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 N1.6

NVIDIA, United States · December 2025

Reasoning VLAOpen weights, NVIDIA OneWay NCAlso written GR00T N1.6, GR00T-N1.6-3B, Isaac GR00T N1.6

Parameters: 3.3 B; 3,286,608,832 parameters in BF16 according to the safetensors metadata of nvidia/GR00T-N1.6-3B on Hugging Face.
GPU memory: 6.1 GB; weights at bf16, computed. Fits a 12 GB card
Inference latency: 44 ms; NVIDIA RTX 4090 with torch.compile, 4 denoising steps, single camera view: 44 ms end to end (22.8 Hz), split as 2 ms data...
Weights: NVIDIA OneWay NC; NVIDIA OneWay Noncommercial License for the weights.

What it is

GR00T N1.6 is the release where NVIDIA started publishing numbers you can check. It swaps the Eagle backbone for a Cosmos Reason 2B variant, doubles the diffusion transformer to 32 layers, drops the N1.5 adapter in favour of unfreezing the top four VLM layers, and predicts state-relative rather than absolute action chunks. Reproducible per-task results appear for SimplerEnv, RoboCasa and LIBERO, along with the first real latency table, at 44 ms on an RTX 4090. Read the per-task columns rather than the averages, because the spread is extreme, from 5 of 100 on cube stacking to 95 of 100 on drawer opening in the same suite. The weights remain noncommercial, which is the reason to move to N1.7 for anything shipping.

Architecture

Backbone: An internal NVIDIA Cosmos-Reason-2B VLM variant, vendored in the code as nvidia/Eagle-Block2A-2B-v2. It supports flexible resolution and encodes images in their native aspect ratio without padding, and it was trained on both general vision-language tasks and embodied reasoning tasks such as next action prediction.
Action head: flow matching diffusion transformer with adaptive layer normalisation, doubled to 32 layers from 16 in N1.5. The post-VLM 4-layer transformer adapter of N1.5 was removed and the top 4 layers of the VLM are unfrozen during pre-training instead. Predicts state-relative action chunks for most embodiments rather than absolute joint angles or end-effector positions.
Action chunk: 16 steps per forward pass
Parameters: 3,286,608,832 parameters in BF16 according to the safetensors metadata of nvidia/GR00T-N1.6-3B on Hugging Face. The repository name and the release blog both say 3B.
Pretraining data: The N1.5 mixture plus several thousand additional hours of teleoperated data from bimanual YAM arms, AgiBot Genie1, simulated Galaxea R1 Pro on the BEHAVIOR suite, and whole-body loco-manipulation with the Unitree G1. NVIDIA describes the base checkpoints as pre-trained on 10k+ hours of robot data. Pre-training ran 300K steps at a global batch size of 16,384. Post-training is typically 10K to 30K steps at batch size 1K or below.
Embodiments: Unitree G1, including whole-body loco-manipulation, AgiBot Genie-1, bimanual YAM arms, Galaxea R1 Pro on the BEHAVIOR-1K suite, Franka Panda (LIBERO_PANDA), WidowX (OXE_WIDOWX) and Google Robot (OXE_GOOGLE)

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 3.3 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	12 GB	Training master weights, rarely used for inference
bf16 or fp16	6.1 GB	The usual way these checkpoints are served
int8	3.1 GB	Quantised, expect some loss of precision on fine motions
int4	1.5 GB	Aggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 12 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	44 ms on NVIDIA RTX 4090 with torch.compile, 4 denoising steps, single camera view: 44 ms end to end (22.8 Hz), split as 2 ms data processing, 25 ms backbone, 17 ms action head. Same configuration on H100 gives 38 ms (26.3 Hz), on RTX 5090 37 ms (27.3 Hz) and on Jetson Thor 105 ms (9.5 Hz).

No VRAM table was published on the N1.6 branch. The 16 GB inference and 40 GB fine-tuning figures in the current repository are stated for N1.7 and must not be carried back to N1.6. N1.6 inference timings were measured on RTX 5090, H100, RTX 4090 and Jetson Thor.

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.

