NVIDIA Isaac GR00T N1.7

NVIDIA Isaac GR00T N1.7 is a 3-billion-parameter, cross-embodiment vision-language-action model developed by NVIDIA for humanoid and manipulation robots. It belongs to the Isaac GR00T family and succeeds GR00T N1.6. The model accepts camera images, a language instruction, robot proprioception, and an embodiment identifier, then predicts a chunk of continuous robot actions.^[1][3][5]

NVIDIA announced N1.7 at GTC on March 16, 2026 and made the weights and code publicly available as an early-access release on April 17.^[1][2][3] Its main changes are a Cosmos-Reason2-2B vision-language backbone based on Qwen3-VL, a relative end-effector action representation shared across human and robot data, pretraining informed by 20,854 hours of EgoScale human egocentric video, and an expanded deployment path for ONNX and TensorRT.^[3][4][7] N1.7 is a downloadable foundation policy that developers adapt to a robot and task. It is not a complete autonomous robot controller or a safety-certified system.^[4][15][16]

Scroll sideways for more →

Release and status

NVIDIA first described N1.7 in a March 16 announcement about its 2026 open-model releases. That announcement called the model an open reasoning VLA for humanoids and said it was commercially viable for real-world deployment.^[1] The downloadable model arrived one month later. On April 17, NVIDIA published an early-access forum notice and a launch article with links to the weights on Hugging Face and the full Isaac-GR00T repository. GitHub records an N1.7 release tag on April 18.^[2][3][4]

The public status labels later diverged. The current repository describes N1.7 as generally available, with production support, a validated feature set, and pull-request contributions. The model card still identifies the checkpoint as GR00T N1.7 EA, while the April forum and launch article also retain the early-access label.^[2][4][5] These sources establish that NVIDIA moved the repository beyond its initial early-access wording, but they do not provide a dated GA announcement. A precise GA transition date therefore cannot be assigned from the published record.

N1.7 is distinct from the GR00T N2 preview that NVIDIA discussed in the March announcement. N2 was described as a later world-action-model architecture with availability planned for the end of 2026; it is not another name for the N1.7 checkpoint.^[1]

Architecture

N1.7 follows the dual-system design used across the GR00T N1 family. Its higher-level component processes vision and language, while a lower-level action model converts the resulting representation and current robot state into continuous controls. NVIDIA presents the two parts as a deliberative System 2 and a fast action-producing System 1.^[3]

Vision-language backbone

System 2 uses nvidia/Cosmos-Reason2-2B, a Cosmos Reason model implemented with the Qwen3-VL architecture. It replaced the Eagle backbone used in N1.6. NVIDIA says this change supports flexible image resolution and preserves the native aspect ratio rather than padding every view to the same shape.^[4][8] The public checkpoint configuration names nvidia/Cosmos-Reason2-2B directly and uses a 2,048-dimensional backbone embedding.^[6]

The backbone encodes RGB observations and a language instruction into features for the action module. It does not itself send motor commands to a robot. The policy also needs structured proprioceptive state, such as joint values or end-effector poses, interpreted through the selected embodiment configuration.^[4][5]

Action head

System 1 is a diffusion transformer trained with flow matching. It begins from a noisy action sample and predicts a velocity field that transforms the sample into an action chunk conditioned on the vision-language features and robot state. The distributed GR00T-N1.7-3B configuration uses four inference timesteps, a 40-step action horizon, a hidden size of 1,024, and maximum state and action dimensions of 132.^[3][6]

The released checkpoint configuration and April launch article describe a 32-layer action transformer with 32 attention heads.^[3][6] A later change list in the current repository says the N1.7 action head changed from 32 layers to 16. That statement conflicts with the public checkpoint artifact, and NVIDIA has not identified a separately versioned replacement checkpoint in the reviewed material.^[4] The 32-layer figure therefore describes the released nvidia/GR00T-N1.7-3B files, rather than an inferred main-branch configuration.

Relative actions and embodiment adapters

N1.7 represents arm motion as relative end-effector deltas instead of only absolute targets. The same wrist-level representation can be calculated for human motion and robot motion, which lets human-video pretraining and robot post-training share an action abstraction. Lightweight embodiment-conditioned adapters handle state and hand-action spaces that differ across robots.^[4][7]

The model's output is a two-dimensional array of continuous values. One dimension represents time within the predicted chunk; the other contains embodiment-specific controls. Those controls can map to joints, relative end-effector motion, grippers, dexterous hands, or latent whole-body actions, depending on the policy configuration and the external controller. The base model does not determine a safe mapping for an unknown robot automatically.^[5][12]

