On July 6, 2026, NVIDIA and Hugging Face announced that Isaac GR00T N1.7 — the latest version of NVIDIA’s humanoid robot foundation model — is now available through LeRobot, Hugging Face’s open-source robotics library.
The licensing change is the headline. GR00T N1.6, released at CVPR 2026 on June 1, was distributed under NVIDIA’s One-Way Noncommercial License. N1.7 ships with commercial licensing, enabling production deployments. That’s the unlock: builders can now ship products with this model.
The architecture also changed in ways that matter for builders doing serious work with humanoid hardware.
What Changed From N1.6 to N1.7
License (The Most Important Change)
N1.6 prohibited commercial use. N1.7 is commercially licensable, validated for production use cases including material handling, packaging, and inspection workflows. If you were waiting to evaluate GR00T for a real product, N1.7 is the version that matters.
Architecture: Cosmos-Reason2-2B Replaces the Old VLM Backbone
GR00T N1.6 used an internal Cosmos-2B variant as its vision-language backbone, with the top 4 layers unfrozen during pretraining. N1.7 replaces this with Cosmos-Reason2-2B, built on Qwen3-VL. This is a meaningful shift: Qwen3-VL brings stronger instruction-following and multimodal reasoning into the backbone.
The action head remains a flow-matching Diffusion Transformer. The dual-system architecture is now explicit:
- System 2 (slow): Cosmos-Reason2-2B — handles task understanding, high-level reasoning, subtask decomposition
- System 1 (fast): 32-layer Diffusion Transformer — outputs continuous action vectors for robot degrees of freedom in real time
Flash attention is now optional in LeRobot’s GR00T integration, which previously caused installation friction. pip install 'lerobot[groot]' now just works.
Training Data: EgoScale and the Scaling Law
The pretraining data increased substantially:
- N1.7 was pretrained on 20,854 hours of human egocentric video across 20+ task categories (from NVIDIA’s EgoScale dataset)
- Previous versions trained primarily on teleoperation demonstrations
NVIDIA published the first dexterity scaling law: going from 1,000 to 20,000 hours of human egocentric video more than doubles average task completion on manipulation benchmarks. This is the first such scaling relationship documented for robot manipulation capability.
The result: better generalization to novel objects and environments with less per-task fine-tuning.
New Capabilities
Finger-level dexterous control. N1.7 supports actuation at individual finger joints, enabling contact-intensive tasks — small parts assembly, precise insertion, object reorientation — that N1.6 couldn’t handle reliably.
Multi-step reasoning without manual decomposition. The Cosmos-Reason2-2B backbone handles task decomposition internally, improving consistency across extended manipulation sequences without requiring builders to manually segment tasks.
The LeRobot Integration
LeRobot is Hugging Face’s open-source robotics library — a PyTorch-native framework for training and deploying robot policies, with datasets, simulation environments, and pre-trained models. As of the July 6 update, LeRobot includes:
- GR00T N1.7 checkpoint loading — load NVIDIA’s published weights directly through the LeRobot API
- NVIDIA Isaac Teleop — NVIDIA’s open-source framework for capturing human demonstrations using external input devices
- Standardized dataset format — Isaac Teleop captures demonstrations in LeRobot’s format, enabling direct fine-tuning
What Isaac Teleop Does
Isaac Teleop solves a recurring problem: capturing high-quality robot demonstration data in a format your training pipeline can use. The framework:
- Accepts input from external control devices (haptic controllers, SpaceMouse, motion capture)
- Outputs standardized trajectories compatible with LeRobot’s dataset schema
- Enables upload to Hugging Face Hub for community sharing and dataset expansion
For builders who need to adapt GR00T N1.7 to a custom robot embodiment or new task category, Isaac Teleop reduces the data collection friction substantially.
Setup
pip install 'lerobot[groot]'
Model weights are available through the Hugging Face collection (nvidia/Isaac-GR00T-N1.7-3B). The GitHub repository (github.com/NVIDIA/Isaac-GR00T) contains the full codebase.
Hardware requirements: NVIDIA Ampere, Hopper, Lovelace, Blackwell, or Jetson. (Same as N1.6 — no new GPU requirement.)
The Bridge NVIDIA and Hugging Face Are Building
NVIDIA has approximately 3 million robotics developers in its ecosystem through Isaac, Omniverse, and Jetson. Hugging Face has approximately 16 million AI developers building with models, datasets, and the Hub. These populations have historically been distinct — robotics builders use ROS, URDF, Isaac Sim; AI builders use transformers, diffusers, datasets.
The GR00T N1.7 + LeRobot integration is a deliberate bridge. NVIDIA’s trajectory data (350,000+ robot demonstrations in LeRobot format), Isaac Teleop, and GR00T weights are now available through the same tooling used for language model fine-tuning. Builders who know how to fine-tune a VLM can now fine-tune a robot policy using the same workflow.
Cosmos 3 (NVIDIA’s physical world model) is listed as “coming soon” to LeRobot, which would add simulation and world-model-based planning to the stack.
What Builders Should Do With This
If you’re working on humanoid or bimanual manipulation problems: N1.7’s commercial license and improved generalization make it the first GR00T version worth evaluating for a production system. The Unitree G1, YAM bimanual manipulator, and AGIBot Genie 1 have published N1.7 validation results.
If you’re an AI builder curious about physical AI: pip install 'lerobot[groot]' + a LeRobot dataset + nvidia/Isaac-GR00T-N1.7-3B is now a working entry point. The same fine-tuning intuitions from language models apply — more demonstration data improves generalization, and the scaling law gives you a rough calibration for how much data you need.
If you need custom embodiment support: Isaac Teleop handles data capture, LeRobot handles formatting and training, and GR00T N1.7 handles the policy. The full loop is now a coherent open-source pipeline for the first time.
Wait if: You need verified third-party benchmarks beyond NVIDIA’s own evaluation, or if your hardware target isn’t in the validated set (Unitree G1, YAM, AGIBot Genie 1). N1.7 is production-licensed but relatively new — independent evaluations will follow over the next few months.
Context: Where This Sits in NVIDIA’s Physical AI Stack
Our CVPR 2026 coverage documented the full stack released on June 1: GR00T N1.6, Alpamayo-R1-10B for autonomous vehicles, OpenShell sandboxed agent runtime, NemoClaw local agent blueprint, and Cosmos 3.
N1.7 is an upgrade within that stack — specifically to the humanoid manipulation model. Alpamayo (AV reasoning), OpenShell (agent sandbox), and NemoClaw (local agent deployment) haven’t announced version updates yet.
The Cosmos 3 integration into LeRobot, when it arrives, would extend the stack further — adding world-model-based environment generation and closed-loop simulation for training.
The commercial license matters most. Every prior GR00T version required noncommercial use agreements, which blocked production deployment. N1.7 removes that gate. Combined with the EgoScale data scaling insight and the LeRobot integration, the path from “interested in physical AI” to “running a robot foundation model in production” just got substantially shorter.