AINEWS
Gemini Robotics 2 drives a humanoid from feet to fingertips and learns a new robot body in a few hours from under 200 examples.
A humanoid robot heard the instruction "put the watering can in the green bin on the bottom shelf" and handled the rest on its own. It walked across a room, picked up the can, stepped aside, bent at the knees, and lowered the can onto a low shelf. The machine was Apptronik's Apollo 2, controlled by Gemini Robotics 2, a vision-language-action model Google DeepMind released to developers on July 30, 2026 through the Gemini API and Google AI Studio.
Earlier DeepMind robot models mostly moved an upper body for tabletop work: pick a cup, slide a block. Gemini Robotics 2 drives the entire machine. It coordinates the legs, torso, arms, and a five-fingered hand with 22 joints. In the watering-can demo, one natural-language instruction produced a chain of walking, balancing, crouching, and object placement inside a human-shaped room. The same model also drives simpler two-fingered grippers on other platforms, which is where the underappreciated shift lives.
The model is one piece of a three-part release. Gemini Robotics 2 turns camera input and text into motor commands. Gemini Robotics ER 2 sits above it as a planning layer: it watches a live video feed, tracks whether the lower model is making progress, and can call tools like Google Search when a step stalls. In one DeepMind video, ER 2 also steers a Boston Dynamics Spot, a four-legged walking robot, to fetch a snack, and coordinates two different robots at once: a wheeled machine and a humanoid, splitting the task between them. The third piece, On-Device 2, runs without an internet connection and can be retuned to an entirely new robot body from fewer than 200 examples in a few hours of training.
That last number is the one that changes the math. Fitting a capable AI controller to a new robot has, until now, meant weeks of bespoke teleoperation data (a human operator driving the robot through tasks while the model watches) and per-hardware tuning by a research team. A few hours of examples on an off-the-shelf machine turns each new robot body into a software target, which is the part of the release that pushes what the industry calls "physical AI" out of the lab demo and toward something a developer can actually try on hardware they didn't build.
Dexterity got the same upgrade. Apollo's 22-joint hand ties knots and seals a ziplock bag in DeepMind's videos, tasks that earlier robot policies could not chain end-to-end. The same control stack also runs on cheaper two-fingered grippers, so a logistics arm and a humanoid can share an upstream model. That shared backbone is what makes the hours-fast body swap possible, because the model has learned general manipulation rather than a single gripper's quirks.
Wired called the release a real step toward "physical AGI." That label is the magazine's, not DeepMind's. Carolina Parada put the company's own line on it more narrowly: "Our goal is to bring AI into the physical world and then build the intelligence layer that can be used by every robot." The demos are short, controlled-room sequences. Apollo 2 is not yet shown tidying a real apartment, and the hours-fast transfer still has to clear the gap between a single lab video and a deployment where a customer unboxes a different machine and gets useful behavior the same afternoon.
The second-order effect shows up in unit economics. A humanoid vendor's bill of materials has always included a hidden line item: the months of teleoperation data and per-platform fine-tuning needed to make a generic model useful on a specific machine. If a new robot body can be onboarded in a few hours from under 200 examples, that line item shrinks, and the model layer starts to look like a shared operating system for robots. That is the bet the release is asking the industry to make.
The watch item is the gap between what is on the API today and what survives contact with a real home: loose rugs, pets, cups left on the wrong shelf. Gemini Robotics 2 gives developers the building blocks for that test. The next few months of third-party demos will show whether the hours-fast body swap keeps its accuracy once the room stops being a stage, and whether "physical AI" gets cheap or stays bespoke.