GR00T N1 Architecture and Artifact Record

GR00T N1 Architecture and Artifact Record

Plain-language model

The original GR00T N1 was a vision-language-action model. It received camera observations, language instructions, robot state, and an embodiment identifier, then generated a sequence of continuous robot actions.

NVIDIA described the March 2025 model as two tightly coupled systems. A vision-language component interpreted observations and instructions. A diffusion or flow-matching action component generated continuous actions. The paper describes joint end-to-end training and post-training for robot-specific tasks.

https://arxiv.org/abs/2503.14734

https://developer.nvidia.com/blog/accelerate-generalist-humanoid-robot-development-with-nvidia-isaac-gr00t-n1/

Original model-card details

The original Hugging Face card identifies image frames, proprioceptive state, language instruction, and embodiment ID as inputs. It identifies a flow-matching transformer that reconstructs continuous actions from noise at inference time.

https://huggingface.co/nvidia/GR00T-N1-2B

The card names a 2B-parameter artifact but also describes a lightweight N1-1B architecture. This inconsistency should be preserved as a source limitation rather than harmonized by inference.

The training record describes cross-embodiment human and synthetic data plus robot sensors, state, and actuator data. Evaluation spans named simulation suites and selected real GR-1 tasks. The reported results belong to those tasks, embodiments, comparators, and evaluation settings.

Artifact-level meaning of open

"Open source," "open model," "open weights," and "open platform" should not be used as synonyms.

Data and adaptation boundary

The launch post describes a data pyramid. Broad human video supplies visual and task semantics. Synthetic data supplies controllable trajectories and variation. Real robot data supplies embodiment-specific state and action evidence.

The available checkpoint does not define a new robot's sensors, kinematics, action space, control rate, calibration, tools, timing, workcell, task tolerance, or risk. Those remain adaptation and system-integration work.

Current-version note

The current N1.7 repository reviewed on July 28, 2026 is a later system. It describes a Cosmos-Reason2-2B or Qwen3-VL-based VLM, relative end-effector actions, a GR00T-specific LeRobot v2 data format, a larger base checkpoint, and updated fine-tuning, evaluation, export, and deployment workflows.

https://github.com/NVIDIA/Isaac-GR00T

Those properties help explain how the project evolved. They must not be presented as features of the original N1 launch.

Safety boundary

The original model card says N1 was not tested or intended for mission-critical applications requiring functional safety. A model output is not a safety function. A public explainer must not present model availability, benchmark success, or repository deployment instructions as approval for a physical application.