Open Source RoboticsRobots, software & learning

NVIDIA Isaac GR00T N1.7

Robotics model

Cross-embodiment ~3B-parameter Vision-Language-Action foundation model pairing a Cosmos-Reason2-2B (Qwen3-VL) VLM backbone with a flow-matching Diffusion Transformer action head, Apache-2.0 reference code, and NVIDIA Open Model License weights. [1] [2] [3]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Humanoid and manipulation robotics teams; Robotics researchers and post-training engineers; Commercial robot developers targeting NVIDIA dGPU or Jetson edge deployment [1] [3]
Parameters
~3B (model card reports 3,000,000,000 [3]
Embodiments
Unitree G1 (UNITREE_G1 / UNITREE_G1_SONIC whole-body), Fourier GR-1 (simulation), Franka Panda (LIBERO_PANDA / DROID), … [1] [3] [4]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

NVIDIA Isaac GR00T N1.7 (NVIDIA/Isaac-GR00T) is a ~3B-parameter Vision-Language-Action (VLA) foundation model designed for humanoid, bimanual, and single-arm robot manipulation. The General Availability (n1.7-release) stack pairs Apache-2.0 training and serving code with five commercially usable Hugging Face checkpoints under the NVIDIA Open Model License, spanning the base nvidia/GR00T-N1.7-3B model (6.44 GB safetensors) and post-trained checkpoints for DROID, LIBERO, SimplerEnv Bridge, and SimplerEnv Fractal. [1] [2] [3]

Model card summary

Architecture

Vision-Language-Action model combining a nvidia/Cosmos-Reason2-2B (Qwen3-VL) vision-language backbone with embodiment-indexed proprioception/action MLPs, 4-layer VL self-attention, and a flow-matching Diffusion Transformer (DiT) action head (32 layers in released GR00T-N1.7-3B config.json; 4 denoising steps by default).

README detailed changelog mentions 16 diffusion layers in code defaults, while pinned nvidia/GR00T-N1.7-3B config.json specifies diffusion_model_cfg.num_layers: 32 and num_inference_timesteps: 4.

[1] [3] [4]

Base model

nvidia/Cosmos-Reason2-2B (derived from Qwen/Qwen3-VL-2B-Instruct)

Verified from config.json model_name and HF API metadata for nvidia/Cosmos-Reason2-2B.

[1] [4]

Parameter count

~3B (model card reports 3,000,000,000; HF safetensors API reports 3,144,016,000 total / 3,455,180,928 BF16 tensor entries for nvidia/GR00T-N1.7-3B)

[3]

Input modalities
rgbproprioceptionlanguage

Variable number of RGB camera frames (native aspect ratio via Qwen3-VL), proprioceptive state vector padded up to max_state_dim: 132, language instruction string, and integer embodiment ID.

[3] [4]

Output modalities
ee-pose-deltajoint-positionsgripper-command

Continuous action chunks (relative EEF pose deltas or embodiment-specific joint/gripper commands; latent whole-body commands for UNITREE_G1_SONIC).

[1] [3] [4]

Action space

Continuous flow-matching action chunks with action_horizon: 40, max_action_dim: 132, max_num_embodiments: 32, and num_inference_timesteps: 4; adopts a shared relative end-effector (EEF) delta action representation across robot and human embodiments.

Verified from nvidia/GR00T-N1.7-3B config.json and README.

[1] [4]

Control frequency

11.7 Hz PyTorch eager / 35.9 Hz TensorRT full pipeline on H100 (1-cam, 4 denoising steps, model card); README reports ~40 Hz on RTX PRO 6000 Blackwell and ~28 Hz on Jetson AGX Thor with TensorRT on batch-1 LIBERO

Upstream-reported inference rates vary by workload (model card 1-camera table lists 10.7 Hz on AGX Thor and 4.6 Hz DiT-only TRT on Orin; README cites ~28 Hz on Thor for batch-1 LIBERO with MAXN + jetson_clocks).

[1] [3]

Trained / supported embodiments
  • Unitree G1 (UNITREE_G1 / UNITREE_G1_SONIC whole-body)
  • Fourier GR-1 (simulation)
  • Franka Panda (LIBERO_PANDA / DROID)
  • WidowX (SIMPLER_ENV_WIDOWX)
  • Google Robot (SIMPLER_ENV_GOOGLE)
  • SO-100 (NEW_EMBODIMENT fine-tuning example)
  • Custom embodiments via NEW_EMBODIMENT up to 132-D state/action

[1] [3] [4]

Training datasets
  • GR00T pretraining mixture (21.6M data points across 13 hybrid human, robot, and simulated datasets per model card)
  • 20K hours of EgoScale human video data
  • Post-training benchmark datasets: DROID, LIBERO, BridgeData V2 (SimplerEnv), Fractal (SimplerEnv)

Full pretraining mixture is not released as a single open bundle; public Physical AI datasets are linked separately on Hugging Face.