Training data and EgoScale

NVIDIA identifies EgoScale as the main human-video pretraining work behind N1.7. The accompanying NVIDIA-led paper trained a VLA on 20,854 hours of action-labeled egocentric human video. The in-the-wild portion covered 9,869 scenes, 6,015 tasks, and 43,237 objects across household, industrial, retail, educational, and other settings. The mixture also included 829 hours from EgoDex, recorded with Apple Vision Pro tracking for more precise wrist and hand signals.^[7]

EgoScale derives relative wrist motion from estimated camera and hand poses. For finger control, it retargets 21 human hand keypoints into the 22-degree-of-freedom joint space of a Sharpa Wave hand. The use of relative wrist motion is meant to remove dependence on global camera movement, while the retargeted joint targets preserve finger articulation.^[7]

The research training procedure had three stages:

Scroll sideways for more →

The aligned dataset contained roughly 30 human trajectories and five teleoperated robot trajectories for each task. Human and robot demonstrations used comparable camera viewpoints, and human motion was captured with Vive trackers and Manus gloves. This smaller stage grounds the human-pretrained representation in robot sensing and control rather than replacing robot data.^[7]

NVIDIA's July 2026 technical blog describes a broader N1.7 pretraining mixture of about 32,000 hours of real demonstrations and human egocentric data, plus about 8,000 hours of simulated rollouts and demonstrations from BEHAVIOR, RoboCasa, and simulated GR-1 environments.^[8] The source does not publish a complete corpus manifest or an exact split of the rounded 32,000-hour figure between human and robot data. The 20,854-hour EgoScale amount can be reported directly, but subtracting it to invent a precise robot-data subtotal would not be supported.

Evaluation

All published N1.7 and EgoScale results in this section were reported by NVIDIA researchers or by the joint NVIDIA and Hugging Face integration team. They had not been independently reproduced in the reviewed sources.

EgoScale real-robot study

The EgoScale paper's main real-robot experiments used a Galaxea R1 Pro with its base and torso fixed, two 7-degree-of-freedom arms, and two 22-degree-of-freedom Sharpa hands. Three RGB cameras supplied a head view and two wrist views. The five evaluation tasks were shirt rolling, card sorting, using tongs to transfer fruit, unscrewing a bottle, and transferring liquid with a syringe.^[7]

Most tasks used 100 teleoperated robot demonstrations. Shirt rolling used 20. The researchers trained each policy with two random seeds and normally ran 10 evaluation trials for each checkpoint. The tongs task used 16 trials across four bottle instances. Initial scenes were aligned with an image-overlay procedure to reduce setup variation.^[7]

The paper reports that its final pretrain-plus-mid-train policy improved average success rate by 54% over a no-pretraining baseline. In the scaling experiment, average task completion rose from 0.30 with 1,000 hours of human pretraining to 0.71 with 20,000 hours. The fitted relationship between human validation loss and the logarithm of training hours had an R-squared value of 0.9983 within the tested 1,000-20,000-hour range.^[7] This fit applies to that experimental range and does not establish indefinite gains beyond it.

For one-shot transfer, the researchers supplied one robot demonstration plus aligned human demonstrations for each new task. The pretrain-plus-mid-train policy reached a reported success rate of 0.88 on shirt folding and 0.55 on bottle unscrewing. The paper also tested transfer to a Unitree G1 with a 7-degree-of-freedom tri-finger hand and reported more than 30 percentage points of absolute improvement across two tasks relative to a baseline without human pretraining. A separate Homie controller handled the G1's lower body, so these trials did not demonstrate that N1.7 alone controlled the complete robot.^[7]

N1.6 comparisons

NVIDIA's July technical article reports N1.7 gains over N1.6 of +10% on DROID-F0, +61% on DROID-F6, +5% on SimplerEnv Bridge, and +2% on SimplerEnv Fractal.^[8] These are first-party comparisons. The article does not provide enough information to treat them as independent estimates of general performance across robots, and they should not be combined with the EgoScale task scores.

LeRobot LIBERO results

The joint integration team evaluated LeRobot post-trained checkpoints on LIBERO, a language-conditioned lifelong robot-learning benchmark whose full suite contains 130 tabletop tasks.^[9][18]

Scroll sideways for more →

The team's published N1.5 average omits the Goal suite, while the N1.7 average includes it. The two average figures therefore do not summarize identical suite sets. The suite-level values are more informative than treating 87% to 96.5% as a strictly like-for-like improvement.^[9]

Embodiments and robot control

N1.7 is called cross-embodiment because a shared backbone and action model can be adapted through embodiment-specific state, action, and normalization definitions. A registered tag identifies how each dataset and robot interface should be interpreted. It does not mean the base model can control every listed robot without the matching checkpoint and controller.^[4][12]

The current repository documents zero-shot base-model tags for DROID, generic X-DOF data, real Unitree G1, and several R1 Pro configurations with Sharpa hands. It documents post-trained paths for DROID, G1 with GEAR-SONIC whole-body control, LIBERO Panda, SimplerEnv Google Robot, and SimplerEnv WidowX. A NEW_EMBODIMENT path lets developers provide a custom modality configuration and post-train on their own data.^[12]

NVIDIA's April launch material also reports validation across locomotion-manipulation, tabletop manipulation, and dexterous bimanual tasks using Unitree G1, Bimanual Manipulator YAM, and AgiBot Genie 1.^[3] These reports show the development team's tested platforms. They are not safety certifications, universal hardware-support guarantees, or evidence of unsupervised commercial deployment.

