Physical AI: foundation models for robots are becoming products
AI · not reviewedWritten with AI from the sources listed, not yet editorially reviewed.
NVIDIA and Google DeepMind now offer models that turn camera images and language directly into robot motion — some already with commercial licences.
The essentials in four points
- At GTC in March 2026, NVIDIA presented Isaac GR00T N1.7 in early access with a commercial licence; the next generation, GR00T N2, was shown only as a preview.
- FANUC, ABB, Yaskawa and KUKA are integrating NVIDIA Jetson modules into their robot controllers; together the four have more than 2 million robots installed (company figures).
- According to trade reports, Google DeepMind released Gemini Robotics 2 at the end of July 2026; it controls legs, torso, arms and fingers with a single model, alongside a version that runs on the robot without a network connection.
- The whole-body and on-device variants are so far available only to selected test partners; commercial licence terms for robot makers had not been published.
What this means for businesses
For businesses, the question shifts from “which robot?” to “which model, which licence, which compute on the robot?”. If you procure robotics today, check whether the controller and edge computer can run such models, and clarify licence terms in writing before the pilot.
Sources · 2
- Robotics 24/7 · 19 Mar 2026NVIDIA GTC 2026: NVIDIA, global robotics leaders look to take physical AI to the real world
- RoboZaps · 31 Jul 2026Google DeepMind’s Gemini Robotics 2 Can Control a Humanoid From Feet to Fingertips
Links open the original source in a new tab. Figures from company statements are labelled as vendor or company figures in the text.
Matching hardware
Cobots in manufacturing
Automate recurring tasks such as machine tending, packing or inspection with collaborative robots.
More from this branch
All in Robotics →The next OROBO Report comes to you.
Planned from December 2026, every Thursday. Free, unsubscribe in one click.