[1] [3]

Interfaces
  • Gr00tPolicy Python API
  • ZMQ PolicyServer / PolicyClient (default port 5555)
  • ONNX / TensorRT deployment pipeline
  • LeRobot groot policy integration
  • GR00T LeRobot v2 parquet + modality.json dataset interface

[1]

Checkpoints

CheckpointFormatSizeAccessPinned revision
nvidia/GR00T-N1.7-3Bsafetensors6.44 GBopen373d34e2723d
nvidia/GR00T-N1.7-DROIDsafetensors6.44 GBopen113eb584e6b4
nvidia/GR00T-N1.7-LIBEROsafetensors132.97 GBopen658540327860
nvidia/GR00T-N1.7-SimplerEnv-Bridgesafetensors33.24 GBopen31733e3dc998
nvidia/GR00T-N1.7-SimplerEnv-Fractalsafetensors33.24 GBopen265afd7f9eec

All 5 GR00T N1.7 repos are ungated on Hugging Face, but loading any of them requires access to the gated nvidia/Cosmos-Reason2-2B backbone. LIBERO (132.97 GB) and SimplerEnv (33.24 GB each) repos bundle multiple suite/step subfolders.

[1] [3]

Is this for you?

Use cases
  • Fine-tuning a 3B cross-embodiment VLA on custom LeRobot v2 datasets with modality.json configs
  • Zero-shot or post-trained inference on DROID, LIBERO, SimplerEnv, and humanoid embodiments
  • Low-latency ZMQ server-client and TensorRT deployment on dGPU, DGX Spark, or Jetson AGX Thor/Orin
  • Whole-body humanoid control paired with GEAR-SONIC via UNITREE_G1_SONIC

[1] [3]

Known limitations
  • Runtime loading requires Hugging Face authentication and approval on the gated nvidia/Cosmos-Reason2-2B VLM backbone (GatedRepoError / 401 otherwise)
  • Video decoding strictly requires torchcodec (0.8.0 on dGPU) and FFmpeg 4-7; decord, pyav, and FFmpeg 8 are unsupported
  • Jetson Orin supports DiT-only TensorRT (not full-pipeline TRT) due to TRT 10.3 backbone limitations, and GB300 (sm_103) does not support torch.compile with pinned Triton 3.3.1
  • Upstream notes 5-6% evaluation variance across runs due to non-deterministic image augmentations

[1] [3]

Safety and limitations
  • Model card explicitly states users are responsible for model inputs, outputs, and implementing safety guardrails and V-model validation before physical robot deployment.
  • Runtime depends on gated nvidia/Cosmos-Reason2-2B VLM weights (HF API license: other); the separate backbone agreement terms were not verified in this review and must be accepted on Hugging Face before loading.
  • Released checkpoints leave strict_vl_padding_mask and strict_action_padding_mask off by default, causing ~2% action shift with batch prompt padding length.

[1] [3]

Compute & fine-tuning

Inference compute
  • Minimum VRAM: 16 GB
  • Documented hardware: NVIDIA H100 80GB, RTX PRO 6000 Blackwell, L40, DGX Spark, Jetson AGX Thor, Jetson Orin
  • Precision: bf16
  • Reported latency: 27.9 ms
  • Edge / consumer capable: Yes

27.9 ms is upstream-reported TensorRT full-pipeline E2E latency on H100 / RTX PRO 6000 Blackwell (85.8 ms PyTorch eager on H100; 93.8 ms TRT on AGX Thor; 216.5 ms DiT-only TRT on Orin; 1 camera, 4 denoising steps).

[1] [3]

Fine-tuning
  • Supported: Yes
  • Methods: Single-GPU and multi-GPU fine-tuning via launch_finetune.py; NEW_EMBODIMENT modality.json adaptation; Multi-dataset mixture training (ds_weights_alpha); LeRobot groot policy fine-tuning
  • Minimum VRAM: 40 GB
  • Recipe: Upstream fine-tuning documentation

README recommends 1+ GPUs with 40 GB+ VRAM (H100 or L40 recommended); standard batch-1 benchmark freezes visual/language backbones while training projector, diffusion model, and VLLN.

[1]

Host computer

Inference requires 1 NVIDIA GPU with 16 GB+ VRAM (e.g., RTX 4090, L40, H100, Jetson AGX Thor/Orin, DGX Spark); fine-tuning recommends 1+ GPUs with 40 GB+ VRAM (H100 or L40 recommended).

