with a modality.json describing video, state, action and language keys. Post-training experiments in the paper used 30, 100 and 300 demonstrations per task, which is the practical range to plan for. Co-training with neural trajectories added between 4.2 and 8.8 points on RoboCasa and 5.8 points on the real GR-1 tasks. Compute-constrained fine-tuning was validated on a single A6000.

Where it helps, where it does not

Strengths

First openly downloadable humanoid-focused foundation model from a major vendor, with training and fine-tuning code in the same repository.
Strong data efficiency. Trained on 10 percent of the real GR-1 teleoperation data it lands within 3.8 points of a Diffusion Policy trained on the full set.
The dual-system split lets a slow VLM at 10 Hz drive a fast DiT action module, which is what makes 16-action chunks feasible at 63.9 ms on an L40.
The data pyramid recipe, using latent actions and an inverse dynamics model to turn action-less human video into a pseudo-embodiment, is the reusable idea in the paper.
Beats BC Transformer and Diffusion Policy on all three simulation benchmarks and in the real world.

Limits

Noncommercial license. The NVIDIA OneWay Noncommercial License blocks any product deployment, which alone disqualifies N1 for most commercial projects.
Absolute simulation numbers are low. 32.1 percent on RoboCasa with 100 demonstrations per task is a research result, not a deployable policy.
Only 88 hours of genuine in-house teleoperation underpin the real-robot layer. The 827 hour figure is generated video, not collected data.
No VRAM figures were published for N1, so hardware planning has to rely on the A6000 mention in the paper.
The 120 Hz action generation rate quoted in the paper is not the rate at which the policy replans. NVIDIA's own later documentation warns explicitly against conflating inference rate, action-execution rate and camera rate.
Three revisions behind. N1.5, N1.6 and N1.7 all supersede it, and only N1.7 is commercially licensable.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the client Drive a real arm now