Isaac-GR00T and LeRobot integration

The native Isaac-GR00T codebase stores demonstrations in a GR00T-specific flavor of the LeRobot v2 format. It adds meta/modality.json to describe how concatenated state and action arrays divide into named fields and which video streams are present. The repository includes tools for open-loop evaluation, simulation benchmarks, post-training, and server-client deployment.^[4]

LeRobot 0.6.0 added a separate native N1.7 policy implementation in July 2026. Hugging Face says it parity-tested the port against NVIDIA's Isaac-GR00T implementation with matching inputs and outputs. N1.7 replaced N1.5 in current LeRobot releases; users who still require the older port must pin LeRobot 0.5.1.^[10]

A joint NVIDIA and Hugging Face tutorial connects Isaac Teleop, an SO-101 arm, and the LeRobot workflow. Demonstrations can be collected with an XR controller or an SO-101 leader arm, saved as a LeRobot Dataset v3 dataset, used to post-train N1.7, and deployed through lerobot-rollout. The authors used an RTX 6000 Pro for their training example.^[9] This workflow demonstrates the integration path. It does not show that model post-training runs on the SO-101 controller or on every Jetson target.

On July 24, NVIDIA Japan promoted the N1.7 and Isaac Teleop integration and said Cosmos 3 support was planned.^[19] That future-tense statement does not establish that Cosmos 3 was already part of the released LeRobot N1.7 policy.

Hardware and inference

NVIDIA recommends one GPU with at least 16 GB of memory for inference and one or more GPUs with at least 40 GB each for fine-tuning. The documented runtime paths cover discrete GPUs on CUDA 12.8, Jetson Orin on CUDA 12.6, and Jetson Thor or DGX Spark on CUDA 13.0. Hardware support depends on the software image, CUDA version, Python version, and available acceleration path.^[4][11]

NVIDIA reports the following end-to-end policy inference rates for four denoising steps and one camera. These measurements cover model replanning, not camera capture, action execution, networking, motor control, or a complete safety loop.^[11]

Scroll sideways for more →

On Orin, the documented TensorRT version cannot accelerate the backbone, so only the action head uses TensorRT. NVIDIA characterizes sub-5 Hz inference as suitable only for slow, non-reactive tasks. Action chunking can let a robot execute controls more frequently than the model replans, but it does not reduce the need to validate latency and closed-loop behavior on the actual machine.^[11]

Weights and licensing

The public Hugging Face checkpoint contains BF16 safetensors, an embodiment map, preprocessing configuration, normalization statistics, and the model configuration. The base checkpoint is about 6.9 GB on the Hub. Loading it also requires access to the separately hosted Cosmos-Reason2-2B backbone.^[4][5]

The software and weights do not use one identical license. NVIDIA's Isaac-GR00T code repository uses the Apache License 2.0. The checkpoint model card points to the NVIDIA Open Model License Agreement. The current agreement, last modified October 24, 2025, says covered models are commercially usable and grants rights to make, sell, distribute, and modify them subject to its conditions.^[4][5][13]

The files distributed with the checkpoint are not fully synchronized with that statement. The bundled LICENSE file contains different NVIDIA license text whose Section 3.3 limits the work and derivatives to noncommercial research or evaluation, except for NVIDIA and its affiliates. The explainability and safety subcards also link to an older noncommercial license.^[14][15][16] NVIDIA's launch article, model-card prose, and current linked Open Model License describe commercial use, but the conflicting bundled text means a downloader should inspect the terms attached to the exact artifact instead of relying on a summary label. This discrepancy is a documentation issue, not a legal conclusion about a particular deployment.

Limitations and safety

NVIDIA states that N1.7 was not tested or intended for mission-critical applications that require functional safety. Its model documentation identifies dynamic surroundings, changes in lighting, environment-specific integration, and limited initial embodiment support as possible sources of underperformance.^[15][16]

A deployer must still supply and validate the robot controller, joint and force limits, collision avoidance, emergency stopping, supervision, and other guardrails required by the application. Post-training on a demonstration dataset does not establish safe behavior outside that data distribution. Benchmark success rates also do not measure rare failures, human-contact safety, cybersecurity, or compliance with industrial or medical standards.^[15][16]

N1.7's published results cover selected laboratory and simulation tasks under specified protocols. They do not establish a general success rate for arbitrary humanoid work, a production deployment record, or functional-safety certification.