[1]

Operating system
  • Linux

x86_64 Linux dGPU and aarch64 Linux (JetPack 7.2 on Jetson Thor/Orin, DGX Spark).

[1] [3]

Dependencies
  • Python 3.12
  • PyTorch 2.7 / CUDA 12.8 (dGPU) or CUDA 13.0-13.2 (Spark/Thor/Orin)
  • transformers 4.57.x
  • flash-attn
  • torchcodec + FFmpeg 4-7
  • nvidia/Cosmos-Reason2-2B (gated Hugging Face backbone)
  • uv / git-lfs

[1] [4]

Skills
  • Python 3.12 and uv environment management
  • PyTorch / CUDA and optional TensorRT export
  • LeRobot dataset formatting and embodiment modality configuration

Inferred from documented uv, CUDA, torchcodec, ZMQ server, and modality.json workflows.

[1] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1] · Model card / documentation [1]

  • Install git-lfs and clone NVIDIA/Isaac-GR00T with --recurse-submodules
  • Install FFmpeg 4-7 (required by torchcodec) and run uv sync --python 3.12 (or platform install_deps.sh on Spark/Thor/Orin)
  • Request access to gated nvidia/Cosmos-Reason2-2B on Hugging Face and authenticate via huggingface-cli login or HF_TOKEN
  • Run standalone_inference_script.py on bundled demo_data/droid_sample or start run_gr00t_server.py over ZMQ
  • Convert custom demonstrations to GR00T LeRobot v2 format with meta/modality.json and run launch_finetune.py

[1]

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
LIBERO (Spatial, Goal, Object, 10-Long)Spatial 97.65% (195/200), Goal 97.5% (195/200), Object 98.45% (197/200), 10-Long 94.35% (189/200)Upstream-reported finetuned N1.7 results (20K steps, batch size 640, state_dropout_prob 0.2); not independently reproduced by OSR.
SimplerEnv Bridge (WidowX, 7 tasks)62.3% average across 7 tasks (vs 56.6% for N1.6)Upstream-reported finetuned nvidia/GR00T-N1.7-SimplerEnv-Bridge results (100 trials/task); includes 2% on widowx_put_eggplant_in_sink; not independently verified by OSR.

[5] [6]

Project notes

Architecture, relative EEF actions, and config caveats

GR00T N1.7 replaces the earlier N1.6 Eagle VLM with nvidia/Cosmos-Reason2-2B (derived from Qwen/Qwen3-VL-2B-Instruct), which encodes native-aspect-ratio RGB camera streams alongside language instructions. Embodiment-indexed MLPs project proprioceptive state (padded up to max_state_dim: 132) and decode continuous actions from a flow-matching Diffusion Transformer (DiT) head. At the pinned nvidia/GR00T-N1.7-3B revision (373d34e2723d), config.json sets action_horizon: 40, max_action_dim: 132, max_num_embodiments: 32, and num_inference_timesteps: 4, adopting a shared relative end-effector (EEF) delta representation trained across robot demonstrations and 20K hours of EgoScale human video. One documentation mismatch to note: the repository's N1.6-to-N1.7 changelog mentions 16 diffusion layers, whereas the released GR00T-N1.7-3B config.json specifies diffusion_model_cfg.num_layers: 32. Released checkpoints also keep strict_vl_padding_mask and strict_action_padding_mask off by default, which upstream notes can shift predicted actions by ~2% depending on batch prompt padding lengths. [1] [3] [4]

Compute requirements and edge deployment setup

Upstream requires 1 GPU with 16 GB+ VRAM for inference (e.g., RTX 4090, L40, H100, DGX Spark, or Jetson AGX Thor/Orin) and recommends 1 or more GPUs with 40 GB+ VRAM for fine-tuning via launch_finetune.py using GR00T LeRobot v2 datasets (meta/modality.json). For Jetson Thor and Orin (JetPack 7.2 / CUDA 13.2, Python 3.12), the documented setup runs bash scripts/deployment/<platform>/install_deps.sh, activates the environment with source .venv/bin/activate && source scripts/activate_<platform>.sh, and invokes plain python or torchrun (avoiding uv run, which re-syncs against the x86_64 pyproject.toml); see scripts/deployment/README.md for per-platform Docker and bare-metal steps. For video decoding, torchcodec is the sole supported backend: the default dGPU installation pins torchcodec==0.8.0 and requires FFmpeg 4–7 (FFmpeg 8 fails to load), whereas JetPack 7.2 Thor/Orin use the cu132 torchcodec 0.15.0 wheel and DGX Spark may build torchcodec from source during install_deps.sh. [1]

