What changed
Google DeepMind introduced Gemini Robotics 2 on July 30 as a family of three models: a vision-language-action model for motor control, an embodied-reasoning model for planning and multi-robot coordination, and an on-device model for local inference. The release claims a shift from upper-body, tabletop tasks toward whole-body control: its Apollo 2 example walks, bends and places a watering can, while the same checkpoint is shown across Apollo 2 and Franka Duo configurations.[1,2]
The decision delta is not that robots are now general-purpose operators. It is that the model layer is being packaged around body transfer and orchestration: ER 2 can watch video, track task progress and hand off actions, while On-Device 2 is said to adapt to new embodiments in a few hours with fewer than 200 examples. That moves the integration question from single-arm task learning toward coordination across bodies and time.[1,3]
The boundary
Availability remains the limiting fact. Google lists Gemini Robotics 2 as private preview and directs developers to an early-access partner program; ER 2 is public through the Gemini API and Google AI Studio, with private preview on the enterprise agent platform. Google’s published figures include 45.7 percent on one whole-body task and 32 to 44 percent on several five-finger tasks, even as the company describes the release as a step toward general-purpose physical AI.[1,2,3]
Axios independently reported the release as an AI-software update, not a customer deployment. No source opened here establishes production use, independent evaluation or field reliability. The next checkpoint is an early-access partner or independent lab publishing repeatable results on a new embodiment with disclosed intervention rules, task duration and failure rates.[1,2,3,4]