Relationship to GR00T-H-N1.7

GR00T-H-N1.7 is a separate post-trained model for surgical robots, not a built-in operating mode of the base N1.7 checkpoint. NVIDIA's healthcare model card says it inherits the N1.7 architecture and adapts it with the Open-H embodiment dataset.^[17] Its healthcare training data, evaluation, and use restrictions require separate treatment; results for GR00T-H should not be attributed to the general N1.7 base model.

References

1. NVIDIA, "NVIDIA Expands Open Model Families to Power the Next Wave of Agentic, Physical and Healthcare AI," March 16, 2026. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Expands-Open-Model-Families-to-Power-the-Next-Wave-of-Agentic-Physical-and-Healthcare-AI/ ↩
2. NVIDIA Developer Forums, "EARLY ACCESS: Isaac GR00T N1.7 - Open Reasoning VLA Model for Humanoid Robotics," April 17, 2026. https://forums.developer.nvidia.com/t/early-access-isaac-gr00t-n1-7-open-reasoning-vla-model-for-humanoid-robotics/366916 ↩
3. NVIDIA, "NVIDIA Isaac GR00T N1.7: Open Reasoning VLA Model for Humanoid Robots," April 17, 2026. https://huggingface.co/blog/nvidia/gr00t-n1-7 ↩
4. NVIDIA, "Isaac-GR00T," GitHub repository and release history, accessed July 25, 2026. https://github.com/NVIDIA/Isaac-GR00T ↩
5. NVIDIA, "GR00T-N1.7-3B Model Card," Hugging Face, accessed July 25, 2026. https://huggingface.co/nvidia/GR00T-N1.7-3B ↩
6. NVIDIA, "GR00T-N1.7-3B Configuration," Hugging Face, accessed July 25, 2026. https://huggingface.co/nvidia/GR00T-N1.7-3B/blob/main/config.json ↩
7. Zheng, R. et al., "EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data," arXiv:2602.16710, February 18, 2026. https://arxiv.org/abs/2602.16710 ↩
8. NVIDIA, "Develop Humanoid Robot Policies End-to-End with NVIDIA Isaac GR00T," NVIDIA Technical Blog, July 7, 2026. https://developer.nvidia.com/blog/develop-humanoid-robot-policies-end-to-end-with-nvidia-isaac-gr00t/ ↩
9. NVIDIA and Hugging Face, "NVIDIA Isaac Teleop and GR00T 1.7 Open VLA Model Available in LeRobot," July 7, 2026. https://huggingface.co/blog/nvidia/nvidia-isaac-teleop-and-gr00t17-in-lerobot ↩
10. Hugging Face, "LeRobot v0.6.0: Imagine, Evaluate, Improve," July 7, 2026. https://huggingface.co/blog/lerobot-release-v060 ↩
11. NVIDIA, "Hardware Recommendations," Isaac-GR00T repository, accessed July 25, 2026. https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/hardware_recommendation.md ↩
12. NVIDIA, "Policy API Guide," Isaac-GR00T repository, accessed July 25, 2026. https://github.com/NVIDIA/Isaac-GR00T/blob/main/getting_started/policy.md ↩
13. NVIDIA, "NVIDIA Open Model License Agreement," last modified October 24, 2025. https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/ ↩
14. NVIDIA, "GR00T-N1.7-3B Bundled License," Hugging Face, accessed July 25, 2026. https://huggingface.co/nvidia/GR00T-N1.7-3B/blob/main/LICENSE ↩
15. NVIDIA, "GR00T-N1.7-3B Explainability Subcard," Hugging Face, accessed July 25, 2026. https://huggingface.co/nvidia/GR00T-N1.7-3B/blob/main/EXPLAINABILITY.md ↩
16. NVIDIA, "GR00T-N1.7-3B Safety and Security Subcard," Hugging Face, accessed July 25, 2026. https://huggingface.co/nvidia/GR00T-N1.7-3B/blob/main/SAFETY_and_SECURITY.md ↩
17. NVIDIA, "GR00T-H-N1.7 Model Card," Hugging Face, accessed July 25, 2026. https://huggingface.co/nvidia/GR00T-H-N1.7 ↩
18. Liu, B. et al., "LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning," arXiv:2306.03310, June 2023. https://arxiv.org/abs/2306.03310 ↩
19. NVIDIA Japan, post about GR00T 1.7 and Isaac Teleop in LeRobot, July 24, 2026. https://x.com/NVIDIAJapan/status/2080564137994633546 ↩