Google DeepMind has released Gemini Robotics 2, a new version of its robotics AI that the company is positioning as a step toward what the field calls "physical AGI" — artificial intelligence that can act competently in the messy real world rather than only in text and images.
According to Wired, whose Will Knight reported the launch, the system bundles several different AI models into a single stack that can control a range of robots, including humanoids. Bloomberg reports that Alphabet's Google DeepMind pitched the model as a fix for machines that struggle with dexterity, allowing humanoid robots to coordinate movements across their entire bodies rather than one limb or gripper at a time.
The architecture matters here. As TechEBlog describes it, the Gemini Robotics ER 2 component sits above the robot's motor control systems, supplying what it calls whole-body intelligence. In plain terms: the low-level machinery that actually moves joints stays where it is, and the AI layers on top as the part that reasons about the task and decides what the body should do. That separation is what lets one model drive different robot hardware instead of being hand-tuned for a single machine.
Wired also flags the obvious caveat. Putting a general-purpose AI model into a physical body carries risks that a chatbot does not — a language model that gets something wrong produces bad text, while a robot that gets something wrong produces force in the real world, near real people and real objects.
Why it matters: the past few years of AI progress have been confined mostly to screens, and a model that can generalize across robot bodies is the clearest sign yet that the same scaling that transformed software is now being aimed at machines that move.