NVIDIA releases Isaac GR00T N1.7
NVIDIA put early access of Isaac GR00T N1.7 — an open vision‑language‑action foundation model aimed at humanoid robots — onto Hugging Face for real‑world deployment. The release is positioned as an open tool for vision-to-action tasks that developers can access directly on the Hugging Face platform. (x.com)
Humanoid robots are getting a new open “brain”: NVIDIA has put Isaac GR00T N1.7 into early access on Hugging Face for developers to download now. (huggingface.co) The release centers on a 3 billion-parameter model, `nvidia/GR00T-N1.7-3B`, plus task-specific checkpoints for DROID, LIBERO and SimplerEnv benchmarks in the same Hugging Face collection. (huggingface.co) GR00T is a vision-language-action model, which means it takes camera views, text instructions and robot state, then outputs continuous motor actions for tasks like grasping, moving objects and hand-to-hand transfers. (huggingface.co) (developer.nvidia.com) NVIDIA introduced the GR00T N1 family on March 18, 2025, as an open foundation model for humanoid robot reasoning and skills, alongside simulation and synthetic-data tools for robot training. (investor.nvidia.com) That setup is aimed at a bottleneck in robotics: collecting enough real-world demonstrations to teach a machine to see, plan and move in cluttered spaces without writing a separate controller for each task. NVIDIA says developers can post-train GR00T models with real or synthetic data for a specific robot, task or environment. (developer.nvidia.com) (huggingface.co) NVIDIA’s earlier N1 design split the job in two parts: a vision-language model that plans from images and instructions, and an action model that turns that plan into precise robot motion. The company described that architecture in 2025 as “System 2” for slower reasoning and “System 1” for faster action. (investor.nvidia.com) By January 2026, NVIDIA said GR00T N1.6 had expanded from tabletop manipulation into full loco-manipulation, combining navigation, perception and whole-body control in a sim-to-real workflow. That version used a 32-layer diffusion transformer and Cosmos Reason components for step-by-step planning. (developer.nvidia.com) The N1.7 materials published this week describe another model refresh: the Hugging Face card keeps the 3B “medium-sized” format, while the accompanying code README says the update adds a new vision-language-model backbone, listed as Cosmos-Reason2-2B or Qwen3-VL, and improved performance. (huggingface.co) (github.com) NVIDIA has been widening the distribution path at the same time. In March 2026, the company said it was integrating Isaac and GR00T with Hugging Face’s LeRobot framework, linking NVIDIA’s robotics developer base with Hugging Face’s larger artificial intelligence community. (nvidianews.nvidia.com) The immediate test for N1.7 is not a benchmark chart but whether robot developers can adapt it to real hardware faster than they could with N1.6. Putting the weights on Hugging Face makes that experiment available to anyone who wants to try. (huggingface.co)
