Isaac GR00T

weights

open(every N1.7 checkpoint is ungated safetensors on the Hub

data

partial(the N1.7 task fine-tunes train on public benchmark sets (Bridge, Fractal, DROID, LIBERO)

code

open(Apache-2.0 training and fine-tuning code in the repository)

license

NVIDIA-Open-Model-License(every N1.7 checkpoint: the base, GR00T-H-N1.7 and the task fine-tunes)

superseded

noncommercial(GR00T-H (the N1.6 surgical release) and the N1.5 Assemble_Trocar fine-tune are under the NVIDIA OneWay Noncommercial License

Every checkpoint in the current N1.7 release, the humanoid GR00T-H-N1.7 included, can be used commercially under NVIDIA's Open Model License, and the training code is Apache-2.0. The pretraining mixture is not released. The older N1.6 GR00T-H and an N1.5 surgical fine-tune are still on the Hub under a non-commercial license; they belong to earlier releases and do not set the score.

• https://developer.download.nvidia.com/licenses/NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf recorded 2026-09-26

  NVIDIA OneWay Noncommercial License, 3.3 Use Limitation: "The Work and any derivative works thereof only may be used or intended for use non-commercially", meaning for research or evaluation.

• https://huggingface.co/api/models?author=nvidia&search=GR00T-N1&sort=createdAt&direction=-1&limit=50 recorded 2026-09-27

  Hub listing newest first: GR00T-N1.7-3B created 2026-02-25, the N1.7 LIBERO, DROID and SimplerEnv fine-tunes 2026-03-17, GR00T-N1.7-ApplePnP-V1 2026-05-20, GR00T-H-N1.7 2026-05-30; GR00T-N1.5-3B_Assemble_Trocar 2026-05-27 carries an N1.5 name; nothing newer than N1.7.

• https://huggingface.co/api/models/nvidia/Cosmos-Reason2-2B recorded 2026-09-27

  Hub API for nvidia/Cosmos-Reason2-2B, the GR00T N1.7 backbone: gated "auto"; license_name nvidia-open-model-license.

• https://huggingface.co/api/models/nvidia/GR00T-N1.7-3B recorded 2026-09-27

  Hub API for nvidia/GR00T-N1.7-3B: gated false; safetensors weight files in siblings.

• https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim/raw/main/README.md recorded 2026-09-26

  PhysicalAI-Robotics-GR00T-X-Embodiment-Sim card: "We provide a set of datasets used for post-training of GR00T N1", CC-BY-4.0.

• https://huggingface.co/nvidia/GR00T-H-N1.7/raw/main/README.md recorded 2026-09-27

  GR00T-H-N1.7 card: license_name nvidia-open-model-license, base_model nvidia/GR00T-N1.7-3B; "It builds on the GR00T N1.7 VLA foundation"; "This model is ready for commercial use."

• https://huggingface.co/nvidia/GR00T-H/raw/main/README.md recorded 2026-09-27

  GR00T-H card, history: "This is the N1.6 non-commercially usable version of GR00T-H. Please see GR00T-H-N1.7 for the newer and commercially usable version"; license_link NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf.

• https://huggingface.co/nvidia/GR00T-N1.5-3B_Assemble_Trocar/raw/main/README.md recorded 2026-09-27

  GR00T-N1.5-3B_Assemble_Trocar card, history: developed based on GR00T N1.5; "This model is ready for non-commercial use"; governed by the NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf.

• https://huggingface.co/nvidia/GR00T-N1.7-3B/raw/main/README.md recorded 2026-09-27

  GR00T-N1.7-3B card: "This model is released under the NVIDIA Open Model License Agreement" and "is ready for commercial/non-commercial use"; the family lists SimplerEnv Bridge, SimplerEnv Fractal, DROID and LIBERO post-trained models on public datasets; "Total number of datasets: 13" in the training mixture.

• https://raw.githubusercontent.com/NVIDIA/Isaac-GR00T/main/README.md recorded 2026-09-27

  Isaac-GR00T README: "We just released GR00T N1.7 General Availability, the latest version of GR00T N1"; N1.6 and N1.5 listed as "Previous releases"; the checkpoint table lists GR00T-N1.7-3B and the LIBERO, DROID, SimplerEnv-Bridge and SimplerEnv-Fractal fine-tunes; the License section gives the code as Apache 2.0 and the model weights as the NVIDIA Open Model License; it says every GR00T checkpoint loads the gated Cosmos-Reason2-2B backbone on first use; launch_finetune.py for fine-tuning on a new embodiment.

• https://ungh.cc/repos/NVIDIA/Isaac-GR00T/files/main recorded 2026-09-26

  Repository tree: gr00t/experiment/launch_train.py, launch_finetune.py, trainer.py and getting_started/finetune_new_embodiment.md.

• https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/ recorded 2026-09-26

  NVIDIA Open Model License: "Models are commercially usable. You are free to create and distribute Derivative Models."

Hugging Face downloads summed over the declared GR00T checkpoints, most of them for the current N1.7 base model.

• https://huggingface.co/api/models?author=nvidia&search=GR00T&sort=downloads&limit=50&expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=lastModified&expand[]=gated recorded 2026-09-26

  Hub listing author=nvidia search=GR00T: GR00T-N1.7-3B 125,319; GR00T-N1.6-3B 29,727; GR00T-H-N1.7 1,639; GR00T-N1.5-3B 1,358; GR00T-H 741; GR00T-N1-2B 463 downloads in 30 days.

GR00T is pretrained across robot bodies and aimed squarely at humanoids, with released data for bimanual arms, the GR1 humanoid and the Unitree G1, and a documented recipe for adapting it to a new robot. It has no separate reasoning and planning model of the kind Gemini Robotics pairs with its action model.

• https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim/raw/main/README.md recorded 2026-09-26

  Post-training data card: cross-embodied bimanual manipulation, humanoid (GR1) tabletop manipulation, robot-arm kitchen manipulation and Unitree G1 loco-manipulation subsets.

• https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/NVIDIA%2FIsaac-GR00T recorded 2026-09-26

  Repository record: "NVIDIA Isaac GR00T N1.7 - A Foundation Model for Generalist Robots."

• https://ungh.cc/repos/NVIDIA/Isaac-GR00T/files/main recorded 2026-09-26

  Repository tree includes getting_started/finetune_new_embodiment.md.