NVIDIA has officially launched the Isaac GR00T Development Platform, an integrated, open-source ecosystem designed to address the fragmentation currently hindering humanoid robotics development. The platform aims to unify the entire pipeline—from data collection and simulation-based training to deployment—allowing developers to transition more efficiently from initial robot bring-up to real-world task execution.
As part of this release, NVIDIA introduced GR00T 1.7, the first commercially usable, open-source Vision-Language-Action (VLA) model for generalized humanoid skills. Distributed under the Apache 2.0 license, the model features a new Cosmos-Reason2-2B backbone, which supports flexible image resolutions and improved long-horizon task reasoning. According to NVIDIA, the 1.7 model demonstrates significant performance gains, including a 61% improvement in DROID-F6 benchmarks compared to previous versions.
The platform integrates several key NVIDIA technologies, including Isaac Lab-Arena for simulation, Isaac Teleop for data capture, and Isaac ROS for on-device inference. By providing a modular workflow, the platform allows teams to train policies using a mix of simulated and real-world demonstration data. This approach is intended to reduce the manual configuration time that typically consumes significant resources during the early stages of robotics projects.
The new toolkit is already seeing adoption across the industry. Major humanoid manufacturers and AI firms—including 1X, Agility, ANYBotics, and Skild AI—are integrating components of the GR00T pipeline into their development cycles. Additionally, academic institutions such as Stanford, CMU, and ETH Zurich are utilizing the workflow to accelerate research into dexterous manipulation and cross-embodiment generalization.
Developers can access the GR00T 1.7 model weights and the platform’s reference workflows via GitHub and Hugging Face, with full documentation provided in NVIDIA’s updated technical guides.