Upstream-reported timing and benchmarks

The model card reports single-camera, 4-denoising-step inference timings across desktop and edge platforms, while the repository README separately cites ~40 Hz on RTX PRO 6000 Blackwell and ~28 Hz on Jetson AGX Thor with TensorRT on batch-1 LIBERO (Thor MAXN + jetson_clocks): [1] [3]

PlatformWorkload / ModeLatencyFrequencyNotes
H100 80GB1-cam, 4 steps — PyTorch eager / Full-pipeline TRT85.8 ms / 27.9 ms11.7 Hz / 35.9 HzModel card
Jetson AGX Thor1-cam, 4 steps — Full-pipeline TRT93.8 ms10.7 HzModel card (README cites ~28 Hz on batch-1 LIBERO)
Jetson Orin1-cam, 4 steps — DiT-only TRT216.5 ms4.6 HzTRT 10.3 does not support backbone engine

For manipulation accuracy, upstream reports finetuned LIBERO success rates of 97.65% (195/200, Spatial), 97.5% (195/200, Goal), 98.45% (197/200, Object), and 94.35% (189/200, 10-Long). Note an internal mismatch in upstream's table where the reported percentages differ slightly from the raw trial ratios (195/200 = 97.5%, 197/200 = 98.5%, 189/200 = 94.5%), which requires upstream clarification before treating either figure as definitive. On SimplerEnv Bridge, upstream reports 62.3% average across seven WidowX tasks (including 2% on widowx_put_eggplant_in_sink). All benchmark and timing numbers are upstream-reported; Open Source Robotics has not run training, fine-tuning, inference, or robot evaluations for this profile. [1] [3] [5] [6]

What we checked

None. Prepared from public upstream repositories and Hugging Face API metadata without hands-on training, fine-tuning, or inference.

Updates

Upstream last updated
Nov 17, 2026 [1]
Latest release
Apr 18, 2026 [1]
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
NVIDIA/Isaac-GR00T@708617f5716c63fc0a327391492ea2b91977ae90; HF nvidia/GR00T-N1.7-3B@373d34e2723d928870157d07dc216a19a624b5bb · combined hardware and software repository

Separate license matrix

Flagship ~3B open-weights cross-embodiment VLA with Apache-2.0 training/deployment code and commercially usable NVIDIA Open Model License weights; included as partial openness because weights use a non-OSI custom license, runtime requires gated nvidia/Cosmos-Reason2-2B access, and the full pretraining mixture is not openly bundled.

LayerAvailability / license
Codeavailable · Apache-2.0
Verified Apache License Version 2.0 in root LICENSE file. [1] [2]
Model weightsavailable · NVIDIA Open Model License (commercial use permitted; custom non-OSI model license)
GR00T N1.7 checkpoints are ungated on HF under NVIDIA Open Model License, but runtime also requires accepting gated access for nvidia/Cosmos-Reason2-2B. [1] [3]
Training datasetsavailable · Mixed (bundled demo_data under repo; public Physical AI / DROID / LIBERO / Bridge datasets have separate licenses; full 21.6M pretraining + 20K-hour EgoScale mixture not released as a single open bundle)
Partial availability: public Physical AI and post-training datasets are available, while the full 21.6M pretraining + 20K-hour EgoScale mixture is not released as a single open bundle. [1] [3]
Documentationavailable · Apache-2.0
[1] [2]

Licensing split and gated backbone requirement

Code and documentation in NVIDIA/Isaac-GR00T are licensed under Apache-2.0 (verified in the root LICENSE file), while the GR00T N1.7 weights are released under the NVIDIA Open Model License, which NVIDIA describes as permitting commercial and non-commercial use under custom non-OSI terms (the full NVIDIA Open Model License text is hosted at NVIDIA's agreement URL rather than in the repository). Separate from GR00T's own licensing, access to the VLM backbone is gated: although all five GR00T-N1.7-* Hugging Face repositories are ungated, every checkpoint loads nvidia/Cosmos-Reason2-2B on first use, failing with GatedRepoError (401) until you accept its terms and authenticate with HF_TOKEN. In Hugging Face API metadata, nvidia/Cosmos-Reason2-2B is tagged license: other (derived from Qwen/Qwen3-VL-2B-Instruct), and its separate license and gated agreement terms were not established from the inspected repository files and must be reviewed directly on Hugging Face before deployment. Finally, while bundled demo_data/ and linked Physical AI, DROID, LIBERO, and Bridge datasets support post-training, the full 21.6M-datapoint pretraining mixture and 20K-hour EgoScale corpus are not released as a single open bundle. [1] [2] [3]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Initial 0.8.0-draft.1 model profile prepared from NVIDIA/Isaac-GR00T@708617f5716c and HF nvidia/GR00T-N1.7-3B@373d34e2723d.