GR00T N1 is NVIDIA's open-source Vision-Language-Action (VLA) foundation model for humanoid robots, announced March 18, 2025. It uses a dual-system architecture: System 2 (a Vision-Language Model for slow reasoning) and System 1 (a diffusion transformer for fast motor control at 200Hz). Nine humanoid robot partners adopted it at launch, including Figure, Agility Robotics, Boston Dynamics, and 1X. Current version is N1.7, available on GitHub.

NVIDIA Physical AI

Isaac, GR00T N1, and Cosmos — how they connect

Q: What is NVIDIA Cosmos?

NVIDIA Cosmos is a world foundation model released in January 2025 (major update March 2025). It generates synthetic training data for robots — specifically photorealistic video of physical scenarios. NVIDIA used Cosmos to generate 780,000 synthetic robot trajectories in 11 hours (equivalent to 9 months of human demonstrations), improving GR00T robot performance by 40%. Cosmos is open-weight under the OpenMDW1.1 license (Linux Foundation).

Q: What is sim-to-real transfer?

Sim-to-real transfer is the process of training an AI model in simulation and deploying it in the real world. The core challenge is the reality gap — simulation is never perfectly accurate, so models trained purely in simulation may fail when deployed on physical hardware. NVIDIA's Isaac Sim and Cosmos address this through photorealistic simulation and synthetic data that closely approximates real-world visual and physical conditions.

NVIDIA built a three-layer platform for Physical AI. Isaac trains robots in photorealistic simulation. GR00T N1 is the open-source foundation model that runs on the robot. Cosmos generates the synthetic training data that makes the whole system scale. Understanding all three is required to understand why NVIDIA is positioned where it is in the humanoid robot race.

The platform

Three layers, one platform

Layer 1

Isaac

Training platform

Isaac Sim 5.0 is a GPU-accelerated robot simulation environment built on NVIDIA Omniverse. Robots train in photorealistic digital twins before touching physical hardware. Isaac Lab provides the RL framework for policy training. Isaac ROS provides CUDA-accelerated ROS2 packages for deployment on AMRs and humanoids.

—Isaac Sim 5.0: photorealistic digital twins on Omniverse

—0.1mm grasping precision achievable in simulation

—Isaac Lab: reinforcement learning framework for robot policy training

—Isaac ROS: CUDA-accelerated ROS2 packages for humanoids and AMRs

—Partners: ABB, FANUC, KUKA, Universal Robots, Boston Dynamics, Figure AI, Fourier Intelligence

Layer 2

GR00T N1

Foundation model

GR00T N1 (announced March 18, 2025) is an open-source Vision-Language-Action (VLA) model. It uses a dual-system architecture: System 2 is a Vision-Language Model that handles slow reasoning and task planning. System 1 is a diffusion transformer that handles fast motor control at 200Hz. The model runs on Jetson Thor hardware. Current version is N1.7, available on GitHub under open weights.

—Announced March 18, 2025 — open source VLA model

—System 2: VLM (slow reasoning, task planning)

—System 1: diffusion transformer (fast motor control at 200Hz)

—Runs on Jetson Thor (T5000)

—Current version: N1.7 — available on GitHub

—9 humanoid partners: Figure, Agility, Boston Dynamics, 1X, Apptronik, and more

Layer 3

Cosmos

World foundation model

Cosmos (January 2025, major update March 2025) generates synthetic training data. Cosmos Predict generates future world states given a current state. Cosmos Transfer converts simulation video to photorealistic video for training data. Cosmos Reason adds physical reasoning capability. NVIDIA used Cosmos to generate 780,000 synthetic robot trajectories in 11 hours — the equivalent of 9 months of human demonstrations — improving GR00T performance by 40%.

—Released January 2025, major update March 2025

—Cosmos Predict: generate future world states

—Cosmos Transfer: sim→photoreal video (synthetic training data)

—Cosmos Reason: physical reasoning layer

—780,000 synthetic trajectories in 11 hours (vs 9 months human demos)

—40% GR00T performance improvement from Cosmos-generated data

—Open-weight: OpenMDW1.1 license (Linux Foundation)

Sim-to-real

Why simulation matters for robot training

Training a robot in the real world is expensive, slow, and dangerous. A robot that drops a glass during training breaks the glass. A simulation that drops a glass costs nothing and runs 10,000 times faster than real time.

The challenge is the reality gap — if the simulation is not realistic enough, the model fails when deployed on physical hardware. It learned physics that does not exist.

Cosmos addresses the reality gap by generating photorealistic training video from simulation. 780,000 trajectories in 11 hours is the number that illustrates the scale advantage: a human demo team collecting the same data would take 9 months.

Adoption

Who uses the NVIDIA Physical AI stack

Humanoid robots (GR00T N1 partners)

—Figure AI

—Agility Robotics (Digit)

—Boston Dynamics

—1X Technologies (NEO)

—Apptronik

—Fourier Intelligence

Industrial robotics (Isaac users)

—ABB

—FANUC

—KUKA

—Universal Robots

—Amazon Robotics

FAQ

Common questions

What is GR00T N1?

An open-source Vision-Language-Action model for humanoid robots, announced March 18, 2025. Dual-system: VLM for reasoning (System 2) + diffusion transformer for 200Hz motor control (System 1). Current version N1.7 on GitHub.

What is NVIDIA Cosmos?

A world foundation model that generates synthetic training data for robots. NVIDIA used it to generate 780,000 robot trajectories in 11 hours (vs 9 months of human demos), improving GR00T performance by 40%. Open-weight under OpenMDW1.1 license.

What is sim-to-real transfer?

Training an AI model in simulation and deploying it in the physical world. The core challenge is the reality gap — NVIDIA addresses this with photorealistic Isaac Sim and Cosmos synthetic data.

Technology hub Edge AI Chips Embodied AI vs AI Agents

Last updated: July 2026