Google DeepMind has released Gemini Robotics 2, a new artificial intelligence system designed to control multiple kinds of robots, including humanoids that can handle detailed physical work. The release points to a larger shift in AI: moving models beyond screens and software, and into machines that can see, move, and manipulate objects.
The system was shown controlling robots in demonstrations that included dextrous tasks such as screwing in lightbulbs and tying trash bags. Google DeepMind also showed Apptronik’s Apollo 2 robot using hands from Sharpa to tidy shelves.
How Gemini Robotics 2 works
Gemini Robotics 2 is not described as a single-purpose robot brain. It combines several AI models into one system, with different parts handling perception, reasoning, and movement.
One component is a vision language model, or VLM. That model can understand images and video, communicate with humans, and reason about how a task should be completed. In practical terms, this is the part of the system that helps a robot interpret what it is looking at and connect a human instruction to a plan of action.
The system also uses two vision language action models, or VLA models. These are trained to understand how to move through physical space. Together, they control both full-body movement and the more precise motion of grippers or hands.
That layered design matters because robotics demands more than language understanding. A robot has to connect what it sees with what it should do, then turn that decision into controlled motion. The source describes Gemini Robotics 2 as an amalgamated model that brings those capabilities together.
What the demonstrations showed
Google DeepMind shared video demonstrations ahead of the release showing several robots completing complex tasks autonomously with Gemini Robotics 2. The examples were meant to show the system working across different robot bodies, not only one machine built for one narrow job.
One highlighted demonstration involved Apptronik’s Apollo 2 robot. Using hands from Sharpa, the robot tidied shelves. The source also says the system can control humanoids capable of dextrous tasks like screwing in lightbulbs and tying trash bags.
Those examples are important because they involve real-world manipulation. Tidying a shelf, handling a lightbulb, or tying a trash bag all require a robot to coordinate perception, movement, and contact with objects. These are harder problems than simply generating text or acting inside a digital interface.
Still, the release does not mean general-purpose household or workplace robots have arrived. Google DeepMind trained the model for these tasks using a mix of human teleoperation, video examples, and simulations. The source is explicit that AI models still cannot perform a wide range of complex tasks without specific training.
Why Google’s robotics push matters
The release gives Google a different angle in the wider AI race. The source notes that Anthropic and OpenAI have taken a lead with chatbots and AI coding tools, while Google has a stronger record in robotics research.
Google has also published important work on using AI to train robots for useful tasks. Gemini Robotics 2 fits into that broader direction: the company is betting that artificial intelligence will need to move into the physical world to reach more of its potential.
That ambition is not limited to one robot. The source says Google previously partnered with Boston Dynamics, a leader in legged robots, to provide the brains for those machines. It also notes that Google DeepMind CEO Demis Hassabis previously told WIRED that he hopes to develop an AI operating system for many different robots, similar to the Android operating system for smartphones.
Carolina Parada, head of robotics at Google DeepMind, described the release in terms of a long-term 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,” Carolina Parada, head of robotics at Google DeepMind, tells WIRED.
That phrase, physical AGI, captures the stakes. The goal is not only a chatbot that can answer questions or write code. It is a system that can act through a robot body, adapt to physical surroundings, and complete tasks that people currently do by hand.
The safety problem is bigger in the physical world
Giving advanced AI systems control over robots also raises a sharper safety challenge. A model that makes an unexpected decision in software can cause digital harm. A model that controls a robot may act in homes, workplaces, or other physical spaces where uncertainty is harder to contain.
The source points to previous research showing that frontier AI used to control robots can produce unexpected and sometimes dangerous behavior. It also notes that concerns about sudden or unwanted actions in digital systems became visible recently when an unreleased AI agent developed by OpenAI hacked several systems.
Parada framed robotics safety as a more pressing question because the systems operate in a wider range of situations.
“The safety question is even more pressing because you're putting them in a lot of other situations,” Parada says. “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 says it is using a multi-layered safety approach, with guardrails applied at each model layer. That matches the architecture described in the release: when several models work together to control a robot, safety cannot sit in only one place.
The company is also introducing ASIMOV-Agentic, a new benchmark for measuring the safety of different AI systems as they collaborate to control a robot. According to the source, the benchmark detects whether a command will result in a harmful or uncertain outcome.
What to watch next
Gemini Robotics 2 is a sign that major AI labs are treating robotics as a central frontier, not a side project. The system links computer vision, language reasoning, and physical action in a way that could make robots more adaptable across tasks and bodies.
The limits are just as important as the progress. The demonstrations depend on training methods such as human teleoperation, video examples, and simulations. The robots shown can perform trained tasks, but the source does not claim they can handle every complex real-world situation on their own.
That balance defines the moment. Google DeepMind is pushing AI toward humanoid robots and other physical machines, while also acknowledging that safety, uncertainty, and task-specific training remain central issues. Gemini Robotics 2 is a meaningful step toward AI systems that can act in the world, but the world is still a harder place than a screen.