LIBERO Object suite, fine-tuned from nvidia/GR00T-N1.6-3B, 20K steps, batch size 640
98.45%
success rateFirst-party, 197 of 200 episodes. Companion suites on the same branch: Spatial 97.65 percent (195/200), Goal 97.5 percent (195/200), Long 94.35 percent (189/200).source
LIBERO Long (LIBERO-10) suite, fine-tuned from nvidia/GR00T-N1.6-3B, 20K steps
94.35%
success rateFirst-party, 189 of 200 episodes.source
RoboCasa Panda Omron Average over 24 kitchen tasks, nvidia/GR00T-N1.6-3B
66.22%
success rateFirst-party. Ranges from CloseDrawer at 100.0 percent down to PnPCounterToMicrowave at 19.0 percent.source
SimplerEnv Bridge (WidowX) Average over 7 tasks, about 100 trials each, GR00T-N1.6-bridge checkpoint
56.6%
success rateFirst-party. Very uneven per task: open_drawer 95 of 100 and put_eggplant_in_basket 89 of 100, but stack_cube only 5 of 100.source
SimplerEnv Fractal (Google Robot) Average over 6 tasks, 100 trials each, GR00T-N1.6-fractal checkpoint
52%
success rateFirst-party. Dragged down by google_robot_open_drawer at 0 of 100 and place_in_closed_drawer at 5 of 100, while pick_coke_can reached 95 of 100.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Demonstrated on bimanual YAM arms, the AgiBot Genie-1 and the Unitree G1, the last including whole-body loco-manipulation with a dedicated published checkpoint for pick-and-place. NVIDIA released fine-tuned checkpoints for DROID and for BEHAVIOR-1K loco-manipulation on the Galaxea R1 Pro. The release blog shows these as video demonstrations without success-rate figures, so the quantitative record for N1.6 is the simulation benchmarks.

Fine tuning it yourself

Fine-tune from the n1d6 branch using GR00T LeRobot format data. Pre-registered embodiment tags are LIBERO_PANDA, OXE_GOOGLE, OXE_WIDOWX, UNITREE_G1 and BEHAVIOR_R1_PRO, with NEW_EMBODIMENT for custom robots. Published recipes use 20K steps at global batch size 640 for LIBERO. Post-training is typically 10K to 30K steps at batch size 1K or below. Expect 5 to 6 percentage points of run-to-run variance.

Where it helps, where it does not

Strengths

First version in the family with published, reproducible public-benchmark numbers on SimplerEnv, RoboCasa and LIBERO, including per-task breakdowns and trial counts.
Real inference timings published across four hardware tiers, so latency planning is possible for the first time. 44 ms on a consumer RTX 4090 with torch.compile.
Cosmos Reason as the backbone brings embodied reasoning, including next-action prediction, into the VLM rather than only general vision-language ability.
Doubling the DiT to 32 layers and switching to state-relative action chunks improved robustness across embodiments.
Broadest embodiment coverage yet, with published checkpoints for Unitree G1 loco-manipulation, AgiBot Genie1, BEHAVIOR-1K on Galaxea R1 Pro, DROID, and both SimplerEnv robots.

Limits

Noncommercial license. Same NVIDIA OneWay Noncommercial terms as N1 and N1.5, which rules out product use. This is the single most important difference against N1.7.
Per-task variance is severe and the averages hide it. On SimplerEnv Bridge the model gets 95 of 100 on open_drawer and 5 of 100 on stack_cube. An average of 56.6 percent tells you almost nothing about a specific task.
Two SimplerEnv Fractal tasks score at or near zero, google_robot_open_drawer at 0 percent and place_in_closed_drawer at 5 percent, indicating whole task families the model cannot do.
No VRAM figures were published for N1.6. The 16 GB and 40 GB numbers in the current repository belong to N1.7.
NVIDIA warns that post-training results vary by as much as 5 to 6 percentage points between runs with identical configuration, seed and dropout, so small differences against these numbers are noise.
The gr1_unified embodiment tag from N1.6 was retired in N1.7, so N1.6 fine-tuning recipes for the GR-1 do not carry forward.

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