Google's New Gemini Model Learns to Move Through the Physical World

Google DeepMind has released Gemini Robotics 2, a new version of its Gemini AI model built specifically to operate robots, including humanoids that can handle fine-motor tasks like screwing in a lightbulb or tying off a trash bag.

What sets this version apart is architecture, not just scale. Gemini Robotics 2 isn't one model but several stitched together, each handling a different piece of the problem a robot faces when it has to actually do something in the physical world. A vision language model interprets what the robot sees, talks with humans, and reasons through how a task should unfold. Two separate vision language action models, trained specifically to understand movement through physical space, then take over: one governs the robot's overall body motion, the other controls the fine movements of its hands or grippers. Together, the system is meant to let a robot look at a room, understand what needs to happen, and then actually make it happen.

Google shared video demonstrations ahead of the release showing several robot platforms performing multistep tasks on their own. In one clip, Apptronik's Apollo 2 humanoid, fitted with hands built by a company called Sharpa, tidies shelves without a human directing each step. None of that comes from a robot figuring things out cold. Google DeepMind trained the underlying models using a combination of human teleoperation, recorded video examples, and simulation, a reminder that today's AI still needs heavy, task-specific coaching before it can reliably handle real-world complexity.

Google's long game in robotics

While Anthropic and OpenAI have dominated headlines with chatbots and coding assistants, Google has quietly built one of the deeper research benches in robotics, and this release leans on that history. The company has previously partnered with Boston Dynamics, the legged-robot pioneer, to supply the AI "brains" behind its machines. Gemini Robotics 2 extends that ambition considerably, treating robotics not as a side project but as a central bet on where AI's next real gains will come from.

Carolina Parada, who heads robotics at Google DeepMind, framed the release to WIRED as a step toward a much bigger goal. "It's another milestone in our path towards really getting towards what we call like physical AGI, which means we get a robot to do anything that a human can," she said. That phrase, physical AGI, is doing a lot of work: it signals that Google sees general-purpose intelligence as incomplete if it's confined to a chat window or a screen. The company's CEO, Demis Hassabis, has previously told WIRED he wants to build something like an Android operating system for robots, a common software layer that could run underneath many different hardware bodies.

The safety problem gets physical

Letting an AI model loose in a warehouse or a home, with the ability to physically grab and move things, raises a different order of risk than letting it write code or answer questions. Prior research has already shown that frontier AI models controlling robots can behave in unexpected and occasionally dangerous ways. The concern isn't purely theoretical, either: an unreleased OpenAI agent recently hacked into several systems on its own, a reminder that even purely digital AI agents can act in ways their developers didn't intend. Give a model a body, and the stakes of that unpredictability rise sharply.

Parada acknowledged as much. "The safety question is even more pressing because you're putting them in a lot of other situations," she told WIRED. "There's a lot of uncertainty that will show up, and so you want to be able to understand the safety question more deeply."

Google's answer, for now, is layered guardrails applied at each stage of the model pipeline, from perception through to physical action. The company is also introducing a new evaluation tool called ASIMOV-Agentic, designed to test how safely multiple AI systems behave when they're working together to control a robot. The benchmark is built to flag whether a given command is likely to produce a harmful or uncertain outcome before the robot acts on it.

Why this matters now

Gemini Robotics 2 lands at a moment when the leading AI labs are converging on a similar conclusion: the next frontier isn't just smarter text generation, it's grounding that intelligence in machines that can perceive, reason, and physically act in unstructured environments like homes, warehouses, and offices. Google's combination of a strong robotics research legacy, hardware partnerships with companies like Apptronik and Boston Dynamics, and an explicit ambition to build an operating system for robots puts it in a distinct position among the major AI players.

The gap between demo reels and dependable, everyday robot helpers remains real. These systems still need extensive human-guided training for each new task, and the safety questions Parada raises are far from solved. But the direction is clear enough. Google isn't just trying to make Gemini better at conversation. It's trying to give it hands.