National Robotics Week: NVIDIA’s latest physical AI for robots, from GR00T models to Isaac Sim and Jetson Thor
NVIDIA is laying out a cohesive stack for physical AI for robots, combining open foundation models, large-scale simulation, and deployment hardware to help teams move faster from R&D to production across industrial manipulation, logistics, and assistance use cases [1][2][3].
Physical AI for robots: NVIDIA’s announcements at a glance
NVIDIA is making open robot foundation models available on Hugging Face and through its Isaac platform so developers can skip costly pretraining and focus on task-specific fine-tuning and deployment [1][2]. The company is also expanding the GR00T N family of vision-language-action models, with GR00T N1.7 entering early access and GR00T N2 advancing a new “world action model” approach aimed at generalist robot policies [1][2]. These releases are paired with Isaac runtime components and the Jetson Thor platform that cover onboard reasoning, planning, and control [1][2].
GR00T N2 and N1.7: capabilities and benchmarks
GR00T N2 builds on DreamZero research and a world action model architecture and currently leads benchmarks such as MolmoSpaces and RoboArena for generalist robot policies. It achieves more than double the task success rate of leading competitors on novel tasks in new environments, signaling improved generalization for unseen scenarios [1][2].
GR00T N1.7, now in early access, targets production-ready systems with generalized skills and dexterous control, with a focus on humanoid platforms and commercial licensing paths [1][2]. For developers tracking NVIDIA GR00T models, the combination of N2’s benchmark leadership and N1.7’s early access positioning is a notable shift toward deployable capability rather than research-only demos [1][2].
Software and simulation stack: Isaac Sim, Omniverse NuRec, and Onshape
NVIDIA’s Isaac platform spans models, data pipelines, simulation, and runtime libraries for end-to-end development. Isaac Sim and Omniverse NuRec 3D Gaussian splatting convert real-world sensor captures into high-fidelity OpenUSD virtual environments for safe, scalable testing and synthetic data generation [1][3]. Teams can iterate robot and environment designs with PTC Onshape integration, then transfer trained policies from simulation to real hardware, tightening the loop between design, validation, and field trials [1][3]. For organizations adopting physical AI for robots, this simulation-to-reality workflow is designed to reduce risk and speed iteration without exposing systems to unsafe conditions during early testing [1][3].
For developers exploring model access beyond vendor portals, Hugging Face provides a distribution channel for robot foundation models, making it easier to evaluate and integrate components alongside existing ML tooling [1][2]. See the platform’s documentation for ecosystem capabilities in model hosting and collaboration via Hugging Face.
Hardware and runtime: Jetson Thor and deployment
The Jetson Thor robotics platform works with the Isaac stack to handle onboard reasoning, planning, and control. This ties the model and simulation layers directly to runtime execution, creating a reference path from trained policies to field deployment, including simulation-to-reality robot deployment workflows [1][2][3]. The approach aligns compute, software, and data so teams can scale from pilots to production with consistent tooling [1][3].
Ecosystem and partners: real-world robots and use cases
A growing set of partners are building humanoid and service robots on NVIDIA’s stack, including NEURA Robotics, Richtech Robotics, AGIBOT, LG Electronics, Boston Dynamics, and Caterpillar, among others [1][2]. These efforts target industrial manipulation, logistics, and home assistance. FieldAI and other partners are also using the tools to scale industrial robotics and synthetic data workflows for physical AI [1][2][3]. The shared pipeline of GR00T models, Isaac Sim robotics simulation, and Jetson Thor aims to cut integration work across this ecosystem [1][2][3].
How to adopt: practical steps for teams
- Access open robot foundation models on Hugging Face via NVIDIA and evaluate baseline capabilities before fine-tuning [1][2].
- Stand up Isaac Sim, generate synthetic datasets with Omniverse NuRec 3D Gaussian splatting, and build high-fidelity OpenUSD scenes from sensor captures [1][3].
- Iterate robot designs using PTC Onshape integration, then train and validate policies in simulation [1][3].
- Deploy to Jetson Thor with the Isaac runtime for onboard reasoning, planning, and control, and execute controlled trials before scaling [1][2][3].
For tool selection, architecture templates, and vendor-neutral guidance, you can also Explore AI tools and playbooks.
Business impact and ROI considerations
The stack is built to reduce pretraining expense, compress development cycles, and expand test coverage through scalable simulation. By aligning models, simulation, and runtime, operators can target faster time to deployment while managing risk with synthetic data and safe testing in virtual environments [1][2][3]. For teams standardizing on a single pipeline for procurement and engineering, physical AI for robots offers a clearer path from prototype to production-grade systems [1][2][3].
Resources, links and next steps
- NVIDIA newsroom briefings on new models, partner robots, and deployment pipelines [1][2]
- NVIDIA’s build guide for simulation-to-reality using Isaac and Omniverse tools [3]
- Model access and collaboration via Hugging Face
Sources
[1] NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots | NVIDIA Newsroom
https://nvidianews.nvidia.com/news/nvidia-releases-new-physical-ai-models-as-global-partners-unveil-next-generation-robots
[2] NVIDIA and Global Robotics Leaders Take Physical AI to the Real World | NVIDIA Newsroom
http://nvidianews.nvidia.com/news/nvidia-and-global-robotics-leaders-take-physical-ai-to-the-real-world
[3] From Simulation to Production: How to Build Robots With AI | NVIDIA Blog
https://blogs.nvidia.com/blog/build-robots-with-ai/
