NVIDIA Isaac GR00T N1.7: Open-Source VLA Model for Humanoid Robots

Discover how NVIDIA Isaac GR00T N1.7 is revolutionizing humanoid robot intelligence with an open-source VLA model.

NVIDIA Isaac GR00T N1.7: Open-Source VLA Model for Humanoid Robots
NVIDIA Isaac GR00T N1.7 VLA model for humanoid robots

On April 17, 2026, NVIDIA announced the release of Isaac GR00T N1.7, an open-source Vision-Language-Action (VLA) model designed for humanoid robots. The model represents a significant advance in robot intelligence, combining visual observations and language instructions to generate continuous robotic actions.

What Is Isaac GR00T N1.7?

Isaac GR00T N1.7 is a VLA model with 3 billion parameters that maps visual observations and language instructions to continuous robotic actions. It uses a cascaded action architecture, separating high-level reasoning from low-level motor control:

  • System 2 (Vision-Language Model): A Cosmos-Reason2-2B backbone processes images and language instructions to produce high-level action tokens, handling task decomposition and multi-step reasoning.
  • System 1 (Diffusion Transformer): A 32-layer Diffusion Transformer takes the VLM's output and the robot's current state, denoising them into precise motor commands in real time.

Inputs include RGB image frames, language instructions and the robot's proprioceptive state, while outputs are continuous-valued action vectors mapped to the robot's degrees of freedom. The model has been validated on a range of tasks, including tabletop manipulation, bimanual manipulation and dexterity tasks across different robots, such as the Unitree G1, the Bimanual Manipulator YAM and the AGIBot Genie 1.

Training on Human Egocentric Video Data

The model was pre-trained on over 20,000 hours of human egocentric video, covering more than 20 task categories, including manufacturing, retail, healthcare and home environments. This approach led to the discovery of a scaling law for robotic dexterity: increasing human video data predictably improves robots' manipulation capabilities, with task completion doubling when data grows from 1,000 to 20,000 hours.

Inference and Deployment

To deploy the model, you can follow these steps:

git clone --recurse-submodules 
cd Isaac-GR00T
bash scripts/deployment/dgpu/install_deps.sh
source .venv/bin/activate

uv run python gr00t/eval/run_gr00t_server.py \
 --embodiment-tag GR1 \
 --model-path nvidia/GR00T-N1.7

# Run the environment loop
from gr00t.policy.server_client import PolicyClient

policy = PolicyClient(host="localhost", port=5555)

obs, info = env.reset()
action, info = policy.get_action(obs, info)

This setup lets the model run on a specific robot, such as the GR1, using the provided policy server.

Implications for Humanoid Robot Development

The release of Isaac GR00T N1.7 offers many opportunities for humanoid robot development:

  • Commercial license: The commercial license allows the model to be deployed in production environments, such as material handling, packaging and inspection.
  • Advanced reasoning: The model is designed for multi-step tasks, improving reliability in complex workflows.
  • Dexterous manipulation: Finger-level control enables contact-rich tasks, such as assembling small parts and handling fragile components.
  • Dexterity scalability: Training on a large dataset of human video predictably improves robots' manipulation capabilities, reducing the need for massive teleoperation.

These features make Isaac GR00T N1.7 a powerful tool for accelerating the development and deployment of humanoid robots across various industrial sectors.

Conclusion

The release of NVIDIA Isaac GR00T N1.7 marks a significant step in the evolution of humanoid robots, offering an open-source VLA model that combines visual observations and language instructions to generate continuous robotic actions. With a commercial license and advanced reasoning and manipulation capabilities, this model is ready to be deployed in production environments, opening new possibilities for automation and human-robot interaction.